See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/209803192

Data driven landslide hazard assessment using geographical information

systems and remote sensing.

Article · January 2002

CITATIONS READS

29 343

1 author:

Mehmet Lutfi Süzen

Middle East Technical University
49 PUBLICATIONS   2,393 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Mehmet Lutfi Süzen on 03 October 2016.

The user has requested enhancement of the downloaded file.

 DATA DRIVEN LANDSLIDE HAZARD ASSESSMENT
USING GEOGRAPHICAL INFORMATION SYSTEMS
AND REMOTE SENSING

A THESIS SUBMITTED TO

THE GRADUATE SCHOOL OF NATURAL AND APPLIED SCIENCES

OF

THE MIDDLE EAST TECHNICAL UNIVERSITY

BY

MEHMET LÜTFİ SÜZEN

IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF

DOCTOR OF PHILOSOPHY

IN

THE DEPARTMENT OF GEOLOGICAL ENGINEERING

JANUARY 2002
Approval of the Graduate School of (Name of the Graduate School)

_____________________
Prof. Dr. Tayfur Öztürk
Director

I certify that this thesis satisfies all the requirements as a thesis for the degree of Doctor
of Philosophy.

_______________________
Prof. Dr. Nurkan Karahanoğlu
Head of Department

This is to certify that we have read this thesis and that in our opinion it is fully adequate,
in scope and quality, as a thesis for the degree of Doctor of Philosophy.

_________________________
Prof. Dr. Vedat Doyuran
Supervisor

Examining Committee Members

Prof. Dr. Nurkan Karahanoğlu ____________________

Prof. Dr. Vedat Doyuran ____________________

Prof. Dr. Reşat Ulusay ____________________

Assoc. Prof. Dr. Vedat Toprak ____________________

Assist. Prof. Dr. Candan Gökçeoğlu ____________________

ii
 ABSTRACT

DATA DRIVEN LANDSLIDE HAZARD ASSESSMENT

USING GEOGRAPHICAL INFORMATION SYSTEMS
AND REMOTE SENSING

Süzen, Mehmet Lütfi

Ph.D., Department of Geological Engineering

Supervisor: Prof. Dr. Vedat Doyuran

January 2002, 196 pages

The purpose of this thesis is to generate and evaluate a concise system of data
driven landslide hazard assessment procedure. The Asarsuyu catchment area and Bolu
Mountain highway pass are selected for this purpose
In this study, three data domains, which are remote sensing products,
geological maps and topographical maps are used. The landslides of four different
historical periods are interpreted using aerial photographs. Land cover is extracted from
satellite images. Thirteen sets of parameter maps are produced from geological and
topographical maps and from remote sensing products. In order to store the information
of these parameter maps in a concise thematic database a 25x25 meter grid is overlaid
to the area. Two different sets of points are defined. First one representing the
properties of slided masses and the other set representing the conditions contributing to
the sliding phenomena. The information falling on these points are stored in five
separate thematic landslide attribute databases, as two main databases are Seed Cells
and Slided Mass Databases, with accessory three other as, Polystats, Fuzzystats and
photo-characteristics databases. Following the creation of the databases, the

iii
information stored are evaluated and preliminary landslide generating decision rules are
extracted.
In conjunction with these preliminary results a hazard assessment procedure,
starting from the least detailed and simple ranging up to most complex statistical
analyses are initiated. Finally a relative accuracy assessment procedure is carried out
using the hazard maps produced. It is found that the most accurate, reliable and realistic
results concerning the landslide hazard assessment of Asarsuyu catchment are
obtained through logistic regression analyses.
The results of landslide hazard assessment of Asarsuyu catchment shows that:
the northern slopes of the Asarsuyu catchment is classified as very low hazard due to
the presence of low population, undisturbed dense forest land cover, being very distant
to E-5 highway and to the major active fault and the favorable lithological conditions.
The southeastern slopes especially the Bolu Mountain Highway pass are definitely on
very high hazard class due to the removal of lateral supports by E-5 highway cut slopes,
close location to active faults, high disturbance of the land cover, high traffic activity
along the highway resulting in extra vibration, and the presence of flyschoidal units

Keywords: Asarsuyu Catchment, Bolu Mountain, Geographical Information Systems,

Landslide Hazard Assessment, Logical regression, Remote Sensing

iv
 ÖZ

COĞRAFİ BİLGİ SİSTEMLERİ VE UZAKTAN ALGILAMA

TEKNİKLERİ KULLANILARAK
VERİ KAYNAKLI HEYELAN AFETİ TAYİNİ

Süzen, Mehmet Lütfi

Doktora, Jeoloji Mühendisliği Bölümü

Tez Yöneticisi: Prof. Dr. Vedat Doyuran

Ocak 2002, 196 sayfa

Bu çalışmanın amacı, kısa ve öz veri kaynaklı bir heyelan afeti tayin yöntemi
geliştirmek ve değerlendirmektir. Bu amaçla uygulama arazisi olarak Asarsuyu su
toplama havzası ve Bolu Dağı otoyol geçişi seçilmiştir.
Bu çalışmada, uzaktan algılama ürünleri, jeolojik haritalar ve topografik
haritalar, olmak üzere üç veri alanı kullanılmıştır. Hava fotoğrafları kullanılarak dört ayrı
dönemin heyelan envanteri çıkarılmıştır. Arazi kullanım haritası için ise, uydu
görüntülerinden yararlanılmıştır. Jeolojik ve topoğrafik haritalardan ve uzaktan algılama
ürünleri kullanılarak 13 ayrı parametre haritası üretilmiştir. Bu parametre bilgilerini bir
veri tabanına aktarabilmek için, 25x25 metrelik bir ağ arazi üzerine oturtulmuştur. İki ayrı
takım nokta belirlenmiştir. Bunlardan ilki kayan kütlelerin özelliklerini, diğeri ise kaymaya
neden olan koşulları içermektedir. Bu noktalarda saklanan bilgiler, beş ayrı veri tabanına
aktarılmıştır. Bunlardan kayan kütle ve kök hücre veri tabanları iki ana veri tabanını
oluşturmaktadır. Diğer üç yardımcı veri tabanı ise, alansal (polystats), şekilsel
(fuzzystats) ve foto-karakteristik veri tabanlarıdır. Veri tabanlarının üretilmesinden sonra
heyelanlar hakkında ilk karar verme kuralları elde edilmiştir.

v
 İlk sonuçlar ışığında en basit ve ayrıntısız analizlerden başlayarak, en
karmaşık istatistiksel yöntemlere kadar uzanan bir heyelan afeti tayin yöntemi
izlenmiştir. En son olarak ise, göreceli hata tayin yöntemi kullanılmıştır. Bunların
ışığında en makul ve gerçeğe en uygun yötemin mantıksal regresyon olduğu sonucuna
varılmıştır.
Heyelan afeti tayininin sonuçlarına gore, düşük nüfus yoğunluğu, bozulmamış
arazi örtüsü, E-5 otoyoluna ile aktif fay hatlarına uzaklığı ve sağlam kaya birimleri ile
Asarsuyu su toplama havzasının kuzey yamaçları çok düşük riskli afet alanı olarak
belirlenmiştir. Güneybatı yamaçları, özellikle Bolu Dağı geçişi ise, E-5 yolunun yarmaları
nedeniyle yok olan yanal destekler, aktif faylara olan yakınlığı, E-5 otoyolundaki yüksek
trafik yoğunluğu ve fliş litolojisinin varlığı nedenyle çok yüksek riskli afet alanı olarak
belirlenmiştir.

Anahtar Kelimeler: Asarsuyu Havzası, Bolu Dağı, Coğrafi Bilgi Sistemleri, Heyelan
Afeti Tayini, Mantıksal Regresyon, Uzaktan Algılama

vi
To my Family

vii
 ACKNOWLEDGMENTS

I would like to express my special thanks to my supervisor Prof. Dr. Vedat

Doyuran, especially for his guidance, patience and motivation. I am sincerely in debt to him
for outrageous guiding and for his continuous support at every stage throughout this study.
To him I am also in great debt for the tedious job of critically reviewing and editing of the
manuscript even in zipped form. You know you are much more than a supervisor for me.

I would like to thank to Prof. Dr. Nurkan Karahanoğlu for his valuable
contributions and showing the way out of the dungeons of statistical world.

I would also like to thank to Assoc. Prof. Dr. Vedat Toprak for his patience,
continuous support and unlimited thrust through every stage of this work.

I would like to thank to Prof. Dr. Reşat Ulusay for his valuable contributions
through the progress report discussions.

I would like to thank to Assist. Prof. Dr. Candan Gökçeoğlu for his valuable
contributions through the manuscript.

I would like to thank to Assoc.Prof.Dr. Bora Rojay especially for his continuous
help in the field studies and his valuable ideas.

I would also like to thank to Middle East Technical University and Department of
Geological Engineering as being my employer, for giving the opportunity to do the research,
and as all stages of this study is carried out with their facilities and financial support.
I would also like to thank to Assist. Prof.Dr. Nuretdin Kaymakçı especially for
his supports in the Netherlands days.

Dr. Arda Arcasoy, Ertan Yeşilnacar, Onur Demir and Gürkan Benekşe are
thanked for their support and fighting on my side against murphy at our “lab”. Very
special thanks go to these guys for the emission of my stress during this study.

Finally, I would express my special thanks to my family for their unbelievable

support, continuous motivation, ultra-patience, and endless trust. Among them I would like
to show my sincere gratitudes to my Beyaz as most of the earth shaking ideas come to
mind in the morning and evening walks with him. I would like to show my very special
thanks and appreciation to my Seda as being the best friend I have ever had even
suffered from my often irritating moods and always more believed to me than I do. You
know, you deserve this title more than I do.

viii
 TABLE OF CONTENTS

ABSTRACT . . . . . . . . . . . . . . . . . . . . . iii
ÖZ . . . . . . . . . . . . . . . . . . . . . . . . v
DEDICATION . . . . . . . . . . . . . . . . . . . . . vii
ACKNOWLEDGMENTS. . . . . . . . . . . . . . . . . . . viii
LIST OF TABLES . . . . . . . . . . . . . . . . . . . . xiv
LIST OF FIGURES . . . . . . . . . . . . . . . . . . . xvi
CHAPTERS
1. INTRODUCTION . . . . . . . . . . . . . . . . . . 1
1.1. Purpose and Scope . . . . . . . . . . . . . . . 1
1.2. Geographical Setting . . . . . . . . . . . . . . . 3
2. BACKGROUND INFORMATION ON LANDSLIDE HAZARD AND
USE OF REMOTE SENSING AND GEOGRAPHICAL INFORMATION
SYSTEMS IN LANDSLIDE HAZARD ASSESSMENT . . . . . . . 5
2.1. Definition of Landslide Hazard and Terminology . . . . . . . . 5
2.1.1. Scale Factor in Analysis . . . . . . . . . . . . . 7
2.1.1.1. National Scale (<1/1.000.000) . . . . . . . . . . 8
2.1.1.2. Regional/Synoptic Scale (< 100.000). . . . . . . . 8
2.1.1.3. Medium Scale (1/25.000 - 1/50.000) . . . . . . . 8
2.1.1.4. Large Scale (> 1/10.000) . . . . . . . . . . . 9
2.1.2. Knowledge Type Used . . . . . . . . . . . . . 10
2.2. Use of Remote Sensing in Landslide Hazard Assessment . . . . . 11
2.3. Geographical Information Systems and Landslide
Hazard Assessment . . . . . . . . . . . . . . . 14
2.3.1. Phases of Natural Hazard Analysis in GIS. . . . . . . . 17
2.3.2. GIS Based Landslide Hazard Zoning Techniques . . . . . 18
2.3.2.1. Trends in Landslide Hazard Zonation . . . . . . . 19
2.3.2.2 Direct Mapping in Landslide Hazard Analysis . . . . . 20
2.3.2.2.1 Landslide Distribution Analysis . . . . . . . . 20
2.3.2.2.2 Heuristic Approach (Geomorphic Analysis) . . . . 21
2.3.2.3. Indirect Mapping in Landslide Hazard Analysis . . . . 21

ix
 2.3.2.3.1. Statistical Methods in Landslide Hazard Analysis . . 21
2.3.2.3.1.1. Bivariate Statistical Methods in Landslide
Hazard Analysis . . . . . . . . . . . 22
2.3.2.3.1.1.1. Landslide Susceptibility Analysis . . . . 24
2.3.2.3.1.1.1.1. Production of the Susceptibility Map . 25
2.3.2.3.1.1.2. Information Value Method . . . . . . 25
2.3.2.3.1.1.3. Weights of Evidence Modelling . . . . . 27
2.3.2.3.1.2. Multivariate Statistical Methods in Landslide
Hazard Analysis . . . . . . . . . . . 31
2.3.2.3.1.2.1. Multiple Regression . . . . . . . . 33
2.3.2.3.1.2.2. Discriminant Analyses . . . . . . . 34
2.3.2.3.2. Knowledge Driven Methods in Landslide
Hazard Analysis . . . . . . . . . . . . 35
2.3.2.3.2.1. Qualitative Map Combination . . . . . . . 35
2.3.2.3.2.2. Favourability Functions . . . . . . . . . 35
2.3.2.3.3. Deterministic Modeling in Landslide Hazard Analysis . 36
2.3.2.3.4 Landslide Frequency Analysis . . . . . . . . 37
2.3.2.4. Accuracy and Objectivity . . . . . . . . . . . 37
2.3.2.5. Evaluation of Methods via Scale Factor . . . . . . . 38
3. GEOLOGY OF THE ASARSUYU CATCHMENT . . . . . . . . . 40
3.1. Geology . . . . . . . . . . . . . . . . . . . 40
3.1.1. Regional Geology and Previous Works . . . . . . . . 40
3.1.2. Stratigraphy . . . . . . . . . . . . . . . . 43
3.1.2.1. Yedigöller Formation . . . . . . . . . . . . 44
3.1.2.2. Kocadere Formation . . . . . . . . . . . . 47
3.1.2.3. Aksudere Formation . . . . . . . . . . . . 47
3.1.2.4. Buldandere Formation . . . . . . . . . . . . 48
3.1.2.4.1. Fındıklıdere Member . . . . . . . . . . . 48
3.1.2.4.2. Çaycuma Formation . . . . . . . . . . . 49
3.1.2.5. Asarsuyu Formation . . . . . . . . . . . . . 49
3.1.2.6. Quaternary Deposits . . . . . . . . . . . . 49
3.2. Tectonism of the Asarsuyu Catchment . . . . . . . . . . 50
3.3. Seismicity of the Asarsuyu Catchment and Environs . . . . . . 51
4. INPUT DATA AND DATA PRODUCTION . . . . . . . . . . . 57
4.1. Geology . . . . . . . . . . . . . . . . . . . 57
4.1.1. Data entry . . . . . . . . . . . . . . . . . 57
4.1.2. Input map generation . . . . . . . . . . . . . . 60
4.2. Elevation . . . . . . . . . . . . . . . . . . . 62

x
 4.2.1. Data entry . . . . . . . . . . . . . . . . 62
4.2.2. Input map generation . . . . . . . . . . . . . . 61
4.3. Infrastructure . . . . . . . . . . . . . . . . . . 68
4.3.1. Data entry . . . . . . . . . . . . . . . . . 68
4.3.2. Input data production . . . . . . . . . . . . . . 68
4.4. Land cover . . . . . . . . . . . . . . . . . . . 71
4.4.1. Data entry . . . . . . . . . . . . . . . . . 71
4.4.2. Input map generation . . . . . . . . . . . . . . 72
4.4.2.1 Georeferencing . . . . . . . . . . . . . . 73
4.4.2.2 Data Processing . . . . . . . . . . . . . . 74
4.4.2.2.1. Principal Component Analysis (PCA) . . . . . . 74
4.4.2.2.2. Decorrelation Stretching . . . . . . . . . . 75
4.4.2.2.3. Vegetation Indices . . . . . . . . . . . . 77
4.4.2.2.3.1. The Tasseled Cap . . . . . . . . . . 78
4.4.2.2.3.2. The Gram-Schmidt . . . . . . . . . . 78
4.4.2.2.3.2. Global Vegetation Index (GVI) . . . . . . 79
4.4.2.2.3.3. Infrared Percentage Vegetation Index (IPVI) . . 79
4.4.2.2.3.4. Modified Normalized Difference Index (MNDI) . . 79
4.4.2.2.3.5. Transformed Vegetation (TVI) . . . . . . 80
4.4.2.4. Normalization . . . . . . . . . . . . . . . 80
4.4.2.5. PCA for Vegetation . . . . . . . . . . . . . 80
4.4.3. Maximum Likelihood Classification . . . . . . . . . . 82
4.4.3.1. Accuracy Assessment . . . . . . . . . . . . 82
4.4.4 Integration of RS, GIS at Database Level . . . . . . . . 83
4.5. Landslide Inventory . . . . . . . . . . . . . . . . 85
4.5.1. Input Data . . . . . . . . . . . . . . . . . 85
4.5.2. Data Production . . . . . . . . . . . . . . . 86
5. LANDSLIDE DATABASES . . . . . . . . . . . . . . . 90
5.1. Topological, Morphometrical database . . . . . . . . . . 90
5.1.1. Polystats Database . . . . . . . . . . . . . . 91
5.1.1.1. 1952 Period . . . . . . . . . . . . . . . 91
5.1.1.2. 1972 Period . . . . . . . . . . . . . . . 95
5.1.1.3. 1984 Period . . . . . . . . . . . . . . . 98
5.1.1.4. 1994 Period . . . . . . . . . . . . . . . 100
5.1.1.5. The Comparison of Four Periods . . . . . . . . . 102
5.1.2. The Fuzzystats Database . . . . . . . . . . . . . 105
5.1.2.1. Form Ratio . . . . . . . . . . . . . . . 105
5.1.2.2. Grain Shape Index . . . . . . . . . . . . . 107

xi
 5.1.2.3. Compactness . . . . . . . . . . . . . . 108
5.1.2.4. Circularity 1 . . . . . . . . . . . . . . . 109
5.1.2.5. Circularity 2 . . . . . . . . . . . . . . . 110
5.1.2.6. Elongation . . . . . . . . . . . . . . . 111
5.1.2.7. The significant changes and evaluation of fuzzy
parameters with relation to Polystats database . . . . 112
5.2. Photo-characteristics Database . . . . . . . . . . . . 114
5.2.1. Massinfo . . . . . . . . . . . . . . . . . . 114
5.2.2. Type . . . . . . . . . . . . . . . . . . . 115
5.2.3. Style . . . . . . . . . . . . . . . . . . . 116
5.2.4. Depth . . . . . . . . . . . . . . . . . . . 116
5.2.5. Distribution of Activity . . . . . . . . . . . . . . 117
5.2.6. Land cover . . . . . . . . . . . . . . . . . 118
5.3. Landslide Attribute Databases . . . . . . . . . . . . . 119
5.3.1. Lithomap . . . . . . . . . . . . . . . . . . 120
5.3.2. Distfault and Faultdens . . . . . . . . . . . . . 121
5.3.3. Elevmap . . . . . . . . . . . . . . . . . . 123
5.3.4. Distance to drainage . . . . . . . . . . . . . . 123
5.3.5. Drainage Density . . . . . . . . . . . . . . . 125
5.3.6. Distridge . . . . . . . . . . . . . . . . . . 125
5.3.7. Aspect . . . . . . . . . . . . . . . . . . 126
5.3.8. Slope . . . . . . . . . . . . . . . . . . . 127
5.3.9. Distsettlement . . . . . . . . . . . . . . . . 128
5.3.10.Distpower & Dist_Roadnetwork . . . . . . . . . . . 128
5.3.11 Dist E-5 Highway . . . . . . . . . . . . . . . 129
5.3.12. Land Cover. . . . . . . . . . . . . . . . . 130
6. HAZARD ASSESSMENT . . . . . . . . . . . . . . . . 132
6.1. Thematic Landslide Attribute spatial distribution analysis (TLASDA) . 132
6.2. Landslide activity analysis (LACTA) . . . . . . . . . . . 135
6.3. Landslide Isopleth analysis (LIA) . . . . . . . . . . . . 138
6.4. Statistical analyses . . . . . . . . . . . . . . . . 140
6.4.1. Bi-variate analyses . . . . . . . . . . . . . . . 147
6.4.2. Multivariate analyses . . . . . . . . . . . . . . 151
6.4.2.1. Factor analysis . . . . . . . . . . . . . . 152
6.4.2.2. Logical Regression . . . . . . . . . . . . . 157
6.4.3. Comparison of two produced hazard maps . . . . . . . 163
6.4.3.1. The comparison of methods via their areas
and corresponding landslide seed cells . . . . . . . . . 163

xii
 6.4.3.2. The comparison of two methods in the spatial domain . . 164
7. DISCUSSION . . . . . . . . . . . . . . . . . . . 170
7.1. Data production . . . . . . . . . . . . . . . . 170
7.2. Data evaluation . . . . . . . . . . . . . . . . 173
7.3. Hazard Analysis . . . . . . . . . . . . . . . . . 176
7.4. Hazard Map Comparison. . . . . . . . . . . . . . . 179
8. CONCLUSIONS . . . . . . . . . . . . . . . . . . 182
REFERENCES . . . . . . . . . . . . . . . . . . . 184
VITA . . . . . . . . . . . . . . . . . . . . . . 196

xiii
 LIST OF TABLES

TABLE

2.1. The number of GRC needed to identify and interpret object of varying
contrast in relation to its background . . . . . . . . . . . . . . 13
2.2. Minimum object size needed for landslide Identification or
Interpretation . . . . . . . . . . . . . . . . . . . . . 14
2.3. Time schedule comparison of phases of landslide hazard
assessment of conventional methods and GIS based methods
based on scale . . . . . . . . . . . . . . . . . . . . 18
2.4. The trends in landslide hazard zonation . . . . . . . . . . . . . 19
2.5. The possible combinations after map crossing . . . . . . . . . . . 28
2.6. The possible combinations after crossing of two binary maps . . . . . . 29
2.7. Classification of Methods based on scale factor . . . . . . . . . . 39
3.1. Studies performed in the study area and its near vicinity . . . . . . . . 41
4.1. RMSE of Ground Control Points . . . . . . . . . . . . . . . 74
4.2. Covariance, Correlation and Transformation Matrices for PCA . . . . . . 76
4.3. Eigenvalues and Associated Percentages . . . . . . . . . . . . 76
4.4. Parameters used in the Gram-Schmidt Transformation . . . . . . . . 79
4.5. Covariance, Correlation and Transformation Matrices of
PCA for Vegetation . . . . . . . . . . . . . . . . . . . 81
4.6. Eigenvalues and Associated Percentages . . . . . . . . . . . . 81
4.7. Error Matrix of the classification . . . . . . . . . . . . . . . 83
5.1. The names and definitions of variables used from
Polystats database . . . . . . . . . . . . . . . . . . 90
5.2. The names, Definitions and formulas of Fuzzy Properties . . . . . . . 92
5.3 The descriptive statistics of Polystats 1952 . . . . . . . . . . . . 93
5.4. Correlations of Polystats variables 1952 . . . . . . . . . . . . . 94
5.5. The descriptive statistics of Polystats 1972 . . . . . . . . . . . . 96
5.6. Correlations of Polystats variables 1972 . . . . . . . . . . . . . 97
5.7. The descriptive statistics of Polystats 1984 . . . . . . . . . . . . 98
5.8. Correlations of Polystats variables 1984 . . . . . . . . . . . . . 99
5.9. The descriptive statistics of Polystats 1994 . . . . . . . . . . . 101
5.10. Correlations of Polystats variables 1994 . . . . . . . . . . . 102
5.11. The results of Paired Samples T test . . . . . . . . . . . . 103

xiv
5.12. The ANOVA table of Polystats variable . . . . . . . . . . . . 105
5.13. The Descriptive Statistics of Fuzzystats database cumulative periods . . 106
5.14. The Result of One-Sample Komogorov-Smirnov Test for Fuzzystats . . 106
5.15. The changes through time in Polystats database . . . . . . . . . 113
5.16. The re-grouping of distribution of activity variable in the photo database . 117
5.17. The nature and ranges of transferred attribute database . . . . . . 119
5.18. The correlation state of DISTFAULT and FAULTDENS variables . . . . 122
5.19 The comparison of whole data and landslide data . . . . . . . . . 122
5.20. The descriptive statistics of the ELEVMAP . . . . . . . . . . . 123
5.21. The descriptive stats of distance to drainage variable group . . . . . 124
5.22. The descriptive statistics of the DRAINAGE DENSITY . . . . . . . 125
5.23. The descriptive stats of distridge . . . . . . . . . . . . . . 126
5.24. The descriptive stats of aspect . . . . . . . . . . . . . . 126
5.25. The descriptive stats of Slope. . . . . . . . . . . . . . . 127
5.26. The descriptive stats of Distsettlement . . . . . . . . . . . . 128
5.27. The descriptive stats of Distpower, Dist road & Distroad+Distpower . . . 129
5.28. The descriptive stats of Dist_E-5 Road . . . . . . . . . . . 130
5.29. The % change of Landcover units . . . . . . . . . . . . . 131
6.1. The two dimensional matrix of LACTA . . . . . . . . . . . . 136
6.2. Two-dimensional matrix of 1952 and 1972 periods . . . . . . . . . 137
6.3. Two-dimensional matrix of 1972 and 1984 periods . . . . . . . . 137
6.4. Two-dimensional matrix of 1984 and 1994 periods . . . . . . . . 137
6.5. The percentiles of seed cells within each variable . . . . . . . . . 143
6.6. Methodological snapshot of used two methods. . . . . . . . . . 147
6.7. Weight values of the all available parameter classes . . . . . . . . 150
6.8. KMO and Bartlett’s test with initial 13 variables . . . . . . . . . 152
6.9. The Anti-Image matrices of initial 13 variables . . . . . . . . . . 153
6.10. The amount of total variance explained via factors . . . . . . . . 154
6.11. The rotated factor matrix and the variable loadings . . . . . . . . 155
6.12. KMO and Bartlett’s test after removal of two variables . . . . . . . 156
6.13. The amount of total variance explained via factors,
after removal of two variables . . . . . . . . . . . . . . 156
6.14. The rotated factor matrix and the variable loadings,
after removal of two variables . . . . . . . . . . . . . . . 157
6.15. The initial assumption of the variables . . . . . . . . . . 160
6.16. The final classification of logistic regression . . . . . . . . . . 160
6.17. The variables and their loadings after logistic regression . . . . . . 161
6.18 The densities of landslides among hazard classes of both methods . . . 164
6.19. The available combinations of re-coding process and their meanings . . 165

xv
 LIST OF FIGURES

FIGURES

1.1. The geographic setting of the study area. . . . . . . . . . . . . 4

1.2. The outline of Asarsuyu catchment with important reference locations . . . . 4
2.1. Graphical representation of hazard, vulnerability and risk. . . . . . . . 6
2.2. An overview of zonation activities . . . . . . . . . . . . . . . 7
2.3. The scales of analysis and minor details . . . . . . . . . . . . . 9
2.4. GIS and its related software systems as components of GIS . . . . . . 15
2.5. The phases of a GIS . . . . . . . . . . . . . . . . . . . 15
2.6. The questions of a well-built GIS should answer . . . . . . . . . . 16
3.1. Regional geological map of the study area and its environs . . . . . . . 40
3.2. The generalized columnar section of the study area . . . . . . . . . 45
3.3. The Geological map the study area . . . . . . . . . . . . . . 46
3.4. Lineament map of the Asarsuyu catchment . . . . . . . . . . . 50
3.5. Epicenter locations of Turkey greater than 5 in magnitude. . . . . . . . 52
3.6. Epicenter locations of mid-west NAFZ greater than 3.5 in magnitude . . . . 53
3.7. Earthquakes occurred in the nearby of the study area . . . . . . . . . 54
3.8. Surface ruptures of the 12 November 1999 Düzce earthquake . . . . . . 55
3.9. Some examples of landslides occurred after 12 Nov.1999 Düzce earthquake. . 56
4.1. LITHOMAP and the attached topological table of the study area. . . . . . 58
4.2. The area distributions of LITHOMAP . . . . . . . . . . . . . . 58
4.3. The FAULTMAP of Asarsuyu catchment . . . . . . . . . . . . 59
4.4. The rose diagram of FAULTMAP; a. Weighted, b Non-weighted . . . . . 60
4.5. The distance calculations in distance raster map production . . . . . . . 60
4.6. Distance raster image (DISTFAULT) produced from FAULTMAP. . . . . . 61
4.7. Fault density (FAULTDENS) of Asarsuyu Catchment. . . . . . . . . . 61
4.8. The color-coded CONTOURMAP . . . . . . . . . . . . . . . 62
4.9. Color coded DEM of Asarsuyu catchment. . . . . . . . . . . . . 63
4.10. Color draped relief model of Asarsuyu catchment. . . . . . . . . . 63
4.11. The drainage system of Asarsuyu Catchment . . . . . . . . . . 64

xvi
4.12. The distance raster of every pixel to the nearest drainage-line
(DISTDRAINMAP) . . . . . . . . . . . . . . . . . . . 65
4.13. The drainage density of Asarsuyu catchment . . . . . . . . . . . 65
4.14. The distance raster of every pixel to the nearest ridge-line
(DISTRIDGEMAP) . . . . . . . . . . . . . . . . . . . 66
4.15. Aspect map of the Asarsuyu catchment . . . . . . . . . . . . . 67
4.16. SLOPE map of Asarsuyu catchment . . . . . . . . . . . . . . 67
4.17. Historical infrastructure databases of Asarsuyu catchment . . . . . . . 69
4.18. Distance rasters and frequency distributions of 1994 period . . . . . . 70
4.19. a) True color composite of Landsat TM 5 (R=3, G=2, B=1), b) False color
composite of Landsat TM 5 (R=5, G=4, B=1) . . . . . . . . . . . 72
4.20. The methodological snapshot of land cover extraction scheme . . . . . 73
4.21. Decorrelation stretching results (R: decor_4, G: decor_3:B: decor_1) . . . 77
4.22. Near Infra Red versus Red band Raster Correlation Graph . . . . . . 79
4.23. Product of Maximum Likelihood Classification . . . . . . . . . . 82
4.24.a) Land-cover map of the study area, b) areal distributions of land cover . . 85
4.25. The landslide attribute database . . . . . . . . . . . . . . . 87
4.26. Landslide inventories of the four time periods. . . . . . . . . . . 88
4.27. Polygon grid midpoints and transferred attributes . . . . . . . . . 89
5.1. The distribution plots of area of 1952 . . . . . . . . . . . . . 93
5.2. Histograms of Polystats database . . . . . . . . . . . . . . 94
5.3. Scatter plots of highly correlated variable pairs of Polystats database (1952) . . 95
5.4. The frequency distributions of variables of 1972 Polystats database . . . . 96
5.5. Scatter plots of highly correlated variable pairs of Polystats database (1972) . . 97
5.6. The frequency distributions of variables of 1972 Polystats database . . . . 99
5.7. Scatter plots of highly correlated variable pairs of Polystats database (1984) . .100
5.8. The frequency distributions of variables of 1994 Polystats database . . . 101
5.9. Scatter plots of highly correlated variable pairs of Polystats database (1994) . 102
5.10. The distributions of Polystats variables in 4 periods . . . . . . . . 104
5.11. The Frequency distribution of cumulative “Form ratio” . . . . . . . 106
5.12. The frequency distributions of “form ratio” in all periods . . . . . . . 107
5.13. The Frequency distribution of cumulative “Grain Shape Index” . . . . 107
5.14. The frequency distributions of “form ratio” in all periods . . . . . . 108
5.15. The frequency distribution of cumulative “Compactness” . . . . . . 108
5.16. The frequency distributions of “compactness” in all periods . . . . . . 109
5.17. The frequency distribution of cumulative “Circularity 1” . . . . . . . 109
5.18. The frequency distributions of “Circularity 1” in all periods . . . . . . 110
5.19. The frequency distribution of cumulative “Circularity 2” . . . . . . . 110
5.20. The frequency distributions of “Circularity 2” in all periods . . . . . . 111

xvii
5.21. The Frequency distribution of cumulative “Elongation” . . . . . . . 111
5.22. The frequency distributions of “Elongation” in all periods . . . . . . 112
5.23.The Frequencies of the mass info variable through time and data table . . 115
5.24.The Frequencies of the type variable through time and data table . . . . 115
5.25.The Frequencies of the style variable through time and data table . . . . 116
5.26.The Frequencies of the depth variable through time and data table . . . 117
5.27. The Frequencies of the distribution of activity variable
through time and data table . . . . . . . . . . . . . . . 118
5.28. The Frequencies of the Land cover variable through time and data table . 119
5.29. The preferred lithologies and their percentages . . . . . . . . . 121
5.30. The frequency Distributions of DISTFAULT and FAULTDENS . . . . 123
5.31. The frequency distributions of ELEVMAP . . . . . . . . . . . 124
5.32. The frequency distributions of Distance to Stream map . . . . . . . 124
5.33. The frequency distributions of Drainage Density . . . . . . . . . 125
5.34. The frequency distributions of Distridge. . . . . . . . . . . . 126
5.35. The frequency distributions of Aspect . . . . . . . . . . . . 127
5.36. The frequency distributions of Slope . . . . . . . . . . . . 127
5.37. The frequency distributions of Distsettlement . . . . . . . . . . 128
5.38. The frequency distributions of Dist_ power+road . . . . . . . . . 129
5.39. The frequency distributions of Distance to E-5 Highway . . . . . . . 129
5.40. The percentage distribution of Landcover units . . . . . . . . . 131
6.1. The Thematic Landslide Attribute spatial distribution of massinfo
attribute of 1994 photo characteristics database . . . . . . . . . 133
6.2. The Thematic Landslide Attribute spatial distribution of type
attribute of 1994 photo characteristics database . . . . . . . . . 133
6.3. The Thematic Landslide Attribute spatial distribution of style
attribute of 1994 photo characteristics database . . . . . . . . . 134
6.4. The Thematic Landslide Attribute spatial distribution of depth
attribute of 1994 photo characteristics database . . . . . . . . . 134
6.5. The Thematic Landslide Attribute spatial distribution of activity
attribute of 1994 photo characteristics database . . . . . . . . . 135
6.6 The mechanism of the isopleth analysis . . . . . . . . . . 139
6.7. The isopleth map of the Asarsuyu catchment . . . . . . . . . . 140
6.8. The snapshot of methodology of percentile method and
reclassified parameter map production . . . . . . . . . . 142
6.9. The percentile maps of morphology of Asarsuyu catchment
with frequency and cumulative histograms . . . . . . . . . . 144
6.10. The percentile maps of lineament and density patterns of Asarsuyu
catchment, with frequency and cumulative histograms . . . . . . . 145

xviii
6.11. The percentile maps of infrastructure and distance to ridge of Asarsuyu
catchment, with frequency and cumulative histograms . . . . . . . 146
6.12 The steps through landslide susceptibility analysis . . . . . . . . 149
6.13. The hazard map and the amounts of landslides in each class as
a result of bivariate analysis . . . . . . . . . . . . . . 151
6.14. Eigenvalues of the factors. . . . . . . . . . . . . . . . 154
6.15. The positions of selected 4430 random landslide free nodes . . . . . 159
6.16. Observed Groups and Predicted Probabilities . . . . . . . . . 161
6.17. The hazard map and the amounts of landslides in each class as
a result of multivariate analysis . . . . . . . . . . . . . 163
6.18. The areal distributions of classified pixels . . . . . . . . . . 165
6.19. The locations of the misclassified pixels . . . . . . . . . . 166
6.20. The locations of the correctly classified pixels . . . . . . . . . . 166
6.21. The locations of the acceptable pixels . . . . . . . . . . . . 167
6.22. The locations of the not acceptable pixels . . . . . . . . . . 168
6.23. The locations of the correctly classified and the acceptable pixels united . 168
6.24. The final hazard map and the infrastructure of Asarsuyu catchment . . . 169
7.1. The elements of data production stage . . . . . . . . . . . . 171
7.2. Snapshot methodology for information transformation . . . . . . . 174
7.3. Components of data Evaluation stage . . . . . . . . . . . . 175
7.4. Components of Hazard Analysis . . . . . . . . . . . . . . 176
7.5. Components of Statistical analyses . . . . . . . . . . . . . 177
7.6. Components of Hazard Map Comparison . . . . . . . . . . . 180

xix
 CHAPTER 1

INTRODUCTION

1.1 Purpose and Scope

Every year, thousands of people all over the world are loosing their lives in
natural disasters, which are quite costly to the world economy both in humilian affairs
and in economical aspects. Natural disasters have major direct and indirect economic
and socio-economic effects in addition to the physical destruction that may occur.
Furthermore, the impact of these disasters are higher in developing countries than in
developed countries due to the fact that the developed countries had already created
disaster mitigation and disaster preparedness programs.
Disasters are natural hazard events in which a natural phenomenon or a
combination of natural phenomena such as earthquakes, mass movements, floods,
volcanic eruptions, tsunamis etc., can cause many loss of lives and damage to the
property. Almost no portion of the earth’s surface is free from the impact of natural
hazards. Due to increasing population densities and uncontrolled or not well-planned
development, especially in developing countries, more and more people are affected by
disasters. High precipitation, the deficiencies of infrastructure and lack of disaster plans
and hazard maps are paid off by human lives in Turkey.
Disasters have both immediate and long-term implications. Plans formulated for
disaster prone areas should therefore cover both these contingencies. It should also be
remembered that a disaster might initiate a chain of severe hazards in addition to the
direct impact damage.
Over the past two decades, many scientists have attempted to assess landslide
hazards and produced maps portraying their spatial distribution. However, up till now
there has been no general agreement on the methods or even on the scope of these
investigations (Brabb, 1984, Carrara, 1983, 1989, 1993).
Despite the methodological and operational differences, all methods proposed
are founded upon a single basic conceptual model. This model requires first the
identification and mapping of a set of geological-geomorphological factors which are

1
directly or indirectly correlated with slope instability. It then involves both an estimate of
the relative contribution of these factors in generating slope failure, and the classification
of the land surface into zones of different hazard degree (Carrara, 1993).
Experience gained from hundreds of surveys carried out in different parts of the
world has demonstrated that well trained investigators are able to detect and correctly
map many or most of the landslides occurring in an area by applying aerial-photograph
interpretation techniques and systematic field checks (Rib and Liang, 1978). However,
old, dormant landslide bodies, are generally intensively modified by farming activity or
covered by dense vegetation and thus, they cannot be readily identified and correctly
classified. This introduces a factor of uncertainty that cannot be readily evaluated and
explicitly incorporated in the subsequent phases of the analysis, being largely
dependent on the skill of the surveyor, and the quality and the scale of aerial
photographs and base-maps used. There, onwards the power of spatial data
manipulation and use of multivariate statistics in a GIS environment arouses.
A further step towards hazard zonation would require the identification of the
conditions leading to the slope failure, their systematic and consistent mapping, and
evaluation of their relative contributions to the mass movements in the area. In fact, the
causes of each slope failure are many, complex and sometimes unknown. Conversely
the geological - geomorphological factors, which are both relevant to the prediction of
landslide hazard that are mappable at an effective cost over a wide region are not as
many yet (Carrara, 1989).
All these methods have significant advantages and drawbacks. The main
advantages of the geomorphological approach lie in the capability of a skilful surveyor to
estimate actual and potential slope failure, taking into consideration a large number of
factors detected in the field or on the aerial-photographs. In addition, local or unique
slope instability conditions can be identified and assessed. The major drawback of the
approach concerns the high subjectivity characterizing all phases of the
geomorphological investigation. The degree of uncertainty associated to the different
phases of the mapping operations cannot be evaluated. Likewise it is difficult or
sometimes impossible to compare landslide hazard maps produced by different
surveyors.
The purpose of this thesis is to generate and evaluate a concise system of
landslide hazard assessment procedure. With the intention of this goal few assumptions
are to be made such as, the rainfall and earthquake are considered as triggering factors
at which the outcomes of the system proposed should be considered with these
triggering events. Furthermore, due to the small size of the area the effects of both
precipitation and earthquake will be uniform through the study area, due to this fact they
are not considered for the analyses.

2
 In landslide hazard analysis it is also aimed that, (a). data set should be easily
accessible and simple; (b). the system should be objective; (c). the system should be
data-driven with minimal expert interaction and (d). the system should be simple and be
easily implemented and used by novice users
The studies are initiated with an extensive review of the previous works carried
on Landslide Hazard Assessment, that is presented in Chapter 2. The definitions,
methodologies, contributions of remote sensing and review of available models in
geographic information systems are considered. The geology of Asarsuyu Catchment
constitutes Chapter 3 which is involved in regional geological setting, identification of
lithological units, the North Anatolian Fault Zone and the earthquakes of the region.
Types of input data and data production stage are given in Chapter 4. Every single
parameter map and the methodologies to create these maps are explained. A total 13
sets of parameter maps are produced and used. The foundations of landslide attribute
databases are also created in this chapter. Following the data production stages
landslide related databases are analyzed in detail in Chapter 5. The comparisons of
polystats and fuzzystats databases in 4 different periods, investigation of photo-
characteristics database and evaluation of the parameter map values of landslide
databases constitute the major issues. The core of this thesis is formed by the
investigation of the available and newly produced hazard assessment procedures,
which are presented in Chapter 6. Finally the results are discussed and conclusions are
given.

1.2. Geographical Setting

The study area is located in the northwestern part of Turkey, in the area
between Bolu and Düzce cities (Figure 1.1). The area is geographically defined as the
catchment of Asarsuyu stream (Figure 1.2). The extents of the coordinates of Asarsuyu
Catchment is defined as 4520500 N, 351500 E in the northwestern edge and 4506750
N, 372500 E in the southeastern edge in Universal Transverse Mercator projection with
36 North zone in European 1950 Mean Datum by International 1909/1924 / Hayford
1910 ellipsoid. The catchment is covered with six 1:25.000 scale topographical maps,
that are G26-a3, G26-b3, G26-b4, G26-c1, G26-c2 and G26-d2 sheets. The catchment
has an ellipsoidal form extending in east-west, having a maximum length of
approximately 20 kilometers and the approximate width is 10 kilometers, also covering
approximately 200 square kilometers.

3
Figure 1.1. The geographic setting of the study area.

Figure 1.2. The outline of Asarsuyu catchment with important reference locations.

4
 CHAPTER 2

BACKGROUND INFORMATION ON LANDSLIDE HAZARD AND USE OF REMOTE

SENSING AND GEOGRAPHICAL INFORMATION SYSTEMS IN LANDSLIDE
HAZARD ASSESSMENT

2.1. Definition of Landslide Hazard and Terminology

Mass movement is defined as “the outward and downward gravitational

movement of earth material without the aid of running water as a transporting agent” by
Crozier (1986), or “the movement of a mass of rock, debris or earth down a slope” by
Cruden (1991). These are the internationally accepted and most widely used definitions
of the phenomenon. Although they are slightly different from each other considering
beyond the scope of inclusion of water, they both point a mass transportation down the
slope in which a hazardous activity for humans can occur.
Rather than dealing with the types, activities and definitions, as they are defined
by the IAEG Commission on Landslides in the 1990’s, a more relational approach is
given by Soeters and van Westen (1996) “Slope instability processes are the product of
local geomorphic, hydrologic and geologic conditions; the modification of these by
geodynamic processes, vegetation, land use practices and human activities; and the
frequency and intensity of precipitation and seismicity”.
Mass movement or slope instability or landsliding are natural denudational and
degradational processes, unless they are affecting human life. Their interference with
ongoing human practices in the terrain makes it a landslide hazard. The general
accepted terminology in the world is followed below by Varnes (1984):

Natural hazard (H): The probability of occurrence of a potentially damaging

phenomenon within a specified period of time and within a given area (Figure 2.1).

Vulnerability (V): The degree of loss a given element or set of elements at risk resulting
from the occurrence of a natural phenomenon of a given magnitude. Scale is 0 (no
change) to 1 (total loss) (Figure 2.1).

5
Specific risk (Rs): The expected degree of loss due to a particular natural phenomenon.
It may be expressed by the product of H and V.

Rs=H*V

Elements at Risk (E): The population, properties, economic activities, including public
services and etc. at risk in a given area.

Total Risk (Rt): The expected number of lives lost, persons injured, damage to property
or disruption of economic activity due to a particular natural phenomenon (Figure 2.1). It
is expressed as
Rt=E*Rs

Figure 2.1. Graphical representation of hazard, vulnerability and risk

Based on the above definitions, hazard and risk information are generally
represented as discrete maps. The discrete classes represent equal probability classes,
which are inturn equal hazard or risk classes. The differentiation of hazard classes and
their groupings are called as “zonation”. The formal definition is as follows: “The term
zonation refers to the division of land into homogenous areas or user defined domains
and the ranking of these areas according to their degrees of actual or potential natural
hazards” (Varnes, 1984).
The natural hazard zoning/mapping constitutes the first and major task of the
earth scientists in natural hazard analysis (Figure 2.2). The zoning of a natural hazard is
the vital part of the study strategy in which the whole strategy will be based on. The
zonation activities are mutually dependent over some factors as shown in Figure 2.2.

6
These factors can be grouped into magnitude properties of the hazard, frequency of the
hazard and the spatial location of the hazard. The next step in hazard mapping is to
show the mapped hazard and to classify the hazard map into some homogenous areas
regarding the equal attributes of the hazard map.

Natural Hazard

Magnitude Frequency Location

Monitoring of Statistical/Historical
the hazard frequency analysis

Dimensions of
the hazard

Hazard Mapping

Zonation into
homogenous areas of

area affected low hazard

similar areas not yet affected high hazard
areas unlikely to be affected

Figure 2.2. An overview of zonation activities

The natural hazard zoning is controlled mainly by two factors, such as: the scale
of the zoning or mapping and the knowledge type used in the hazard zoning.

2.1.1. Scale Factor in Analysis

Before starting to any data collection, an earth scientist working on a hazard

analysis project should have to answer a number of interrelated questions:

1. What is the aim of the study?

2. What scale and with what degree of precision must the result be presented?
3. What are the available resources in the form of money, data and manpower?

7
 As the aim of the study would be previously defined, the scale and the precision
are the first parameters to be defined prior to the start of the project. Hence, the scale
factor should have to be determined at the first glance as it controls the type of the input
data, nature of the analysis and the output data of the project. The outcome precision is
also dependent on the scale chosen. Although the precision is dependent on scale, it is
also an independent parameter regarding the nature of the project. The necessary
adjustments should be made with the scale until the output precision and the desired
precision fulfills the project conditions. The resource analysis will be conducted after the
aim and scale is fixed.
The following scales of analysis, which were presented in the International
Association of Engineering Geologists (IAEG) Monograph on engineering geological
mapping (IAEG, 1976) can also be distinguished in general natural hazard zonation
(Figure 2.3).

2.1.1.1. National Scale (<1/1.000.000)

The national scale analysis is used only to outline the problem, give an idea
about the hazard types and affected hazard prone areas. They are prepared generally
for the entire country and the required map detail is very low, even in the best case
giving only data based on records in the form of an inventory. The degree of the hazard
is assumed to be uniform. These kinds of maps are generally prepared for agencies
dealing with regional (agricultural, urban or infrastructure) planning or national disaster
prevention / hazard assessment agencies.

2.1.1.2. Regional/Synoptic Scale (< 1/100.000)

The scale is still so small to use in any quantitative method, but these maps are
used for regional planning and in early stages of region wise planning activities. The
areas to be investigated are still large in the order of thousands of square kilometers
and the map detail is low again. Only simple methods are used with qualitative data
combination and the zoning is primarily based on regional geomorphological Terrain
Mapping Units / Complexes (TMU) or dependent on regional geological units.

2.1.1.3. Medium Scale (1/25.000 - 1/50.000)

These hazard maps are made mainly for agencies dealing with intermunicipal
planning or companies dealing with feasibility studies for large engineering works. The
areas to be investigated will have areas of several hundreds of square kilometers. At

8
this scale considerably more detail is required than at the regional scale. These maps
do serve especially for the choice of corridors for infrastructure construction or zones for
urban development. Statistical techniques are dominantly used in this scale.

1.National Scale
<1:1.000.000
Inventory
Entire Country
Based on Records

2. Regional Scale
<1:100.000
Regional Planning
Large Areas
Simple Methods

3. Medium Scale
1:25.000 to 1:50.000
Local Planning
Areas up to 200 sqkm
Statistical Methods

4. Large Scale
> 1:10.000
Detailed Planning
Small areas
Stability analysis

Figure 2.3. The scales of analysis and minor details

2.1.1.4. Large Scale (> 1/10.000)

These hazard maps are produced generally for authorities dealing with detailed
planning of infrastructure, housing or industrial projects or with evaluation of risk within a
city or within a specified project area. They cover very small areas hence the
deterministic hazard analyses become available to be used. The detail level of the maps
is set into maximum. They are based on physical numerical models that require
extensive data collection in the field and laboratory surveys.

9
2.1.2. Knowledge Type Used

Prediction of landslide hazard for areas not currently subject to landslide hazard
is based on the assumption that hazardous phenomena that have occurred in the past
can provide useful information for prediction of future occurrences. Unlike general
educational geological phrases in this case “present is not key to the past but present
and past are the keys of future”, of which the real value of engineering and its futuristic
approaches are represented. Therefore, mapping these phenomena and the factors
thought to be of influence is very important in hazard zonation. In relation to the analysis
of the terrain conditions leading to slope instability, two basic methodologies can be
recognized (van Westen, 1993):

1. The first mapping methodology is the experience-driven applied-geomorphic

approach, by which the earth scientist evaluates direct relationships between
landslides and their geomorphic and geologic settings by employing direct
observations during a survey of as many existing landslide sites as possible.
This is also known as direct mapping technology.

2. The opposite of this experience-based, or heuristic approach is the indirect

mapping methodology, which consists of mapping a large number of
parameters considered to potentially affect landsliding and subsequently
analyzing (statistically) all these possible contributing factors with respect to the
occurrence of slope instability phenomena. In this way the relationships
between the terrain conditions and the occurrence of the landslides may be
identified. On the basis of the result of this analysis, statements are made
regarding the conditions under which slope failures occur.

Another division of techniques for assessment of slope instability hazard was

given by Hartlen and Viberg (1988), who differentiated between relative and absolute
hazard assessment techniques.

1. Relative hazard assessment techniques differentiate the likelihood of

occurrence of mass movements for different areas on the map without giving
exact values.
2. Absolute hazard maps display an absolute value for the hazard such as a factor
of safety or a probability of occurrence.

10
 Furthermore the hazard assessment techniques can also be divided into three
broad classes based on use of statistical methods (Carrara, 1983; Hartlen and Viberg,
1988; Soeters and van Westen, 1996).

1. White box models: based on physical models (slope stability and hydrologic
models), also referred to as deterministic models
2. Black box models: not based on physical models but strictly on statistical
analysis
3. Gray box models: partly based on physical models and factual data and partly
on statistics.

2.2. Use of Remote Sensing in Landslide Hazard Assessment

The phenomenon, landslide is affecting the earth’s surface, hence it also falls in
to the research and application areas of both aerial and space born remote sensing.
The nature of this phenomenon as it is occurring at the surface of earth lets earth
scientists to exploit this fact using remotely sensed data. On the other hand, again the
nature of this phenomenon limits the applications, as being dynamic and sometimes
being quite small in terms of conservative remote sensing language. Furthermore they
reveal very small information when they are observed in planar 2-D, however, they
contain large amounts of data when explored in 3-D. Basing on this fact the use of
stereo-remote sensing products seems to be indispensable, which reveals the true
morphodynamical features of the landslides. These information are providing the
diagnostic information regarding the type of the movement (Crozier, 1973). The general
application fields of remote sensing in landslide business are monitoring the change of
landslide activities through time (change detection) and mapping out where the hazard
occurs.
Plenty of researchers have tested the usage of remote sensing products
through the last 30 years. Two major groupings could be made upon the investigation of
these researches. These are aerial photography and space-born sensor images.
Numerous applications have been carried out which generally define the
landslide areas. Chandler and Moore (1989), Chandler and Brunsden (1995) and
Fookes et al. (1991) give excellent applications for photogrammetry. For single
landslides and for smaller areas, a monitoring scheme is best applied with this
technique with large accuracies. In opposition, the applicability of this technique limits
the extents of the interest area as the larger areas could be accomplished by classical
aerial photographical studies easily. For stereographical aerial photography Rengers
(1986), Sissakian (1986) and Mollard (1986) could be counted as single application

11
manuscripts. However, studies with landslides and aerial photographs are as old as the
applications of first stereographical aerial photographs, resulting in plenty of textbooks
and textbook sections.
The landslide information extracted from the remotely sensed products is mainly
related with the morphology, vegetation and the hydrological conditions of the slope.
The slope morphology is best examined with stereographical coverages. Generally the
identification of the slope instabilities are indirect, they are identified by associated
elements with slope instability process. The advantages of aerial photographs can be
listed as:

1. They provide quite older coverages before digital world starts

2. The flight coverages are flexible for new missions
3. The spatial and temporal resolution are very high
4. Stereoscopic coverage
5. Most of the geoscientist are familiar
6. Every country have at least one full coverage of their land due to military
reasons

The disadvantages are as follows:

1. Low spectral resolution

2. The nature of photograph as hardcopy
3. Presence of distortions in the images
4. Absence of coordinate information
5. Orthorectification is needed to remove distortion and add coordinate information
6. The resultant map is dependent to the experience of interpreter

The applications with space born images are quite new compared to the others.
Furthermore, they are generally defining the landslides indirectly by mapping out other
parameters such as land cover. Some examples from the literature could be said of
Gagon (1975); Mc Donalds and Grubbs (1975); Sauchyn and Trench (1978); Stephens
(1988); Huang and Chen (1991) and Vargas (1992).

In comparison to the aerial photographs, the advantages of satellite images are:

1. Getting the bigger picture

2. Larger spectral range
3. Easily accessible
4. No distortion
5. Only georeference is needed to transfer the coordinates

12
The disadvantages are:

1. Low spatial resolution

2. More expensive than aerial photographs of the same resolution
3. Limited stereo capability
4. Limited number geoscientists are familiar

Although there are plenty of disadvantages of aerial photographs, they are the
most frequently used medium in landslide projects as they have cheaper high-resolution
images. The spatial resolution nearly controls everything in landslide hazard
assessment. The comparison of spatial resolution of photographic and non-
photographic remote sensing product requires the concept of Ground Resolution Cell
(GRC) defined by Strandberg (1967) and introduced to landslides first by Rengers et al.
(1992). Strandberg (1967) suggested that the formula for GRC is in relation to scale as:

GRC=S/1000R

Where GRC is ground resolution cell in meters, S in the scale number of

photograph and R is the resolution of photographic system (line pairs/mm, normally 40
in conventional systems)
Soeters and van Westen (1996) figured out the necessary minimum number of
GRC’s namely the pixels in the images, in order to identify and interpret the landslides,
which is presented in Table 2.1. They also exploited this information and created a
comparison table of photographic images with non-photographic ones that is also
presented in Table 2.2.

Table 2.1. The number of GRC needed to identify and interpret object of varying
contrast in relation to its background (Soeters and van Westen, 1996).

The number of GRC

For identification For interpretation

Extreme contrast 20-30 40-50

High contrast 80-100 120-140

Low Contrast 1000-1200 1600-2000

13
Table 2.2. Minimum object size needed for landslide Identification or Interpretation
(Soeters and van Westen, 1996).

2
Size m needed for

High Contrast Low Contrast

GRC size (m) Identification Interpretation Identification Interpretation

Landsat MSS ~80 160000 288000 7040000 11520000

Landsat 5 TM 30 22500 40500 990000 1620000

Spot Multispectral 20 10000 18000 440000 720000

Spot Panchromatic 10 2500 4500 110000 180000

Aerial Photographs

1:50000 1 25 45 1100 1800

1:15000 0,3 6,5 11,5 300 450

Basing on the above facts, aerial photography is still indispensable in the

landslide hazard zonation activities. However, remote sensing is depicting an important
role in landslide hazard assessment, though, this role is not the primary role in the
game.

2.3. Geographical Information Systems and Landslide Hazard Assessment

A GIS is defined as a “powerful set of tools for collecting, storing, retrieving at

will, transforming, and displaying spatial data from the real world for particular set of
purposes” (Burrough, 1986). A more specific definition is given by Bonham-Carter
(1996) as follows: “a geographic information system, or simply GIS, is a computer
system for managing spatial data. The word geographic implies that the locations of the
data items are known, or can be calculated, in terms of geographical coordinates. The
word information implies that the data in GIS are organized to yield useful knowledge,
often as colored maps and images, but as also statistical graphics, tables and various
on-screen responses to interactive queries. The word system implies that a GIS is made
up from several interrelated and linked components with different functions. Thus, GIS
has functional capabilities for data capture, input, manipulation, transformation,
visualization, combination, query, analysis, modeling and output.”
These international valid definitions of GIS directly oppose to the belief that GIS
is only a CAD software or only a drawing tool. CAD can only constitute a small portion of
the whole integrated system, in which an ideal GIS and its possible components are
shown in Figure 2.4.

14
Figure 2.4. GIS and its related software systems as components of GIS (modified from
Bonham-Carter, 1996)

Generally a GIS consists of the following phases (Figure 2.5).

Figure 2.5. The phases of a GIS

A GIS if based over the former components should answer the following
questions (Figure 2.6):

More and more the products of mapping and inventory are being stored in data
banks for their ultimate retrieval or combination with data from other sources. Often they

15
are incorporated is GIS or LIS (Land Information Systems) which serve as a base for
programmable data manipulation and selective information extraction for planning and
project assessment.

Figure 2.6. The questions of a well-built GIS should answer.

The development of GIS and LIS systems is of considerable interest in the

context of satellite surveying, change detection and monitoring. The flexibility of digital
data processing, combined with quick input of new data (possible from updating on the
basis of satellite remote sensing records) offers new possibilities to the surveyor,
cartographer and planner.
It is clear that in a rapidly developing society, change detection is of great
importance. In modern society, mapping suffers from high rate of change: change in
land use in rural and urban areas; change in requirements for maps and inventories;
change in concepts in the various disciplines of earth and social sciences, leading to
different interpretations of the same data and change in the economical and technical
factors on which mapping methods were based.
In order to refine the discussion around landslide hazard one can say that, the
occurrence of slope failure depends generally on complex interactions among a large
number of partially interrelated factors. Analysis of landslide hazard requires evaluation
of the relationships between various terrain conditions and landslide occurrences. An
experienced earth scientist has the capability to mentally assess the overall slope
conditions and to extract the critical parameters. However, an objective procedure is
often desired to quantitatively support the slope instability assessment. This procedure

16
requires the evaluation of the spatially varying terrain conditions as well as the spatial
representation of landslides. A GIS allows for the storage and manipulation of
information concerning the different terrain factors as distinct data layers and thus
provides an excellent tool for slope stability hazard zonation.

The advantages of GIS for assessing landslide hazard include the followings:

1. A much larger variety of hazard analysis techniques become attainable.

Because of the speed of calculation, complex techniques requiring a large
number of map overlays and table calculations become feasible.
2. It is possible to improve models by evaluating their results and adjusting the
input variables. Users can achieve maximum results by a process of trial and
error, running the models several times, whereas it is difficult to use these
models even once in the conventional manner. Therefore, more accurate
results can be expected.

The disadvantages of GIS for assessing landslide hazard include the followings:

1. A large amount of time is needed for data entry. Digitizing is especially time
consuming
2. There is a danger in placing too much emphasis on data analysis as much as
the expense of data collection and manipulation based on professional
experience. A number of different techniques of analysis are theoretically
possible, but often the necessary data are missing. In other words, the tools are
available but cannot be used because of the lack or uncertainty of the data.

2.3.1. Phases of Natural Hazard Analysis in GIS

The following phases can be distinguished in the process of a hazard analysis

using GIS (van Westen, 1993); there is a logical order in the sequence though
sometimes they may be overlapping. The time schedule of these phases is listed in
Table 2.3.

1. Choice of the working scale and the methods of analysis which will be applied
2. Collection of existing maps and reports with relevant data
3. Interpretation of images and creation of new input maps
4. Design of the database and definition of the way in which the data will be
collected and stored.

17
 5. Fieldwork to verify the photo-interpretation and to collect relevant quantitative
data
6. Digitizing of maps and attribute data
7. Validation of the entered data
8. Manipulation and transformation of the raw data to a form which can be used
in the analysis
9. Analysis of data for preparation of hazard maps.
10. Evaluation of the reliability of the input maps and inventory of the errors
which may have occurred during the previous phases
11. Final production of hazard maps and adjoining report

Table 2.3. Time schedule comparison of phases of landslide hazard assessment of

conventional methods and GIS based methods based on scale (numbers are in
percents of the total project time) (van Westen, 1993).

Regional Scale Medium Scale Large Scale

PHASES
Conventional GIS Based Conventional GIS Based Conventional GIS Based
Methods Methods Methods Methods Methods Methods

1. Choice of scale and methods <5 <5 <5 <5 <1 <5

2 Collection of existing data <5 <5 <5 <5 8 8

3. Image Interpretation 50 50 30 30 10 20

4. Database design 0 <5 0 <5 0 <5

5. Fieldwork <5 <5 7 7 10 20

6. Data Entry 0 20 0 30 0 15

7. Data Validation 0 <5 0 5 0 5

8. Data Manipulation 0 <5 0 5 0 5

9. Data Analysis 30 10 48 10 61 10

10. Error Analysis 0 <5 0 <5 0 <5

11. Final Map Production 10 <5 10 <5 10 <5

2.3.2. GIS Based Landslide Hazard Zoning Techniques

An ideal map of slope instability hazard should provide information on the

spatial probability, temporal probability, type, magnitude, velocity, runout distance and
retrogression limit of the mass movements predicted in a certain area (Hartlen and
Viberg, 1988). A reliable landslide inventory defining the type and activity of all
landslides, as well as their spatial distribution, is essential before any analysis of the
occurrence of landslides and their relationship to environmental conditions are
undertaken. Even the inventory of historical periods are of great use in the final
analyses. The differentiation of slope instability according to type of movement is

18
important, not only because different types of mass movement will occur under different
terrain conditions, but also because the impact of slope failures on the environment has
to be evaluated according to type of failure.
2.3.2.1. Trends in Landslide Hazard Zonation

A large amount of research on hazard zonation has been done in last 30 years,
as the consequence of an urgent demand for slope instability hazard mapping.
Overviews of the various slope instability hazard zonation techniques can be found in
Hansen (1984), Varnes (1984), Hartlen and Viberg (1988). The general trends in
landslide hazard zonation are given in Table 2.4. The distribution analyses and
qualitative analyses are generally used for very large areas with very low detail such as
national hazard maps. The deterministic and frequency analyses are used generally for
very small areas such as specific large engineering projects like dams, nuclear power
plants, highway strips, open pit mine slopes and spoils. Monitoring and laboratory
analyses are indispensable for these analyses. Good reviews of these initial
deterministic methods can be found in Lambe and Whitman (1969), Hoek and Bray
(1981), Graham (1984), Bromhead (1986) and Anderson and Richards (1987). The
limited GIS examples of these methods could be cited as Ward et al. (1982), Okimura
and Kawatani (1986), Mulder and van Asch (1988), Mulder (1991) and Hammond et al.
(1992). The statistical analyses have the most flexibility in scale and in data type and will
be investigated in detail in the following sections.

Table 2.4. The trends in landslide hazard zonation (van Westen, 1993).

Type of landslide hazard analysis Main characteristics

Direct mapping of mass movement features resulting in a
A. Distribution analysis map, which gives information only for those sites where
landslides have occurred in the past.
Direct, or semi-direct, methods in which the geomorphological
map is re-numbered to a hazard map, or in which several
B. Qualitative analysis maps are combined into one using subjective decision rules,
based on the experience of the earth scientist.
Indirect methods in which statistical analyses are used to
C. Statistical analysis obtain predictions of the mass movement hazard from a
number of parameter maps.
Indirect methods in which parameter maps are combined in
D. Deterministic analysis slope stability calculations.
Indirect methods in which earthquake and/or rainfall records or
hydrological models are used for correlation with known
E. Landslide frequency analysis landslide dates, to obtain threshold values with a certain
frequency.

19
2.3.2.2 Direct Mapping in Landslide Hazard Analysis

2.3.2.2.1 Landslide Distribution Analysis

The most straightforward approach to landslide hazard zonation is a landslide

inventory, based on any or all of the following; aerial photo interpretation, ground
survey, and a database of historical occurrences of landslides in an area. The final
product gives the spatial distribution of mass movements, which may be represented on
a map either as affected areas to scale or point symbols (Wieczorek, 1984).
Such mass movement inventory maps are the basis for most other landslide
hazard zonation techniques. However, they can be used as an elementary form of
hazard map because they display the location of a particular type of slope movement.
They provide only information for the period shortly preceding the date that aerial
photographs are taken or the fieldwork conducted. They provide no insight into temporal
changes in mass movement distribution. Many landslides occurred some time before
photographs are taken my have become undetectable. Therefore, a refinement is the
construction of landslide activity maps, which are based on multitemporal aerial photo
interpretation (Canuti et al., 1979, 1985, 1986). Landslide activity maps are
indispensable to study the effects of temporal variation of a factor such as land use on
landsliding. Landslide distribution can also be shown in the form of a density map
(Wright et al, 1974). The resulting density values are interpolated and used as landslide
isopleths. They can also be used to cite out the current situation of the landslide density
per terrain mapping unit or catchment or a predefined geological unit. This method may
also be used to test the importance of each individual parameter for predicting the
occurrence of mass movements. If the method is used to test the importance of specific
parameter classes, the user decides, on the basis of his/her field experience, which
individual maps or combination of parameter maps will be used. The method is most
appropriate at medium or large scales. At the regional scale the construction of a mass
movement distribution or activity map is very time consuming and too detailed for
procedures of general zoning.
However, the selection of terrain mapping unit and the conversion of continuous
parameter maps into discrete parameter maps involve a quite large subjectivity into the
analysis. Furthermore, this analysis should have to be done for each parameter map
and for different parameter classes. The effects of the separate parameters with respect
to each other are not implemented, hence it is still expert dependent who will be on the
charge to define the parameter classes and the parameter maps to be used. The
analysis is similar to general bivariate analyses, but does not end up with a hazard

20
score of any sampling frame in the area. Basically, a simple density per kilometer or per
sample area will provide much more objective results about the factual data.

2.3.2.2.2 Heuristic Approach (Geomorphic Analysis)

In heuristic methods the expert opinion of the earth scientist making the survey
is used to classify the hazard. These methods combine the mapping of mass
movements and their geomorphologic setting as the main input factor for hazard
determination.
The basis of geomorphic analysis was outlined by Kienholz (1977), who
developed a method for producing a combined hazard map based on the mapping of
“silent witnesses” (Stumme Zeugen). The geomorphic method is also known as the
direct mapping method. The hazard is determined directly either in the field or by photo
or satellite image interpretation by the earth scientist. The process is based on individual
experience and the use of reasoning by analogy. The decision rules are, therefore,
difficult to formulate because they vary from place to place, yielding as unformalized
applicable rules that vary from polygon to polygon. This method is totally subjective and
dependent on the skill and experience of the earth scientist. However, GIS serves as an
undeniable tool for reproduction and querying the entered data. This method can be
applied at all scales in a relatively short period. Some examples of geomorphic analyses
can be found in Carrara and Merenda (1974), Brunsden et.al. (1975), Stevenson (1977),
Malgot and Mahr (1979), Kienholz (1977,1978,1980,1984), Kienholz et al. (1983,1988),
Grunder (1980), Ives and Messerli (1981), Rupke et al. (1987,1988), Perrot (1988),
Hermelin (1990,1992), Hearn (1992) and Seijmonsbergen (1992). A weighting scheme
is also present in this type of analysis, however this weighting scheme is also quite
subjective and “blind weighting” is suggested for this type of weighting by Gee (1992).

2.3.2.3. Indirect Mapping in Landslide Hazard Analysis

2.3.2.3.1. Statistical Methods in Landslide Hazard Analysis

Aiming at a higher degree of objectivity and better reproducibility of the hazard

zonation, which is important for legal reasons, statistical techniques have been
developed for assessment of landslide hazard.
In statistical landslide hazard analysis the combinations of factors that have led
to landslides in the past are determined statistically, and quantitative predictions are
made for areas currently free of landslides but where similar conditions exist.
Furthermore, overlying of parameter maps and calculation of landslide densities form

21
the core of the analysis. Most of the analyses are based on the relationship between the
landslide densities per parameter class compared with the landslide density over the
entire area. Each method has its own specific rules for data integration required to
produce the total hazard map. Two different statistical approaches are used in landslide
analyses: bivariate and multivariate approaches.
Although the statistical techniques can be applied at different scales, their use
becomes quite restricted at the regional scale, where an accurate input map of landslide
occurrences may not be available and where most of the important parameter cannot be
collected with appropriate accuracy. At large scales, different factors will have to be
used, such as water table depth, soil layer sequences and thicknesses. These data are
very difficult to obtain even for relatively small areas. Therefore, the medium scale is
considered most appropriate for this technique.

2.3.2.3.1.1. Bivariate Statistical Methods in Landslide Hazard Analysis

In this method, overlay of parameter maps and calculations of landslide

densities form the core of the analysis, the importance of each parameter, or specific
combinations of parameter can be analyzed individually. Using normalized values
(landslide density per parameter class in relation to the landslide density over the whole
area), a total hazard map can be made by addition of the weights for individual
parameters. The weight values can also be used for design decision rules, which are
based on the experience of the earth scientist. It is also possible to combine various
parameter maps into a map of homogenous units, which is then overlaid by the
landslide map to give a density per unique combination of input parameters.
It should be stressed that the selection of parameters has also an important
subjective element in this method. The following GIS procedures are used (van Westen,
1993).

1. Classification of each parameter map into a number of relevant classes.

2. Combination of the selected parameter maps with landslide map via map
overlay
3. Calculation of weighting values based on the cross table data
4. Assignment of weighting values to the various parameter maps or design of
decision rules to be applied to the maps, and classification of the resulting
scores in a few hazard classes.

As it is seen from the procedure list the first and the last item contains quite
large subjectivity, it is not clear to how to divide the parameter maps into classes and

22
how many classes should there be? Furthermore, the division of the final hazard map
into hazard classes inherits the same problem. This problem limits these methods, as
the start and the end directly depends on the expert, which means a degradation in the
final hazard map and also limits the reproducibility of the hazard maps under different
conditions.
The medium scale is most appropriate for this type of analysis. The method is
not detailed enough to apply at the large scale, and at the regional scale the necessary
landslide occurrence map is difficult to obtain.
Bivariate statistical analysis deals with one dependent variable (in this case the
occurrence of mass movements) and one independent variable. The importance of each
factor is analyzed separately. Specific combinations of variables can also be tested by
treating the combination map as a new variable. The methods are based on the
assumption that the important factors leading to mass movements can be quantified by
calculating the density of mass movements for each variable class. However, the new
parameter map production as crossing the available parameter maps in fact carries this
bivariate analysis procedure into somewhat multivariate domain; as the factor analysis
in the multivariate domain also bases its core on new parameter maps with different
factor loadings from the initial parameter maps.
In bivariate statistical analysis, each factor map is combined with the landslide
distribution map, and weighting values based on landslide densities are calculated for
each parameter class. Several statistical methods have been applied to calculate
weighting values; these have been termed the landslide susceptibility method (Brabb,
1984; van Westen, 1992, 1993), Information value method (Kobashi and Suzuki, 1988;
Yin and Yan, 1988), weight of evidence modeling method (Spiegelhalter, 1986;
Bonham-Carter, 1996). Furthermore, there still exist not enough exploited methods as
Bayesian combination rules, certainty factors, Dempster and Shafer belief method and
fuzzy logic. The three of the methods will be further investigated in the following
sections, which are landslide susceptibility analyses, information value analyses and
weights of evidence methods. The first two of them depend on the density of landslides
in parameter classes; even though they result in approximately same hazard maps, the
calculation schemes are different from each other. The weights of evidence method
utilize the usage of binary dumb variables, and again based on the probability of
occurrence of landslides in parameter classes. However, the usage of binary dumb
variables turns a simple GIS in to a small scale chaos especially in landslide hazard
projects. As this method is first created to assess the locations of ore deposits, the
variables of its initial version were well defined and directly depend on factual data,
where all of the answers of “what if?” questions were known.

23
2.3.2.3.1.1.1. Landslide Susceptibility Analysis

A simple and useful method in statistical analysis to determine the importance

of different variables for the occurrence of mass movements is the use of pair wise map
crossing. In order to evaluate the importance of the individual maps, a cross between
these maps and a landslide occurrence map is prepared. For each variable class and
landslide type, two types of densities can be calculated.

1. Area density: the density expressed as the number of pixels with landslides divided
by the total number of pixels within the variable class. This can be displayed as a
percentage or permillage contents.

Npix ( SX i )
Darea = 1000
Npix ( X i )

where Darea : Areal density per millage

Npix( SX i ) : number of pixels with mass movements within variable
class Xi.
Npix ( X i ) : number of pixels within variable class Xi.

2. Number density: the density expressed as the number of landslide occurrences per
square kilometer of the area of the variable class.

1 * 106
Dnumber = Number ( SX i )
Area ( X i )

where DNumber : Number Density (Number/km2)

Area ( X i ) : Area in square meters of variable class Xi.
Number ( SX i ) : Number of mass movements within variable class Xi.

To evaluate the influence of each variable, weighting factors should have to be

introduced, which compare the calculated density with the overall density in the area.
The formula for the density-based area is:

Warea = 1000
Npix ( SX i )
− 1000
∑ Npix( SX i )
Npix ( X i ) ∑ Npix( X i )

24
and for the density based on number/km2

1 * 106 1 * 106
Wnumber = Number ( SX i ) − ∑ Number ( SX )
Area ( X i ) ∑ Area ( X ) i
i

2.3.2.3.1.1.1.1. Production of the Susceptibility Map

The weight values for the variable classes are added to produce a hazard map.
With the number of input maps and different combinations of type and activity, a number
of different susceptibility maps can be made. The optimal combination of variables is
generally a problem, however selection of a small set of maps incorporate the most
relevant variables. Two methods have been applied in literature. First method is
selection of maps based on field experience in which the variables that are considered,
on the basis of field experience, to be relevant for the occurrence of mass movements
are selected and summed. The other method is called the stepwise map combination,
adding the various input maps one by one. After the addition of another map, the
resulting scores are analyzed by crossing with the map showing active landslides. The
percentage of pixels with landslides and a total score larger than zero is calculated
(correctly classified pixels). If this percentage decreases after the addition of another
map, such a map is rejected. If the percentage increases, the map is included (van
Westen, 1993). However, the sequence of this summation changes everything which is
a major drawback of this method and needs to be justified.

2.3.2.3.1.1.2. Information Value Method

The use of a combination of numerical variables (such as slope angle values)

and alphanumerical variables (such as lithological variables) in a statistical analysis is
generally problematic. This can be solved by treating each variable class as a separate
variable, which can only one of the two states: present (1) or absent (0). It can be
determined whether a variable class is present or absent. The information value method
can be applied both to land units as well as on a pixel basis. The hazard information
method, developed by Yin and Yan (1988) is based on the following simple formula for
calculating the information value Ii for variable Xi:

Si N i
I i = log
S N

25
 where:
Si: the number of land units or pixels with mass movements and the presence of
variable Xi,
Ni: The number of land units or pixels with variable Xi
S: The total number of land units or pixels with mass movements
N: The total number of land units or pixels.

The degree of a hazard for a land unit or pixel j is calculated by the total
information value Ij

m
I j = ∑ X ij I j
i =0

where:
m : number of variables,
Xij : 0, if the variable Xi is not present in the land unit or pixel j and 1, if the
variable is present.

For the assessment of precision of the classification, Yin and Yan (1988)
presented the following equation

Mi  M − Mi 
A= 3 1 − 
Ni  N − Ni 

in which:
A: precision of the predicted result
N: total number of terrain units (catchments in this case) in this area
N i: total number of units with landslides
M: number of terrain units predicted as unstable
Mi: number of terrain units predicted as unstable which have landslides.

The information value method applied on a pixel basis is in fact very similar to
the susceptibility determination. The only difference is that in the information value
method the log value of the quotient of class density over map density is entered,
whereas in the susceptibility method the difference in densities was used. The
information values are always smaller than the weight values.

26
2.3.2.3.1.1.3. Weights of Evidence Modelling

This method was developed at the Canadian Geological Survey (Agterberg et

al., 1990; Bonham-Carter et al, 1990) and was applied to the mapping of mineral
potential. Sabto (1991) applied the method for landslide hazard analysis. The method
consists of reducing each set of landslide-related factors on a map to a pattern of a few
discrete states. In its simplest form, the pattern for a feature is binary, representing its
presence or absence within a pixel. According to Bonham-Carter et al. (1990), the first
step is determining the prior probability of landslides, which is given by the density of
pixels with landslides within the study area.

Npix ( slide)
Pprior =
Npix (total )

in which
Pprior : prior probability,
Npix (slide) : the number of pixels with a landslide occurrence,
Npix (total) : the total number of pixels in the map

for mathematical reasons it is more convenient to use the odds (O):

Pprior Npix ( slides)

O prior = =
1 − Pprior Npix (total ) − Npix ( slides)

Considering the relationship between a binary variable map (bi) and a landslide map
(S), the following combinations are possible:

Bi : Npix( Bi ) Npix( total )

Bi : Npix( total ) − Npix( Bi ) Npix( total )

four combinations of Bi and S are possible in the map: Bi I S , Bi I S , Bi I S ,

Bi I S .

The conditional probability of choosing a pixel with a landslide, given that the
cell contains pattern Bi, is:

27
 { }
P S Bi =
Bi I S
Bi

and the three other conditional probabilities are:

{ }
P S Bi =
Bi I S
Bi
{ }
P S Bi =
Bi I S
Bi
{ }
P S Bi =
Bi I S
Bi

According to Bayes rule:

P{Bi S} P{ S}
{ } = { P{B} }
P B S P{ S }
{ }
P S Bi =
P{Bi }
, P S Bi
i

+
Bonham-Carter et al. (1990) defined positive and negative weights ( Wi and

Wi − ), which combine these conditional probabilities:

Wi + = log e
P{Bi S }
Wi − = log e
{
P Bi S }
{ } { }
and
P Bi S P Bi S

In GIS the method can be implemented rather easily. It is considered as the

simple crossing of a binary landslide map with a binary variable map. The four possible
resulting combinations are given Table 2.5.below.

Table 2.5. The possible combinations after map crossing

Variable Class represented as binary pattern

LANDSLIDES
1 (present) 0 (absent)
Present 1 Npix1 Npix2
Absent 0 Npix3 Npix4

28
The weights of evidence can be written in numbers of pixels as follows:

Npix1 Npix2
Npix1 + Npix2 Npix1 + Npix2
Wi + = log e Wi − = log e
Npix3 Npix4
Npix3 + Npix4 Npix3 + Npix4

If more binary maps are used, the weights can be added, provided that the
variable maps are conditionally independent with respect to landslide occurrence. The
logarithm of the posterior odds can be calculated as follows:

{ }
n
log e O S B1k I B2k I B3k ........ Bnk = ∑ Wi k + log e O prior { S}
i =1

and the posterior probability as:

O
P{ S } =
(1 + O)

The contrast C=W+ -W- gives a useful measure of the correlation between the
variable map and the landslide occurrence. C becomes zero when a map has a
distribution which is spatially independent of the points.
The main assumption for univariate statistical methods is that the maps should
be conditionally independent. To test this independence a pairwise test can be executed
(Bonham-Carter et al, 1990). All possible pairs of variable maps should be evaluated
separately. The pairwise test includes the calculation of observed and expected
frequencies of landslides. Therefore, the maps are crossed pairwise, and the resulting
cross map is then crossed again with the mass movement map. The combinations
obtained from crossing two binary maps and a landslide map is given in the Table 2.6.

Table 2.6. The possible combinations after crossing of two binary maps
possible combinations of binary maps
Landslides
Bi I B2 B1 I B2 Bi I B 2 Bi I B 2
Present Npix1 Npix 2 Npix 3 Npix 4
Absent Npix 5 Npix 6 Npix 7 Npix 8

29
 Using the weight of evidence modeling, the logarithm of the odds for each
unique overlap of two variable classes, is calculated by:

( )
log e O S B1 B2 = W1+ + W2+ + log e O( S )

( )
log e O S B1 B2 = W1+ + W2− + log e O( S )

log e O( S B B ) = W
1 2 1
−
+ W2+ + log e O( S )

( )
log e O S B1 B2 = W1− + W2− + log e O( S )

The predicted number of pixels in each unique overlap can be calculated using:

mi = Pi Npix i

in which :
mi : the number of predicted landslides for the overlap of two classes
Pi : The calculated probability for the overlap of the two classes
Npixi : The number of pixels in each overlap (for Bi I B2 this will be Npix1 +
Npix5)

The conditional independence is tested with the following formula:

8
mi
G = −2∑ X i log
2

i =1 xi

in which
xi: the number of mass movement occurrences for the overlap of two classes
(for Bi I B2 this will be Npix1)

The function G2 has a χ2 distribution with 2 degrees of freedom (Bonham-

Carter et al., 1990). On the basis of the result of the χ2 test the selection of the variable
maps is made. The weight of evidence values are added and the posterior probability is
calculated. After classification of the posterior probability, the expected number of
landslide occurrences per probability class is calculated for each class and compared

30
with the observed number of occurrences per probability class. The expected frequency
per class is given by:

f i( e) = Pi Npix i

in which
f i( e) : expected number of occurrences per probability class i

Pi : the probability per class i

Npix i : the area (in pixels) of probability class i.

By crossing the predictor map with the mass movement map the actual number

of mass movements can be calculated, and the χ2 test can be applied

(f )
2
i ( 0)
− f i( e)
χ2 = ∑
f i( e)

where, f i ( 0) is the observed frequency of landslides.

2.3.2.3.1.2. Multivariate Statistical Methods in Landslide Hazard Analysis

Multivariate statistical analyses of important causal factors controlling landslide

occurrence may indicate the relative contribution of each of these factors to the degree
of hazard within a defined land unit. The analyses are based on the presence or
absence of stability phenomena within these units (van Westen, 1993).
Multivariate statistical analysis models for landslide hazard zonation were
developed in Italy, mainly by Carrara (1983, 1988) and his colleagues (Carrara et al.,
1990, 1991, 1992). In their applications, all relevant factors are sampled either on a
large-grid basis or in morphometric units. For each of the sampling units, the presence
or absence of landslides is also determined. The resulting matrix is then analyzed using
multiple regression or discriminant analysis. With these techniques good results can be
expected in homogenous zones or areas with only a few types of slope instability
processes. When complex statistics are applied, as was done by Carrara (1983, 1988)
and his colleagues (Carrara et al., 1990, 1991, 1992) or by Neuland (1976) or by
Kobashi and Suzuki (1988), subdivision of the data according to the type of the
landslide should be also made as well. Therefore, large data sets are needed to obtain

31
enough cases to produce reliable results. The use of complex statistics implies
laborious efforts in collecting large amounts of data, because these methods do not use
selective criteria based on professional experience. Multivariate statistical analyses of
important factors related to landslide occurrence give the relative contribution of each of
these factors to the total hazard within a defined land unit. The analyses are based on
the presence or absence of mass movement phenomena within these land units, which
may be catchment areas, interpreted geomorphic units, or other kinds of terrain units.
The following GIS procedures are used to evaluate multivariate statistics in
landslide hazard zonation:

1. Determination of the list of factors that will be included in the analysis. As many
input maps are of alphanumeric type, they must be converted into numerical
maps. These maps can be converted to presence/absence values for each
landunit or presented as percentage cover or the parameter classes can be
ranked according to increasing mass movement density. By overlaying the
parameter maps with the land-unit map, a large matrix is created.
2. Combination of the land unit map with the mass movement map via map
overlay and dividing the stable and unstable units into two groups.
3. Export of the matrix to a statistical package for subsequent analysis.
4. Importation of the results per land-unit into the GIS and recoding of the land
units. The frequency distribution of stable and unstable classified units is
checked to see whether the two groups are separated correctly.
5. Classification of the map into a few hazard classes.

Two types of multivariate analyses have been conducted in the literature

extensively, multiple regression and discriminant analyses. There exists plenty of other
statistical methods, such as logistic regression or analysis of the parameter maps prior
to bivariate analyses by factor analyses. However, these methods require more than
entry level statistics and the data manipulation should be done very carefully, as within
these methods data manipulation is not a speculative event. Although the multiple
regression and discriminant analyses constitute some part of the landslide hazard
analysis literature some real big drawbacks are introduced, as the data used for these
analyses should have to be distributed normally, which is quite impossible when dealing
with natural data. Especially when the data sets of distance to some object is used.
Several normality conversion tables could have been used in order to convert the data
into normal distribution such as log-log or log-normal coversions, however, these
conversions do inherit some critical biases to the natural distribution of the data. Some
authors have tried to exploit the data via using dummy binary variables but this had

32
increased the complexity of the data structure and limits the flexibility of the statistical
system. Examples of these dummy variables could be seen in Carrara et al. (1990,
1991, 1992) and in Chung et al. (1993). On the other hand, the use of binary logical
regression, which is free of data distribution issues, are not so well exploited in the
literature only few examples in the last few years are observed, such as: Atkinson and
Massari (1998), Dai et al. (2001) and Lee and Min (2001). It is also seen from this fact
that logical regression is quite new in this area. In the next sections, only multiple
regression and discriminant analyses will be introduced, the application of logical
regression and tidbits will be explained in the application chapters.
Although these techniques can be applied at different scales, their use becomes
quite restricted at the regional scale, where an accurate input map of landslide
occurrences may not be available, and where most of the important parameters cannot
be collected with satisfactory accuracy. At large scales, different factors will have to be
used (such as water-table depth, soil layer sequences and thickness). These data are
very difficult to obtain even for relatively small areas. Therefore, the medium scale is
considered most appropriate for these sets of techniques.

2.3.2.3.1.2.1. Multiple Regression

The most common and well-known multivariate statistical method used in earth
sciences is multiple regression. It is used to correlate landscape factors and mass
movements, according to the following linear equation.

Y = b0 + b1 X 1 + b2 X 2 +..........+ bn X n

The dependent variable Y represents the presence (1) or absence (0) of a mass
movement. It can also be expressed as the percentage of a terrain unit covered by
landslides. The variables X1-Xn are the independent variables, such as slope class,
geological units, etc. the symbols b0-bn are the partial regression coefficients. The
standardized partial regression coefficients, which are the partial regression coefficients
expressed in units of standard deviation, indicate the relative contribution of the
independent variables to the occurrence of landslides (Davis, 1986). The following
statistics are used to evaluate the result of a calculation.

R2 : amount of variance accounted for by the model. It adjusts for the

number of independent variables in the regression
SE: standard error of estimate. The square root of the residual mean square
error. It measures the unexplained variability in the dependent variable.

33
 MEA: absolute mean error. The average of the absolute values of the
residuals, which is the average error one can expect in a prediction.

The use of terrain units for the sampling of variables in multiple regression analysis
is welcomed with a number of problems.

1. Sampling method
2. Size of terrain unit
3. Resultant maps
4. Sample areas / Prediction areas
5. Complexity of the study areas

In order to avoid these cited problems, generally a pixel based approach is used,
even in this approach the data requirements of normal distribution could not be
achieved. A series of assumptions are made about the assumption of the data normality
which in fact degrades the efficiency of the whole system.

2.3.2.3.1.2.2. Discriminant Analyses

A second type of multivariate analysis is discriminant analysis. The objective of

the analysis is to find the best discrimination between two groups: units or pixels with
and those without mass movements. The analysis results in a discriminant function:

Ds = B0 + B1 X 1 + B2 X 2 +........ Bn X n

where Xi are the values of the variables and Bi the calculated coefficients. Before any
further analysis can be performed, the success of the formula in separating the two
groups must be tested. For this purpose three tests can be used.

1. the variability between the two groups and within the groups, and the total
variability of the data, are calculated. The ratio of the variability between the two
groups and the variability within the groups is called the eigenvalue. It should be
maximized for a good discriminant function.
2. the ratio of the variability between the two groups and the total variability is
called “Wilk’s λ ”. A small value indicates strong variation between groups and
less variation within groups. A Wilk’s λ of 1 indicates that there is equally great

34
 variation within groups as between groups (i.e. that the function does not
discriminate)

3. the χ2 test to determine if the two groups are significantly different.

Furthermore, as the slope stability depends on several factors acting at the
same time, some efforts have been directed towards the acquisition of simply and
quickly determined parameters. Stevenson (1977) using scored factors proposed a
method to evaluate relative landslide risk in clayey slopes.
Discriminant analysis provides a more accurate stability assessment. A
classical work using statistical techniques is that from Jones et al. (1961) on landslides
in Pleistocene terrace deposits of Colombia river. A total of 160 slump-earthflow
movement and additional 160 stable slopes were considered. Qualitative and
quantitative factors influencing sliding were searched. A final analysis using the
discriminant - function method was performed considering as influencing factors:
original slope (X1), submergence percentage (X2), terrace height (X3) and groundwater
(X4).

2.3.2.3.2. Knowledge Driven Methods in Landslide Hazard Analysis

2.3.2.3.2.1. Qualitative Map Combination

To overcome the problem of the “hidden rules” in geomorphic mapping, other

qualitative methods based on qualitative map combination have been developed. In
qualitative map combination, the earth scientist uses the expert knowledge of an
individual to assign weighting values to a series of parameter maps. The terrain
conditions at a large number of locations are summed according to these weights,
leading to hazard values that can be grouped into hazard classes. The problem with this
method is in determining the exact weighting of the various parameter maps. Often
insufficient field knowledge of the important factors prevents the proper establishment of
the factor weights, leading to unacceptable generalizations (Soeters and van Westen,
1996).

2.3.2.3.2.2 Favourability Functions

In order to achieve the minimum common factors of expert knowledge

dependency and the information derived from the original data, the geographical
database is designed under the envelope of some data and expert knowledge
dependent functions. This is also used to decrease the subjectivity of expert knowledge
and not to deal with the redundant information yielding from pure statistical analyses. In
favourability analyses, the data layers are first divided into a number of expert designed
classes such as geological, geomorphological or slope classes and etc. For data

35
integration (numeric and alphanumeric databases) each layer is transformed into a
number between α and β, where α and β are known constants such as 0 and 1 or -1
and +1. This transformation is the basic step of forming a probabilistic favourability
function from a class to the interval α and β. After completion of this stage, some
previously defined decision rules are applied to let the expert earth scientist decide
about the factor probability and favourability of the current situation such as standard
probability measures, certainty factor, Dempster-Shafer belief method and fuzzy logic
interpretation (Soeters and van Westen, 1996; van Westen, 1993).

2.3.2.3.3. Deterministic Modeling in Landslide Hazard Analysis

The methods described so far give no information on the stability of a slope as

expressed in terms of its factor of safety, in order to obtain this information these kinds
of slope stability deterministic models are necessary.
Despite problems related to collection of sufficient and reliable input data,
deterministic models are increasingly used in hazard analysis of larger areas, especially
with the aid of GIS techniques, which can handle the large number of calculations
involved in determination of safety factors over large areas. Deterministic methods are
applicable only when the geomorphic and geologic conditions are fairly homogeneous
over the entire study area and the landslide type is simple. The advantage of these
white box models is that they are based on slope stability models, allowing the
calculation of quantitative values of stability (safety factors). The main problem with
these methods is the degree of simplification which is required in the acceptance limits
of the assumptions. A deterministic method, usually applied for translational slides is the
infinite slope model. These deterministic models generally require the use of ground
water simulation models. Stochastic methods are sometimes used to select input
parameters for the deterministic models (Mulder and van Asch, 1988; Mulder, 1991;
Hammond et al. 1992).
The result is a map showing the average safety factor for a given magnitude of
groundwater depth and seismic acceleration. The variability of the input data can be
used to calculate the probability of failure in connection with the return period of
triggering events. Generally the resulting safety factors and probability factors should
not be used as absolute values unless the analysis is done in a small area where all the
parameters are well known. Normally they are only indicative and can be used to test
different scenarios of slip surfaces and groundwater depths. The method is applicable
only at large scales over small areas. At regional and medium scales, the required
detailed input data, especially concerning groundwater levels, soil profile, and
geotechnical descriptions, usually cannot be provided.

36
2.3.2.3.4 Landslide Frequency Analysis

The probability of mass movement occurrence at a certain place within a certain

time period can only be determined when a relationship can be found between the
occurrence of landslides and the frequency of triggering factors, such as rainfall or
earthquakes. The most promising technique is the calculation of antecedent rainfall,
which is the accumulated amount of precipitation over a specified number of days
preceding the day on which a landslide occurred (Crozier, 1986).
The method is most appropriate at medium and large scales. At regional scale,
it may be difficult to correlate known landslides at one location with rainfall records from
a different location in the area. The spatial component is usually not taken into account
in this analysis and therefore the use of GIS is not crucial, however GIS can be used to
analyze the spatial distribution of rainfall.

2.3.2.4. Accuracy and Objectivity

The most important question to be asked in each landslide hazard study relates
to its degree of accuracy. The terms accuracy and reliability are used to indicate
whether the hazard map makes a correct distinction between landslide free and
landslide prone areas. The accuracy of landslide prediction depends on a large number
of factors the most important of which are:

1. accuracy of the models

2. accuracy of the input data
3. experience of the earth scientist
4. size of the study area

The context of accuracy is a fatal section in disaster management, as the wrong

decision of landslide free areas will cause loss of lives, which discloses to the aim of
hazard and risk assessment. This fatal section can be checked out by some statistical
analyses and trying to find out the possible error component, furthermore the error
component should have to be put in all of the maps produced and the knowledge that
are made public accessible.
Related to the problem of assessing the accuracy of hazard maps is the
question of their objectivity. The terms objective and subjective are used to indicate
whether the various steps taken in the determination of the degree of hazard are
verifiable and reproducible by other researchers or whether they depend on the
personal judgement of the earth scientist in charge of the hazard assessment.

37
 Objectivity in the assessment of landslide hazard does not necessarily result in
an accurate hazard map. For example, if a very simple but verifiable model is used or if
only a few parameters are taken into account, the procedure may be highly objective but
produce an inaccurate map. On the other hand, subjective studies, such as detailed
geomorphic slope stability analyses, when made by experienced geomorphologists may
result in very accurate hazard maps. Yet, such a good, but subjective assessment may
have a relatively low objectivity because its reproducibility will be low. This means the
same evaluation made by an other expert will probably yield another result, which can
have clearly undesirable legal effects (Soeters and van Westen, 1996; van Westen,
1993).
.
2.3.2.5. Evaluation of Methods via Scale Factor

Any hazard evaluation involves a large degree of uncertainty. Prediction of

natural hazards such as landslides, which are caused by interaction of factors which are
not always fully understood and sometimes unknown, confronts earth scientists with
especially large problems. However, the use of statistics indeed will increase the
accuracy of the input data, this minor improvement will reduce the degree of uncertainty
in the assessment. On the other hand the use of multivariate statistics in GIS will yield in
assembling factor maps, that could not be done only on statistical packages. This
assemblage is routing the earth scientists, to use GIS which confronts the user to more
complex and more variable dominated platforms. More onwards, the used models have
to be improved by the availability of huge amount of data and availability of adequate
type and method of handling. Based on this knowledge, the available methodologies in
landslide hazard zonation can be classified and rated as follows in correspondence to
the analysis of scale factor (Table 2.7).

38
Table 2.7. Classification of Methods based on scale factor (Soeters and van Westen,
1996, van Westen, 1993).(the first number indicates the feasibility 1:Low, it would take
too much time and money to gather sufficient information in relation to the expected
output; 2: Moderate: a considerable investment would be needed, which only
moderately justifies the output; 3: good, the necessary input data can be gathered with a
reasonable investment related to the expected output. The second number indicates the
usefulness 1:of no use, 2: of limited use, 3: useful).

Regional Medium Large Usefulness of GIS in

Method
Scale Scale Scale the analysis
Landslide Distribution Analysis 2/3 3/3 3/3 Intermediate
Landslide Density Analysis 2/3 3/2 3/1 Intermediate/high
Landslide Activity Analysis 1/3 3/3 3/3 Intermediate/high
Landslide Isopleth analysis 2/3 3/2 3/3 High
Geomorphological Landslide Hazard analysis 3/3 3/3 3/3 very low
Qualitative Landslide Hazard Analysis 3/3 3/2 3/1 high
Landslide Susceptibility Analysis 1/3 3/3 3/2 high
Information Value Method 1/1 3/3 3/2 high
Weights of Evidence Method 1/1 3/3 3/2 high
Multivariate Statistical Analysis 1/2 3/2 3/2 high
Deterministic Landslide Hazard Analysis 1/1 1/2 2/3 high
Antecedent Rainfall Analysis 2/2 3/3 3/2 Very low

39
 CHAPTER 3

GEOLOGY OF THE ASARSUYU CATCHMENT

3.1. Geology

3.1.1. Regional Geology and Previous Works

The study area is located in northern Central Anatolia, within the North
Anatolian Fault Zone (NAFZ), and it is mainly comprised of Paleozoic
metamorphic/intrusive basement and its covering Mesozoic and Tertiary flyschoidal
deposits (Figure 3.1) in the Pontide Tectonic unit of Ketin 1966.

Figure 3.1. Regional geological map of the study area and its environs (modified from
Tokay,1964).

40
 The study area is hosting the main state highway connecting the capital city
Ankara to the biggest metropolitan area İstanbul. The presence of the Anatolian
highway and the problems associated with the construction of Bolu Tunnels attracted
the attention of the scientists and hence a number of studies had been carried out
concerning the geology, NAFZ and engineering geological problems of this region.
Hereforth, a brief summary will be given about the previous works of the study area in
table form (Table 3.1). The citations would be made in the appropriate parts in the
following sections with detailed explanations when needed.

Table 3.1. Previous studies performed in the study area and its near vicinity

Research
Researcher Research Area Remarks
Field
Blumenthal
Regional First definitions of tectonic units
(1948)
First sub-division of Paleozoic
Ketin (1955) Akçakoca-Düzce
Massifs
Uysallı (1959) Bolu-Merkeşler Coal Resources
Abdülselamoğlu Definition of Paleozoic and
Mudurnu-Göynük
(1959) mesozoic formations
Bolu-Gerede-
Ketin (1967) Paleozoic Units
Mengen-Yıgılca
First defined the age of the
Northern Slopes of
Batum (1968) conglometares that overlie Bolu
Asarsuyu Valley
massif as Sillurian
Mudurnu-
Gözübol (1978) Geology/NAFZ
Dokurcun-Abant
G e o l o g y

Görmüş (1980) Yığılca Stratigraphy/Tectonics

Geological evolution, relations of
Yılmaz et al.
Abant-Dokurcun metamorphic rocks and the
(1981)
ophiolitic rocks in area.
Görmüş (1982a) Yığılca Stratigraphy
Görmüş (1982b) Yığılca Tectonic and evolutionary model
Ereğli-Yığılca-Bolu-
Kaya (1982) Stratigraphy-Tectonics
Mengen
Kaya & Dizer
Mengen Coal Resources
(1981-1982a)
Kaya & Dizer Stratigraphy of Mesozoic and
North Bolu
(1981-1982b) Tertiary sequences
Cerit (1983) Mengen Geology
Serdar & Demir Bolu-Mengen-
Petroleum Resources
(1983) Devrek
Öztürk et al. Stratigraphy of north and south of
Abant-Yeniçağa
(1984) NAFZ

41
Table 3.1. (continued)

Aydın et al. Çamdağ-

Complete stratigraphic outline.
(1987) Sünnicedağ
Stratigraphy of Mesozoic and
Kaya et al. (1986) Yığılca
Tertiary sequences
Cerit (1990) Bolu Massif Geology and tectonics
Erendil et al.
Bolu Massif Economical resources
(1991)
G e o l o g y

Yalçın & Cerit Investigated the metamorphism by

Bolu Massif
(1991) using clay mineralogy
Microfabric study in Paleozoic
Koral et al. (1994) Asarsuyu
rocks
Gedik & Alkaş
Bolu Region Carbondioxide potential
(1996)
Sözen et al. Geochemistry/Economical
Düzce and Devrek
(1996) resources
Güler (1999) Bakacak Geology of gypsum
Ustaömer &
Bolu Massif Geochemistry – Evolution model
Rogers (1999)
Orkan et al. Elementary landslide susceptibility
Highway Route
(1977) map
Canik (1980) Bolu Hydrogeology
Aktimur et al. Landuse pattern and NAFZ related
Bolu
(1983) problems
Bolu Mountain Engineering geology of Highway
Dalgıç (1994 a,b)
Highway Pass pass and Bolu Tunnel
Stability and nature of the
Dalgıç et al(1995) Asarsuyu
Yumrukaya landslide
Engineering

Dalgıç & Gözübol

Bolu Tunnel Stability problems in Bolu Tunnel
(1995)
Enginering geology of the
Astaldi (1995) Asarsuyu Asarsuyu region for Bolu Tunnel
construction
lithology and the fracture pattern in
Dalgıç (1997) Bolu Tunnel
the Bolu Tunnels
Simşek & Dalgıç consolidation properties of the
Düzce
(1997) clays at Düzce
Aydan & Dalgıç Prediction of Deformation inside
Bolu Tunnel
(1998) the Bolu tunnel
slope stability problems in the
Dalgıç (1998,a) Asarsuyu
Asarsuyu valley
Selection of crushed Rock
Dalgıç (1998,b) Asarsuyu
Quarries

42
Table 3.1. (continued)
Bolu Mountain Investigated two landslides in the
Işın (1999)
Highway Pass highway by insitu instumentation
Deformations in the Bolu Tunnels
E n g i n e e r i n g
Unterberger and
Bolu Tunnel after the 12. Nov 1999 Düzce
Brandl (2000)
earthquake
Interdisciplinary study concerning
the geological, engineering
Aydan et al. (2000) Düzce geological, seismological and
geotechnical aspects of Düzce
earthquake
Engineering report regarding both
Sucuoğlu et al.
Regional the Marmara and Düzce
(2000)
earthquakes
Ketin (1969) Regional First definition
Ambraseys (1970) Regional Characteristic features
Tokay (1973) Gerede-Ilgaz Characteristics
Mudurnu-
Gözübol (1978) Dokurcun- Structural properties of the NAFZ
Abant
Şengör & Canıtez Characteristics + evolutionary
Regional
(1982) model
Abant-
N A F Z ( T e c t o n i c s )

Öztürk et al. (1984) Paleo and Neo-tectonic structures

Yeniçağa
Nurlu(1993) Bolu-Sapanca RS
Closing-up Structures-effect of
Neugebauer (1994) Abant
bends in NAFZ
Pattern of faulting in the area, also
Bolu Tunnel –
deformation in the area gathered by
Astaldi (1995) Asarsuyu
GPS measurements and evaluated
valley
together with microseismic data.
Adapazarı –
Neugebauer (1995) Kinematics of NAFZ
Bolu
Şaroğlu et al. Gerede-
(1995) Eskipazar
Neugebauer et al.
Abant-Sapanca Overstep in NAFZ
(1997)
Slip distributions of Düzce
Akyüz et al. (2000) Düzce
Earthquake
Gölcük-
Taymaz (2000) Sapanca- Seismotectonics
Düzce

3.1.2. Stratigraphy

The basement of the Asarsuyu catchment is composed of pre-Devonien

Yedigöller Formation, an assemblage of metadiorite-metagranite and amphibolite–
amphibolite gneiss (Figure 3.2). Following in the sequence İkizoluk Formation and its
Çatak member forms the Devonien which tectonically overlies the underlying Paleozoic

43
basement. Upper Cretaceous is characterized by Abant Complex and Elmalık Granite,
which again have a tectonic contact with underlying sequence. Starting from Upper
Cretaceous to Pliocene a carbona-clastic sequence (flysch) is seen in the area, which is
represented by Atyayla, Bayramışlar, Fındıcak and Apalar Formations. Plio-Quaternary
period is represented by Asarsuyu formation, which is a combination of terrace
conglomerates and alluvial fill deposits in the Asarsuyu valley itself. Finally, Quaternary
alluvial fill and slope debris cover unconformably the whole sequence in the Asarsuyu
catchment.

3.1.2.1. Yedigöller Formation

Yedigöller Formation is the core of the Bolu massif and composed of Paleozoic
magmatic rocks. The formation has widespread outcrops in the eastern and northern
slopes of the Asarsuyu catchment (Figure 3.3) The formation is first named by Aydın et
al. (1987).
The main lithologies of the formation are amphibolite, gneiss (metagranite),
metadiorite and meta-quartzdiorite. Also very small unmappable aplite, andesite, basalt
and diabase dikes are common. Some of these are mylonitized by regional paleo and
neotectonic events. The unit is in general affected from the regional metamorphism, and
show cataclastic deformation in the vicinity of the fault zones. This formation can be
divided into two as one having greenschist metamophism and having no metamorphism.
The greenschist facies of this formation generally have a medium degree of weathering
resulting in a thin residual soil cover (1 - 3 meters).
The lower boundary of the formation cannot be seen in the catchment; however
its upper boundary is tectonic with the overlying Devonian İkizoluk Formation, which is
not observed in the study area.
This formation comprises the southwest part of the Bolu Massif. Blumenthal
(1948) assigned Paleozoic age to the metamorphics in the area. Ketin (1967) indicated
that the age of the assemblage of amphibolite, gneiss and basic rock in Bolu-Gerede-
Mengen-Yığılca area is Cambrian. Erendil et al. (1991) assigns the name Bolu Granitoid
to the formation and cites that the age should be Pre-Late Ordovician. While Canik
(1980), Aydın et al. (1987) and Cerit (1990) suggest Pre-Cambrian age. Based on these
literature arguments, the age of the Yedigöller Formation is assigned as Pre-Devonian.

44
 Alluvium Formation
Member
System

Stage
Lithology Description

Pliocene Quaternary
active alluvium in the Asarsuyu
valley, the terrace deposits
and the slided masses

c
Unconformity
i
alternation of clayey silt,

Çaycuma Asarsuyu
sandy silt and sand
with cobbles
o
z

Unconformity
alternation of turbiditic
sandstone and siltstone,
o

Pre-Devonian U.Ordovisien M.Sillurian Cretaceous Paleocene Eocene

calcareous mudstone,
mudstone and marl
with gypsum intercalations
n
e
C

beige, white turbiditic

limestones, greenish
grey sandstone, siltstone,
marl and limestone
Aksudere Buldandere
Findiklidere

alternations, greenish
Mesozoic

purplish conglomerates,
mudstones and tuffs.

Unconformity
-L.Sillurian -M.Devonian

phyllites, shale, recrystallized

limestone, dolomitic limestone
and marls
P a l e o z o i c

Kocadere

massive to thick bedded

purplish grey conglomerates
and sandstones

Tectonic contact
Yedigöller

amphibolite, gneiss
(metagranite), metadiorite,
meta-quartzdiorite
aplite, andesite, basalt
diabase dikes

Figure 3.2. The generalized columnar section of the study area, (not-to-scale) (modified
from Erendil et al. ,1991).

45
Figure 3.3. The geological map the study area (modified from Erendil et al. ,1991).

46
3.1.2.2. Kocadere Formation

This formation generally crops out as a rim to the Yedigöller Formation. The
outcrops of this formation is seen only in the eastern part of the catchment at northern
slopes of Asarsuyu river. The formation is defined by Erendil et al. (1991).
Kocadere Formation is composed of massive to thick bedded purplish grey
polygenetic conglomerates and sandstones (greywacke). Limited mudstone and
siltstone bands are also observed. Conglomerates are poorly sorted, and their
components are derived from underlying metamorphics, and magmatic rocks, the matrix
is composed of chlorite, quartz and serrisite (Erendil et al., 1991).
The lower boundary of the formation is tectonized and obscured. Its upper
boundary is conformable with the Aksudere Formation.
The equivalent of this formation, the Kurtköy Formation of Istanbul (Kaya, 1973)
is given an age of Ordovician. The other possible equivalents of this formation,
“Hamzafakılı sandstones” (Tokay, 1952), Işığandere Formation (Görmüş, 1982 a, b) and
“purple arkoses and conglomerates” (Batum, 1968) are aged as Late Silurian. Based on
these ages, Erendil et al. (1991) assigns Late Ordovician-Early Silurian age to Kocadere
Formation.

3.1.2.3. Aksudere Formation

Aksudere Formation is seen in the northern slopes of Sarıçökek village and at

the eastern parts of the catchment at southern part of Asarsuyu river near Aşağıbakacak
settlement and in between the two major landslides near Bakacak village. The formation
is defined by Erendil et al. (1991).
Aksudere formation is represented by phyllites, shale, recrystallized limestone,
dolomitic limestone and marls. The phyllites show significant schistosity and the main
constituents are quartz, serrisite muscovite, chlorite, epidote and feldspars. They also
host some embedded massive recrystallized limestone blocks. Phyllites and
recrystallized limestones alternate with shales, siltstones, sandstones, dolomitic
limestones and marls as thin to medium beds. The uppermost levels of this formation is
characterized by the dominance of recrystallized limestones. The formation forms a
gentle rolling topography, as it is composed of easily erodible lithologies. The main
colors observed in this formation are dark grey, beige, bluish-greenish grey and brown.
The lower boundary is conformable with Kocadere Formation and its upper
boundary is also conformable with Kırdoruk Formation. However, the Kırdoruk
Formation does not outcropping in the Asarsuyu catchment. According to Erendil et al.
(1991), the observed thickness of Aksudere Formation is about 1500-2000 meters.

47
 Görmüş (1982, a) stated that the equivalent of lower facies of Aksudere
Formation is Kocadere formation which is Lower Devonian. The Devonian aged Kartal
formation of Kaya (1973) can also be correlated with Aksudere formation. Dalgıç (1994,
a,b) grouped the Kocadere and Aksudere formations under İkizoluk formation and
assigned an age of Devonian. Furthermore, he subdivides the uppermost section of
Aksudere formation, the recrystallized limestones as Çatak Member in İkizoluk
formation. Erendil et al. (1991) assigned an age of Middle Silurian to Middle Devonian.

3.1.2.4. Buldandere Formation

Erendil et al. (1991) uses this formation name to cover all of the Upper
Cretaceous-Paleocene cover units, that overlie the Paleozoic Bolu Massif.

3.1.2.4.1. Fındıklıdere Member

Fındıklıdere member comprises the middle levels of Buldandere Formation

outside of the Asarsuyu catchment. The member is first defined by Erendil et al. (1991).
Fındıklıdere member has widespread outcrops in the southern slopes of the catchment.
The lithology of the Fındıklıdere member consists of beige-white turbiditic
limestones, greenish grey sandstone, siltstone, marl and limestone alternations,
greenish purplish conglomerates, mudstones and massive to medium to thick bedded
gypsums. The gypsum occurrences are generally observed near the summit of Bolu
Mountain and near Dipsizgöl and Yukarıaçma villages. There are also some rock
quarries opened for gypsum production in these localities. Gypsums outcrops have
significant karstic dissolution features, such as siphons going up to tens of meters in
depth. The association of these layers resembles an Upper Cretaceous flyschoidal
sequence.
The lower boundary of the member in the study area is unconformable with the
Aksudere Formation, while its upper boundary is conformable with Eocene Çaycuma
Formation. The thickness of the formation is estimated by Erendil et al. (1991) as 600
meters.
For the age of this member Erendil et.al. (1991) give a long list of fossils and
assign Campanian - Ilerdian age. Similar formations are named and aged as:
Cretaceous Flysch (Abdülselamoğlu, 1959), Gökveren Formation – Eocene (Gözübol,
1978 and Yılmaz et al., 1981), Sarıkaya Formation – Upper Cretaceous/Paleocene
(Görmüş, 1980), Akveren Formation – Campanian/Lower Paleocene (Aydın et.
al.,1987), Fındıcak Formation – Upper Cretaceous/Paleocene (Dalgıç, 1994, a,b).

48
3.1.2.4.2. Çaycuma Formation

This formation seems like the continuation of the cover units of the Bolu Massif
and Fındıklıdere member. It crops out at south of the Kaynaşlı town at Akmeşe and
Türbe hills, also in the north of Yörükler village. The formation is first named by Saner et
al. (1979). The main lithologies of the formation is alternation of turbiditic sandstone and
siltstone, calcareous mudstone, mudstone and marl with gypsum intercalations.
The lower boundary of this formation is conformable with Fındıklıdere member,
while its upper boundary is not seen in the Asarsuyu catchment. The age of this
formation is assigned by Erendil et al. (1991) as Eocene and is supported by both
literature and fossil records. Dalgıç (1994) defined another member in the Asarsuyu
catchment as Açma member, for which the age of Late Eocene was assigned and the
dominant lithologies are defined as alternations of argillaceous limestones, gypsums,
calcareous mudstones and clayey gypsums, which are similar to Fındıklıdere member of
Erendil et al. (1991).

3.1.2.5. Asarsuyu Formation

Minor differences and the spatial orientation of the outcrops led Dalgıç (1994,
a,b) to differentiate this formation from Quaternary alluvium. The Asarsuyu Formation
crops out near Elmalık village and near the Asarsuyu portal of the Bolu Tunnels. Main
lithologies of this formation include alternation of clayey silt, sandy silt and sand with
cobbles.
The Lower boundary is unconformable with the older formations and its upper
boundary is again unconformable with Quaternary alluvium. The age of Plio-Quaternary
is based on palynological data by Astaldi (1990).

3.1.2.6. Quaternary Deposits

The Quaternary deposits are characterized by the active alluvium in the

Asarsuyu valley, the terrace deposits in the northern margins of the Kaynaşlı plain and
the slided mass of the to huge landslides in the southern slopes of Asarsuyu valley.

49
3.2. Tectonism of the Asarsuyu Catchment

The Düzce-Bolu region is one of the most tectonically active regions of Turkey.
During complex tectonic evolution of the area, the catchment had been subjected to
different regimes and different deformations which are reflected in the current outcrops
and topography. The active faults, belonging to the North Anatolian Fault Zone, and the
older thrust faults are the major large scale structural features and form the main
tectonic elements of this catchment. The thrust faults are believed to act in the
compressive regime after Late Cretaceous and ceased in Late Miocene (Yılmaz et al.
1981; Dalgıç, 1994) and classified as Paleotectonic features. The deformation of the
region took the form of lateral strike-slip faulting after Late Miocene period in the
Neoectonic regime. The deformation is still active as the North Anatolian Fault Zone and
its associated faults take the act in hand and reshape the catchment (Aydan et al.,
2000). During aerial photographical interpretation studies these faults are mapped and
digitized. Due to the scope of this study both the thrusts and other active faults are
treated together, as large discontinuities in the catchment.

Figure 3.4. Lineament map of the Asarsuyu catchment

50
 The earthquake activity of this region is directly controlled by the presence and
activity of North Anatolian Fault zone and its associated fault segments. The North
Anatolian Fault is the northern margin of Anatolian block that is escaping from the
collusional zone between Arabian and Eurasian plates into the Eastern Mediterranean
Sea (Şengör et al., 1985). The North Anatolian Fault is one of the most well known and
most studied faults in Turkey, which has been defined some 50 years ago. The length of
this fault is approximately 1200 kilometers and it is not a single fault, which in turn,
worths to consider it as a fault zone (Sucuoğlu et al., 2000). The width of this zone is
about 1 kilometer is widening up to 5 kilometers in the west (Demirtaş and Yılmaz,
1996). The total displacement made by this fault is reported as 20-25 kilometers by
Neugebauer (1994) at Abant lake environs. Global positioning system studies (GPS)
have been ongoing in the area since 1988 and the lateral slip on the Düzce fault is
estimated as 7.5±1.5 mm/year (Astaldi, 1995)
The main fault strand of the North Anatolian Fault splays into two strands in the
west of Bolu; the southern strand goes through the Lake Abant and then the Mudurnu
valley. The northern strand steps to the north, forming the southern boundary of the
Düzce basin being called the Düzce fault. In the west end of the Düzce basin, thereafter
it is named the Hendek fault. Şaroglu et al. (1987, 1992) who published "Active Faults of
Turkey" showed no minor connecting fault between the main trace of the North
Anatolian fault and the Düzce fault. Barka and Erdik (1993) reported that there are two
secondary faults, the Elmalık and Aşagı Bakacak faults that occured between the main
trace and the Düzce fault as connecting faults.

3.3. Seismicity of the Asarsuyu Catchment and Environs

A total of 623 earthquakes, that are greater than 5 in instrumental magnitude

have been recorded in the NAFZ in the period of January 1900 to March 2001 (Figure
3.5) having a total of 98 quakes greater than 6 in magnitude. However, the seismic
evaluation of the whole fault zone is out of scope in this study, so the concentration will
be turned on the study area and its vicinity. In order to compare the activity in the west
and middle parts of the NAFZ and the rest of Turkey the earthquakes of the
western/middle is counted and shown in Figure 3.6. 151 earthquakes out of 623 are
counted, which represent nearly 25% of total earthquakes in Turkey greater than 5, in
the study area. When zoomed into the study and environs, there are a total of 2596
earthquakes, of which 410 of them are greater than 3.5 in magnitude. Their spatial
locations with their magnitude distributions are shown in Figure 3.7. Four earthquakes
greater than 7 occurred in the area, chronologically 26-05-1957 Abant, 07-22-1967
Mudurnu, 17-08-1999 Marmara and 12-11-1999 Düzce earthquakes. The surface

51
rupture pattern of the 12-11-1999 Düzce earthquakes is shown in Figure 3.8.
Furthermore some small landslides have occurred in the are immediately after the
earthquake, some examples of these are presented in figure 3.9.
The historical records show that on the western zone of the NAFZ, 3 major
earthquakes are recorded (Ambraseys and Finkel, 1995). The 25-05-1719 earthquake
occurred east of Marmara sea resulted in 6000 casualties, and 4/5 of İzmit city is
demolished. The second quake occurred at 2-09-1754 in the vicinity of İzmit Bay and
approximately 1000 casualties was recorded. A third quake was recorded again in the
east of Marmara Sea affecting the whole coast line of the Marmara region at 22-05-1766
resulting in 5000 casualties.

Figure 3.5. Epicenter locations of Turkey greater than 5 in magnitude (Magnitudes in Ms)

52
Figure 3.6. Epicenter locations of mid-west NAFZ greater than 3.5 in magnitude

53
Figure 3.7. Earthquakes occurred in the nearby of the study area.

54
Figure 3.8. Surface ruptures of the 12 November 1999 Düzce earthquake (Aydan et al.,
2000), Displacements are in meters.

55
Figure 3.9. Some examples of landslides occurred after 12 November 1999 Düzce
earthquake.

56
 CHAPTER 4

INPUT DATA AND DATA PRODUCTION

In this section the sources of input data and the steps in input data production
are explained, as the data entry and production is the most cumbersome and time
consuming steps of any kind of a Geographical Information System. The brief analysis
of the distribution of the data produced are given in this chapter, however, the detailed
analysis of the landslide database and the parameter maps are left to be explained in
the next chapter. All of the input data and the produced data sets are valid for the
common landslides except the two huge landslide bodies in the study area.

4.1. Geology

4.1.1. Data entry

The geological map of the region is compiled from the available reports,
publications, dissertation thesis and an analogue map is prepared. The compiled
analogue map is transformed into digital image, via digitizing and editing using TNT
MIPS. A database concerning the lithology names and the ages given by the previous
researchers are attached. Polygon topology is also built, validated and attached as an
internal table (Figure 4.1). The resultant map is called LITHOMAP.
In the compiled geological map 11 lithologies exist. Even though, there are 7
units in the generalized columnar section in Figure 3.2, some of the lithological units
show significant heterogeneity, so that they are classified as different lithological units.
For example Yedigöller Formation is divided into two as one showing greenschist
metamorphism and the other one not; Gypsum occurrences in the Fındıklıdere Member
is separated, also the Quaternary alluvium has a distinct talus/landslide deposit unit
which has to be separated from the normal valley filling alluvium. As a result, the post-
classified 11 units are covering a total of nearly 1196.5 square kilometers. The largest
areas are represented by Fındıklıdere member (22%), Yedigöller Fm. (16.67%),
Aksudere Fm. (15.24%), Greenschist facies of Yedigöller (15.17%) of which in total

57
represents nearly the 70% of the area. The remaining 7 units represent only the 30%
and within this 30% nearly half of it is represented by the combination of Quaternary
terrace deposits and the two large landslides in the area (Figure 4.2)

Figure 4.1. LITHOMAP and the attached topological table of the study area.

25.00

20.00

15.00
P e rce nt

10.00

5.00

0.00
Gr.s c his t Slope
A s ars uyu Y edigöller A ks udere Cay c uma Fındıklıdere Koc adere
Y edigöller Mengen Fm A lluv ium Gyps um Depos its /La
Fm Fm Fm Fm Mb Fm
Fm nds lide

% 0.29 15.17 16.67 15.24 3.56 0.03 8.34 0.62 14.29 22.17 3.62

Form a tion Na m e

Figure 4.2. The area distributions of LITHOMAP.

58
 Following this, the faults of the area are compiled from literature In addition to
this, the photo lineaments are merged to create a final fault map. For the photo
lineaments Landsat TM 5 image of 1993, the stereo panchromatic aerial photographs of
1994 (1:25.000 scale), 1984 (1:15.000 scale), 1972 (1:25.000 scale) and 1952 (1:35.000
scale), and the digital elevation model of 1994 with 25-meter grid spacing are used. The
snapshots of the satellite remote sensing and elevation data will be presented in
following sections. The thrusts, lineaments, strike slip faults and earthquake
faults/ruptures are considered as potential water sources and triggering sources so they
are all together grouped as photo lineaments of the Asarsuyu catchment without any
genetic discrimination. The resulting map is called the FAULTMAP including all of the
mapped faults and photo-lineaments in the area (Figure 4.3). 126.284 kilometer length
of fault lines are observed in the study area representing 118 fault line segments. The
average length of fault-lines is 1,079.35 meters. The directional analysis is done based
on “Weighted Segments” method in which the method uses the direction of each line
segment and creates the Rose diagram from the total length of all segments in each
direction. (Figure 4.4.a). The main directions are generally confined to in east-west
trend, which is also conformable with the main trend of the North Anatolian Fault zone in
the study area. The non-weighted rose diagram of the fault lines also has a major trend
in E-W direction, while another small trend is seen as N-S. The N-S trending fault lines
comprise shorter fault lines than those of E-W trending ones (Figure 4.4.b).

Figure 4.3. The FAULTMAP of Asarsuyu catchment.

59
 a b

Figure 4.4. The rose diagram of FAULTMAP; a. weighted, b. non-weighted.

4.1.2. Input map generation

The LITHOMAP is used without any post-processing as it is, only a vector

topology validate process is carried out to rebuild and check the polygon topology and
integrity of the vector map. This is also necessary to assess the “one-to-one implied”
attachment style of the database attached. However, the FAULTMAP is transferred into
raster format, to a distance raster, as to represent the distance of every pixel to the
nearest line in the map. This procedure calculates not the aerial orthogonal distance
(oD), but the true surface distance (tD) (Figure 4.5) with the aid of a detailed Digital
Elevation Model. The calculated distances are assigned as a 16-bit digital value to the
every pixel in the raster image produced (Figure 4.6). The distance raster map is called
DISTFAULT. The DISTFAULT to fault-lines input parameter map shows a logarithmical
distribution as the frequency decreases when the distance from the fault lines increases.

Figure 4.5. The distance calculations in distance raster map production.

60
Figure 4.6. Distance raster image (DISTFAULT) produced from FAULTMAP

Another input map produced from FAULTMAP is the lineament density in 1-

kilometer search radius. In order to calculate the density the number of occurrences of
fault lines in a radius of 1 kilometer search distance is calculated for every pixel of 25
meters. A TNT MIPS script is written for this calculation. This density map calculated for
each pixel (Figure 4.7) is called FAULTDENS. The maximum value encountered in the
FAULTDENS map is 768 and the minimum is 0 meters. The distributions mean is 195.9
having a standard deviation of 154.6

Figure 4.7. Fault density (FAULTDENS) of Asarsuyu Catchment

61
4.2. Elevation

4.2.1. Data entry

The digital elevation data of Asarsuyu catchment is gathered from General

Command of Mapping Turkish Army. This data is composed of contour lines and points
with elevation information in ARC/INFO E00 format (Figure 4.8), and then converted into
TNT-MIPS vector format. The previously built line topology concerning the elevation
information is rebuilt and validated. This map is called the CONTOURMAP.

Figure 4.8. The color-coded CONTOURMAP.

4.2.2. Input map generation

Rather than using a discrete elevation map such as the imported contour lines,
it is more advantageous to work with a continuous map, regarding this advantage, the
contour data is converted to a color-coded continuous map (Digital Elevation Model –
DEM) by “profiling algorithm” in TNT-MIPS (Figure 4.9a). The Profiles method uses
pairs of input elevation values on opposite sides of each output raster cell and performs
a multi-directional linear interpolation procedure to create a surface raster. Edge cells
are processed first; the search for cell pairs is restricted to cells parallel to the edge. For
other cells the process searches in eight different directions and uses the closest pair of
values (including edge cell values) to assign an interpolated value to an output cell. For
all of the parameter maps the resolution of the DEM produced will create a reference
layer and here forth 25 meters is taken as the working resolution of this project. Hence,

62
all of the raster inputs are fixed into 25 meters spatial resolution and the DEM is used as
reference raster for these.
In order to increase visual interception of the DEM it has been converted into a
relief map and presented in Figure 4.10. The produced DEM (ELEVMAP) will be used
as the elevation input data for the elevation attributes of landslides.

Figure 4.9. Color coded DEM of Asarsuyu catchment

Figure 4.10. Color draped relief model of Asarsuyu catchment (illumination 045o,
vertical exaggeration x3).

63
 Following the accurate modeling of the topography in the Asarsuyu catchment,
a watershed analysis is carried out to extract the microcatchments, the ridge points and
the drainage-lines (Figure 4.11). The process begins by evaluating the elevation raster
for depressions and constructs watershed polygons based on the depressions
recognized. A vector object is generated comprising polygons that encompass the
watersheds and pour point locations. The watershed process uses the Deterministic-8
(D8) algorithm (Jenson and Domingue, 1988) for flow path determination. This algorithm
computes the terrain slopes between the central cell and each of its eight neighbors.
Flow direction is then defined as the direction to the neighbor, either adjacent or
diagonally, with steepest downward slope. The extracted drainage-lines are compared
with the actual streams in the topographical sheets and corrected, if necessary. Also the
Strahler order of each stream segment is assigned into a separate table manually to be
used in the generation of other drainage dependent parameter maps and to assess the
drainage conditions in the area.
The distances of every pixel regarding the drainage-lines are calculated
(DISTDRAINMAP) and presented in Figure 4.12 The minimum distance of pixels is 1
meter and the maximum is 452.95 meters. The distribution has a mean of 97.56 with
standard deviation of 73.14.

Figure 4.11. The drainage system of Asarsuyu catchment, including the microcatchments
(MICROCATCHMAP), drainage-lines (DRAINAGEMAP), and ridges (RIDGEMAP). (Color
coding is arbitrary).

64
Figure 4.12. The distance raster of every pixel to the nearest drainage-line
(DISTDRAINMAP).

The drainage lines are also used to calculate the kilometer square density of
drainage lines in the whole study area. To maintain the 1 square kilometer search
distance 564 meter search radius with 250 meter offset is used (Figure 4.13). The vector
of drainage lines are converted to point data with a distance of 25 meters. The analysis
show that the maximum count is achieved at 352, hence the maximum drainage length
in the area is nearly 9 kilometers in 1 square kilometer area.

Figure 4.13. The drainage density of Asarsuyu catchment

The outputs of the watershed analysis are outputs for calculation of the nearest
distances to the ridges of each pixel (Figure 4.14). The minimum value for the distance

65
to ridges is 1.62 meters and the maximum is 658.97 meters. The distribution has a
mean of 138.72 meters with standard deviation of 99.86.

Figure 4.14. The distance raster of every pixel to the nearest ridge-line
(DISTRIDGEMAP).

The topographical modeling of the Asarsuyu catchment also serves as an input

to calculate the topographical derivatives such as aspect and slope maps of the
catchment, which are very crucial and contain vital information for the landslide
development in the area.
The aspect is a measure of slope orientation and is calculated in compass
degrees as the azimuth from the North. Due to the raster data format itself and the
computational limits the aspect distribution has sensitivities in principal directions and in
every 22.5 degrees, which is a drawback of the algorithm used. Although the model
creates reliable elevation data and slope calculations, the small size of the unique
elevation facets prefers to be oriented in 16 principal directions which are 22.5 degrees
apart from each other. The produced and color-coded ASPECT map and its frequency
distribution is presented in Figure 4.15. The ASPECT map has a range between –1 and
359, -1 representing the flat lying areas and 0 as the north, any other value is the
azimuth measurement from North. The minimum value –1 and the maximum is 359
degrees. The distribution has a mean of 182.45 with standard deviation of 110.

66
Figure 4.15. Aspect map of the Asarsuyu catchment (ASPECT).

The slope is the measure of surface steepness and is calculated in degrees.

The produced and color-coded SLOPE map and its frequency distribution is presented
in Figure 4.16. The SLOPE has a range between 0 and 90, 0 representing the flat lying
areas and 90 as the vertical; any other value is the compass measurement from
horizontal. Hence the minimum value of the data is 0 and the maximum is 56 degrees.
The distribution has a mean of 16.97 degrees with standard deviation of 10.46.

Figure 4.16. SLOPE map of Asarsuyu catchment.

67
4.3. Infrastructure

The presence of several infrastructure elements such as houses, power lines

and road network might contribute to the evolution of landslides in the area. The
infrastructure has a mutual relationship with the landslide hazard as either it causes the
slide or it is affected from the slide. The causes can be explained as: near the houses
there is a groundwater surcharge either by sanitary disposals or by small-scale
backyard irrigation. For the construction of power line poles, the forest under and in the
vicinity of the pylon is cut, so land cover changes. For the road building both the cut
slopes, the land cover change and the economical activity near the roads, due to
highway tourism, attract people. However, all of these built structures are then directly
be faced with the hazard, as the elements at risk possessing vulnerability to landslide
hazard.

4.3.1. Data entry

For the construction of historical infrastructure database, the necessary features

are digitized from the 1:25.000 scale topographical maps of 1952, 1972 and 1994 maps
(Figure 4.17). Although the map of 1972 is published in 1977, it is referred as 1972
maps, as it had been prepared from 1972 aerial photographs. No maps are produced
from 1984 photographs as they were taken for forestry applications

4.3.2. Input Data Production

For the landslide database only the distance rasters of 1994 period is produced as it
will yield the maximum development stage in the study area (Figure 4.18). The distance
raster of settlement of 1994 has a minimum value of 0 ranging to a maximum of 6093.46
meters (Figure 4.18.a). The mean of this distribution is 1258.46 meters with a standard
deviation of 1320.73. The area is divided in to two in a NW-SE trend, as all of the
settlements are located south of this dividend. This could be attributed to the presence of
continuously decreasing background noise in the frequency histogram of
DISTSETTLEMENT. The power lines and the road network is merged as genetically they
are affecting the topography in a similar way such as, land cover disturbance is needed to
construct both and in order to construct the power lines a new temporary road is opened.
The distance raster is then constructed from the merged vector coverage. The distance
raster of merged coverage of 94 has a minimum value of 0 ranging to a maximum of
2312.68 meters (Figure 4.18.b), with a mean of 246,1 and a standard deviation of 302.099.
The influence of E-5 highway is decided to be considered as a separate parameter, so a
vector containing only E-5 highway is formed (Figure 4.18.c). The distance raster of E-5
highway of 94 has a minimum value of 0 ranging to a maximum of 8366.8 meters. The
mean of this distribution is 2467.49 with a standard deviation of 1771.85.

68
 a

c
Figure 4.17. Historical infrastructure databases of Asarsuyu catchment a) 1952, b)
1972, c) 1994

69
 a

Figure 4.18. Distance rasters and frequency distributions of 1994 period a) distance
raster of settlement of 1994, b) distance raster of road network and power lines of 1994,
c) distance raster of E-5 road of 1994

70
4.4. Land cover

A landslide hazard assessment should not only depend on the production of

landslide inventory map and its analysis. A complete hazard assessment system will
require also the assessment of external factors leading to instability rather than the
topography and material derived properties. The land cover distribution is one of the
external factors that can easily be mapped and monitored in time if needed with the aid
of satellite images (Soeters and van Westen, 1996). In landslide hazard assessment
projects and in environmental and engineering studies, accurate and up to date
information about land cover on a regional scale is often resembles vital information for
help of decision rule generation. Basically a land cover map is obtained by classifying
remotely sensed images. Typically this is performed by the spectral analysis of
individual pixels and their association with other neighboring pixels. The results of
classification depend largely on the type of area, land-cover type, and image acquisition
date. However, the results of the classification are directly affected by spectral confusion
of land-cover types and mixed pixels (Kam, 1995). The vegetation cover and soil
moisture conditions produces distinctive spectral responses in the electromagnetic
spectrum, that gives the opportunity to the classifier to classify them easily. However,
landslides also produce subtle changes in the health of the vegetation, altering the
natural state of the surface and underground drainage conditions, so the soil moisture.
The size of the classified areas regarding the spatial resolution of the Landsat TM5
satellite is often too small to allow interpretation of individual slope instabilities. This
restriction limits the use of spectral data, but incorporates in conjunction with other factor
maps in which they together provide convergent information for slope instabilities.

4.4.1. Input Data

A multispectral Landsat TM5 image of Bolu, acquired on 1990, was used for this
study (Figure 4.19).Other materials used to extract the land cover of Asarsuyu
catchment include:
1. Aerial photographs, panchromatic black-and-white, for 1953 (1:35.000), 1972
(1:25.000), 1984 (1:15.000) and 1994 (1:25.000).
2. Topographic maps for 1994 (scale: 1:25.000) at which the infrastructure
(buildings and roads) of 1994 are digitized from. The data produced are stored
in a GIS database for further integration processes.

71
 a

b
Figure 4.19. a) True color composite of Landsat TM 5 (R=3, G=2, B=1), b) false color
composite of Landsat TM 5 (R=5, G=4, B=1).

4.4.2. Input Map Generation

Following several visits to the study area and detailed aerial photography
interpretations, a land-cover classification scheme was developed. The classification
scheme comprises six land-cover classes that display all the major land-covers
encountered in this area, related to considering the landslide hazard assessment
procedure. First class is the dense forest area. This is the oldest coverage detected in
the study area, since it was observed on 1950’s aerial photographs as a very dense
forest. Second one is the young-age-forest area. In 1970’s, a vast number of forest fires,
removing the dense forest cover occurred in the study area, also with very intense
human deforestation activities (forest industry). The reforestation studies were carried
out after 1980’s. Third class is the mixed group of barren areas, grasslands and

72
agricultural areas. Fourth one is soil moisture class, which should be on the bare land
with significant amount of water in the surface. This could also be said of some form of
wetland where the groundwater is very near to the surface. While the remaining two
(road and settlement) come from GIS 1994 database.
The extraction of land-cover map from Landsat TM5 imagery and subsequent
post classification process consisted of six major steps (Figure 4.20).

Landsat TM5
Multispectral

STEP 1 Georeference of Landsat TM5 imagery

Decorelation Stretching Principal Component Vegetation Indices

Transformation Analysis
Greenness Component of Tasseled Cap Trans.
DC1 PC1 Greenness Component of Gram-Schmidt Trans.
STEP 2 . . GVI
IPVI
. .
. . MND
TVI
DC6 PC6

Normalization of Vegetation Indices STEP 3

PCA for Vegetation

Greenness_Tasseled
Maximum Likelihood Greenness_Schmidt PC1_Vegetation
GVI .
Classifier IPVI .
STEP 5 DC1
MNDVI
TVI
.
STEP 4
DC2 DC4
PC1_Vegetation PC8_Vegetation
Inverse_PC2
PC2_Vegetation

Classification Result

GIS database
STEP 6 Land-cover Map
Road
Settlement

Figure 4.20. The methodological snapshot of land cover extraction scheme

4.4.2.1 Georeferencing

The procedure to rectify the Landsat digital data sets to the national coordinate
system involved the following steps:
1. The determination of the ground control points (GCPs) from 1:25000
topographic maps of 1994 and from the digital image data (Table 4.1)
2. The computation of a least-squares solution for a third-order polynomial
equation required to register the image data sets to the national coordinate
system
3. The re-sampling of the data sets to a 30-m pixel resolution using nearest
neighbor interpolation on the Landsat TM data.

73
 The image was geometrically corrected based on the national grid reference. A
total of twenty-six control points extracted from topographical map were employed in the
georeference. The third degree polynomial equation with nearest neighbor interpolation
was employed in the process.

Table 4.1. RMSE (Root Mean Square Errors) of Ground Control Points

Ground Control Points vs. RMSE (m)

1 23.307 10 11.928 19 17.683

2 18.364 11 11.927 20 17.061

3 16.473 12 18.524 21 5.677

4 20.913 13 18.294 22 20.310

5 3.837 14 9.985 23 16.644

6 11.759 15 19.089 24 15.329

7 17.060 16 17.278 25 17.297

8 10.500 17 17.636 26 15.014

9 21.728 18 15.095 MEAN RMSE 15,3

4.4.2.2 Data Processing

4.4.2.2.1. Principal Component Analysis (PCA)

Extensive inter-band correlation is a problem frequently encountered in the

analysis of multispectral image data. The Principal Components process uses the
principal components statistical technique for reducing dimensionality of correlated
multispectral data. The presence of correlations among the bands of a multispectral
image implies that there is redundancy in the data, which means some information, is
being repeated. It is the repetition of information between the bands that is reflected in
their inter-correlations (Mather, 1999, pp.126-137). As all of the bands show different
electromagnetic signature of the same feature and due to the orientation and range of
the spectral band the information transferred is generally repetitive.
A correlation matrix showed that Band 1-Band 2, Band 1-Band 3 and Band 2-
Band 3 were highly correlated (Table 4.2). This seemed to be a major deficiency of the
Landsat TM5 imageries encountered elsewhere. In order to overcome this problem, a
principal component analysis (PCA) was used to transform the highly correlated Landsat
TM5 data into statistically independent orthogonal axes on which the original satellite
data were reprojected.
The results of the PCA showed that first (52.7 %) and second (41.8 %) principal
components accounted for 94.6 % of the variance within the entire Landsat TM5 data

74
set except the thermal 6th band. Components 3, 4, 5 and 6 respectively accounted for
4.2%, 0.75 %, 0.35 % and 0.12 % of the remaining variance, respectively. To extract
vegetated areas and reduce the amount of computation, inverse of PC 2 (255 – PC 2)
was used in the PCA for vegetation stage, which was obtained by looking at various
combinations of color composite images. The resulting eigenvalues of the principal
components are presented in Table 4.3.

4.4.2.2.2. Decorrelation Stretching

Decorrelation stretching is a color enhancement technique that is based on a

principal component transformation of correlated multispectral image data which
enhances the color display of highly correlated raster sets, such the first three Landsat
TM bands (Table 4.2). Decorrelation stretching was performed by using reverse
transformation of principal components. This enhancement exaggerates the differences
in spectral properties between surface materials to a greater degree than is possible
using conventional contrast enhancement of the original bands. Furthermore, to
enhance the color in highly correlated images, there is a need to selectively exaggerate
the least correlated portion of the spectral data, which is to decrease the correlation.
Decreasing the correlation of spectral data corresponds to exaggerating the color
saturation without changing the distribution of hues (or relative color composition).
The Decorrelation Stretching process involves three fundamental steps. First, a
principal-component transformation is applied with the rows and columns of the
eigenvector matrix transposed. Second, contrast equalization is applied by a Gaussian
stretch, so that histograms of all principal components approximate a Gaussian
distribution of a specified variance. Third, a coordinate transformation that is the inverse
of the principal component rotation is applied so that the data are projected in their
original spectral channels, using eigenvectors as weightings for each principal
component. This inverse operation maximizes the spectral separability of different
surface types in the restored spectral channels. The decorrelation stretched images
created by this process can also be used as components for making color composites.
Gillespie et al. (1987) explained in detail the Decorrelation Stretching technique and its
theoretical and mathematical underpinnings.
In theory, band 1 is sensitive to water and soil moisture while band 2 and
especially band 4 are sensitive to vegetation that is why related decorrelation
components of these original bands represent specific features. As a result of the
contrast stretching the Landsat 5 TM image of Asarsuyu catchment, subtle variations in
surface materials are more easily discriminated by using decorrelated raster set. DC 1 is
sensible to moisture in the area while DC 4 and DC 2 represent different non-common
properties of vegetation through the whole original raster image set (Figure 4.21).

75
Table 4.2.Covariance, Correlation and Transformation Matrices for PCA

Variance / Covariance Matrix

Raster Band1 Band2 Band3 Band4 Band5 Band7
Band1 64.5906 42.3019 68.3945 -36.2414 69.6276 59.9829

Band2 42.3019 30.8834 48.569 -19.1387 53.3876 43.4149

Band3 68.3945 48.569 82.3578 -50.2956 82.4668 73.3751

Band4 -36.2414 -19.1387 -50.2956 306.3435 110.5237 -13.1521

Band5 69.6276 53.3876 82.4668 110.5237 253.4512 117.3291

Band7 59.9829 43.4149 73.3751 -13.1521 117.3291 82.456

Correlation Matrix
Raster Band1 Band2 Band3 Band4 Band5 Band7
Band1 1 0.9471 0.9377 -0.2576 0.5442 0.8219

Band2 0.9471 1 0.963 -0.1968 0.6034 0.8603

Band3 0.9377 0.963 1 -0.3166 0.5708 0.8904

Band4 -0.2576 -0.1968 -0.3166 1 0.3966 -0.0828

Band5 0.5442 0.6034 0.5708 0.3966 1 0.8116

Band7 0.8219 0.8603 0.8904 -0.0828 0.8116 1

Transformation Matrix
Axis Band1 Band2 Band3 Band4 Band5 Band7
PC 1 0.1028 0.0787 0.1174 0.1969 0.3442 0.16

PC 2 0.1497 0.0981 0.1835 -0.3974 0.0336 0.1376

PC 3 0.2276 0.1459 0.178 0.1763 -0.2548 -0.0174

PC 4 0.2783 0.0278 -0.1369 -0.0587 0.1519 -0.3465

PC 5 -0.243 0.1774 0.2511 -0.0119 0.0685 -0.2481

PC 6 -0.0707 0.4846 -0.3089 -0.0174 -0.0201 0.0983

Table 4.3. Eigenvalues and Associated Percentages

Axis Eigenvalues Percentages Cumulative

PC 1 432,3325 52,7182 52,7182

PC 2 343,1456 41,8428 94,561

PC 3 34,5831 4,217 98,778

PC 4 6,1278 0,7472 99,5252

PC 5 2,8701 0,35 99,8752

PC 6 1,0233 0,1248 100

76
Figure 4.21. Decorrelation stretching results (R: decor_4, G: decor_3:B: decor_1).

4.4.2.2.3. Vegetation Indices

The reflectance spectrum of foliage shows a low reflectance (~0.05) in the

visible part of the spectrum while solar irradiance is maximum. Light is absorbed by
vegetation for photosynthesis. In the near infrared (NIR), foliage has a high reflectance
(~0.5), with a very rapid transition between red and NIR regions at ~750 nm. This is
completely different from the reflectance spectrum of the 'background material against
which the leaves are usually observed. Soil reflectance gradually increases at higher
wavelengths over the same region, though its absolute reflectance varies with soil-type
and moisture content (wet soil being darker than dry soil). So the ratio or difference
between two spectral bands on either side of 750 nm will give a measure of the quantity
of foliage present. Those bands are usually chosen centered in the red part of the
spectrum at 660 nm and in the near infrared at 870 nm. Under the light of this fact,
remotely sensed spectral bands can tell user something useful about vegetation by
calculating various indices. Under the light of this fact, remotely sensed multispectral
bands can tell user something useful about vegetation by calculating various indices. A
vegetation index is a number that is generated by some combination of remote sensing
bands and may have some relationship to the amount of vegetation in a given image
pixel. Therefore, six different vegetation indices were calculated to extract different types
of vegetation cover located in the area. These are Greenness component of Tasseled
Cap transformation, Greenness component of Gram Schmidt, GVI, IPVI, MNDVI and TVI.

77
4.4.2.2.3.1. The Tasseled Cap

This index is first defined by Kauth and Thomas (1976) in which it rotates the
MSS data such that the majority of information is contained in two components or
features that are directly related to physical scene characteristics (Lillesand and Kiefer,
1994, pp. 577-579). The index is then reapplied to TM data by Crist and Cicone (1984).
The transformation for the six nonthermal Landsat TM bands (1-5 and 7) computes
three index values: First one is greenness, which is strongly related with the amount of
green vegetation, the second, brightness which is the weighted sum of the all input
bands and defined in the principal variation in the soil reflectance and thirdly wetness is
related to canopy and soil moisture. Most of the variability in soil and vegetation
conditions contained in the six TM bands is expressed in these three dimensions. Each
of these indices is computed cell by cell as a weighted sum (linear combination) of the
input band values. The computation has the form:

Greenness =-0.24717 * TM1 - 0.16263 * TM2 - 0.40639* TM3 + 0.85468 * TM4 + 0.05493 * TM5 -0.11749 * TM7
Brightness =0.33183 * TM1 + 0.33121 * TM2 + 0.55177* TM3 + 0.42514 * TM4 + 0.48087 * TM5 + 0.25252 * TM7
Wetness =0.13929 * TM1 + 0.22490 * TM2 + 0.40359 * TM3 + 0.25178 * TM4 - 0.70133 * TM5 - 0.45732 * TM7

4.4.2.2.3.2. The Gram-Schmidt

The Gram-Schmidt transformation computes the Gram-Schmidt

orthogonalization indices to delineate bare soil from vegetation in TM imagery. Input
consists of a pair of raster objects; the spectral band of one object should be red and the
other photo-infrared. Output also consists of a pair of raster objects: Greenness and Soil
Brightness. The Gram-Schmidt transformation computes Bright Soil coefficients for both
red and infrared rasters based on the parameter values entered for Dark Soil and Bright
Soil (Figure 4.22). The operation computes Green Vegetation coefficients for the red
and infrared rasters from the parameter values entered for Dark Soil and Green
Vegetation (Table 4.4). These coefficients are used in the following formulas to produce
two output rasters, one that displays the amount of green vegetation and another that
displays the amount of soil brightness.

Greenness = 0.5 + (red coeff1 * red cell value) + (NIR coeff1 * NIR cell value)
Soil Brightness = 0.5 + (red coeff2 * red cell value) + (NIR coeff2 * NIR cell value)

78
Table 4.4.Parameters used in the Gram-Schmidt Transformation

Cell value of Red Band Cell value of NIR Band

Dark Soil 19 14

Bright Soil 43 22
Greenness 10 100

Figure 4.22. Near Infra Red versus Red band Raster Correlation Graph

4.4.2.2.3.2. Global Vegetation Index (GVI)

GVI stands for Green Vegetation Index (Terrill, 1994). There are several
GVI’s. The basic way these are devised by using two or more soil points to define a soil
line. The distance of the pixel spectrum in band space from the soil line along the
“greenness” axis is the value of vegetation index.

GVI = -0.2848*TM1-0.2435*TM2-.5436*TM3+0.7243*TM4+0.0840*TM5-0.1800*TM7

4.4.2.2.3.3. Infrared Percentage Vegetation Index (IPVI)

IPVI is the Infrared Percentage Vegetation Index first defined by Crippen (1990).
IPVI is functionally equivalent to NDVI and RVI, but it only ranges in value from 0 to1. It
also eliminates one mathematical operation per image pixel which is important for the
rapid processing of large amounts of data (Terrill, 1994).

IPVI=(NIR)/(NIR+red)= 0.5*(NDVI+1)

4.4.2.2.3.4. Modified Normalized Difference Index (MNDI)

The Modified Normalized Difference operation creates an output raster object

known as the ND or green biomass raster object. The operation shows a measure of the
difference between the values of two input bands RED and NIR. The normalized

79
difference is modified by the values of Path Radiance for Band A, Path Radiance for
Band B, and Minimum Euclidean Distance fields. The expression for the operation is as
follows:

Modified ND = Scale * (NIR - b) * 254 / (NIR + RED - a - b) + 1

(a and b are path radiances for RED and NIR, respectively).

4.4.2.2.3.5. Transformed Vegetation (TVI)

The TVI operation computes the Transformed Vegetation Index from a pair of
input raster objects. The spectral band of one object should be RED and the other NIR.
The expression for the operation is as follows (Lillesand and Kiefer, 1996):

TVI = 100 * ( sqrt[(NIR-Red) / (NIR+Red)] + 0.5)

4.4.2.4. Normalization

The values of pixels in all vegetation indices fall within a range of –1 to +1.
Therefore, they were all transformed into 0-255 range to be put into PCA. For
normalization process, the following formula was applied:

Normalized DN = [(Input DN – Min DN) / (Max DN – Min DN)] * Scale Factor

Scale Factor = 255

4.4.2.5. PCA for Vegetation

The land cover class vegetation can be made up of several land use classes. It
can be natural vegetation, forestry or agricultural vegetation. Within these classes
vegetation can be separated based on species, biomass, diseases and other. To obtain
different species distribution and health conditions, there is a need to obtain
components, which represent different vegetation types and their properties. However,
having 8 different factors resulted as most of the data as being redundant, conserving
the species information and removing the redundant information PCA was applied to
normalized vegetation indices, DC 4 and inverse of PC 2 (Table 4.5).
The results of the PCA for vegetation showed that the first principal component
accounted for 75.6 % of the variance within the entire data set. Components 2, 3, 4, 5,
6, 7 and 8 respectively accounted for 20.6 %, 2.09 %, 0.73 %, 0.41 %, 0.28 %, 0.13 %
and 0.06 % of the remaining variance (Table 4.6.). Thus 96.3 % of the total variance the
eight components were explained by the first two principal components.

80
Table 4.5. Covariance, Correlation and Transformation Matrices of PCA for Vegetation

Variance / Covariance Matrix

Raster DC 4 Greenness_T Greenness_G GVI IPVI MODIFIED_ND PC 2 INVERSE TVI
DC 4 1622.3652 587.3567 650.7342 419.4895 616.597 638.939 403.7972 395.832

Greenness_T 587.3567 297.3846 309.1295 237.7942 481.9862 516.4237 231.7356 331.5238

Greenness_G 650.7342 309.1295 337.117 242.5585 456.0567 487.8237 232.8455 311.6548

GVI 419.4895 237.7942 242.5585 238.0185 510.2078 486.4798 225.6202 307.4057

IPVI 616.597 481.9862 456.0567 510.2078 1372.325 1326.2998 499.0305 836.5095

MODIFIED_ND 638.939 516.4237 487.8237 486.4798 1326.2998 1460.6291 491.7419 921.0824

PC 2 INVERSE 403.7972 231.7356 232.8455 225.6202 499.0305 491.7419 237.6426 311.0036

TVI 395.832 331.5238 311.6548 307.4057 836.5095 921.0824 311.0036 607.6737

Correlation Matrix
Raster DC 4 Greenness_T Greenness_G GVI IPVI MODIFIED_ND PC 2 INVERSE TVI
DC 4 1 0.8456 0.8799 0.6751 0.4132 0.4151 0.6503 0.3987

Greenness_T 0.8456 1 0.9763 0.8938 0.7545 0.7836 0.8717 0.7799

Greenness_G 0.8799 0.9763 1 0.8563 0.6705 0.6952 0.8227 0.6886

GVI 0.6751 0.8938 0.8563 1 0.8927 0.8251 0.9487 0.8083

IPVI 0.4132 0.7545 0.6705 0.8927 1 0.9368 0.8738 0.916

MODIFIED_ND 0.4151 0.7836 0.6952 0.8251 0.9368 1 0.8347 0.9777

PC 2 INVERSE 0.6503 0.8717 0.8227 0.9487 0.8738 0.8347 1 0.8184

TVI 0.3987 0.7799 0.6886 0.8083 0.916 0.9777 0.8184 1

Transformation Matrix
Axis DC 4 Greenness_T Greenness_G GVI IPVI MODIFIED_ND PC 2 INVERSE TVI
PC 1_Veg 0.1532 0.0889 0.089 0.0801 0.1888 0.1964 0.0791 0.1245

PC 2_Veg 0.3895 0.0673 0.0981 0.0101 -0.1571 -0.1652 0.0046 -0.108

PC 3_Veg -0.0082 -0.0509 -0.0469 0.1409 0.2723 -0.2066 0.1077 -0.1664

PC 4_Veg -0.1458 0.134 0.2029 0.1399 -0.111 -0.0752 0.153 0.0383

PC 5_Veg -0.0493 0.1118 0.2109 -0.0354 0.1411 -0.0987 -0.3414 0.0114

PC 6_Veg -0.0492 0.0071 0.1321 0.025 -0.0335 0.2873 -0.0233 -0.4425

PC 7_Veg -0.0235 0.0658 0.1165 -0.407 0.1026 -0.0525 0.1955 -0.0367

PC 8_Veg 0.0307 -0.522 0.3237 0.0099 -0.0033 0.0008 0.0248 0.0849

Table 4.6. Eigenvalues and Associated Percentages

Axis Eigenvalues Percentages Cumulative %
PC 1_Veg 4670,2763 75,6546 75,6546

PC 2_Veg 1274,0529 20,6386 96,2932

PC 3_Veg 129,2741 2,0941 98,3873

PC 4_Veg 45,2019 0,7322 99,1196

PC 5_Veg 25,0813 0,4063 99,5259

PC 6_Veg 17,5736 0,2847 99,8105

PC 7_Veg 7,7659 0,1258 99,9363

PC 8_Veg 3,9298 0,0637 100

81
4.4.3. Maximum Likelihood Classification

Supervised classification began with the selection of training areas for each of the
land-cover classes. Spectral signatures were generated from the training-area. The
signature files were employed in a maximum-likelihood classifier to lace all pixels in one
of four land-cover classes. In the classification four different data were used. These are:

i. Two decorrelation components (DC 1 and DC 2)

ii. Two PCA-vegetation components (PC 1-vegetation and PC 2-vegetation)

These four data sets represent different unique spectral signatures of features in four-
dimensional spectral space. DC 1 was selected for extracting soil moisture. DC 2 was
selected to differentiate soil and vegetation types, which have spectrally same response.
Because study area is dominantly covered by vegetation. PC 1-Vegetation and PC 2-
Vegetation are also added, which are used to indicate different spectral responses of
vegetation, into supervised classification (Figure 4.23).

Figure 4.23. Product of Maximum Likelihood Classification

4.4.3.1. Accuracy Assessment

Classification accuracy was determined by comparing a sample of classified

pixels with ground information derived from aerial photographs and field data
(Congalton, 1991). To determine the accuracy of classification, approximately 29
reference areas (226 pixels) were selected as reference data for the comparison of
ground information with the classification result. These pixels had to be pure rather than
mixed pixels to ensure that the correct land cover was identified for each pixel. As with
the training pixel, they were chosen with the aid of 1:15000, 1:25000 and 1:35000 scale

82
aerial photographs of the study area. For each pixel, the ground information determined
from the aerial photographs (field checking when necessary) was compared with the
classification results by means of confusion matrices. The total number of correct pixels
in a category is divided by the total number of pixels of that category as derived from
ground information (i.e., the column total). This accuracy measure indicates the
probability of a reference pixel being correctly classified. This accuracy measure is often
called “producer’s accuracy”, because the producer of the classification is interested in
how well a certain area can be classified. On the other hand, if the total number of
correct pixels in a category is divided by the total number of pixels that were classified in
that category, then this result is a measure of commission error. This measure, called
“user’s accuracy” or reliability, is indicative of the probability that a pixel classified on the
map or image actually represents that category on the ground (Lunetta et al., 1991).
User’s accuracy of dense forest is 100 per cent, which is perfect. Although
user’s accuracy of both young-age forest and soil moisture looks low, there are several
factors, which creates complexity to differentiate these classes. For example, even by
looking at aerial photographs, it is also difficult to differentiate the young-age forest from
old and dense forest at 1990’s. For soil moisture 7 pixels out of 56 is classified as mixed
class, which is not very abnormal since they are associated with each other (Table 4.7.
In the mixed group 6 out of 83 pixels were misclassified as forest as the mixed group
itself by definition also contains some scattered very small patches of dense trees.
However, the overall accuracy (Kappa Coefficient) of the classification is 92.48 %, which
indicates that the classification is within the desired confidence level.

Table 4.7 Error Matrix of the Classification

4.4.4 Integration of RS, GIS at Database Level

In the database level of integration, analysis are performed with the integrated
vector features, database and raster products. The integration of remote sensing
information into a GIS occurs naturally in a Raster format (or Raster GIS) because both
data structures are approximately the same. Integration into a vector system requires

83
somewhat more effort, updating vector information by using an image as a backdrop for
vector editing (Faust et al., 1991).
Road network and building layers were first converted into raster format. Result
of maximum likelihood classification was used as a reference data in this conversion. In
raster map of building, DN of 100 represents buildings while DN of 250 indicates other
features located in the area. In raster map of road, DN of 200 represents road while DN
of 250 indicates other features. DN of Land cover obtained from supervised
classification are indicated as follows:

1 : Forest Area
2 : Young-age Forest Area
3 : Soil Moisture
4 : Mixed Group

Logical and Relational Raster Overlay Analysis:

if ( Building_Layer = = 100 )
{
Landcover = 5 ( DN of 5 shows the Buildings in the product of this process)
}
else if ( Road_Layer = = 200 )
{
Landcover = 6 ( DN of 6 shows the Road in the product of this process)
}
else
{
Landcover = Classification_Result
}
Finally, the product of this process includes the additional two classes (Buildings
and Road) (Figure 4.24). This is also called integration results to GIS / Attribute Update
(Archibald, 1987). As a result the majority of the area (44%) is covered by dense forest,
%33 is mixed class with agriculture grassland and 16% is covered by young forest
(Figure 4.24.b)
It is noticeable that the classification accuracy is significantly higher for
vegetated areas than for other classes. For example, forest area and mixed group were
most accurately classified at levels of 100 per cent and 91.57 percent, respectively.
Although soil moisture and young-aged forest areas classes showed more
errors of commission, forest and mixed group represented more errors of omission. The
errors of commission were largely caused by confusion between young-age forest and
old - dense forest, soil moisture and mixed group, while the errors of omission resulted
from confusion between old-dense forest and mixed group, mixed group and soil
moisture Indeed, confusion between mixed group and soil moisture has the highest
percentage in the total classification error. This is because within the mixed classes
there may be green vegetation and/or agricultural fields, which reflect same spectral
signature with soil moisture.

84
 2%

Dense Forest
33%
44% Young Forest
Moisture
Mixed
Settlement
5%
16%

Figure 4.24.a) Land-cover map of the study area, b) areal distributions of land cover.

4.5. Landslide Inventory

In this section the creation of landslide attribute databases are explained. Four
landslide attribute databases are created using the photo interpretation of the stereo
panchromatic aerial photographs.

4.5.1. Input Data

The input data for landslide attribute databases and the inventories of four
different time periods consist of 4 sets of analogue stereo panchromatic aerial
photographs. The time periods and scales of the photographs are as follows: 1952
(1:35.000), 1972 (1:25.000), 1984 (1:15.000) and 1994 (1:25.000).

85
4.5.2. Data Production

The landslide occurrences of the Asarsuyu catchment are mapped in four

different time periods. The geomorphological and morphometrical attributes of the
landslides are mapped and they are transferred into each years corresponding
topographical map in order to minimize the georeference residual errors in a considerable
amount. Furthermore, a relational database is constructed and attached into the landslide
map (Figure 4.25). The attribute database of the each landslide consists of 7 attributes of
which they are named as Massinfo, Type, Style, Depth, Landcover, and Distribution of
activity and Possible Cause.
“Massinfo” attribute is the morphology of the slope instability seen in the
photograph; it has two terms that can be used “Scarp&Path” and “Scarp&Body”.
“Scarp&Path” is used to define the landslides where scarp is clearly visible and the
direction of movement is easily estimated although there are no signs of body. On the
other hand, the latter “Scarp&Body” is assigned to the landslides where it’s scarp and
body is clearly seen.
“Type” attribute is the class of the landslide according to the Varnes (1978)
classification. The available items that can be used in the database are confined only to
“slide”, “flow” and “flow/slide (complex)” due to the occurrence in the study area. The
“slide” and “flow” is straightforward and self-explanatory, however the “flow/slide
(complex)” is used for the slope instabilities that are started as a slide and then the
displaced mass is flowed out remaining only a visible scarp.
“Style” attribute is defined by “the way in which different movements contribute to
the landslide” (Varnes, 1978) and the definitions used in this study is based accordingly on
the report of Unesco Working Party on World Landslide Inventory, 1993. The terms that
can be used for the study area from the available terms are “successive”, “single” and
“multiple”. The “single” slides consist of a single movement of displaced material; the
“multiple” movements are landslides with repeated development of the same type of
movement. The definition of “successive” is “A successive movement is identical in type to
an earlier movement but in contrast to a multiple movement does not share displaced
material or a rupture surface with it”.
“Depth” attribute is assigned as the relative measure obtained from the
interpretation of the aerial photographs, which is significantly dependent on the experience
of the interpreter.
“Distribution of activity” attribute is self explanatory as is assesses the activity
distribution of a landslide. The available terms are “advancing”, “retrogressive”, “diminishing”
and “moving” and the definitions used in this study is based accordingly on the report of
Unesco Working Party on World Landslide Inventory, 1993. The “retrogressive” is defined as
the movement is continuing by the extension of the rupture surface in the direction opposite

86
to the movement of displaced material”. If the rupture surface is extending in the direction of
movement it is said to be “advancing”. The term “Diminishing” is used for a landslide whose
displacing material is decreasing in volume. Landslides whose displaced material continue
to move but rupture surfaces show no visible changes can be simply described as “moving”.
The “land cover” attribute is directly taken from the aerial photographical
interpretation, dependent on the study area characteristics, and five classes have been
defined. These are “bare land”, “grass land”, agricultural fields”, “dense forest” and “young
forest”.

Figure 4.25. The landslide attribute database

The produced historical landslide inventory maps are presented in Figure 4.26. For
every landslide inventory map a polygon topology is built and validated, in order to yield area
(Polystats) and shape dependent (Fuzzystats) attribute databases. The photo interpretation
database, Polystats and Fuzzystats databases are explored in detail for every period in the
next chapter.
After compiling the four years inventory databases, all of them are merged together
to form a single training resource for statistical analyses in order to figure out the parameters
controlling the instability. If a slide occurs in any one of the periods it is included in the
merged landslide map. Following the merge process a buffer of 100 meters is calculated
around the landslide polygon and manually edited just to be buffered in the crown and flank
areas of the polygons. The buffer is calculated in order to get the pre-failure surface
attributes of the slides. The decision rule for the boundaries of the buffer line is as follows: If
the distance between the slide boundary and microcatchment divide line is smaller than 100
meters then use the microcatchment divide line, if the distance is larger than 100 meters
then use the 100 meter buffer line for the seed zone generation. After finishing the merge
and buffering processes, a polygon grid of 25 meter pixel size is overlaid over the landslide
polygons. The midpoints of this polygon grid are calculated and the attributes of landslides
and the previously produced input maps in this chapter are transferred as separate attribute
tables to these points (Figure 4.27). These separate attribute tables are then merged to
construct a huge relational database concerning all of the parameters, which are going to be

87
used for later statistical model creation. Each parameter map will now be treated as a new
variable in this database.

d
Figure 4.26. Landslide inventories of the four time periods, a. 1952, b. 1972, c. 1984, d.
1994

88
Figure 4.27. Polygon grid midpoints and transferred attributes (white points are merged
polygons, black ones are buffered zone)

89
 CHAPTER 5

LANDSLIDE DATABASES

This chapter will deal with the characteristic features and the descriptive
statistics of the landslides mapped. The aim of this data exploration is to explore the
training data, which will be used in the further analyses and to characterize the
landslides and their evolutions in all of the four periods starting from 1952 to 1994.

5.1. Topological, Morphometrical database

In this section a brief description of the morphometrical characteristics of the

landslides will be given. In order to do this, the vector topology dependent statistical
attribute table/database (Polystats) and the fuzzy properties attribute table/database
(Fuzzystats) will be used. The topological database is consisted of 11 variables
originally, however 4 of them are valid if the polygon is hosting an island polygon, 2 of
them are related with the spatial position of the polygon, which will be stressed in the
following chapters, and 1 of them is a derived variable from Fuzzystats database,
therefore, only 4 variables from the Polystats database will be used in order to trim the
redundancy. The Polystats variables and their definitions are presented in Table 5.1.

Table 5.1. The names and definitions of variables used from Polystats database.

NAME DEFINITION
Area The area of a polygon.
BoundLen The total length of a polygon's boundary.
MaxDim The length of the longest diagonal from and to all polygon angles.
Roughness is a measure of the irregularity of a polygon's perimeter.
Roughness
It is calculated as: (MaxDim * BoundLen) / Area.

In the Fuzzystats table, 12 parameters are available; the names definitions and
formulas are given in Table 5.2. These fuzzy properties are reflecting the measures of
shape, that in the scope of investigation. A shape is a difficult property to measure or
define precisely and mathematically. It is difficult, if not impossible, to construct a

90
measure which is unique to a single shape. There have been attempts to characterize a
variety of shapes including simple forms, like sand grain shapes, to very complex forms
as are indicative of fossilized organisms. From a larger or more regional perspective the
evaluation of shapes can be applied to drainage basins, coral atolls, salt diapirs, oil
fields, or structural traps. Based on these, if the similarly shaped objects are present in a
vector object, characteristic shape measure values associated with a known shape
should be defined. All of the available variables are given in Table 5.2, but “thinness
ratio”, “Normalized Dispersion”, “Simplicity”, Shore line Development” will not be
implemented for analyses as they are directly dependent of the digitizing process and
even some of them undefined for the landslide polygon vector object, such as there are
no sliver polygons, or the distances in between boundary vertices are equal. Also the
“correlation” and “orientation” variables are not implemented as “orientation” yields a bi-
directional value which is not applicable to landslides, as it is not concerning the
orientation of the landslide but the orientation of the polygon without any thematic and
genetic information is concerned. This means that the value is only dependent on the
orientation of a physical long axis, which may not be in the same direction with the
landslide or the downslope gradient. Moreover, the “correlation” is a dummy variable
that is used for the calculation of “orientation” variable, yet it is also skipped. Only 6 of
the Fuzzystats variables out of 12 are explored to find and evaluate shape dependent
differences or similarities in the area from 1952 to 1994.

5.1.1. Polystats Database

5.1.1.1. 1952 Period

33 landslides have been mapped from 1:35.000 scale panchromatic stereo

aerial photographs. The minimum area recorded is 3175,15 square meter and the
maximum is 817501,36 square meters, having a mean of 79687,86 and the standard
deviation of the distribution is 153440,75 (Table 5.3). As for the sake of reader’s agility,
the general definitions and rules of thumbs of skewness and kurtosis parameters will be
introduced briefly below.
Skewness is “A measure of the asymmetry of a distribution”. The normal
distribution is symmetric, and has a skewness value of zero. A distribution with a
significant positive skewness has a long right tail. A distribution with a significant
negative skewness has a long left tail. As a rough guide, skewness values more than
twice its standard error is taken to indicate a departure from symmetry.
Kurtosis is ”A measure of the extent to which observations cluster around a
central point”. For a normal distribution, the value of the kurtosis statistic is 0. Positive

91
kurtosis indicates that the observations cluster more and have longer tails than those in
the normal distribution and negative kurtosis indicates the observations cluster less and
have shorter tails.

Table 5.2. The names, definitions and formulas of fuzzy properties

Name Definition Formula

Measures shape with a maximum value 2
Form Ratio FR = Area / (Long Axis )
approximating one for squares
Measures shape with a minimum value
Grain Shape approximating two for long skinny polygons,
GSI = Perimeter / Long Axis
Index pi for circles, four for squares, and may be
very large value for fractals.
Measures shape with a maximum value
Compactness C= 2 * sqrt (pi * Area) / Perimeter
approximating one for circles.
Thinness 2
Measures the shape of sliver polygons. TR= 4 * pi * Area / (Perimeter )
Ratio
Measures shape, reflecting the element's
2
Circularity 1 similarity to a circle, with a maximum value C1= sqrt (Area / (pi * MaxRadius )),
approximating one for circles.
Measures shape, reflecting the element's
Circularity 2 similarity to a circle, with a maximum value C2= sqrt (MinRadius / MaxRadius)
approximating 1.0 for circles.
Normalized Measures shape presenting a value
ND= pi * WeightedRadiusSum / Area)
Dispersion approximating one for circles.
Measures a shape's simplicity, used. in
separation of polygons with the same shape 2
Simplicity S= MeanSegmentLength / Area
but with a different number of boundary
vertices
Measures polygon generalization, which is
Shore Line used in cartography for studying the
SLD= Perimeter / (2 * sqrt (pi * Area))
Development relationship between measured distance and
map scale.
Cor=the correlation coefficient between
Measures the complexity and integrity of the
Correlation the X and Y coordinates of the
polygon boundary shape.
boundary
Orientation Measures a polygon's orientation. O= arccos (Cor)
Elongation Measures polygon proportions. E= ShortAxis / LongAxis,

Based on the definitions, above the area variable of 1952 has a long right tail,
and it clusters around smaller values with longer tails than a normal distribution (Table
5.3 and Figure 5.1.a). The distribution type is not crucial here and will not be
implemented as for the original landslide polygons the population size is quite small
such as N=33 in this case. However, just to test the validity of skewness and kurtosis
rules, a Q-Q plot of normality test is applied and seen that the distribution is far away
from a normal distribution (Figure 5.1.b).

92
Table 5.3 The descriptive statistics of Polystats 1952.
Descriptive Statistics of Variables of Polystats Database (1952)
AREA BOUNDLEN MAXDIM Roughness
Valid 33 33 33 33
N
Missing 0 0 0 0
Mean 79687,8645907315 1081,3042428379 398,8217684336 7,8582125715
Std. Error of Mean 26710,6068250663 183,9006305842 53,4735059243 ,7545948866
Median 29514,2572021500 798,1903720200 293,1430513500 6,8409510700
Mode 3175,15747070a 220,57400666a 84,03677303a 4,86070260a
Std. Deviation 153440,7542336397 1056,4286931289 307,1819047120 4,3348175988
Variance 23544065059,78822000 1116041,5836660140 94360,7225824886 18,7906436150
Skewness 4,044 2,924 1,875 4,247
Std. Error of Skewness ,409 ,409 ,409 ,409
Kurtosis 17,847 9,625 3,334 21,019
Std. Error of Kurtosis ,798 ,798 ,798 ,798
Range 814326,21105957 5210,71663773 1247,00648479 24,81103996
Minimum 3175,15747070 220,57400666 84,03677303 4,86070260
Maximum 817501,36853027 5431,29064439 1331,04325782 29,67174256
Sum 2629699,53149414 35683,04001365 13161,11835831 259,32101486
a. Multiple modes exist. The smallest value is shown

30
Normal Q-Q Plot of AREA
FREQUENCY

500000
25
400000
Expected Normal Value

20 300000

200000
15
100000

10 0

Std. Dev = 153440,8 -100000

5
Mean = 79688 -200000
0 N = 33,00 -300000
-400000 0 400000 800000
25

12 0

22 00

32 00

42 00

52 00

62 00

72 00

82 00
00

50

50

50

50

50

50

50

50

-200000 200000 600000 1000000

00

AREA Observed Value

a b

Figure 5.1. The distribution plots of area of 1952: a) Frequency distribution, b) Normal
Q-Q Plot.

The boundary lengths are ranging from 220.57 meters to a maximum of

5431.29 meters. The mean is 1081.30 meters and the standard deviation is 1056.42.
The skewness and kurtosis suggest that the distribution is not normally distributed and
having long right tails yielding in an asymmetrical histogram with a clustering behavior
(Figure 5.2.a). The MaxDim is ranging from 84.03 meters up to 1331.04 meters with a
mean of 398.82 meters and a standard deviation of 307.18. It has also a long right tail
with values clustering around mean values (Figure 5.2.b). The roughness has a
minimum of 4,86 and a maximum of 29.67 with a mean of 7.85 and standard deviation
of 4.33. It has also long right tail and an asymmetrical distribution (Figure 5.2.c).

93
 16 10 20

FREQUENCY

FREQUENCY

FREQUENCY
14
8
12 15

10
6

8 10
4
6

4
2 5
Std. Dev = 1056,43 Std. Dev = 307,18
2 Std. Dev = 4,33
Mean = 1081 Mean = 399
Mean = 8
0 N = 33,00 0 N = 33,00
0 N = 33,00
0

10

20

30

40

50

50

25

45

65

85

10

12
5 10 15 20 25 30
00

00

00

00

00

50

50
BOUNDLEN MAXDIM ROUGHNESS

a b c

Figure 5.2. Histograms of Polystats database, a) “BoundLen”, b) “MaxDim”, c)

“Roughness”.

The correlations of the variables in the Polystats table are calculated and found
that the “MaxDim” and “BoundLen” within themselves has the best correlation coefficient
(0.956), also the “area” has good correlations with “BoundLen” and “MaxDim”, 0.841
and 0.817, respectively. However, none of the variables are correlatible with
“roughness”, as the highest correlation coefficient of “roughness” is 0.423 with “MaxDim”
(Table 5.4). Their scatterplots are presented in Figure 5.3.

Table 5.4. Correlations of Polystats variables 1952.

AREA BOUNDLEN MAXDIM ROUGHNESS
Pearson Correlation 1,000 ,841** ,817** -,012
Sig. (2-tailed) , ,000 ,000 ,947
AREA Sum of Squares and Cross-products 753410081913,223 4363532624,387 1231861404,435 -255540,974
Covariance 23544065059,788 136360394,512 38495668,889 -7985,655
N 33 33 33 33
Pearson Correlation ,841** 1,000 ,956** ,243
Sig. (2-tailed) ,000 , ,000 ,173
BOUNDLEN Sum of Squares and Cross-products 4363532624,387 35713330,677 9931185,608 35649,475
Covariance 136360394,512 1116041,584 310349,550 1114,046
N 33 33 33 33
Pearson Correlation ,817** ,956** 1,000 ,423*
Sig. (2-tailed) ,000 ,000 , ,014
MAXDIM Sum of Squares and Cross-products 1231861404,435 9931185,608 3019543,123 18023,079
Covariance 38495668,889 310349,550 94360,723 563,221
N 33 33 33 33
Pearson Correlation -,012 ,243 ,423* 1,000
Sig. (2-tailed) ,947 ,173 ,014 ,
ROUGHNESS Sum of Squares and Cross-products -255540,974 35649,475 18023,079 601,301
Covariance -7985,655 1114,046 563,221 18,791
N 33 33 33 33
**. Correlation is significant at the 0.01 level (2-tailed).
*. Correlation is significant at the 0.05 level (2-tailed).

94
 A 1600 1600
Linear Regression Linear Regression
5000 Linear Regression
maxdim = 98.13 + 0.28 * boundlen
boundlen = 619.78 + 0.01 * area maxdim = 268.53 + 0.00 * area R-Square = 0.91
A A
R-Square = 0.71 R-Square = 0.67
A A
4000 A
1200 1200
A A
boundlen

maxdim

maxdim
3000
800 800
A A A

2000 A A
A A A
A A
A AA A
A
A A A
A A A
A A
AA 400 AA
A
400 AA
A
A
A
A A
1000 A
AA AA AAA
AA AA AA
AA
A AAA A
AA
A
A A
A AA
A
AA
A A
A A
A
A
A A
A A
A
A
A A A
A A

0 250000 500000 750000 0 250000 500000 750000 1000 2000 3000 4000 5000

area area boundlen

a b c

Figure 5.3. Scatter plots of highly correlated variable pairs of Polystats database, a)
”area-BoundLen”, b) “area-MaxDim”, c) “BoundLen MaxDim”.

The correlations are quite logical as the “area” is closely related with the
perimeter (“BoundLen”) of a polygon, so with the maximum diagonal (“MaxDim”). As the
“area” increases the two other parameters should have to be increased also. On the
other hand, “roughness” is a derived component from the above three variables, but it
has very low correlation, which implies that the irregularity of the polygons are not
affected with how long or how large the polygon is.

5.1.1.2. 1972 Period

45 landslides have been mapped from 1:25:000 scale panchromatic stereo

aerial photographs. The previous period’s vector data set was used as a precursor for
the interpretation, every landslide in the 1952 database checked in the 1972 period. A
significant land-use change is observed in the study area; quite large areas are
deforested either by human abuse of the forests as an economical source or by
extensive forest fires.
The “area” variable has a minimum value of 2027.52 square meters, and the
maximum is 786179.14 square meters (Table 5.5.). The distribution has a mean of
64906.05 and a standard deviation of 129606.99. Similar to 1952 period the “area”
variable is asymmetrical and has long right tail. Although only 5 out of 45 case create
this long right tail, similar results with less positive skewed distribution is obtained when
the last 5 is omitted in the histograms (Figure 5.4).
The “BoundLen” variable is ranging from 176.39 to 5502.14 meters with a mean
of 964.23 and a standard deviation of 950.44 (Table 5.5). The kurtosis and skewness
values suggest a long right tail with clustering around smaller values. The “MaxDim”
variable has a minimum of 65.51 meters and the maximum of this variable is 1328.03
meters. The variables distribution suggests a less positively skewed and less clustered

95
distribution rather than “area” and “BoundLen” variables with a mean of 357.15 and a
standard deviation of 277.49 (Figure 5.4). For the “roughness” variable the minimum
maximum range is defined as 4.9 and 30.68 with a mean of 8.11 and standard deviation
of 4.24. The distribution has again a right long tail and clustering around small values.

Table 5.5. The descriptive statistics of Polystats 1972.

Descriptive Statistics of Variables of Polystats Database (1972)

AREA BOUNDLEN MAXDIM ROUGHNESS

Valid 45 45 45 45
N
Missing 0 0 0 0
Mean 64906.0579833980 964.2372695904 357.1536664922 8.1114281029
Std. Error of Mean 19320.6706300324 141.6841362389 41.3659686927 .6332609391
Median 21517.4001464800 684.0998115000 264.8689724500 6.6746764100
Mode 2027.52832031a 176.39683459a 65.51217920a 4.90223315a
Std. Deviation 129606.9986989081 950.4460798904 277.4913538558 4.2480435223
Variance 16797974111,73878000 903347.7507790310 77001,4514647170 18.0458737673
Skewness 4.520 3.268 2.102 3.792
Std. Error of Skewness .354 .354 .354 .354
Kurtosis 23.029 12.706 4.614 18.215
Std. Error of Kurtosis .695 .695 .695 .695
Range 784151.61450196 5325.74997547 1262.52586739 25.77959581
Minimum 2027.52832031 176.39683459 65.51217920 4.90223315
Maximum 786179.14282227 5502.14681006 1328.03804659 30.68182896
Sum 2920772.60925291 43390.67713157 16071.91499215 365.01426463
a. Multiple modes exist. The smallest value is shown

30 30
FFREQUENCY

FREQUENCY

20 20

10
10

Std. Dev = 129607.0

Std. Dev = 950.45
Mean = 64906
Mean = 964
0 N = 45.00
0 N = 45.00
25000 225000 425000 625000 825000
0

50

10

15

20

25

30

35

40

45

50

55

125000 325000 525000 725000

0

00

00

00

00

00

00

00

00

00

00

AREA BOUNDLEN

a b
15 25
FREQUENCY

FREQUENCY

20

10

15

10

Std. Dev = 277.49

Std. Dev = 4.25
Mean = 357
Mean = 8
0 N = 45.00
0 N = 45.00
10

30

50

70

90

11

13

10

12

14

16

18

20

22

24

26

28

30
0

00

00

MAXDIM ROUGHNESS

c d

Figure 5.4. The frequency distributions of variables of 1972 Polystats database, a) area,
b) BoundLen, c) MaxDim, d) roughness.

96
 The correlations are also similar with the period of 1952 as the maximum
correlation is in between “MaxDim” and “BoundLen” (0.957) and the other two best
correlations are in between pairs “area”-“BoundLen” (0.837) and “area”-“MaxDim”
(0.815). Likewise, the roughness has the minimum correlation with the other variables
(Table 5.6 and Figure 5.5).

Table 5.6. Correlations of Polystats variables 1972, the best correlations are shown in
bold.
AREA BOUNDLEN MAXDIM ROUGHNESS
Pearson Correlation 1.000 .837** .815** -.046
Sig. (2-tailed) . .000 .000 .763
AREA Sum of Squares and Cross-products 739110860916.506 4536214108.6 1290093036.0 -1117680.682
Covariance 16797974111.739 103095775.19 29320296.273 -25401.834
N 45 45 45 45
Pearson Correlation .837** 1.000 .957** .188
Sig. (2-tailed) .000 . .000 .215
BOUNDLEN Sum of Squares and Cross-products 4536214108.564 39747301.034 11102333.528 33464.256
Covariance 103095775.195 903347.751 252325.762 760.551
N 45 45 45 45
Pearson Correlation .815** .957** 1.000 .349*
Sig. (2-tailed) .000 .000 . .019
MAXDIM Sum of Squares and Cross-products 1290093036.019 11102333.528 3388063.864 18084.486
Covariance 29320296.273 252325.762 77001.451 411.011
N 45 45 45 45
Pearson Correlation -.046 .188 .349* 1.000
Sig. (2-tailed) .763 .215 .019 .
ROUGHNESS Sum of Squares and Cross-products -1117680.682 33464.256 18084.486 794.018
Covariance -25401.834 760.551 411.011 18.046
N 45 45 45 45
**. Correlation is significant at the 0.01 level (2-tailed).
*. Correlation is significant at the 0.05 level (2-tailed).

A
1600 1600
Linear Regression
5000 Linear Regression Linear Regression
boundlen = 565.88 + 0.01 * area
R-Square = 0.70 maxdim = 243.86A + 0.00 * area maxdim = 87.82 + 0.28 * boundlen A
R-Square = 0.66 R-Square = 0.92
A A
4000 1200 1200
A
A A
boundlen

maxdim

maxdim

3000
800 A 800 A
A

A A A
2000 A AA
A
A A AA AA
A A A
AA AA
A
AA 400 A A 400 AA
A A A AA
AA
A A A
A
1000 AAA AAA
AAA A
A
A A
A
AA A AA
A
AA
A AA A
AA
A A
A
A
A
A A
AA
A
AA
A
A
A A A
AA
A A
A A
A
A
A A
A A
A
A
A
A
A A A
0
0 250000 500000 750000 0 250000 500000 750000 0 1000 2000 3000 4000 5000

area area boundlen

a b c

Figure 5.5. Scatter plots of highly correlated variable pairs of Polystats database (1972),
a) ”area”-“BoundLen”, b) “area”-“MaxDim”, c) “BoundLen”-“MaxDim”.

97
5.1.1.3. 1984 Period

Only 39 landslide bodies are mapped from 1:15.000 scale panchromatic stereo
aerial photographs of 1984 period. Although the scale of the photograph is much more
adequate for landslide monitoring, the number of landslides mapped is decreased
probably due to intensive re-forestation studies in the study area.
The “area” variable ranges in between 4815.04 and 956223.4 square meters
(Table 5.7). The distribution is again having a long right tail and a mean 74395.84 and a
standard deviation of 161246.42 (Figure 5.6). The “BoundLen” variable has a minimum
of 278.82 and a maximum of 5412,4 meters. It is again positively skewed and the
kurtosis value indicates a clustering in smaller values. The mean and the standard
deviation is 1047.86 and 984.93, respectively. The “MaxDim” of landslides are ranging
between 113.96 and 1307.1 meters with a long right tail and clustering in the smaller
values. The mean is found to be 393.5 meters and the standard deviation is 285.31.

Table 5.7. The descriptive statistics of Polystats 1984

Descriptive Statistics of Variables of Polystats Database (1984)

AREA BOUNDLEN MAXDIM ROUGHNESS

Valid 39 39 39 39
N
Missing 0 0 0 0
Mean 74395.8434682997 1047.8647222372 393.5067514408 8.8682533931
Std. Error of Mean 25820.0918495585 157.7151385261 45.6874778206 .6695358135
Median 24466.1716308600 739.9429591700 327.1816415600 7.8408232500
Mode 4815.04296875a 278.82556722a 113.96302751a 5.14394527a
Std. Deviation 161246.4219189557 984.9307244129 285.3182075415 4.1812498149
Variance 26000408581.6658900000 970088.5318925580 81406.4795546676 17.4828500148
Skewness 4.755 3.096 2.049 3.245
Std. Error of Skewness .378 .378 .378 .378
Kurtosis 24.886 11.056 4.288 14.631
Std. Error of Kurtosis .741 .741 .741 .741
Range 951408.35791016 5133.58286545 1193.14480709 24.13985703
Minimum 4815.04296875 278.82556722 113.96302751 5.14394527
Maximum 956223.40087891 5412.40843267 1307.10783460 29.28380230
Sum 2901437.89526369 40866.72416725 15346.76330619 345.86188233
a. Multiple modes exist. The smallest value is shown

The correlations of the variables of 1984 database are similar with the period of
1952 and 1972. However, except “BoundLen-MaxDim” pair they exhibit slightly lower
correlation coefficients, as the maximum correlation is again in between “MaxDim” and
“BoundLen” (0.955) and the other two best correlations are in between pairs “area”-
“BoundLen” (0.799) and “area”-“MaxDim” (0.792). On the other hand, the “roughness”
variable is showing the minimum correlation with the other variables, and in this period
the correlation coefficients are even lower than that of the previous two periods (Table
5.8 and Figure 5.7).

98
 30 30

FREQUENCY

FREQUENCY
20 20

10 10

Std. Dev = 161246.4

Std. Dev = 984.93
Mean = 74396
Mean = 1048
0 N = 39.00
0 N = 39.00
25

12

22

32

42

52

62

72

82

92

10
0 1000 2000 3000 4000 5000 6000
00

50

50

50

50

50

50

50

50

50

25
500 1500 2500 3500 4500 5500 6500
0

00

00

00

00

00

00

00

00

00

00
0
AREA BOUNDLEN

a b
15 20
FREQUENCY

FREQUENCY
15

10

10

Std. Dev = 285.32

Std. Dev = 4.18
Mean = 394
Mean = 9
0 N = 39.00
0 N = 39.00
10

30

50

70

90

11

13

4 6 8 10 12 14 16 18 20 22 24 26 28 30
0

00

00

MAXDIM ROUGNESS

c d

Figure 5.6. The frequency distributions of variables of 1972 Polystats database, a) area,
b) BoundLen, c) MaxDim, d) roughness.

Table 5.8. Correlations of Polystats variables 1984, the best correlations are shown in
bold.
AREA BOUNDLEN MAXDIM ROUGHNESS
Pearson Correlation 1.000 .799** .792** -.129
Sig. (2-tailed) . .000 .000 .435
AREA Sum of Squares and Cross-products 988015526103.304 4822526244.705 1384862450.239 -3299801.888
Covariance 26000408581.666 126908585.387 36443748.691 -86836.892
N 39 39 39 39
Pearson Correlation .799** 1.000 .955** .128
Sig. (2-tailed) .000 . .000 .438
BOUNDLEN Sum of Squares and Cross-products 4822526244.705 36863364.212 10201484.903 20001.389
Covariance 126908585.387 970088.532 268460.129 526.352
N 39 39 39 39
Pearson Correlation .792** .955** 1.000 .300
Sig. (2-tailed) .000 .000 . .064
MAXDIM Sum of Squares and Cross-products 1384862450.239 10201484.903 3093446.223 13585.801
Covariance 36443748.691 268460.129 81406.480 357.521
N 39 39 39 39
Pearson Correlation -.129 .128 .300 1.000
Sig. (2-tailed) .435 .438 .064 .
ROUGHNESS Sum of Squares and Cross-products -3299801.888 20001.389 13585.801 664.348
Covariance -86836.892 526.352 357.521 17.483
N 39 39 39 39
**. Correlation is significant at the 0.01 level (2-tailed).

99
 A
1600 1600

5000 Linear Regression

Linear Regression Linear Regression
maxdim = 289.23 + 0.00 * area
boundlen = 684.74 + 0.00 * area R-Square = 0.63 maxdim = 103.52 + 0.28 * boundlen
A
R-Square = 0.64
A A R-Square = 0.91 A
4000 A 1200 1200

A A
boundlen

maxdim

maxdim
3000

800 800
A A A

2000 AA A A
A
A A A
A A
A A A
AA AA
A
AA
AAA AA
A
A
AA 400 A 400 A
AA
A A A
1000 AA
A
A
A A
A A
A
A A A
A
A A A
A
A
A
A AA AAA
A
A A
A AA
A
A
A
A
A A
A
A A
A
A
A
A A A
A
A A AA
A A

0 250000 500000 750000 0 250000 500000 750000 1000 2000 3000 4000 5000

area area boundlen

a b c

Figure 5.7. Scatter plots of highly correlated variable pairs of Polystats database 1984,
a) ”area”-“BoundLen”, b) “area”-“MaxDim”, c) “BoundLen”-“MaxDim”.

5.1.1.4. 1994 Period

From the 1:25.000 scale panchromatic stereo aerial photographs of 1994 period
37 landslide bodies are mapped. The re-forestation studies seem to work well, however,
the road-cuts seem to cause and/or trigger the landslides in the 1994 period, which is
recorded in the aerial-photographs.
The range of the “area” variable is in between 5098.61 and 921243.32 square
meters respectively. The data is again clustered in smaller values and have a long right
tail with positive skewness and kurtosis values (Table 5.9. and Figure 5.8). The
distribution has a mean of 73854.92 and 159508.47 as standard deviation. The
“BoundLen” variable has a minimum of 289.25 and maximum of 5556.74 meters with
1043.56 as mean and 1012.49 as the standard deviation. The distribution is again right
long tailed and clustered in smaller values. The “MaxDim” variable is found to have a
minimum of 118.26 and 1325.77 meters as maximum value. It has again a long right tail
and clustering with a mean of 390.85 and a standard deviation of 286.96. The
“roughness” variable has a minimum of 4,95 and a maximum of 22.4 with a mean of
8.49 and a standard deviation of 3.24. The distribution is again long right tailed and
clustered in small values.
The correlations of the variables of Polystats database of 1994 period is similar
to that of 1952, 1972 and 1984 as the highest correlatible variables are “boundlen”-
“maxdim”, “area-MaxDim” and “area”-“BoundLen” pairs in decreasing correlation
coefficient order. Conformably the “roughness” variable has the least correlation with the
other remaining three variables (Table 5.10 and Figure 5.9).

100
Table 5.9. The descriptive statistics of Polystats 1994

Descriptive Statistics of Variables of Polystats Database (1994)

AREA BOUNDLEN MAXDIM ROUGHNESS

Valid 37 37 37 37
N
Missing 0 0 0 0
Mean 73854,9253061662 1043,5660012157 390.8584145365 8,4932245689
Std. Error of Mean 26223,0310271408 166,4534916783 47.1766174767 .5342227747
Median 23967,4675293000 722,0782377300 327.5310183200 7,9804270400
Mode 5098,61462402a 289,25587197a 118,26133385a 4.95063299a
Std. Deviation 159508,4705627398 1012,4970622184 286.9641610937 3,2495502770
Variance 25442952181,2644300000 1025150,3010008170 82348.4297521847 10,5595770025
Skewness 4.640 3.222 2.113 2.480
Std. Error of Skewness .388 .388 .388 .388
Kurtosis 23.500 11.823 4.520 8.704
Std. Error of Kurtosis .759 .759 .759 .759
Range 916144,71289063 5267,48713772 1207,50785113 17,48556595
Minimum 5098,61462402 289,25587197 118,26133385 4.95063299
Maximum 921243,32751465 5556,74300969 1325,76918498 22,43619894
Sum 2732632,23632815 38611,94204498 14461,76133785 314,24930905
a. Multiple modes exist. The smallest value is shown

30 30
FREQUENCY

FREQUENCY

20 20

10
10

Std. Dev = 159508.5

Std. Dev = 1012.50
Mean = 73855
Mean = 1044
0 N = 37.00
0 N = 37.00
0

10

20

30

40

50

60

70

80

90

50

10

15

20

25

30

35

40

45

50

55

60
00

00

00

00

00

00

00

00

00

00

00

00

00

00

00

00

00

00

00

00
0

0
0

AREA BOUNDLEN

a b
12 15
FREQUENCY

FREQUENCY

10

8 10

4
5

2
Std. Dev = 286.96
Std. Dev = 3.25
Mean = 391
Mean = 8
0 N = 37.00
0 N = 37.00
10

30

50

70

90

11

13

10

12

14

16

18

20

22
0

00

00

MAXDIM ROUGHNESS

c d

Figure 5.8. The frequency distributions of variables of 1994 Polystats database, a) area,
b) BoundLen, c) MaxDim, d) roughness.

101
Table 5.10. Correlations of Polystats variables 1994, the best correlations are shown in
bold.

AREA BOUNDLEN MAXDIM ROUGHNESS

Pearson Correlation 1.000 .803** .815** -.113
Sig. (2-tailed) . .000 .000 .504
AREA Sum of Squares and Cross-products 915946278525.519 4669481948.475 1343207587.405 -2115075.769
Covariance 25442952181.264 129707831.902 37311321.872 -58752.105
N 37 37 37 37
Pearson Correlation .803** 1.000 .955** .120
Sig. (2-tailed) .000 . .000 .481
BOUNDLEN Sum of Squares and Cross-products 4669481948.475 36905410.836 9990958.352 14163.130
Covariance 129707831.902 1025150.301 277526.621 393.420
N 37 37 37 37
Pearson Correlation .815** .955** 1.000 .269
Sig. (2-tailed) .000 .000 . .107
MAXDIM Sum of Squares and Cross-products 1343207587.405 9990958.352 2964543.471 9043.903
Covariance 37311321.872 277526.621 82348.430 251.220
N 37 37 37 37
Pearson Correlation -.113 .120 .269 1.000
Sig. (2-tailed) .504 .481 .107 .
ROUGHNESS Sum of Squares and Cross-products -2115075.769 14163.130 9043.903 380.145
Covariance -58752.105 393.420 251.220 10.560
N 37 37 37 37
**. Correlation is significant at the 0.01 level (2-tailed).

A
1600 1600
Linear Regression Linear Regression
5000 Linear Regression
boundlen = 667.05 + 0.01 * area maxdim = 282.55 + 0.00 * area
maxdim = 108.35 + 0.27 * boundlen
R-Square = 0.65 R-Square = 0.66 A
A
A R-Square = 0.91 A

4000 A 1200 1200

boundlen

maxdim
maxdim

A A

3000

800 800
A A
A
2000 A A A
A
AA A
A
A
A AA
AA
A A A
AA A A
A
A A A 400 AAA
1000 AA 400 A
AA A A
AA
AA AA A
A
A
A
A A
A
A
A A A
A
A
A
AA
A
A AA
A A
A
A
A
A
A A
A AA
A
A A
A
A AA
A
A
A A
A A
A
A AA
A
A
A
A A
0 250000 500000 750000
0 250000 500000 750000 1000 2000 3000 4000 5000
area
area boundlen

a b c

Figure 5.9. Scatter plots of highly correlated variable pairs of Polystats database (1994),
a) ”area-BoundLen”, b) “area-MaxDim”, c) “BoundLen MaxDim”.

5.1.1.5. The Comparison of Four Periods

Three major events have occurred through period 1952 and 1994 in the area,
which are revealed from aerial photographical interpretation. First one is the sudden de-
forestation in the area in between the 1952-1972 period. This might either be due to an
extensive forest fire or human abuse of the forest resources for economical raw material
needs. The change in land cover should affect the states of landslides in the area and
this is supported by the increase of number of landslides in the area from 33 to 45. On
the other hand, not only the generation of new landslides but, it is also expected that
present landslides could also re-activate, yielding to an increase in the “area”, “MaxDim”
and “BoundLen” variables. Although there are slight variations in the histograms (Figure

102
5.10), the variables are not showing a significant response to this assumption and the
statistical analyses of the 3 variables could not say that these variables in these periods
differ significantly in between 1952 and 1972 periods (Table 5.11). The Paired Samples
t-Test of the all-available variables in the Polystats database fails to reject the null
hypothesis of “The two population means are equal”. Also the intensive reforestation
which is observed in 1984 period decreases the mappable landslide number from 45 to
39. However, the variables in the database are still ignorant to this. On contrary the
present route of the E-5 highway is constructed at these years and there exist some
slope instabilities related to this road, which might be on the charge to balance the effect
of reforestation, yet no net change could be observed in the histograms. Furthermore,
more intensive reforestation and road cut of E-5 in the 1994 period is recorded,
stabilization in the forest areas and reactivation or generation in the road cuts had
created a balance giving out a number of 37 slides in 1994 period. Furthermore, some
statistical tests are used to consider whether the differences are significant or not, such
as One-way ANOVA. The analyses show that the differences are not significant (Table
5.12). The landslides are open to external factors such as land cover change or
geomorphological regime changes and they sometimes give strong responses. Here in
this case, the landslides give a statistically insignificant response. The number of
landslides change, the minimum-maximum, and the mean of the variables change
slightly. A classification based on these attributes will be quite fictitious or it will force the
nature to fit into a generated model.

Table 5.11. The results of Paired Samples t test.

Paired Samples Test

Paired Differences
95% Confidence Interval of the Difference Sig.
Mean Std. Deviation Std. Error Mean Lower Upper t df (2-tailed)
Pair 1 A_52 - A_72 3133.2352943428 225138.3699281057 39799.2170203623 -78037.8029747261 84304.2735634117 .079 31 .938
Pair 2 A_72 - A_84 -2274.5032497102 226608.6488145063 37254.2323797402 -77829.5884501831 73280.5819507627 -.061 36 .952
Pair 3 A_84 - A_94 -22118.4051513676 176248.2498681428 28975.0337926642 -80882.4973644682 36645.6870617330 -.763 36 .450
Pair 4 B_52 - B_72 6.8854877403 1551.3268994336 274.2384426067 -552.4275034747 566.1984789553 .025 31 .980
Pair 5 B_72 - B_84 -11.4809617414 1476.2485762631 242.6937709486 -503.6867426775 480.7248191948 -.047 36 .963
Pair 6 B_84 - B_94 -69.1765357922 1375.9556543506 226.2057161523 -527.9429917081 389.5899201238 -.306 36 .762
Pair 7 M_52 - M_72 9.1444219503 444.0748629193 78.5020867312 -150.9616395081 169.2504834087 .116 31 .908
Pair 8 M_72 - M_84 -7.4818299246 414.7648201123 68.1869163963 -145.7713060134 130.8076461642 -.110 36 .913
Pair 9 M_84 - M_94 -21.4812258095 378.1292478052 62.1640654097 -147.5557939434 104.5933423245 -.346 36 .732
Pair 10 R_52 - R_72 -2.8773702812E-02 5.0865587919 .8991850537 -1.8626737106 1.8051263049 -.032 31 .975
Pair 11 R_72 - R_84 -.2011592211 6.0301049122 .9913431409 -2.2116962981 1.8093778559 -.203 36 .840
Pair 12 R_84 - R_94 .4110326595 5.5999202025 .9206212103 -1.4560736943 2.2781390132 .446 36 .658
Pair 13 A_52 - A_84 27069.3210870145 187088.1695223885 34741.4009083855 -44095.2126496878 98233.8548237167 .779 28 .442
Pair 14 A_52 - A_94 -8360.2636152007 261756.4594128563 49467.3211193707 -109858.8226564464 93138.2954260449 -.169 27 .867
Pair 15 A_72 - A_94 1592.8329171311 225210.0548926682 38067.4472200363 -75769.5276795204 78955.1935137826 .042 34 .967
Pair 16 B_52 - B_84 76.4396940662 1552.4303504508 288.2790789233 -514.0732300303 666.9526181627 .265 28 .793
Pair 17 B_52 - B_94 -78.3791612904 1749.3247757647 330.5913084969 -756.6964965475 599.9381739667 -.237 27 .814
Pair 18 B_72 - B_94 29.5674699980 1454.6627507336 245.8828825955 -470.1266680919 529.2616080879 .120 34 .905
Pair 19 M_52 - M_84 19.4612184045 434.1497285018 80.6195806536 -145.6805063749 184.6029431839 .241 28 .811
Pair 20 M_52 - M_94 -33.3368587636 494.8430475471 93.5165458442 -225.2169613225 158.5432437954 -.356 27 .724
Pair 21 M_72 - M_94 7.5023579734 412.3526156054 69.7003135055 -134.1457214458 149.1504373926 .108 34 .915
Pair 22 R_52 - R_84 -.8524458559 6.7143750055 1.2468281373 -3.4064575169 1.7015658052 -.684 28 .500
Pair 23 R_52 - R_94 -.7490438543 5.6189166838 1.0618754416 -2.9278322901 1.4297445816 -.705 27 .487
Pair 24 R_72 - R_94 .2064057986 5.9090675511 .9988147165 -1.8234299248 2.2362415219 .207 34 .838

103
 a_52 a_72 a_84 a_94
30

20
Count

10

0
200000 400000 600000 800000 200000 400000 600000 800000 200000 400000 600000 800000 200000 400000 600000 800000

Value Value Value Value

b_52 b_72 b_84 b_94

15

10
Count

0
1000 2000 3000 4000 5000 1000 2000 3000 4000 5000 1000 2000 3000 4000 5000 1000 2000 3000 4000 5000

Value Value Value Value

m_52 m_72 m_84 m_94

10

6
Count

250 500 750 1000 1250 250 500 750 1000 1250 250 500 750 1000 1250 250 500 750 1000 1250

Value Value Value Value

r_52 r_72 r_84 r_94

12

8
Count

0
10 20 30 10 20 30 10 20 30 10 20 30

Value Value Value Value

Figure 5.10. The distributions of Polystats variables in 4 periods. (A_52: “Area” in 52

period, b_72 “BoundLen” in 1972, m_84: “MaxDim” in 1984, r_94: “roughness” in 1994),
each column represents single period and each row represents single parameter.

104
Table 5.12. The ANOVA table of Polystats variable.
ANOVA

Sum of Squares df Mean Square F Sig.

Between Groups 4505636342.854 3 1501878780.951 .066 .978
AREA Within Groups 3396482747458.552 150 22643218316.390
Total 3400988383801.406 153
Between Groups 300699.730 3 100233.243 .101 .959
BOUNDLEN Within Groups 149229406.760 150 994862.712
Total 149530106.489 153
Between Groups 44889.490 3 14963.163 .180 .910
MAXDIM Within Groups 12465596.681 150 83103.978
Total 12510486.171 153
Between Groups 21.763 3 7.254 .446 .720
ROUGHNESS Within Groups 2439.812 150 16.265
Total 2461.575 153

5.1.2. The Fuzzystats Database

As defined in section 5.1. of this chapter these parameters are used to see the
changes in the shapes of the landslides. The previous database was directly concerned
with the spatial distribution, orientation and topological (area related) properties of the
landslide; this database is generated from the Polystats database regarding the
topology of the vector object. In order to assess the shape, 6 variables are extracted
from Fuzzystats database. In this section a variable based data exploration will be
carried out rather than period basis like in the previous section. Before starting to
explore the variables separately, the descriptive statistics (Table 5.13) and the results of
one-sample Kolmogorov-Smirnov Normality test for the cumulative 4 periods (Table
5.14) is given. These state that the normality theory and normality assumptions are valid
for the parameters which are going to be used.

5.1.2.1. Form Ratio

The form ratio theoretically approximates to 1 when the shape is a perfect

square, to Π/4 for perfect circles and to lower values near 0 for long skinny polygons.
The distribution of “Form Ratio” suggests that the landslides generally have a long but
narrow outline as expected; since majority of them have flow component in the slide
style. The minimum and the maximum of the variable is 0,082 and 0,74; very few of
them resemble a circle as the value approximates to 0.7. The distribution mean is 0,40
and standard deviation is 0,14 (Table 5.13) yielding in a conclusion that landslides are
not circles and are not skinny long polygons but their shape is in between them, namely
they are ellipsoidal in shape. The positive near 0 skewness value and negative 0 near
kurtosis value also suggest no departure from normal distribution with shorter tails but a
slightly longer right tail and clustering around a fixed value (mean) (Figure 5.11). The
past periods distributions are similar to each other, and their mean are ranging between

105
0,38 to 0,42 (Figure 5.12) yielding that the land cover change in the time period and its
effect is not adequate to change the shape of the landslides, if a reactivation is present
or new generation of landslides are seen the shape is still ellipsoidal.

Table 5.13. The Descriptive Statistics of Fuzzystats database cumulative periods.

FormRatio GSI Compactness CIRCUL1 CIRCUL2 Elongation

Valid 154 154 154 154 154 154
N
Missing 0 0 0 0 0 0
Mean ,4020 2,7111 ,8101 ,6070 ,5971 ,6811
Std. Error of Mean ,0117 ,0216 ,0085 ,0093 ,0114 ,0150
Median ,3878 2,6829 ,8295 ,6178 ,5998 ,6838
Mode ,0822a 2,1626a ,4279a ,2943a ,1258a ,3333a
Std. Deviation ,1450 ,2684 ,1052 ,1158 ,1412 ,1856
Variance ,0210 ,0721 ,0111 ,0134 ,0199 ,0344
Skewness ,2576 ,3895 -1,2066 -,4422 -,4961 -,1608
Std. Error of Skewness ,1955 ,1955 ,1955 ,1955 ,1955 ,1955
Kurtosis -,4440 -,2282 1,9392 -,1388 ,3179 -,9013
Std. Error of Kurtosis ,3886 ,3886 ,3886 ,3886 ,3886 ,3886
Range ,6622 1,3215 ,5319 ,5382 ,7344 ,7396
Minimum ,0822 2,1626 ,4279 ,2943 ,1258 ,2604
Maximum ,7444 3,4841 ,9597 ,8325 ,8602 1,0000
Sum 61,9041 417,5088 124,7518 93,4758 91,9520 104,8850
a. Multiple modes exist. The smallest value is shown

Table 5.14. The result of One-Sample Komogorov-Smirnov Test for Fuzzystats

database.

FormRatio GSI Compactness CIRCUL1 CIRCUL2 Elongation

N 154 154 154 154 154 154
a,b Mean ,4020 2,7111 ,8101 ,6070 ,5971 ,6811
Normal Parameters
Std. Deviation ,1450 ,2684 ,1052 ,1158 ,1412 ,1856
Absolute ,0752 ,0698 ,1063 ,0887 ,0631 ,0689
Most Extreme
Positive ,0752 ,0698 ,0779 ,0362 ,0408 ,0444
Differences
Negative -,0633 -,0474 -,1063 -,0887 -,0631 -,0689

Kolmogorov-Smirnov Z ,9329 ,8656 1,3186 1,1012 ,7826 ,8548

Asymp. Sig. (2-tailed) ,3489 ,4419 ,0618 ,1768 ,5726 ,4580

a. Test distribution is Normal.

b. Calculated from data.

30
FREQUENCY

20

10

Std. Dev = ,15

Mean = ,4
0 N = 154,00
,0

,2

,4

,6

,8

FormRatio
Figure 5.11. The frequency distribution of cumulative “Form ratio”.

106
 fr52 fr72 fr84 fr94
10

6
Count

0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9 0,1 0,2 0,3 0,4 0,5 0,6 0,7 0,8 0,9

FORM RATIO FORM RATIO FORM RATIO FORM RATIO

Figure 5.12. The frequency distributions of “form ratio” in all periods.

5.1.2.2. Grain Shape Index

In theory, the “Grain Shape Index” (GSI) variable approaches to Π for circles, 4
for squares and 2 for long skinny polygons. Larger values might be expected when the
polygon is of fractal nature; however, there are no fractals in the study area so the
maximum is 3,344 and the minimum is 2,178 (Table 5.13). This range states that the
shape of the landslides in the region range from long skinny polygons to circles. The
mean of the distribution is 2,71 and the standard deviation is 0,27. The distribution is
slightly positive skewed and the kurtosis value is near to 0 yielding in there is no strong
tendency to cluster around some values. The cumulative frequency distribution of “Grain
Shape Index” is shown in Figure 5.13. When each period is explored individually it is
seen that means are converging around 2,7. Slight changes occur through time and no
net trend changes could be seen in “Grain Shape Index” (Figure 5.14).

60
FREQUENCY

50

40

30

20

10 Std. Dev = ,27

Mean = 2,7
0 N = 154,00
2,

2,

2,

3,

3,

4,
0

GSI

Figure 5.13. The frequency distribution of cumulative “Grain Shape Index”.

107
 gsİ52 gsİ72 gsİ84 gsİ94
10

6
Count

2,2 2,4 2,6 2,8 3,0 3,2 3,4 3,6 3,8 2,2 2,4 2,6 2,8 3,0 3,2 3,4 3,6 3,8 2,2 2,4 2,6 2,8 3,0 3,2 3,4 3,6 3,8 2,2 2,4 2,6 2,8 3,0 3,2 3,4 3,6 3,8

GSI GSI GSI GSI

Figure 5.14. The frequency distributions of “form ratio” in all periods.

5.1.2.3. Compactness

The “compactness” variable measures the shape of the polygon with a

maximum value approximating one for circles. The mean of the cumulative distribution
of all periods is 0,8 and the standard deviation is 0,1. The negative skewness value and
the positive higher kurtosis value suggest a distribution to have a long left tail and more
clustering than a normal distribution (Figure 5.15). However, the non-parametric tests
are just on the limit to consider it as a normal distribution (Table 5.14). Although the
distributions mean itself has a converging value at 0.8, namely more circular than
elongated, it conflicts with the form ratio when investigated in periodical basis. An
increase in the long skinny polygons in 1972 period in the “form ratio” histogram is
observed (Figure 5.12), in the same period in compactness histogram (Figure 5.16) it is
observed that an increase in the near circular region.

30
FREQUENCY

20

10

Std. Dev = ,11

Mean = ,8
0 N = 154,00
,4

,5

,6

,7

,8

,9

Compactness

Figure 5.15. The frequency distribution of cumulative “Compactness”.

108
 com52 com72 com84 com94

10

8
Count

0,40 0,60 0,80 1,00 0,40 0,60 0,80 1,00 0,40 0,60 0,80 1,00 0,40 0,60 0,80 1,00

COMPACTNESS COMPACTNESS COMPACTNESS COMPACTNESS

Figure 5.16. The frequency distributions of “compactness” in all periods.

5.1.2.4 Circularity 1

The “Circularity1” variable measures shape, reflecting the element’s similarity to a circle,
with a maximum value approximating to 1 for circles. The minimum is represented with
0,3 and the maximum is 0.83. This variable has a mean of 0,6 and a standard deviation
of 0,11 (Figure 5.17). The negative skewness value implies a slightly longer left tail and
the small negative kurtosis value implies less clustering. In the period analysis, it seen
that the mean values decrease from 0,63 to 0,58, departuring from circularity field
(Figure 5.18).

60
FREQUENCY

50

40

30

20

Std. Dev = ,12

10
Mean = ,6
0 N = 154,00
,1
,2
,3
,4
,5
,6
,7
,8
,9
1,
0

CIRCULARITY 1

Figure 5.17. The frequency distribution of cumulative “Circularity 1”.

109
 cİr152 cİr172 cİr184 cİr194

10

8
Count

0,00 0,20 0,40 0,60 0,80 0,00 0,20 0,40 0,60 0,80 0,00 0,20 0,40 0,60 0,80 0,00 0,20 0,40 0,60 0,80

Circularity 1 Circularity 1 Circularity 1 Circularity 1

Figure 5.18. The frequency distributions of “Circularity 1” in all periods.

5.1.2.5. Circularity 2

Likewise “Circularity1” variable this variable also measures the circularity of the
element with a maximum value approximating to 1. In the calculation of “Circularity 1”
only area and maximum radius of the polygon is implemented, however in the
calculation of “Circularity 2“ both the minimum radius and the maximum radius is used
rather than the area (Table 5.2). Although the calculation scheme is different than
“Circularity 1”, the mean and standard deviation of “Circularity 2” are similar as
respectively 0.6 and 0.14 (Figure 5.19). However, the decrease in circular elements is
more pronounced in 1972 period (Figure 5.20).

50
FREQUENCY

40

30

20

10
Std. Dev = ,14
Mean = ,6
0 N = 154,00
,1
,2
,3
,4
,5
,6
,7
,8
,9
1,
0

CIRCULARITY 2

Figure 5.19. The frequency distribution of cumulative “Circularity 2”.

110
 cİr252 cİr272 cİr284 cİr294

6
Count

0,00 0,20 0,40 0,60 0,80 0,00 0,20 0,40 0,60 0,80 0,00 0,20 0,40 0,60 0,80 0,00 0,20 0,40 0,60 0,80

Circularity 2 Circularity 2 Circularity 2 Circularity 2

Figure 5.20. The frequency distributions of “Circularity 2” in all periods.

5.1.2.6. Elongation

The “Elongation” variable measures the ratio of the polygons short axis to long
axis. The variable theoretically has a maximum value of 1 and a minimum as a
convergent value to 0. The cumulative distribution has a minimum of 0,26 and a
maximum approaching to theoretical maximum. The mean of the cumulative distribution
is 0,68 and the standard deviation is 0,18. The negative skewness and kurtosis values
imply that the distribution has a low tendency to cluster and has a slightly longer left tail
(Figure 5.21). The effect of de-forestation is clearly seen on the 1972 histogram as a
new peak around 0,6 is formed when compared into 1952 period (Figure 5.22). In the
1984 period the elongation decreases probably due to intensive re-forestation studies
and in 1994 period a very minor increase can be attributed to the new road-cuts.

20
FREQUENCY

15

10

5
Std. Dev = ,19
Mean = ,7
0 N = 154,00
,0
,1
,2
,3
,4
,5
,6
,7
,8
,9

ELONGATION

Figure 5.21. The frequency distribution of cumulative “Elongation”.

111
 el52 el72 el84 el94
10

6
Count

0,00 0,20 0,40 0,60 0,80 0,00 0,20 0,40 0,60 0,80 0,00 0,20 0,40 0,60 0,80 0,00 0,20 0,40 0,60 0,80

Elongation Elongation Elongation Elongation

Figure 5.22. The frequency distributions of “Elongation” in all periods.

5.1.2.7. The significant changes and evaluation of fuzzy parameters with relation
to Polystats database.

The effect of de-forestation, re-forestation and road construction is sometimes

well represented in the frequency histograms of the fuzzy variables and sometimes due
to the formulation they are very ignorant to the real fact. Although statistically the
Polystats show no significant change in any of the periods there are some minor
changes, which could lead some clues for the exploration of Fuzzy parameters. A brief
table showing the changes in the Polystats parameters is given in Table 5.15. Although
the 1952-1972 period in the table states that all the parameters are decreasing, the
photo interpretation yields that this is because of newly generated landslides are
relatively very small in area, in maxdim and in boundlen. Furthermore, in 1972-1984
periods, it is seen that the newly generated landslides are either disappearing or getting
stable; nearly all the variables are converging to the values in 1952. It is obvious that
when the effects of smaller landslides are taken off the distribution all of the variables
should increase as seen in the table also. In 1984-1994 periods, the larger landslides
remain relatively stable and some larger slides (larger than that of de-forestation stage)
are developed in relation with the increased activity in the highway. Another striking
feature is seen in the Table 5.15. As the mean values of all three variables show the
same response in each period, such as decrease, increase and again decrease. This
could be only attributed to generation of smaller slides (due to de-forestation) so the
mean value is pulled down, vanishing of smaller landslides (reforestation) so the mean
values are increased and again the mean values are pulled down (road activity) as
newer slides are generated or the older small ones are reactivated.

112
Table 5.15. The changes through time in Polystats database

1952-1972 1972-1984 1984-1994

Intensive de-forestation Re-forestation and Re-forestation and
Parameter / Period
Minimal road activity Minor road activity Intensive road activity
Number of landslide S T T
Area

Min T S S
Max T S T
Mean T S T
Boundlen
Min T S S
Max S T S
Mean T S T
Maxdim

Min T S S
Max T T S
Mean T S T

The effects of deforestation are clearly seen in the “Form Ratio” variable as it
has an increase in 1972 period in near 0.3-0.4, giving rise to an increase in log skinny
polygons. However, the compactness variable, which is sensitive to circularity of the
polygon, shows an increase that can be attributed to a larger increase in area than the
perimeter. For the direct circularity measuring variables (“Circularity1” and
“Circularity2”), a pronounced decrease is observed in the circular side rather than an
increase in the lower (noncircular) values. However, an abrupt increase in elongation is
observed in 1972 period. It is hard to say only looking to elongation graph of 1972 that
the landslides are getting longer, but it could be concluded that the ratio of short axis to
long axis is getting clustered around the mean, though the “Grain Shape Index” is
having an increase in the values around 2.5 and 2.8. Based on all of the evidences
reflected in Polystats and Fuzzystats variables the effects of de-forestation could be said
to generate or re-activate flow type slope instabilities, of which the last word is left to be
justified after the evaluation of photo characteristics database variables as they would sit
on the true fact observations.
The effect of re-forestation activities is seen in 1984 and 1994 periods
cumulatively. In the images of 1984 reforestation is started few years ago and yet the
land cover is still not mature, in 1994 the land cover is very similar to that of 1952,
mainly covered with dense forests. These are reflected as a decrease from both sides of
the “Form Ratio” graph, clustering into mean values from 1984 to 1994, giving rise to
more ellipsoidal shape. Namely the increase in elongation in 1984 is balanced with the

113
road cuts and dense forests in 1994. The “Grain Shape Index” responds to these
changes as cutting off the extreme values and resulting in more homogenous
distribution. This is probably due to the decrease in long axis and so the perimeter.
Although the reforestation seems to be efficient, some outlier peaks are observed in the
2.4-2.5 field of “Grain Shape Index” 1994, which could be attributed to the landslides
occurring by the road cuts. Also for the compactness variable the increase in 1972
period is trimmed in both 1984 and 1994 periods giving rise to a decrease in perimeter
and area. Both of the circularity variables show slight increments in the circular side and
decreases in the elongated side. A very sharp increase in the “Circularity 2” 1994 could
easily be attributed to newly generated landslides caused by road cuts. The “elongation”
graph of 1972 is trimmed down to mean values in the 1984 period especially in the 0,6-
0,8 range and a very slight increment in 1994 period in 0,5 group. As a result in 1984
and 1994 periods, the newly generated flow type instabilities are either vanished or
stabilized but other new or regenerated larger slides (larger than that of de-forestation)
occur due to the road activities in the area.

5.2. Photo-characteristics Database

The photo characteristics of the landslides have been mapped and recorded in
each year’s relational database, which was also linked to the polygon landslide map.
Each of the mapped seven attributes is investigated in this section in period basis. This
exploration in the data probably will not yield in finding the direct causes, but it will be
helpful for delineating the general scheme and general trends of the evolutions of slope
instabilities in the Asarsuyu catchment and also will be used in conjunction with the
Polystats and Fuzzystats databases. The definitions of the variables used in
photocharacteristics database and the available items are presented in the previous
chapter.

5.2.1. Massinfo

The “Massinfo” variable as described in the previous chapter in section 4.5.2 is

the morphology seen in the photograph. The available items for this variable is
Scarp&Path and Scarp&Body. The frequencies of the above mentioned items are
shown with the data table in Figure 5.23. It is seen that there is a sharp increase in the
1972 period, following a continuous decreasing trend in both the number of landslides
and in the S&B item.

114
 50

FREQUENCY
45
40
39
37
34
30
33 1952 1972 1984 1994
27 s&b 22 34 27 25
25
20 22 s&p 11 11 12 12
# of slide 33 45 39 37
S&B
10 11 11 12 12
S&P

0 # OF SLIDE
1952 1972 1984 1994

PERIODS

Figure 5.23.The frequencies of the mass info variable through time and data table

5.2.2. Type

The “Type” variable has significant drawbacks when a time dependent study is
carried on, such as if the slide has started as a slide in the photo year; continued its
activity as a flow in the period between two photo periods and it is seen in the next
photo set as diminished, only the pathway of the flow and the scarp is seen. What code
will be assigned to this slide is a big dilemma between the landslide investigators. A
significant increase is seen at the 1972 period both in the slides and in the flows. After
1972 period although the number of slides decreases from 19 to 13 the percentage
decreases only 9 % from 42% to 33% (relative to the number of total landslides) as the
number of landslides decreases (Figure 5.24). Most of the landslides present in 1952
and 1972 period are diminished in 1984 and 1994 periods. The further analysis of the
types will be handled in detail in further sections in the analysis chapter.

30
FREQUENCY

26
25 25

20

19

17 1952 1972 1984 1994

15

13
Flow 17 25 26 25
12
10
Slide 15 19 13 12
FLOW Flow/slide 1 1 0 0
FLOW_SLI

0 SLIDE
1952 1972 1984 1994

YIL

Figure 5.24.The frequencies of the type variable through time and data table

115
5.2.3. Style

The “Style” variable is in fact a similar variable to distribution of activity variable,

like the very generalized form of it. It deals with the mass of the landslide if it is coherent
and displays the same activity through its life span or is it sporadic, as it slides and
stabilizes. This stabilization does not require the stabilization of the slided mass indeed
the scarp of the first instability stabilizes. This small difference should have to be
stressed as it the definition stabilizes the whole landslide then a variable called
“distribution of activity” could only be attributed to multiple slides (Figure 5.25). Generally
the landslides in the area are characterized by single landslides, however the number of
multiple slides have a very slight increasing trend in the newer periods, with probable
contribution other factors.

40
FREQUENCY

30

20
1952 1972 1984 1994
single 27 38 30 29
multiple 6 7 9 8
10

SINGLE

0 MULTIPLE
1952 1972 1984 1994

YEAR

Figure 5.25.The frequencies of the style variable through time and data table

5.2.4. Depth

The “Depth” variable is totally dependent to the experience of the interpreter

and the position of the parallax of the stereo-pair. A unified approach have been tried to
fill the database column, as after the interpretation of the photographs, a second pass
only for depth variable is conducted. The degree of depth is relatively chosen,
dependent of the photo characteristics and the characteristics of the displaced mass.
The landslides in the area are characterized dominantly by shallow slides (Figure 5.26).
The quantity of deeper landslides decreases through time, where the shallow ones
remain approximately same.

116
 40

FREQUENCY
30

20
1952 1972 1984 1994
deep 13 13 9 8
shallow 20 32 30 29
10

DEEP

0 SHALLOW
1952 1972 1984 1994

YEAR

Figure 5.26.The frequencies of the depth variable through time and data table

5.2.5. Distribution of Activity

“Distribution of activity” is probably the most important parameter in evaluating

the history of change in the landslides through time. Also the previously explained
parameters such as “type” and “style” could be well understood with the help of this
activity parameter. However, during photo interpretation there are some trivia situations
such as the slide is diminishing via moving its slided mass or the slide is both enlarging
and widening. Based on these special cases, the distribution of activity is re-grouped as
activity originated in or by the scarp area yielding further development of the original
slide, activity originated and confined in the slided mass and stabilization of the slide
(Table 5.16). In the photo interpretation if no sign of activity is observed the slides gain
an attribute as dormant.

Table 5.16. The re-grouping of distribution of activity variable in the photo database

Activity confined in scarp area

Advancing
Enlarging
Widening
Combinations of the above
Activity confined in slided mass
Moving
Diminishing
Combinations of these
Stabilized slides
Dormant/inactive

As this analysis deals with the two periods of time the number of the slides used
in this analysis is taken as the lowest amount of slide of the two years. Furthermore if
the slide does not have an attribute as either scarp related or mass related the stable
attribute is given. The striking fact in the graph presented in Figure 5.27 is the increase

117
in scarp in 1972-1984 periods that might be attributed to the new re-routing of E-5
highway. The mass related variable has an increasing trend with a very small slope
through time and this might again be attributed to the construction of highway in the
study area. When stayed out of the between variable changes, the starting period has a
significant discrepancy within the variables themselves as the difference of mass related
activity is started as 18, which means there are 18 cases which were not active in the
1952 period and activated in 1972. This is one of the clear evidence of de-forestation.

50
FREQUENCY

40

30
scarp mass Stabile valid count year
20
SCARP
6 18 13 37 1994-1984
MASS 9 16 14 39 1984-1972
10
STABILE 4 14 15 33 1972-1952
0 VALID_#
1972-1952 1984-1972 1994-1984

YEAR

Figure 5.27. The frequencies of the distribution of activity variable through time and
data table

5.2.6. Land cover

A “Land cover” attribute scheme is constructed with the help of field and remote
sensing studies, and aerial photographical interpretation. The major elements for the
scheme are dense forest, bare land, agricultural area, grassland and forest. There are
also some associations present within the landslide polygon so possible combinations of
the above major elements are also accepted. Although the landslide polygons do not
reflect all of the area, they could be used as random sampled information source, so the
below conclusions could have been derived.
The striking fact in the graph is the abrupt decrease of dense forest from 1952
to 1984 period and flawless increase in 1994 period. On the other hand, the increase in
forest in 1984 period is the reason for increase of dense forest in 1994 period and
shows the efficiency of re-forestation studies started at 80’s. The only change in the land
cover variable, except the former items, is the decrease in bare land amount, which
might be an indicative of increased human activity (Figure 5.28).

118
 25

FREQUENCY
20

Dense Forest (DF)

15
Bare Land (BL)

Agriculture (A)

10 Dense Forest / Agri

GR - A

GRASS (GR)
5
BL - GR

DF - GR

0 FOREST
1952 1972 1984 1994

YEAR

denseforest bare agri DF/Agri GR/Agri Grass Bare/GR DF/GR Forest Year
22 5 3 2 2 8 1 2 2 1952
8 20 3 2 2 8 1 1 2 1972
5 15 3 2 2 9 1 1 9 1984
22 5 3 2 2 8 1 2 2 1994

Figure 5.28. The frequencies of the land cover variable through time and data table

5.3. Landslide Attribute Databases

All of the parameter maps produced are crossed with the landslide inventory
map and the landslide buffer map so that the attributes in each parameter map could be
transferred to the inventory maps. The inventory map will have 4 attribute databases, of
which the three of them are investigated in the previous sections of this chapter. The
last attribute database, but the most important one is the transferred attribute database.
It consists of (13) previously produced variables; the nature and the ranges are given in
Table 5.17.

Table 5.17. The nature and ranges of transferred attribute database

Variable Definition Nature Range (Min-Max) (unit)

LITHOMAP Outcropping material Nominal -
DISTFAULT Distance to Fault line Ratio 0 – 4791 (m)
2 2
FAULTDENS Density of Fault line in km Ratio 0 – 178 (# / km )
Elevation above Mean Sea
ELEVMAP Ratio 220 – 1580 (m)
Level
DISTDRAIN Distance to drainage lines Ratio 0 – 452 (m)
2
DRAINAGE DENS Drainage Density Ratio 12-352 (#/km )
DISTRIDGEMAP Distance to ridges Ratio 0 – 658 (m)
ASPECTMAP Aspect of the slopes Ratio -1 – 359 (degrees)
SLOPEMAP Amount of Slope Ratio 0 – 56 (degrees)
DISTSETTLEMENTMAP Distance to settlement Ratio 0 – 6093 (m)
Distance to powerlines and Ratio 0 - 2312 (m)
DISTPOWER+ROAD
roadnetwork
DISTE-5MAP Distance to E-5 highway Ratio 0 – 8366 (m)
LANDCOVERMAP Type of the land cover Nominal -

119
 All of the attributes are transferred to the central grid nodes of 25 meters by 25
meters. The resulting data set is consisted of 5493 rows of data. The aim of this section
is to characterize the landslides by their attributes, to characterize the seed attributes
and to compare the original available topography and other attributes with the slided
masses attributes where needed. However a note should be added as the area of the
landslides occur here as a natural weighting parameter, the larger the landslides the
larger number of grid nodes they will have, this argument can be carried without any
tension unless the number 8 landslide is omitted. The number 8 has no activity through
all of the periods. This landslide is known as Kom Landslide and believed to be a huge
dormant slide, creating its own stable parameters (probably being the oldest landslide in
the dormant slide in the area), which should not be mixed with the rest. Also this
landslide is represented with 1346 nodes out of 5493 nodes, nearly 24% percent of the
total nodes. In order to remove the effect of this dormant landslide the nodes of his slide
are not used so the remaining node population is decreased down to 4147 nodes.
For all of the parameters a comparison will be given in the proceeding sections
with the seed data distributions. The seed data is extracted using a 100-meter buffer
and the microcatchment boundaries. The decision rule for defining the seed zone is: If
the distance between the slide boundary and microcatchment divide line is smaller than
100 meters then use the microcatchment divide line, if the distance is larger than 100
meters then use the 100 meter buffer line for the seed zone generation. Following the
seed node generation, the same attributes as the slided mass is also transferred, and
the number of the seed nodes is 4430. The seed cells of the Kom landslide is included,
because it is coherent with the similar landslides in the valley and due to the seed cell
selection decision rule the number of the seed cells have already been decreased
significantly. This inclusion stands for the sake of the factual data preservation.

5.3.1. LITHOMAP

The “LITHOMAP” of the slided mass gives out some preferred conditions like
the sum of the first three most preferred lithologies are approximately 92.9% which is
quite homogeneous and remarkable (Figure 5.29). Landslides exist in 8 lithologies, but
the area has 11 lithologies. The landslide missing lithologies are green schist facies of
Yedigöller Formation and Asarsuyu formation. There are no landslides seen in the
gypsum, this is due to the fact that the huge landslides are not taken into account in this
analysis.
In the LITHOMAP of seed cells, the situation slightly changes, though the
preferred conditions are still the same. The cumulative total of the preferred first three
lithologies are 89,5 % which is 3,4% lower than that of slided mass. Moreover the

120
distributions in the preferred lithologies are also changing. While percentage of
Aksudere formation is increasing the percentage of Yedigöller (21,4 % to 18,8 %) and
Fındıklıdere (43.4 % to 32,6 %) formations are decreasing. Also in the remaining 10% of
the data some slight changes have been observed as the percentage of Quaternary is
decreased as expected and Çaycuma formation is increased.

Yedigöller
Yedigöller (GS) Yedigöller (GS)
18,8%
,7% 1,5%

Yedigöller
20,7%

Fındıklıdere Fındıklıdere

43,4% 32,6%

Aksudere

38,1%

Aksudere
Talus
28,1%
2,3%
Talus
Caycuma Kocadere
1,3%
,7% 3,0%
Kocadere
Quaternary
2,0%
,2%
Quaternary
Caycuma
3,2%
3,5%

a b

Figure 5.29. The preferred lithologies and their percentages: a) slided Mass, b) the
seed cells.

Although the preferred lithologies represent 92.9% in slided mass and 89.5 % in
seed cells they can only range up to 54.08 % in the study area (Figure 4.2). The
remaining approximately 45 % of the available lithologies in the study area is only
reflected to slided mass and seed cells as approximately 10%, which are direct
indications that they are not preferred by landslides.

5.3.2. DISTFAULT and FAULTDENS

The distance to fault and the density of fault variables are two highly negative
correlated variables (Table 5.18). In theory, the distance to fault should be decreasing
when the density of the fault lines are increasing. The negative correlation is much
better represented in the seed cells rather than the slided mass, as indicating the seed
cells are much more sensitive to faults. The comparison of minimum, maximum and the
mean values of both the slided mass, seed cells and the whole data is given in Table
5.19.
It is seen that the maximum value of the landslide data can go up to 1441 and
the seed cells have a maximum of 1517 that is nearly one third of the maximum of the
whole data (Figure 4.5), which means there is redundant information in the whole data,
not relevant to landslides. On the other hand the maximum of the FAULTDENS are
nearly the same both in observed and in whole data (Figure 5.30). This is because of

121
the spatial orientation of both the landslides and the fault lines. A polynomial distribution
is seen in the frequency histogram of FAULTDENS as having peaks around 50, 225 and
600. These could be attributed to landslides very distant to faults, landslides near to one
set of faults and landslides near to two or more sets of faults. This could easily be seen
when the thematic maps of the fault density is investigated (Figure 4.6). Hence, it can be
said that landslides are not much far away than 1500 meters to the fault lines and the
preference order related to density is: near to one set of fault, very far away from faults
and the least preferred as crisscrossing of two or more sets of fault lines, which is valid
also for the seed cells.

Table 5.18. The correlation state of DISTFAULT and FAULTDENS variables: a) slided
mass, b) seed cells.

Correlations Correlations

D_FAULT DENSFAY D_FAULT DENSFAULT

Pearson Correlation 1,000 -,648** Pearson Correlation 1,000 -,703**
D_FAULT Sig. (2-tailed) , ,000 D_FAULT Sig. (2-tailed) , ,000
N 4147 4147 N 4430 4430
Pearson Correlation -,648** 1,000 Pearson Correlation -,703** 1,000
DENSFAY Sig. (2-tailed) ,000 , DENSFAULT Sig. (2-tailed) ,000 ,
N 4147 4147 N 4430 4430
**. Correlation is significant at the 0.01 level (2-tailed). **. Correlation is significant at the 0.01 level (2-tailed).
a b

Table 5.19. The comparison of whole data and landslide data

Slided mass data Seed cells Whole Data

DISTANCE TO

Min 1 0 0
FAULT

Max 1411 1517 4791

Mean 361,47 408,52 725,887
St.Dev 286,44 339,66 827,725

Min 0 0 0
DENSITY
FAULT

Max 671 705 768

Mean 229,1 235,56 195,193
St.Dev 163,58 161,03 154,61

122
 1000 1000

FREQUENCY

FREQEUNCY
800 800

600 600

400 400

200 Std. Dev = 286,44 200 Std. Dev = 339,66

Mean = 361 Mean = 409
0 N = 4147,00 0 N = 4430,00
0

20

40

60

80

10

12

14

0
20
40
60
80
10
12
14
0

00

00

00

0
0
0
0
00
00
00
DISTFAULT DISTFAULT
a b
600 600
FREQUENCY

FREQUENCY
500 500

400 400

300 300

200 200

100 Std. Dev = 163,58 100 Std. Dev = 161,03

Mean = 229 Mean = 236
0 N = 4147,00 0 N = 4430,00
0

10

20

30

40

50

60

70

10

20

30

40

50

60

70
0

0
FAULTDENS FAULTDENS
c d

Figure 5.30. The frequency distributions of DISTFAULT (a. slided mass, b. seed cells)
and FAULTDENS (c. slided mass, b. seed cells).

5.3.3. ELEVMAP

Although the available elevation data in the study area ranges between 220 and
1580 meters, the landslides are observed only between 295 and 1095 meters and their
seed cell values ranges between 300 and 1130 meters. This is probably at higher
elevations more sound and intact rocks are present. The descriptive statistics of the
three groups are given in Table 5.20. The majority of the data is distributed in the range
of 350 to 800 meters. An imperfect double peak is observed in the distribution, resulting
that some means of preference is present and concentrated around 700 and 400
meters, but either near/equal preferences exist or the preferences are indistinguishable.

Table 5.20. The descriptive statistics of the ELEVMAP.

Slided mass data Seed cells Whole Data

ELEVATION

Min 295 300 220

Max 1050 1130 1580
Mean 594,62 661,84 679,96
St.Dev 177,191 179,21 265,06

123
 500 600

FREQUENCY
500
400

400

FREQUENCY
300

300

200
200

100 Std. Dev = 177,91 Std. Dev = 179,21

100
Mean = 595 Mean = 662
0 N = 4147,00 0 N = 4430,00
30
40
50
60
70
80
90
10
11

30
40
50
60
70
80
90
10
11
0
0
0
0
0
0
0
00
00

0
0
0
0
0
0
0
00
00
ELEVATION ELEVATION
a b

Figure 5.31. The frequency distributions of ELEVMAP, a. Slided mass, b. Seed cells

5.3.4. Distance to drainage

The distance to drainage variable is the distance to all drainage lines without
any ordering in between the drainage lines. The descriptive statistics of the drainage
related parameter maps are presented in Table 5.21.

Table 5.21. The descriptive stats of distance to drainage variable group

Landslide data Seed Cells Whole Data

DISTANCE TO
DRAINAGE

Min 0 0 0
Max 342 397 452,948
Mean 87,16 118,3 97,56
St.Dev 66,29 76,99 73,144

The frequency histograms of the drainage related parameters of the landslide

data are given in Figure 5.32. The shapes of the cumulative histogram of both the
landslide data and the whole data (Figure 4.13.a, b) and the first order strahler order
streams are very similar to each other, due to the abundance of first order streams.

800 600
FREQUENCY

500
600
400
FREQUENCY

400 300

200
200
Std. Dev = 66,29 100 Std. Dev = 76,99
Mean = 87 Mean = 118
0 N = 4147,00 0 N = 4430,00
0

50

10

15

20

25

30

35

0
50

10
15

20
25

30
35

40
0

0
0

0
0

0
0

DISTDRAINAGE DISTDRAINAGE
a b

Figure 5.32. The frequency distributions of Distance to Stream map, a. Slided Mass, b.
Seed Cells.

124
5.3.5. Drainage Density

The drainage density data in the area is ranging from 12 to 352 drainage lines in
one square kilometer. The seed cells of the landslides are characterized by a range of
55 to 287, no significant preference is observed as the mean and standard deviation of
the whole data and seed cells are similar. The difference in range does not change the
distribution parameters, only a fraction of the whole data is represented with same
weightings. The drainage density attribute frequencies of slided mass and the seed cell
is presented in Table 5.22.

Table 5.22. The descriptive statistics of the DRAINAGE DENSITY.

Slided mass data Seed cells Whole Data

DRAINAGE

Min 58 55 12
DENSITY

Max 292 287 352

Mean 176.33 160 160.12
St.Dev 42.53 37 50.22

500 800
FREQUENCY

FREQUENCY

400
600

300

400

200

200
100
Std. Dev = 42,53 Std. Dev = 37,01
Mean = 176 Mean = 161
0 N = 4147,00 0 N = 4430,00

a b
60
80
10
12
14
16
18
20
22
24
26
28
60
80
10
12
14
16
18
20
22
24
26
28

0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
0

Figure 5.33. The frequency distributions of Drainage Density, a. Slided mass, b. Seed
cells.

5.3.6 DISTRIDGE

All of the distance to ridge parameter distributions are similar to each other
(whole data & landslide related data), which is an indication of no significant distribution
free preference is present. However, a genetic preference is seen in the distributions as
both the whole and landslide data exhibit clustering around smaller values. This
situation can be summarized as both the area and the landslides do tend to prefer
nearer distances to ridges, namely to the microcatchment divide lines. They tend to be
located in the upper parts of the slopes. Although the distributions are similar, here are

125
slight changes like, the Distance to Ridge of seed cells seems to have a more sharp
peak around 25-100 meters of distance than that of the slided mass data as expected
(Figure 5.34). The descriptive statistics and the comparison of the data are shown in
Table 5.23.

550 800
FREQUENCY

FREQUENCY
500
700
450
600
400

350 500
300
400
250

200 300

150
200
100 Std. Dev = 77,02 Std. Dev = 82,12
Mean = 133 100 Mean = 114
50
0 N = 4147,00 0 N = 4430,00
0
50
10
15
20
25
30
35
40

50

10

15

20

25

30

35
0
0
0
0
0
0
0

0
DISTTEPE DISTRIDGE
a b

Figure 5.34. The frequency distributions of Distridge, a. Slided mass, b. Seed cells.

Table 5.23. The descriptive stats of distridge

Slided mass data Seed cells Whole Data

DISTANCE TO

Min 4 4 1
RIDGE

Max 377 364 658,97

Mean 132,58 113,88 138,172
St.Dev 77,02 82,12 99,86

5.3.7. Aspect

Although the aspect distribution of the whole data have systematical errors
arising from the gridding model, no significant direction is seen. However, in the slided
data and in the seed cells, the north facing slopes exhibit a very distinctive preference
(Figure 5 35). The descriptive statistics and the comparison of the data is shown in table
5.24.

Table 5.24. The descriptive stats of aspect

Slided mass data Seed cells Whole Data

Min -1 -1 -1
ASPECT

Max 358 358 359

Mean 156,36 168,37 182,45
St.Dev 137,07 126,25 110,08

126
 700

FREQUENCY
600

500

400

300

200

Std. Dev = 126,25

100
Mean = 168
0 N = 4430,00

0
45
90
13
18
22
27
31
36
5
0
5
0
5
0
ASPECT
a b

Figure 5.35. The frequency distributions of Aspect, a. Slided mass, b. Seed cells.

5.3.8.Slope

The slope distribution in the study area has a broader distribution than the
landslide data (Figure 4.16), also this is seen in the descriptive statistics of the slope of
landslide related data as the means are similar to each other (Table 5.25). The landslide
related data (Figure 5.36) is much more clustered than the original available data, due to
the fact that a very small portion of the landslide related data are associated with alluvial
slopes that are gentler than the rest of the data.

Table 5.25. The descriptive stats of Slope

Slided mass data Seed cells Whole Data

Min 0 1 0
SLOPE

Max 52 51 56
Mean 21,44 22,41 16,97
St.Dev 8,88 9,12 10,46

600 600
FREQUENCY

FREQUENCY

500 500

400 400

300 300

200 200

100 Std. Dev = 8,88 100 Std. Dev = 9,12

Mean = 21 Mean = 22
0 N = 4147,00 0 N = 4430,00
0
5
10
15
20
25
30
35
40
45
50

0
5
10
15
20
25
30
35
40
45
50

SLOPE1 SLOPE
a b

Figure 5.36. The frequency distributions of Slope: a. Slided mass, b. Seed cells.

127
5.3.9. Distsettlement

The distances to settlement parameter histogram has a significant background

noise, reflecting the irrelevant data present in the whole study area (Figure 4.18.b). This
irrelevance is also seen in the compared descriptive statistics, as the maximum values
of the landslide related data is quite less than that of the whole data (Table 5.26). The
magnitude of irrelevance is much better seen when the slided data and the seed cell
histograms are explored (Figure 5.37). On the other hand both of the landslide related
histograms show a bimodal distribution having one peak around 150-200 meters and
the other around 1800 meters apart from the individual settlements. The first one could
easily be attributed to the effect of households and the other should be attributed to
external other factors.

Table 5.26. The descriptive stats of Distsettlement

Slided mass data Seed cells Whole Data

DISTANCE TO
SETTLEMENT

Min 2 4 0
Max 2218 2629 6093,46
Mean 700,04 699,45 1258,46
St.Dev 639,12 651,25 1370,20

700 800
FREQUENCY

FREQUENCY

600 700

600
500
500
400
400
300
300
200
200
Std. Dev = 639,12 Std. Dev = 651,25
100 100
Mean = 700 Mean = 699
0 N = 4147,00 0 N = 4430,00
0
20
40
60
80
10
12 0
14 0
16 0
18 0
20 0
22 0

0
20
400
600
800
100
1200
1400
1600
1800
2000
2200
2400
2600
0
0
0
0
0
0
0
0
0
0
00

00

DISTSETTLEMENT DISTSETTLEMENT
a b

Figure 5.37. The frequency distributions of Distsettlement, a. Slided mass, b. Seed

cells.

5.3.10. Distpower & Dist-Roadnetwork

The distance to power lines and road network has also significant low frequency
noise in the histogram shown in Figure 4.18.h. The means also suggest a preference in
smaller values, which are nearer to power lines and road network. Furthermore, the
difference in maximums are quite striking as the whole data range up to 5566 meters

128
but the slided data and seed cells range only up to 700 meters approximately (Table
5.27). The nearly 5000 meters of difference is the source of background noise in the
whole data histogram. On the other hand no significant preference is seen in the
frequency histograms of the slide related data, only a general clustering is present in the
smaller values, which are nearer to power lines and road network. This could be
attributed to the fact that the land cover of the region is disturbed during the construction
of power lines and the road network. The descriptive stats of the distance to power lines
and road network parameter map are presented in (Table 5.27). The frequencies of
merged vector of power lines and road network are presented in Figure .5.38.

Table 5.27. The descriptive stats of Distpower, Dist road & Distroad+Distpower

Slided mass data Seed cells Whole Data

DISTANCE TO

NETWORK
POWER &

Min 0 0 0
ROAD

Max 547 683 2312,68

Mean 184,53 183,78 246,121
St.Dev 119,31 136,88 302,099

600 400
FREQUENCY

FREQUENCY

500
300
400

300 200

200
100
100 Std. Dev = 142,29 Std. Dev = 126,28
Mean = 176 Mean = 180
0 N = 4430,00 0 N = 4147,00
0
50
10
150
200
250
300
350
400
450
500
550
600
650

0
50
10
150
200
250
300
350
400
450
500
550
600
0

DIST_POWER + ROAD DIST_POWER + ROAD

a b

Figure 5.38. The frequency distributions of Dist-power+road a. Slided mass, b seed

cells.

5.3.11 Dist E-5 Highway

The E-5 highway is probably the most important and active highway among the
Turkey’s national highways, which connects İstanbul and Ankara. The high traffic activity
of this road creates more vibration and local load than the other roads in the area, based
on this the road itself is extracted from the cumulative road network in the area as a
separate parameter. Both of the slided mass and seed cell frequency histogram of the
distance to E-5 highway reveals important information. The bimodal distribution of these
histograms suggests that the landslides prefer two conditions. One group is caused by
the presence of E-5 highway as the frequency of seed cells show a large peak in the

129
values smaller than 400 meters. The other group is very distant to the highway, in the
order of kilometers, at which the effect of vehicle vibration would be minimal, resulting in
the conclusion that the landslides should be caused by other parameters.

Table 5.28. The descriptive stats of Dist_E-5 Road

Slided mass data Seed cells Whole Data

DISTANCE TO

Min 0 0 0
E-5

Max 3101 3112 8366,8

Mean 1254,62 1217,83 2467,49
St.Dev 737,07 899,59 1771,85

600 600
FREQUENCY

FREQUENCY
500 500

400 400

300 300

200 200

100 Std. Dev = 737,07 100 Std. Dev = 899,59

Mean = 1255 Mean = 1218
0 N = 4147,00 0 N = 4430,00
0
40
80
12
16
20
24
28
32

0
40
80
12
16
20
24
28
32
0
0
00
00
00
00
00
00

0
0
00
00
00
00
00
00
DISTE5 DISTE5
a b

Figure 5.39. The frequency distributions of Distance to E-5 highway, a. Distroad, b.

Diste-5

5.3.12. Land cover

The 93% percent of the land cover units in the study area is represented with
three units (Dense Forest 44%, Mixed zone 33%, Young forest 16%) after the maximum
likelihood classification of Landsat TM5 (Figure 4.23). Although the cumulative
percentage of the same land cover units does not change in the slided mass and seed
cells, the proportions of the Landcover units significantly change (Figure 5.40). For
example the three units are represented as 93.9 % in the slided mass data and 95.1% in
the seed cells. The main differences in the whole data and landslide data is observed as
an increase in mixed zone and forest unit (Table 5.29), on the other hand there is a
sharp decrease in the dense forest as expected. In other words, landslides are less
associated with dense forest unit.

130
Table 5.29. The % change of Landcover units

Whole data Slided mass Seed cells

Mixed 33% 42,5% 38,2%

Dense forest 44% 26% 28,1%

Forest 16% 25,4% 26,8%

Total 93% 93,9% 95,1%

Road Road
,9% 3,1%
Settlement Settlement
1,1% 1,0%

Dense Forest Dense Forest

26,0% 26,8%

Mixed Mixed

42,5% 38,2%

Forest
Forest
25,4%
Soil Moisture
Soil Moisture 28,1%
2,7%
4,1%

a b

Figure 5.40. The percentage distribution of Landcover units, a.) slided mass, b) seed
cells.

131
 CHAPTER 6

HAZARD ASSESSMENT

In this chapter various methods are considered to analyze the presented data in
Chapters 4 and 5, in order to assess the landslide hazard using Geographical
Information Systems.

6.1. Thematic Landslide Attribute spatial distribution analysis (TLASDA)

This is the simplest type of analysis that only reflects the current situation as
where the landslides have occurred in a selected period. Such a map only shows the
outlines of the individual landslides not the general scheme of landslide hazard in the
area. However, such thematic maps are useful for gathering the information about the
frequency, type, and depth. No direct information is present for the possible causes of
the landslides, as the map does not contain parameter information. Nevertheless, this
type of analysis provides a visual input showing the severity of the hazard with regard to
the attribute of the landslides.
19 combinations of attribute maps are available for the cumulative of 4 periods,
however only the maps of the last period (1994) are presented here in this chapter. All
of the maps used in the analysis are converted to binary vector maps using their
attributes in order to give more visual perception. The “Massinfo” attribute has two
available items as “scarp & body” and “scarp & path”. Although the map shows no clear
preferences, the map reveals that the larger the landslides, they possess “scarp & body”
attribute (Figure 6.1).
The “Type” attribute has two available items as “flow” and “slide”. The flow type
shows close relation with the E-5 highway as most of the flow dominated landslides
occur in the Bolu mountain pass of the E-5 highway (Figure 6.2).
The “Depth” and “Style” attributes show similar spatial preferences. The “Style”
attribute has two available items as “single” and “multiple” and the “Depth” attribute has
two available items as “shallow” and “deep”. The single type landslides are generally
smaller landslides (Figure 6.3 and Figure 6.4).

132
Figure 6.1. The Thematic Landslide Attribute spatial distribution of massinfo attribute of
1994 photo characteristics database, where black areas show Scarp & Body and grey
areas show Scarp & Path attributes.

Figure 6.2. The Thematic Landslide Attribute spatial distribution of type attribute of 1994
photo characteristics database, where black areas show Flow and grey areas show
Slide attributes.

133
Figure 6.3. The Thematic Landslide Attribute spatial distribution of style attribute of
1994 photo characteristics database, where black areas show Single and grey areas
show Multiple attributes.

Figure 6.4. The Thematic Landslide Attribute spatial distribution of depth attribute of
1994 photo characteristics database, where black areas show Shallow and grey areas
show Deep attributes.

The “Distribution of Activity” attribute has two available items as “Scarp related
activities” and “mass related activities”. Although not significant, the landslides near to
E-5 highway resembles more activity than the rest of the landslides (Figure 6.5).

134
Figure 6.5. The Thematic Landslide Attribute spatial distribution of activity attribute of
1994 photo characteristics database, where black areas show Scarp, grey areas show
Mass attributes, white polygons with black border show stable attribute. Polygons which
do not have any attribute in this field are not shown.

Due to the low level of this analysis, only some general statements regarding
the whole area could be generated. This analysis was first developed for regional
analysis, even for national scale analyses. Based on the very coarse resolution of the
analysis, the statements are very general and their validity is suspicious. After the
investigation of the above maps the severity of the landslide hazard is found to be
dominant in the southern slopes of the Asarsuyu catchment, in the northern slopes
hardly any landslide occurs. The eastern part which constitutes the higher parts of the
catchment is also landslide free. Landslides are generally occurring near to the E-5
highway and near to the greater landslides. Although these distribution maps could be
overlaid with other available maps, this was not done, as parameter versus attribute
information will be used in other analyses.

6.2. Landslide activity analysis (LACTA)

The TLASDA and the analyzed data in the previous chapter do not yield any
information about the trend of activity changes. They do not provide any information
regarding the individual landslides. This analysis is consisted of a two-dimensional
matrix resulted from non-graphical querying of the constructed GIS database. The
“Distribution of Activity” column of the photocharacteristics database is used in
constructing this matrix. First the value fields in the “Distribution of Activity” column is
reclassified in to 5 new value items, in coherence with section 5.2.5. Two new value

135
items are introduced to the matrix as “present” and “absent”, in order to consider the
total number of the landslides. The details and probable meaning of the crossed items
are presented in Table 6.1.

Table 6.1. The two dimensional matrix of LACTA

Activity in older photo set

Scarp Body Present Absent Dormant
The activity
No change, still
migrates from Reactivation
Scarp activity confined New landslide New landslide
body to scarp in scarp area
in scarp area
area
Activity in younger photo set

The activity New landslide, New landslide,

No change, still
migrates from Reactivation development development
Body activity confined
scarp to body in body area stage is not stage is not
in body area
area recorded recorded
No activity can No activity can New landslide, New landslide,
be classified but be classified but development development
Present No change
still landslide is still landslide is stage is not stage is not
present present recorded recorded
Landslide Landslide Landslide Landslide
Absent becomes becomes becomes No change becomes
invisible invisible invisible invisible
New landslide,
development
Dormant Stabilization Stabilization Stabilization No change
stage is not
recorded

Although the row named “absent” in the “activity in younger photo set” could be
attributed to some errors of photo interpretation. Once there occurs a slide there should
be its relicts. It should be noted that either the land cover or some human activities had
viped out the relict features of the landslide.
In the matrix of 1952 and 1972 (Table 6.2) all of the value items are clustered
around “present” and “absent” as 1952 period is the first photo set, hence no activity
information before 1952 could be found. The increase in number of landslides is
reflected in the matrix in “absent – present” pair, as 13 new landslides. 14 slides are
reactivated showing activation in the body area, 14 landslides are still present but no
signs of activity could be seen. 4 landslides possess activity in the scarp area, and only
one slide is vanished.
In 1972-1984 matrix the striking result is the “body-body” pair, it constitutes 10
out of 47 slides (Table 6.3). This could be attributed to the reactivation in the slided
mass area of older slides in the study area. Two new slides are recorded in 1984 period
and 8 landslides are vanished.

136
Table 6.2. Two-dimensional matrix of 1952 and 1972 periods.

Activity in 1952 photo set

Scarp Body Present Absent Dormant Σ
4
Scarp - - - - 4

Activity in 1972 photo set

(8.7%)

14
Body - - - - 14
(30.43%)

14 13
Present - - - 27
(30.43%) (28.26%)

1
Absent - - 1 - 2
(2.17%)

Dormant - - - - - -
Σ - - 33 14 -

Table 6.3. Two-dimensional matrix of 1972 and 1984 periods.

Activity in 1972 photo set

Scarp Body Present Absent Dormant Σ
3 1 4 1
Scarp - 9
(6.38%) (2.13%) (8.51%) (2.13%)
Activity in 1984 photo set

1 10 5
Body - - 16
(2.13%) (21.28%) (10.64%)

12 1
Present - - - 13
(25.53%) (2.13%)

3 5
Absent - - - 8
(6.38%) (10.64%)

1
Dormant - - - - 1
(2.13%)

Σ 4 14 27 2 -

In 1994 period two more landslides are vanished, The main activity is continuing
in body areas (“body-body” pair) (Table 6.4). On the other hand, 10 landslides are
observed as present reflecting that no significant activity could be recorded.

Table 6.4. Two-dimensional matrix of 1984 and 1994 periods.

Activity in 1984 photo set

Scarp Body Present Absent Dormant Σ
5 1 2
Scarp - - 7
(12.82%) (2.56%) (5.13%)
Activity in 1994 photo set

3 14
Body - - - 17
(7.69%) (35.9%)

10
Present - - - - 10
(25.64%)

1 1
Absent - 8 - 10
(2.56%) (2.56%)

1 1
Dormant - - - 2
(2.56%) (2.56%)

Σ 9 16 13 8 1

137
 When all of the matrices are explored together, it is evident that the main activity
is confined to body related activities, probably some forms of flow type. However, it
could not be denied that the scarp activities are also in an up trend, probably notifying
that the area is becoming geomorphologically mature, so the landslides. As a de-
forestation re-forestation cycle has been observed in the area, probably the area is
reshaping itself due to the changing geomorphological constraints. In the deforested
period new landslides and intense surface processes (soil creep, erosion, etc.) took
place changing the kinematical dynamics, yielding in the creation of new down gradient
forces of nature. Although the number of landslides decreases through 1972 to 1994,
the activity states even become more instable, not generating new slides but activating
the older ones.

6.3. Landslide Isopleth analysis (LIA)

This analysis is the most basic quantitative analysis that could be applied to
landslide hazard evaluation, and is extensively used in any kind of spatial and attribute
dependent data. The backbone of this analysis is that it consists of a counting circle.
The circle is moved through the map with a constant offset and at each location the
points falling into the counting circle is counted and recorded in the center of the
counting circle. Generally a counting circle of 1 km2 area is used as a convention to call
the density at each point as per km2. The offset amount is generally selected as the half
radius of the counting circle. In this case to achieve the 1 km2 area a radius of 564 meter
is used and the offset is taken as 250 meters (Figure 6.6).
The resulting text file is then linearly interpolated and a continuous density
surface is formed using ordinary kriging. This raster file is then converted into vector via
contouring using cubical convolution algorithm. The isopleth intervals in contouring are
selected as 20 percent (Figure 6.7).
The quality and resolution of the isopleth map is strongly dependent on the
circle size and the offset distance. The larger the size of the search circles the lower the
resolution is and the greater the generalization. However, small circle sizes are creating
redundant information as when size gets smaller and smaller the resultant map gets
similar to the original landslide inventory map. On the other hand, the offset size of the
circles also affects the final product. When offset distance exceeds the radius of the
counting circle the reason for doing an isopleth map vanishes as the data turns into
discrete rather than continuous data.

138
Figure 6.6. The mechanism of the isopleth analysis

The resultant isopleth map of Asarsuyu catchment with 564 meter radius and
250 meter offset reveals information quite conformable with the non-spatial results of the
previous chapter. In order not to create duplication the overlay of every parameter map
is not shown here, only significant results are listed below.

1. The E-5 road has cut through the maximum density areas of the
isopleth
2. Few settlements are located in the 80-100% interval of the isopleth, but
other intervals have a dispersed layout.
3. The geological units show significant preferences, as the greenschist
facies of Yedigöller formation has no landslides, and the flysch
sequence has the most landslides.
4. Strong relationship with the fault density is seen especially in the lower
density intervals

139
Figure 6.7. The isopleth map of the Asarsuyu catchment, each contour level represents
20 % landslide density.

6.4. Statistical analyses

Two types of statistical analyses are planned to be carried out in the Asarsuyu
catchment for the estimation of future landslide hazards. The aim to use statistical
methods is to increase the objectivity of the assessment, and to let the data derive its
own decision rules.
To achieve this goal, as explained in earlier chapters, the nodes of seed cells
are used as decision rule generator or training samples. 4430 seed cells are introduced
to a database containing all of the available variables in the system.
However, a major impediment is compromised in the nature of the data as all of
the bi-variate methods used are designated for some form of categorical data not for
continuous data sets. This arouses from the fact that the all-available bi-variate methods
base themselves to the landslide density or abundance in certain parameter classes. If
the continuous data is used as it is, the densities will be calculated for the whole map
and not even a single natural preference in the area will be utilized for hazard
assessment. Consequently, a continuous to discrete categorical conversion seems to be
indispensable. Some efforts have been carried out to categorize some continuous data
in the literature in the last century, unfortunately any single example in landslide hazard
assessment or in geosciences are affirmed. Some authors of mathematical, medical and
statistical experience have proposed methods of conversion. Generally these methods
depend on the optimum bin width classification of the histograms of various parameters,
and further they do not have a spatial dimension. The earliest published rule for
selecting the bin width appears to be that of Sturges (1926). This proposal is more of a

140
number-of-bins rule rather than a bin width rule itself, but essentially amounts to
choosing the bin width.

Range of Data
ĥ =
1 + log2 n
where n is the sample size

However, Scott (1979, 1992) showed that this bin width leads to an over
smoothed histogram, especially for large samples and proposed an unbiased estimation
of a probability density function, which is achieved when:

W = 3.49 σ N−1 / 3

Where W is the width of the histogram bin, σ is the standard deviation of the
distribution and N is the available samples. This estimator worked well for Gaussian
distributions, where it led to overlay large bin widths and hence over smoothing.
Friedman and Diaconis (1981) suggested a more simple method:

W = 2(IQR )N−1 / 3

Where W is the width of the histogram bin, IQR is the inter quartile range (the
th
75 percentile minus the 25th percentile) and N is the sample size. Numerical
comparisons by Emerson and Hoaglin (1983) of the Scott and Freedman-Diaconis (FD)
rules showed the FD rule led to narrower bin widths, although in practical applications
the two rules will often lead to the same choice of interval width (Izenman, 1991).

All of the methods cited above has a number of disadvantages which are listed below:

• Over smoothed class divisions

• Distribution dependent
• Valid for only one parameter map
• Cannot be applied to multi-modal distributions (multi-modality overrides the
assumptions)

Due to the cited disadvantages of former methods a new method is proposed to

classify the continuous data sets into categories, which could be called as “percentile
method” (Figure 6.8). The core of this method is dependent on equal frequencies. The
frequency domain is quite free of distribution parameters and could easily be implemented in
bimodal or multi-modal distributions. The stages of this method is as follows. First the
percentiles of each variable of seed cells are found and recorded; secondly the whole

141
Figure 6.8. The snapshot of methodology of percentile method and reclassified
parameter map production. The areas with a star are the nature’s own decision rule and
not taken into calculations as there is no landslide in these area.

142
parameter map is classified accordingly to the seed cells percentile limits (Figure 6.8.). The
resultant classes of the parameter map have same landslide density but the areas of the
percentile classes on the whole parameter map are not equal to each other. This constitutes
the natural weighting of each parameter class. Furthermore, the minimum and maximum
values of the parameter map are dependent of the seed cells data base, such as if the
whole area has elevations of the magnitude as 200 meters as the minimum and the
minimum of the seed cells are 300 meter, the area lying in between 200-300 meters are not
taken into consideration as the nature itself creates her first decision rule as not having any
landslides in that range.
All of the 13 parameters are analyzed and out of 13 excluding the 2 categorical
ones (geological map and the land cover map) the remaining 11 parameters are re-
classified according to the seed cell percentile values. All of the percentile maps are
shown in Figures 6.9, 6.10 and in 6.11 with their frequency and cumulative histograms
overlaid with percentile ranges.

Table 6.5. The percentiles of seed cells within each variable.

ASPECT SLOPE ELEV D_DRAIN D_E5 D_SETTLEDENS_FAULT D_FAY D_RIDGE D_PRO DENSDR
Valid 4430 4430 4430 4430 4430 4430 4430 4430 4430 4430 4430
N
Missing 0 0 0 0 0 0 0 0 0 0 0

Mean 168,37 22,41 661,84 118,30 1217,83 699,45 235,56 408,52 113,88 176,22 160,68

Median 180,00 22,00 680,00 108,00 1206,00 459,00 209,00 306,00 96,00 145,00 157,00

Mode 45 22 750 25 1 153 0 114 13 9 150

Std. Deviation 126,25 9,12 179,21 76,99 899,59 651,25 161,03 339,66 82,12 142,29 37,01

Variance 15939,83 83,13 32116,12 5927,73 809255,17 424130,59 25930,03 115368,76 6743,79 20245,18 1369,43

Range 359 50 830 397 3112 2625 705 1517 360 683 232

Minimum -1 1 300 0 0 4 0 0 4 0 55
Maximum 358 51 1130 397 3112 2629 705 1517 364 683 287
10 9 11 420 25 75 105 12 73 14 17 119
20 27 15 474 49 211 167 110 127 39 41 136
30 45 17 530 68 391 230 157 180 52 70 144
40 83 20 620 87 834 322 189 239 75 107 150
Percentiles 50 180 22 680 108 1206 459 209 306 96 145 157
60 225 25 730 130 1607 609 242 386 122 189 165
70 284 28 770 155 1857 804 286 478 152 241 174
80 315 30 820 183 2114 1397 357 675 191 302 187
90 333 34 880 221 2427 1825 481 977 238 382 208

143
 100

80

Cumulative Percent
60

40

20

0
-1 27 55 104 158 200 241 282 315 348
13 41 72 135 174 220 261 300 333

600

500

400

300

200

Frequency
100

0
0 40 80 120 160 200 240 280 320 360
20 60 100 140 180 220 260 300 340

a
100

90

80

70

Cumulative Percent
60

50

40

30

20

10
0
1 5 9 13 17 21 25 29 33 37 41 45 49
3 7 11 15 19 23 27 31 35 39 43 47 51

1200

1000

800

600

400

Frequency
200

0
0 5 10 15 20 25 30 35 40 45 50

b
100

90

80

70
Cumulative Percent

60

50

40

30

20

10
0
300 415 487 560 628 704 762 854 927 1045
365 455 524 610 674 725 806 886 990

600

500

400

300

200
Frequency

100

0
300 400 500 600 700 800 900 1000 1100
350 450 550 650 750 850 950 1050 1150

Figure 6.9. The percentile maps of morphology of Asarsuyu catchment, with frequency
and cumulative histograms, a) Aspect, b) Slope, c) Elevation

144
 100

90

80

70

60

Cumulative Percent
50

40

30

20

10
0

0
60
11
179
238
297
357
416
475
534
595
688
762
820
907
983
109
1168
1245
1316
42
500

400

300

200

Frequency
100

0
10
20
30
40
50
60
70
80
90
10
11 0
12 0
13 0
14 0
15
0
0
0
0
0
0
0
0
0
0
0
0
0
00
00
100

90

80

70

Cumulative Percent
60

50

40

30

20

10
0
0 67 135 201 267 333 400 470 539 611
34 102 168 234 300 366 437 504 576 662

600

500

400

300

200

Frequency
100

0
50
10
15
20
25
30
35
40
45
50
55
60
65
70
0
0
0
0
0
0
0
0
0
0
0
0
0
100

90

80

70
Cumulative Percent

60

50

40

30

20

10
0
0 36 72 108 144 180 216 252 290 328
18 54 90 126 162 198 234 271 308 357

500

400

300

200
Frequency

100

0
0 80 160 240 320 400
40 120 200 280 360
c
100

90

80

70
Cumulative Percent

60

50

40

30

20

10
0
55 80 104 128 152 176 200 224 248 277
68 92 116 140 164 188 212 236 261

800
FREQUENCY

600

400

200

d
60

80

10

12

14

16

18

20

22

24

26

28
0

Figure 6.10. The percentile maps of lineament and density patterns of Asarsuyu
catchment, with frequency and cumulative histograms, a) Distance to fault , b) Fault
density, c) Distance to Drainage d) Drainage density

145
 100

90

80

70

Cumulative Percent
60

50

40

30

20

10
0
4 44 78 112 146 180 214 248 282 317
26 61 95 129 163 197 231 265 300 336

700

600

500

400

300

Frequency
200

100

0
0 40 80 120 160 200 240 280 320 360

a
20 60 100 140 180 220 260 300 340

100

90

80

70

60

Cumulative Percent
50

40

30

20

10
0

4
87
16
248
339
411
504
582
678
768
855
952
116
1413
1673
1700
1808
1918
2049
2262
37
800

600

400

Frequency
200

0
20
40
60
80
10
12
14
16
18
20
22
24
26
0
0
0
0
00
00
00
00
00
00
00
00
00
100

90

80

70

Cumulative Percent 60

50

40

30

20

10
0
0 58 116 174 232 290 348 409 469 550
29 87 145 203 261 319 377 439 502 623

600
FREQUENCY

500

400

300

200

100

c
0
50
10
15
20
25
30
35
40
45
50
55
60
65
0
0
0
0
0
0
0
0
0
0
0
0

100

90

80

70

60
Cumulative Percent

50

40

30

20

10
0
0
10
216
326
447
636
834
108
1103
1342
1540
1607
1729
1858
1979
2196
2242
2479
2526
2893
01

500

400

300

200
Frequency

100

d
0
20
40
60
80
10
12 0
14 0
16 0
18 0
20 0
22 0
24 0
26 0
28 0
30 0
0
0
0
0
0
0
0
0
0
0
0
0
0
0
00

Figure 6.11. The percentile maps of infrastructure and distance to ridge of Asarsuyu
catchment, with frequency and cumulative histograms, a) Distance to ridge , b) Distance
to Settlement, c) Distance to Power and road network d) Distance to E-5 Highway.

146
6.4.1. Bi-variate analyses

In bi-variate analyses, as explained in Chapter 2, the core of the analysis is to

get the densities of landslide occurrences in each parameter map and in each
parameter map’s classes, and to get some data driven weights based on the class
density and the landslide density.
Two of the previously mentioned methods have been utilized in this study as:
landslide susceptibility and information value method. In the landslide susceptibility
method as the nodes of seed cells are representing 25x25 meter area in the map, the
area density method is used. A brief recapitulation of the methods is given in Table 6.6.

Table 6.6. Methodological snapshot of used two methods

LANDSLIDE SUSCEPTIBILTY INFORMATION VALUE

Npix ( SX i ) Si N i
Darea = 1000 I i = log
Npix ( X i ) S N

in which Darea = Areal density per millage In which

Si: the number of land units or
Npix( SX i ) = number of pixels with mass pixels with mass movements and the
movements within variable class Xi. presence of variable Xi,
Ni: The number of land units or
Npix ( X i ) = number of pixels within variable class pixels with variable Xi
Xi. S: The total number of land
units or pixels with mass movements
To evaluate the influence of each variable, weighting factors N: The total number of land units
should have to be introduced, which compare the calculated or pixels.
density with the overall density in the area. The formula for the
density-based area is:
The degree of a hazard for a land unit or
pixel j is calculated by the total information
value Ij

Warea = 1000
Npix ( SX i )
− 1000
∑ Npix( SX i )
Npix ( X i ) ∑ Npix( X i ) m
I j = ∑ X ij I j
i =0

in which
m = number of variables,
Xij = 0 if the variable Xi is
not present in the land unit or pixel j and 1
if the variable is present.

147
 Simply, each area of class in the reclassified parameter maps (Figures 6.9,
6.10, 6.11) is divided by the landslide density in this class. As previously noted, the
landslide densities are bound by the percentile method as approximately to %10 in each
class. Hence the natural weights of each parameter class is solely dependent on its
aerial coverage in the parameter map. Furthermore, it can be called as a weight
standardization module as it is clear that weights of any arbitrary parameter classes are
equal in concept but of course different in their values due to the areas of classes.
However, they are still not comparable. In order to achieve the standardization of both
the parameter classes in its parameter map and the natural weights of the parameter
classes with respect to other classes, from each single parameter value the value of the
sums of all of the weights of parameter classes in each map is subtracted. Hereforth,
the weight of sixth percentile of slope class has a comparable weight with the weight of
the third percentile of elevation as an example. On the other hand, a minor step to
reduce the computational difficulties are made as the weight values of each class are
normalized using the minimum weight in its parameter class. The example of this
procedure is given in Figure 6.12 for both landslide susceptibility analysis and for
information value method.
The steps shown in Figure 6.12 are carried out for all parameter maps of
Asarsuyu catchment. The values of all parameter classes are shown in Table 6.7, which
are then used to construct the hazard map.
The exploration of the weights yielded in expected values as well as some
surprising weights, such as; in the fault density parameter the weights are decreasing
when the density of fault lines decrease, of which it is expected that the rock units would
stay intact enough in the absence of landslides. However, distance to fault has a
surprising result as the first percentile (0 to 73 meter distance) gets 0 weight, which is
probably due to the fact that some other factors in combination control the activity in
those areas. Also the first percentile of distance to E-5 highway (0 to 75 meter distance)
gets the maximum weight, which overrides some physical factors in the area. Another
surprising fact was seen in the categorical parameters as the DSA class gets higher
weight value than the flyschoidal sequence, which is due to the fact that they have the
same amount of landslide but DSA is outcropping in a very restricted area.
Merely the same steps are encountered with the information value, however,
after the analysis of the resultant weights, it was decided not to use as the method is
based on log differences which do compress the mean values and speculating the
extreme or outlier values. Although the weight values are different and more sensitive to
the extremes, the signs of the values were similar to that of landslide susceptibility
analyses, which is an affirmative condition for the validity of the method used.

148
Figure 6.12 The steps through landslide susceptibility analysis

149
Table 6.7.Weight values of the all available parameter classes

Parameter Classes
Parameter map 10 20 30 40 50 60 70 80 90 100
fault density 0,26 0,00 9,65 23,87 37,26 32,81 26,93 41,21 33,73 34,72
elevation 3,88 27,74 18,68 5,11 17,34 25,69 34,35 19,67 16,97 0,00
d to ridge 0,00 2,37 3,42 4,79 4,82 7,52 7,67 9,14 11,06 22,16
d to settle 0,00 19,38 27,19 19,18 14,61 21,92 23,68 4,20 18,66 6,30
d to pr 5,53 3,65 2,10 2,54 6,90 9,19 11,70 14,30 15,55 0,00
d to fault 0,00 9,67 15,76 17,16 17,02 16,11 17,62 3,89 4,79 4,76
d to e5 99,56 46,53 31,07 2,82 8,25 7,11 22,66 22,09 15,82 -0,01
d to drain 0,00 0,35 4,44 7,67 8,72 9,50 11,46 16,00 19,30 19,83
drainage dens -0,01 5,89 20,14 26,17 22,53 22,06 18,18 12,04 10,15 1,59
slope 0,00 8,28 18,15 24,11 29,40 29,92 30,04 32,92 28,57 30,51
aspect 23,33 32,02 26,71 18,51 -0,01 3,68 3,52 11,20 15,72 22,84

quat BGY DSA tkbf talus bolugran Eocay sok tkb pliomen asarsuyu gypsum

geology 0,67 2,08 58,04 35,95 3,69 28,73 22,68 19,60 0,00 0,00 0,00 0,00

dense forest forest soil moisture mixed settlement road

landcover 4,69 29,03 0,00 14,43 0,16 124,04

After calculation of the weights, the weigh values are assigned to the initial
parameter maps. Following this all of the 13 parameters maps are added up to create
the hazard map. No extra weighting procedure for the parameter maps are used in the
summation process as the classes have been normalized and they received their
natural weightings from the data itself.
The resultant hazard map is then reclassified into 4 hazard zones (very low, low,
high, very high) using the hazard maps distribution parameters. The mean value of the
hazard map is taken as the pivot point and classes are assigned to the + and - one
standard deviations of the distribution. The resultant map and the landslide amounts in
these hazard classes are given in Figure 6.13. It is seen that 48 % of the total area is
classified as high and high hazard class and within these classes 93.3 % of seed cells
are encountered. It should also be noted that the two giant landslide bodies (Bakacak
and Bülbülderesi slides) are not taken into consideration at this stage, in order to show
the real distribution of the hazard classes and to see if any information could be
obtained from form the analyses within these landslide bodies. On the other hand 52%
of the study area is classified as low and very low hazard, which in turn hosts only 6.7 %
of total seed cells in the area. This distribution also validates that the classification is
quite reasonable.

150
 DN Value Hazard class % area covered % landslide
0-93 Very low 14,45 0,5
93-146 low 37,56 6,2
146-199 High 31,20 26,7
199-481 Very High 16,79 66,6

Figure 6.13. The hazard map and the amounts of landslides in each class as a result of
bivariate analysis. (red is very high hazard, orange is high hazard, green is low hazard
and blue is very low hazard). The grey polygon in the figure is the huge landslide body.

6.4.2. Multivariate analyses

The multivariate statistical analyses of the important causal factors for landslide
occurrence may indicate the relative contribution of each of these factors to the degree
of hazard within a defined land unit. The analyses are based on the presence or
absence of stability phenomena within these units (van Westen, 1993). Two major
trends are recorded in the literature as the standard multiple regression and discriminant
analyses. However, many of the authors who use these methods do override the
necessity of the data to be normally distributed. It could have been said of an
assumption, though this assumption is a vital one controlling the applicability of these
methods. Instead of the common literature two other multivariate techniques are utilized
in this study as factor analysis to understand better the interrelations of the parameter
maps, and the logical regression analysis to figure out a multivariate dominated hazard
map.

151
6.4.2.1. Factor analysis

Factor analysis is similar to principal components analysis in that it is a

technique for examining the interrelationships among a set of variables (Afifi and Clark,
1998). Both of these techniques differ form multiple regression analyses, as there does
not need to be a dependent and a series of independent variables. The factors obtained
in a factor analysis are selected mainly to explain the interrelationships among the
original variables. The major emphasis is placed on obtaining easily understandable
factors that convey the essential information contained in the original set of variables.
In the initial stage all of the 13 parameters of the seed cells are included in the
factor analysis, and principal axis factoring method with varimax rotation is selected as
the factor analysis method. A number of tests should have to be performed for the
validity of factor analysis with the given variables so a KMO-Bartlett test is conducted
(Table 6.8). The Kaiser-Meyer-Olkin Measure of Sampling Adequacy is a statistic which
indicates the proportion of variance in the variables which is common variance, i.e.
which might be caused by underlying factors. High values (close to 1.0) generally
indicate that a factor analysis may be useful with the available data. If the value is less
than 0.50, the results of the factor analysis probably won’t be very useful. In this case it
is nearly just over the limit as the value is 0.593. Bartlett’s test of sphericity indicates
whether the correlation matrix is an identity matrix, which would indicate that the
variables are unrelated. The significance level gives the result of the test. Very small
values (less than .05) indicate that there are probably significant relationships among
the variables. A value higher than about .10 or so may indicate that the data are not
suitable for factor analysis. Based on these critical values, factor analysis seems not to
yield a very successful result.

Table 6.8. KMO and Bartlett’s test with initial 13 variables.

KMO and Bartlett's Test

Kaiser-Meyer-Olkin Measure of Sampling Adequacy. ,593

Approx. Chi-Square 13926,75

Bartlett's Test of Sphericity df 78
Sig. ,000

The next step in the analyses is to figure out which variables are not fitting to the
model, which is done by exploring the anti-image matrices. The anti-image matrices
contain the negative partial covariances and correlations. They can give an indication of
correlations that aren’t due to the common factors. Small values indicate that the

152
variables are relatively free of unexplained correlations. Most or all values off the
diagonal should be small (close to zero). Each value on the diagonal of the anti-image
correlation matrix shows the Measure of Sampling Adequacy (MSA) for the respective
item. Values less than 0.5 may indicate variables that do not seem to fit with the
structure of the other variables. The anti image matrix of the initial 13 variables are
presented in Table 6.9, and it is seen that distance to drainage and distance to ridge are
the two variables that do not fit into the structure of the remaining variables. As this is
the first iterative pass of the system it was decided to include these parameters in order
to see the effects of these variables after and before their removal.

Table 6.9. The Anti-Image matrices of initial 13 variables.

Anti-image Matrices
ASPECT SLOPE ELEV landcover D_DRAIN DISTE5 DSETTLE DEN_FAU D_FAY D_RIDGE D_PRO DEN_DR Geocode
ASPECT ,914 -4,622E-02 4,994E-02 -9,060E-02 -8,535E-02 -7,857E-02 -7,283E-03 2,491E-02 3,996E-02 5,290E-02 3,701E-02 9,134E-02 -7,265E-04
SLOPE -4,622E-02 ,819 2,417E-02 1,966E-02 -1,785E-02 2,930E-02 -7,419E-04 -2,450E-02 7,915E-02 -,247 -5,592E-02 -3,079E-02 -,127
ELEV 4,994E-02 2,417E-02 ,713 1,565E-02 -,117 ,140 -,236 2,330E-02 7,471E-03 -6,583E-02 1,806E-02 -2,870E-02 -3,853E-02
landcover -9,060E-02 1,966E-02 1,565E-02 ,632 -3,500E-02 ,124 ,108 -2,170E-02 2,745E-02 -3,650E-02 ,153 1,891E-02 -5,119E-02
D_DRAIN -8,535E-02 -1,785E-02 -,117 -3,500E-02 ,788 -3,453E-02 8,459E-02 -4,981E-02 3,904E-02 ,158 -7,490E-02 ,263 -5,707E-02
DISTE5 -7,857E-02 2,930E-02 ,140 ,124 -3,453E-02 ,423 -,226 9,600E-02 -4,663E-02 1,803E-02 -4,477E-03 2,776E-02 -5,863E-02
Anti-image
DSETTLE -7,283E-03 -7,419E-04 -,236 ,108 8,459E-02 -,226 ,372 -,100 6,159E-02 7,264E-03 -3,253E-02 7,537E-02 -3,981E-02
Covariance
DENS_FAULT 2,491E-02 -2,450E-02 2,330E-02 -2,170E-02 -4,981E-02 9,600E-02 -,100 ,411 ,237 8,426E-02 -5,751E-02 1,746E-02 ,146
D_FAULT 3,996E-02 7,915E-02 7,471E-03 2,745E-02 3,904E-02 -4,663E-02 6,159E-02 ,237 ,393 4,922E-02 -,128 ,152 3,576E-02
D_RIDGE 5,290E-02 -,247 -6,58E-02 -3,650E-02 ,158 1,803E-02 7,264E-03 8,426E-02 4,922E-02 ,829 3,522E-02 8,372E-02 3,546E-03
D_PRO 3,701E-02 -5,592E-02 1,806E-02 ,153 -7,490E-02 -4,477E-03 -3,253E-02 -5,751E-02 -,128 3,522E-02 ,819 -1,981E-02 -6,681E-02
DENS_DRAIN 9,134E-02 -3,079E-02 -2,87E-02 1,891E-02 ,263 2,776E-02 7,537E-02 1,746E-02 ,152 8,372E-02 -1,981E-02 ,707 -9,759E-02
Geocode -7,265E-04 -,127 -3,85E-02 -5,119E-02 -5,707E-02 -5,863E-02 -3,981E-02 ,146 3,576E-02 3,546E-03 -6,681E-02 -9,759E-02 ,847
a
ASPECT ,545 -5,340E-02 6,184E-02 -,119 -,101 -,126 -1,249E-02 4,064E-02 6,669E-02 6,075E-02 4,275E-02 ,114 -8,254E-04
a
SLOPE -5,340E-02 ,611 3,163E-02 2,733E-02 -2,222E-02 4,980E-02 -1,345E-03 -4,224E-02 ,140 -,300 -6,825E-02 -4,047E-02 -,153
a
ELEV 6,184E-02 3,163E-02 ,485 2,332E-02 -,156 ,255 -,459 4,305E-02 1,412E-02 -8,561E-02 2,362E-02 -4,043E-02 -4,957E-02
a
landcover -,119 2,733E-02 2,332E-02 ,786 -4,961E-02 ,240 ,222 -4,260E-02 5,511E-02 -5,043E-02 ,213 2,829E-02 -6,997E-02
a
D_DRAIN -,101 -2,222E-02 -,156 -4,961E-02 ,433 -5,984E-02 ,156 -8,757E-02 7,020E-02 ,196 -9,323E-02 ,352 -6,987E-02
a
DISTE5 -,126 4,980E-02 ,255 ,240 -5,984E-02 ,630 -,570 ,230 -,114 3,047E-02 -7,609E-03 5,080E-02 -9,800E-02
Anti-image a
DSETTLE -1,249E-02 -1,345E-03 -,459 ,222 ,156 -,570 ,540 -,256 ,161 1,309E-02 -5,895E-02 ,147 -7,096E-02
Correlation
a
DENS_FAULT 4,064E-02 -4,224E-02 4,305E-02 -4,260E-02 -8,757E-02 ,230 -,256 ,579 ,590 ,144 -9,913E-02 3,241E-02 ,247
a
D_FAULT 6,669E-02 ,140 1,412E-02 5,511E-02 7,020E-02 -,114 ,161 ,590 ,618 8,626E-02 -,225 ,289 6,201E-02
a
D_RIDGE 6,075E-02 -,300 -8,56E-02 -5,043E-02 ,196 3,047E-02 1,309E-02 ,144 8,626E-02 ,480 4,273E-02 ,109 4,231E-03
a
D_PRO 4,275E-02 -6,825E-02 2,362E-02 ,213 -9,323E-02 -7,609E-03 -5,895E-02 -9,913E-02 -,225 4,273E-02 ,696 -2,604E-02 -8,019E-02
a
DENS_DRAIN ,114 -4,047E-02 -4,04E-02 2,829E-02 ,352 5,080E-02 ,147 3,241E-02 ,289 ,109 -2,604E-02 ,589 -,126
a
Geocode -8,254E-04 -,153 -4,96E-02 -6,997E-02 -6,987E-02 -9,800E-02 -7,096E-02 ,247 6,201E-02 4,231E-03 -8,019E-02 -,126 ,546

a. Measures of Sampling Adequacy(MSA)

The amount of the total variance explained with the initial 13 variables are
shown in Table 6.10. This table gives the amount of cumulative variance explained with
the initial solution and initial rotation of the factor analyses. As in the first steps of factor
analyses the decision rule for acceptance of the new factors are defined as “if a new
factor is to be created it should have an effect of at least equal to that of an initial
variable”. Dependent on this rule eigenvalues of smaller than 1 are not included in the
factor analysis, as a result, with the initial 13 variables only the first 4 factors are taken
into consideration and they could only explain the 57% of the total variance observed
(Figure 6.14).

153
 Scree Plot
3,0

2,5

2,0

1,5
Eigenvalue

1,0

,5

0,0
1 2 3 4 5 6 7 8 9 10 11 12 13

Factor Number
Figure 6.14. Eigenvalues of the factors.

Table 6.10. The amount of total variance explained via factors

Total Variance Explained

Extraction Sums of Squared
Initial Eigenvalues Loadings otation Sums of Squared Loading
% of Cumulative % of Cumulative % of Cumulative
Total Variance % Total Variance % Total Variance %
1 2,616 20,121 20,121 2,176 16,738 16,738 1,985 15,269 15,269
2 2,035 15,652 35,774 1,713 13,175 29,913 1,806 13,892 29,161
3 1,518 11,679 47,453 ,895 6,883 36,797 ,850 6,536 35,697
4 1,247 9,590 57,043 ,563 4,334 41,130 ,706 5,433 41,130
5 ,992 7,633 64,676
6 ,987 7,590 72,266
Factor 7 ,918 7,060 79,326
8 ,694 5,337 84,663
9 ,576 4,432 89,095
10 ,523 4,025 93,119
11 ,439 3,378 96,497
12 ,242 1,863 98,360
13 ,213 1,640 100,000
Extraction Method: Principal Axis Factoring.

The next step is to figure out what variables are responsible for the selected 4
factors, this is done by analyzing the factor matrix after varimax rotation. The Table 6.11
reports the factor loadings for each variable on the components or factors after rotation.

154
Each number represents the partial correlation between the item and the rotated factor.
The bold numbers represent the maximum correlations within these factors. Based on
these maximum correlations, it could be concluded that the first factor is composed of
human activities (Distance to settlement, Distance to E-5 highway, Distance to power
lines and road network and the land cover), the second one is attributed to primarily to
lineament pattern and minor to settlement and elevation, the third factors constituents
are the drainage system and its resultant morphodynamic attributes (distance to
drainage, drainage density and aspect), and the last factor the fourth one is attributed to
the material properties and drainage dependent morphodynamic items (Slope, Distance
to ridge and geology).
Upon the completion of this initial iterative pass, a step by step variable removal
scheme is applied based on the rules of Anti-Image Matrices. The best solution is found
after the second pass with the removal of distance to drainage and distance to ridge
parameters. After the removal of these two variables the Kaiser-Meyer-Olkin Measure of
Sampling Adequacy is increased from 0.0593 to 0,618, which is seen in Table 6.12.

Table 6.11. The rotated factor matrix and the variable loadings

Rotated Factor Matrix a

Factor
1 2 3 4
D_SETTLE ,863 ,387 4,536E-02 5,018E-02
D_E5 ,726 -,220 ,130 -2,1E-02
landcover -,635 7,658E-02 5,137E-02 6,173E-02
D_PRO ,322 -,143 8,227E-02 -3,8E-02
DENS_FAULT -,123 ,834 -1,0E-02 -,114
D_FAY ,182 -,825 ,126 -,140
ELEV ,274 ,302 4,949E-02 9,436E-02
D_DRAIN -3,7E-02 9,862E-02 ,676 -,114
DENS_DRAIN -,190 ,203 -,505 9,406E-02
ASPECT 1,652E-02 -1,0E-02 ,273 1,308E-02
SLOPE -6,8E-02 ,199 -4,5E-03 ,579
D_RIDGE -7,0E-02 -2,0E-02 -,125 ,481
Geocode ,187 -,150 1,867E-02 ,262
Extraction Method: Principal Axis Factoring.
Rotation Method: Varimax with Kaiser Normalization.
a. Rotation converged in 5 iterations.

155
Table 6.12. KMO and Bartlett’s test after removal of two variables.

KMO and Bartlett's Test

Kaiser-Meyer-Olkin Measure of Sampling Adequacy. ,618

Approx. Chi-Square 12215,87

Bartlett's Test of Sphericity df 55
Sig. ,000

Furthermore, the amount of the total variance explained with the variables are
also increased from 57 % to 63 %, of which is shown in Table 6.13. Although 63% is still
very low for such kind of analyses, further removal of the variables would probably yield
in the degradation of the model success, hence the removal scheme is ended up with
11 variables.

Table 6.13. The amount of total variance explained via factors, after removal of two
variables.

Total Variance Explained

Extraction Sums of Squared
Initial Eigenvalues Loadings Rotation Sums of Squared Loadings
% of Cumulative % of Cumulative % of Cumulative
Total Variance % Total Variance % Total Variance %
1 2,590 23,543 23,543 2,180 19,817 19,817 1,926 17,509 17,509
2 2,033 18,485 42,028 1,711 15,557 35,374 1,865 16,953 34,462
3 1,212 11,014 53,043 ,638 5,800 41,174 ,657 5,977 40,439
4 1,125 10,225 63,267 ,424 3,857 45,031 ,505 4,592 45,031
5 ,921 8,373 71,640
Factor 6 ,831 7,551 79,190
7 ,740 6,727 85,918
8 ,630 5,729 91,647
9 ,449 4,080 95,727
10 ,249 2,263 97,990
11 ,221 2,010 100,000
Extraction Method: Principal Axis Factoring.

The final factor loadings are seen in Table 6.14. Due to the different variables
encountered into the analyses, the rotation scheme and the factor loading scheme is
slightly changed, resulting in a more stable model. The stability of the model is also
seen in the generic differentiation of the factors and their responsible variables. Such as
the elevation parameter has promoted to the first factor which is more meaningful, as
the second factor is solely remained for the effects of lineaments. It is obvious that the
presence of lineaments are not controlled by elevation, on the other hand the presence
of settlement, position of the highway, land cover and the power lines and the road
network are dependent of elevation. Consequently, the first factor is dominated by
human activities. As noted above the second factor is solely dependent on lineament
pattern and density in the area. In the third factor geological units and slope is promoted

156
from 4th factor and drainage components and aspect fall into 4th factor, which in turn
more logical as the author believes that the material properties control the slope and
also the presence of landslides; hence third factor is attributed to material properties.
The remaining two variables aspect and the density of drainage is responsible from the
fourth factor and could be considered as the morphodynamical factor of the Asarsuyu
catchment.

Table 6.14. The rotated factor matrix and the variable loadings, after removal of two
variables
Rotated Factor Matrix a
Factor
1 2 3 4
D_SETTLE ,870 ,322 2,762E-02 ,198
D_E5 ,669 -,272 ,103 ,293
landcover -,658 ,146 1,409E-02 9,911E-02
D_PRO ,338 -,200 3,538E-02 -9,5E-02
ELEV ,316 ,274 3,784E-02 -4,1E-02
D_FAY ,121 -,909 -4,9E-02 3,709E-02
DENS_FAULT -2,0E-02 ,758 -,218 -6,0E-02
SLOPE -5,4E-02 ,249 ,225 -3,9E-02
Geocode ,140 -,112 ,714 2,305E-02
ASPECT -4,3E-02 -2,7E-03 3,079E-02 ,500
DENS_DRAIN -,187 ,274 ,177 -,321
Extraction Method: Principal Axis Factoring.
Rotation Method: Varimax with Kaiser Normalization.
a. Rotation converged in 6 iterations.

6.4.2.2. Logical Regression

Logical regression allows forming a multivariate regression relation between a

dependent variable and several independent variables which might affect the probability
of the searched situation. If the searched variable is a dichotomous outcome the best
method with free of predictor variable type is seem to be logistic regression (Afifi and
Clark, 1998; Atkinson and Massari, 1998; Dai et al. 2001; Lee and Min, 2001).
Binomial (or binary) logistic regression is a form of regression which is used
when the dependent is a dichotomy and the independents are continuous variables,
categorical variables, or both. Multinomial logistic regression exists to handle the case of
dependents with more classes. Logistic regression applies maximum likelihood
estimation after transforming the dependent into a logit variable (the natural log of the
odds of the dependent occurring or not). In this way, logistic regression estimates the
probability of a certain event occurring. Note that logistic regression calculates changes

157
in the log odds of the dependent, not changes in the dependent itself as Ordinary Least
Squares (OLS) regression does.
The logistic model can be written in its simplest form as:

1
P=
1 + e−z

where P is the probability of an event occurring. P is the estimated probability of

landslide occurrence. As z varies from -∞ to +∞, the probability varies from 0 to 1 on an
s shaped curve. And where z is defined as:

Z=B0+B1X1+B2X2+B3X3+…………….+BnXn

Where B0 is the intercept of the model, n is the number of independent variables, …Bi
(I=1,2,3,…..,n) is the slope coefficient of the model and Xi(I=1,2,3,…..,n) is the
independent variable.

In an extended form the equation of logistic regression could be written as:

1
Probability of belonging to population I (logit) = B0+ B1X1+ B2X2+ B3X3+…………….+ BnXn
1+ e

The advantage of the logistic regression over simple multiple regression is that,
through the addition of an appropriate link function to the usual linear regression model,
the variables may be either continuous or categorical, or any combination of both types.
In general the advantage of logistic regression modeling over the other multivariate
statistical techniques, including multiple regression analysis and discriminant analyses,
is that the dependent variable can have only two values – a dichotomous outcome -,
and that predicted values can be interpreted as probability because they are constrained
to fall into an interval between 0 and 1 (Kleinbaum, 1991). Logistic regression has many
analogies to Ordinary Least Squares (OLS) regression: logit coefficients correspond to
B coefficients in the logistic regression equation, the standardized logit coefficients
correspond to beta weights, and a pseudo R2 statistic is available to summarize the
strength of the relationship. Unlike OLS regression, however, logistic regression does
not assume linearity of relationship between the independent variables and the
dependent, does not require normally distributed variables, does not assume
homoscedasticity, and in general has less stringent requirements. The success of the
logistic regression can be assessed by looking at the classification table, showing

158
correct and incorrect classifications of the dichotomous, ordinal, or polytomous
dependent. (Afifi and Clark, 1998; Wrigley, 1984)
However, in a strict sense, it is not a probability because the extrinsic
parameters triggering the landslides such as the rainfall and earthquake vibration are
not accounted for. It might be appropriate to term it as landslide susceptibility based on
the intrinsic physical parameters.
In order to carry out the logical regression the total number of seed cells (4430)
are used. Moreover, 4430 random sample nodes are selected from the landslide free
areas of Asarsuyu catchment, that are presented in Figure 6.15. Upon the selection of
these random nodes, the values of the parameter maps are then transferred to the
database of the random data set. Following the creation of the random data set
database, the seed cells and the random set database is merged and a new column of a
binary variable indicating the presence and absence of the landslides are added. This
stage is repeated 4 times in order to have 4 different sets of random points, which in
turn would let the user to see if there is any convergence in the success of logistic
regression analyses.

Figure 6.15. The positions of selected 4430 random landslide free nodes.

The initial assumption of the variables by the logistic regression is shown in

Table 6.15.

The system test reveals that the variables and the system constructed are valid.
A Hosmer-Lemeshow test and Cox & Snell R square and Nagelkerke R square values
are obtained and the statistical package supports that the system is still valid with these
variables.

159
Table 6.15. The initial assumption of the variables.

Classification Table a,b

Predicted

LANDBIN Percentage
0 1 Correct
0 0 4430 ,0
LANDBIN
Step 0 Observed 1 0 4430 100,0
Overall Percentage 50,0
a. Constant is included in the model.
b. The cut value is ,500

After the validation of the system, logistic regression is applied to the data set
and the resultant classification table is presented in Table 6.16. The system in overall
has a success of classifying 77.3% of the pixels correctly, which is quite acceptable. The
77.3% is the highest classification success among the 4 different random sets, though
the remaining three were oscillating around 75% with plus minus 1 %.

Table 6.16. The final classification of logistic regression

Classification Table a

Predicted

LANDBIN Percentage
0 1 Correct
0 3387 1019 76,9
LANDBIN
Step 1 Observed 1 985 3445 77,8
Overall Percentage 77,3
a. The cut value is ,500

The loadings of the variables after logistic regression is presented in Table 6.17.
Based on these values the logistic regression equation is compiled as follows:

Z= 0,773046364 + (0,130082590 * GEOCODE) – (0,004154044 * DRAINAGE

DENSITY) – (0,000897442 * DISTANCE TO POWERLINES AND
ROAD NETWORK) – (0,004813297 * DISTANCE TO RIDGE) +
0,000212306 * DISTANCE TO FAULT) – (0,000525944 * DENSITY
OF FAULT) – (0,00114028 + DISTANCE TO SETTLEMENT) –
(0,001257937 * DISTANCE TO E-5 HIGHWAY) + (0,001047155 *
DISTANCE TO DRAINAGE) – (0,120726833 * LANDCOVER) +
(0,002760724 * ELEVATION) + (0,52982916 * SLOPE) –
(0,00296956 * ASPECT)

160
Table 6.17. The variables and their loadings after logistic regression
Variables in the Equation

95,0% C.I.for EXP(B)

B S.E. Wald df Sig. Exp(B) Lower Upper
ASPECT -,000296956 ,000 1,540 1 ,215 1,000 ,999 1,000
SLOPE ,052982916 ,003 318,496 1 ,000 1,054 1,048 1,061
ELEVATION ,002760724 ,000 232,298 1 ,000 1,003 1,002 1,003
LANDCOVER -,120726833 ,024 25,276 1 ,000 ,886 ,846 ,929
D_DRAIN ,001047155 ,000 6,883 1 ,009 1,001 1,000 1,002
D_E5 -,001257937 ,000 932,863 1 ,000 ,999 ,999 ,999
a D_SETTLE -,000114028 ,000 4,944 1 ,026 1,000 1,000 1,000
Step 1
DENS_FAULT -,000525944 ,000 5,854 1 ,016 ,999 ,999 1,000
D_FAY ,000212306 ,000 7,705 1 ,006 1,000 1,000 1,000
D_RIDGE -,004813297 ,000 210,251 1 ,000 ,995 ,995 ,996
D_PRO -,000897442 ,000 33,496 1 ,000 ,999 ,999 ,999
DENS_DRAIN -,004154044 ,001 31,743 1 ,000 ,996 ,994 ,997
GEOCODE ,130082590 ,017 59,496 1 ,000 1,139 1,102 1,177
Constant ,773046364 ,229 11,390 1 ,001 2,166
a. Variable(s) entered on step 1: ASPECT, SLOPE, ELEV, LANDCOVE, DISTDERE, DISTE5, DSETTLE, FAYDENS, DFAY,
DTEPE, DPRO, DENSDR, JEOCODE.

The observed groups and the predicted probabilities of these groups are
presented in Figure 6.16. It is clearly seen that in the values larger than the cutoff value,
are under dominance of binary variable 1 which is landsliding, the opposite of this
argument is valid also for the values lower than cut off value as safe pixels. In the cut off
value it is seen that both probabilities are nearly same, with a little emphasis on
landslides side. Furthermore, the values smaller than 0.25 still have some landslide
pixels which support the hazard classification scheme as very low to low hazard classes
in lower values than cut off value.

Figure 6.16. Observed groups and predicted probabilities

161
 The logit of this z function is calculated for all of the pixels of Asarsuyu
Catchment. The end members of the classification scheme is fixed as 0 being the no
hazard class and 1 as the total hazard class. The foundation of the classification is
based on these end members, hence the class boundaries are as follows: 0-0.25 very
low hazard, 0.25-0.5 low hazard, 0.5-0.75 high hazard and 0.75-1 very high hazard. This
re-classification of the hazard map is shown as the landslide hazard map of Asarsuyu
catchment (Figure 6.17.)
The hazard map produced from logical regression results in more homogenous
zones than that of bivariate analyses (Figure 6.13), especially in the end members of the
zonations; in very low and very high hazard classes. The low and very low hazard
classes constitute 72.31% of the area with corresponding 22.23% of the total landslide
seed cells. On the other hand, the rest of the area is classified as high and very high
hazard that yield in 27.68% of the area with corresponding 77.77% of the total landslide
seed cells. Based on these numbers, the multivariate analysis results gives out a more
comprehensive hazard classification in which the details of the comparisons of the
produced two hazard maps will be further investigated in the next section.
It should also be noted that the grey polygons in the both produced hazard
maps do belong to the two large landslide bodies. As previously mentioned, the
boundaries of landslide polygons are digitized from the aerial photographs of 4 different
periods. The union of these 4 polygons represents the final landslide polygon. The
hazard class of this huge polygon is assigned as very high hazard based on field
information and the aerial photographical interpretation. It was obvious from the
photographs that these giant landslide polygons are formed by the aggradations of
smaller landslides. The evidence was a relict landslide scarp at which the slided body of
this relict scarp is acting as the host of newer landslides. Also the current slope
morphology of these polygons suggests that there should be 4 or 5 different landslide
associations. Furthermore, the work carried out by Işın (1999) in the Bolu mountain
proved the slope movement by insitu monitoring. Based on these conditions it is
concluded that these landslide polygons will behave in residual shear strength
conditions, rather than peak conditions as that of the seed cells, consequently they will
be put in to very high hazard class in the final hazard map.

162
 DN Values Hazard class % area covered % landslide
0-0,25 Very low 46,86 3,97
0,25-0,5 low 25,45 18,26
0,5-0,75 High 15,58 29,48
0,75-1 Very High 12,10 48,28

Figure 6.17. The hazard map and the amounts of landslides in each class as a result of
multivariate analysis. (red is very high hazard, orange is high hazard, green is low
hazard and blue is very low hazard). The grey polygon in the figure is the giant landslide
body

6.4.3. Comparison of two produced hazard maps

Two hazard maps are produced from bivariate analysis and logical regression
analyses. Both of them produced acceptable results, as both of them classify the
majority of the seed cells in high or very high hazard classes. However, they have to be
analyzed in order to reveal which method is more successful and which method is more
accurate. Therefore, two comparison schemes are developed and presented in the
following sections.

6.4.3.1. The comparison of methods via their areas and corresponding landslide
seed cells.

It has been shown that both methods classify less than half of the study area as
high or very high hazard in conjunction more than two thirds of the seed cells in these
areas, which could be said of a success. However, when the class areas are normalized
with the landslide seed cell counts some important issues have aroused. In order to
normalize the areas, the area percent values are divided with the landslide seed cell
percent values, that is called as the seed cell area index (SCAI) density of landslides
among the classes, which is presented in Table 6.18.

163
Table 6.18 The densities of landslides among hazard classes of both methods.

Area % Seed % SCAI

Bivariate Very Low 14,45 0,5 28,8945
Low 37,56 6,4 5,8684
High 31,20 26,8 1,1642
Very High 16,79 66,8 0,2514

Logical Very Low 46,86 3,97 11,79

Low 25,45 18,26 1,3939
High 15,58 29,48 0,5284
Very High 12,10 48,28 0,2506

The logic behind SCAI lies in correct classification of seed cells within very
conservative areal extent. As a result, it is desired that the high and very high hazard
classes should have very small SCAI values and low and very low hazard classes to
have higher SCAI values.
When the SCAI values of the two methods are compared it is found that the
hazard map (LHM) produced from logical regression analyses has lower SCAI values
than that of bivariate version. Only in very low hazard class bivariate hazard map (BHM)
has a better result. In low and high hazard classes LHM has a clear superiority, however
in very high hazard class they are quite close to each other but the logical one has a
slight advantage as less area are classified as very high hazard and also its SCAI is
slightly lower. Although the BHP has a high SCAI value which is desirable for the very
low hazard class, the area classified is only 14.45 % of the total area which is very low
for settlement planning purposes. The system should be a little more flexible rather than
a mechanical rigid system, as considering the acceptable risk of the dwellers in the
area. The 3.44% increase of landslide seed cells in the very low hazard class in the
LHM ended up in 34.28 % of extra area with minimal hazard, which could be accepted.

6.4.3.2. The comparison of two methods in the spatial domain

The SCAI in the previous section does not reveal any information about the
change of hazard score within a pixel. In order to achieve the pixel basis changes or
mismatches both of the hazard maps are first re-classified into known numerical values.
The bivariate hazard map is classified as 1,2,3,4 starting from very low hazard and
ending up with very high hazard, correspondingly. The logical regression hazard map is
classified by the 10 times multiplication of new class numbers of the bivariate hazard
map, accordingly the class values are 10,20,30,40. After this re-coding process the two
maps are added up. The available outcomes are presented in Table 6.19.

164
Table 6.19. The available combinations of re-coding process and their meanings

As can be seen in Table 6.19, some combinations result in misclassified pixels.

These are dominated by the absence of hazard score in both or in one of the hazard
maps. The amount of these pixels compared to the whole classified area is merely 1.12
%, which is negligible, and shown in Figure 6.18 as the summation of 0,1,2,3,4,10,20,30
and 40 class id’s. Although it is negligible, the spatial locations of these error prone
pixels should be investigated (Figure 6.19). Generally they fall into the borders of the
hazard map, which indicate a small acceptable resampling error.

Area Distributions of Possible Classifications

25

20

15
Km2

10

0
0 1 2 3 4 10 11 12 13 14 20 21 22 23 24 30 31 32 33 34 40 41 42 43 44
Class ID

Figure 6.18. The areal distributions of classified pixels.

165
Figure 6.19. The locations of the misclassified pixels

If the two hazard maps converge in the same hazard classes after summation,
such as they possess 11, 22,33 and 44 hazard ID’s they are called as correctly
classified pixels. The areal extent of these pixels are constituting 36,67 km2 and 34.16 %
of the total area (Figure 6.18). The locations of the correctly classified pixels are shown
in Figure 6.20.

Figure 6.20. The locations of the correctly classified pixels (the numbers in the legend
indicate the pixel counts).

Another pixel association is called the acceptable classification when the hazard
classes in both of the hazard maps are differing from each other by one rank in the
hazard classification scheme. Such as a change from very low hazard to low hazard, or
change from high hazard to very high hazard is acceptable. This association is indicated
in the summation map by the following id’s 12,21,23,32,34 and 43. The area covered

166
with this association is 48,69 km2 corresponding to 45.36 % of the total area. The
locations of acceptable pixels are shown in Figure 6.21.

Figure 6.21. The locations of the acceptable pixels (the numbers in the legend indicate
the pixel counts).

The not acceptable pixels are defined as the difference of hazard ranks should
be more than one rank. Such as changing from very low hazard to very high hazard is
not acceptable. These pixels re-represented in the summation as 13, 14, 24, 31, 41 and
42. The area covered by these not acceptable pixels are 20.79 km2 with corresponding
19.6 % of the area. The locations of these pixels are shown in Figure 6.22.
Upon the investigation of the not accepted pixels of the two methods, it is seen
that the bivariate hazard map is overestimating the hazard classes relative to the logical
hazard map. As shown in the legend of Figure 6.22, six pixel values are present and the
first three of them has greater occurrences than the remaining three. The first three was
13, 14 and 24, indicating that they belong to high and very high hazard class in bivariate
map and low to very low hazard class in logistic regression map. Although this is a
relative comparison, it can be said that LHM is underestimating the hazard, however,
the classification scheme fits well in the remaining three pixel values. If LHM was
underestimating the hazard the remaining three pixel counts should be more that of the
observed values. A further investigation is made in order to find the reasons why BHM
overestimates the hazard with the aid of initial parameter maps, percentile maps and the
hazard maps. It is found that most of the errors are dependent on the percentile division
of the parameter maps. Such as the first percentile of distance from E-5 has the most
weight among the other all percentiles, which is the result of the faint E-5 highway trace
in the western part of Figure 6.22. Also the fault density and distance to fault percentile
maps are responsible to the mismatch of the two hazard maps in the areas shown with

167
arrows in Figure 6.22. On the other hand, the good correlation in the high and very high
hazard classes of both methods should not be underestimated.

Figure 6.22. The locations of the not acceptable pixels (the numbers in the legend
indicate the pixel counts).

In order to see where both hazard methods have the same or acceptable
classifications, the acceptable pixels and the correct pixels are added up, which
represents an acceptable classification of nearly %80 of the area. This result is
presented in Figure 6.23.

Figure 6.23. The locations of the correctly classified and the acceptable pixels united
(the numbers in the legend indicate the pixel counts).

After a comparison of Figure 6.23 with other hazard and error maps, and taking
into consideration about the reasons of erroneous pixels the hazard map produced from
logical regression analysis is decided to be used for further analyses. The final hazard
map with the infra structure overlaid is shown in Figure 6.24.

168
Figure 6.24. The final hazard map and the infrastructure of Asarsuyu catchment.

169
 CHAPTER 7

DISCUSSION

The aim of this chapter is to endure step by step through the landslide
assessment procedures while evaluating the possible errors and their reasons.
Furthermore, a comprehensive flowchart will be produced while discussing and
evaluating these stages.

7.1. Data Production

In harmony with the purpose and scope of the project three data domains were
used to generate the needed data. These domains are remote sensing, topographical
maps and existing geological maps.
The working size of the pixels for all of the products are selected as 25 by 25
meters. By rule of thumb, 1 map millimeter equivalent of the base map is adequate.
Finer resolution would yield in larger errors as the locations of the pixels would be
suspicious, moreover, the products other than topographic map should have to be
resampled rigorously.
The first step through any hazard assessment scheme is the evaluation of
current situation and preferably having information about the past states of the hazard.
In order to achieve this goal, remote sensing technology is used, unless there are no
archive information and no monitoring stations.
The spatial resolution of the satellite images is becoming as good as
comparable to the aerial photographs, however, the prices of the same resolution
satellite products are incomparable to conventional aerial photographs. The aerial
photographs are still the cheapest solution, in addition the military bloom of the cold war
resulted in presence of at least one full coverage of the country almost in all nations
across the world, which are hosting the true and non speculative information about the
past situation of the hazards. On the other hand, the presence of distortions in the aerial
photographs makes the information extraction stage quite painful, as the interpretation
should have to be transformed into plain cartographic coordinates. Recent

170
developments in the computer based photogrammetry reduce the time required for
these transformations but requires another investment item through the institute’s
budget. Nevertheless, with the aid of these systems the horizon of the researchers
would be quite broadened as they will be able to derive the topographical and
morphological decision rules of the period at which the aerial photographs are taken.
In this research, the spatial positions of the landslides in 4 different periods were
extracted from the aerial photographs and cordially transformed into base maps
manually (Figure 7.1). The interpretation is acquainted with a photo checklist regarding
the observable characteristics of the landslides. This checklist in the further stages
would be converted to a database attached to every single landslide.
This base map transformation, when done by experienced interpreters, is quite
comparable with the results obtained from digital products. Yet, the manual
transformation is always remaining as a large debate over one of the principal inputs of
the hazard assessment procedures. This transformation inherits a minor subjectivity to
the system depending on the experience of the interpreter.

Figure 7.1. The elements of data production stage.

Though the aerial photographs are considered as the most capable and
cheapest item within the remote sensing products, the great spectral resolution of
satellite products still have superiority over the panchromatic aerial photographs.

171
However, their restricted superiority also restricts their usages as they could now only
be used for land cover extraction or for gathering regional geological information. The
new trends of researches point out that probably within a decade the prices of the
satellite and historical aerial photographs would equalize. Furthermore, the researches
in active sensors (radar systems) and laser-dominated systems (Lidar) are flourishing,
which are promising activities for the next decade beyond the hazard assessment point
of view. Landsat TM 5, due to its spectral superiority and price advantage, is used to
extract the land cover information in Asarsuyu catchment, which would be treated as a
parameter map in further stages (Figure 7.1).
The second and the most speculative domain is the geological map domain, as
the maps have been prepared for general geological needs without considering the
special needs of the hazard assessment procedures. Further groupings regarding the
material are made rather than its stratigraphic content and age. These maps were
compiled from existing literature, so coordinate mismatches and information
inadequacies are frequent in the published maps. These geological unit mismatches
also inherits a speculative perspective to the map prepared which is also dependent on
the experience of the geologist. The lack of coordinate system in the compilation map
creates a positional error, which should be eliminated by further resampling of the
geological maps. Consequently the lineament information has same deficiencies, as
though it was both compiled from existing maps and refined after aerial photographical
interpretation (Figure 7.1)
The third and relatively the least speculative domain is the topographical map
domain, assuming that there are not any errors present in the elevation information. In
this thesis 11 variables out of 13 variables are derived from topographical maps (Table
5.16). The morphology, and infrastructure information are all derivatives from these
topographical maps (Figure 7.1). The topographical maps have minimal processing
errors as they have been digitized from the hardcopy templates by the Mapping
Headquarters. The nature of the information present in the topographical maps
generally consisted of vector coverages, as the representations of contour lines in point,
line or polygon form. This form is not the most suitable form to be used in a GIS, yet it
has to be converted into continuous maps. When the vector data has z attributes
(topographical contours), this conversion is achieved through gridding operations. On
the other hand if the vector object has no z attributes (infrastructure information), either
the density of the feature or the nearest distance of the pixels to that feature is
calculated to form a continuous map. A new approach is taken in this thesis, as taking
the true distances to the objects rather than taking the map distances, which inturn
duplicates the importance of the accuracy of the elevation model, as true distance is the
function of DEM (Figure 4.5).

172
 In order to achieve standardization, through the three domains, all of the maps
were resampled by using a reference raster as the DEM. This resampling is vital to fix
the extents and the centers of the pixels.
The products of aerial photograph and geological map domains have quite large
subjectivities, when created by novice users, which should then have to be treated very
carefully. The topographic map domain remains more objective than the others, as the
errors might rise only from mechanical procedures, which could be validated more easily
without thematic information about the maps.
Although the sequences presented here is quite objective the data domain and
data type selection is still dependent on the available data and also dependent on the
experience of the user, in which the assumption is that the selected maps will represent
adequately the sliding conditions. Extreme caution should be applied in data set
selection in order not produce redundant information. Although the distance to drainage
and drainage density pair and distance to fault and fault density pairs seems to create
duplicate information, they do possess different information. Distance to drainage maps
are controlling the chance of any stream to erode the toe part of the landslide, while the
drainage density yields in hydrological properties of the lithologies. Similarly the distance
to fault map has information about the seismic activity and if not sealed would yield in
increase in surface water with springs, while the fault density is a direct representation
of how crushed the rock units are in the area. As noted above maps like distance to nth
Strahler order, or distance to epicenters are not used in order not to create redundant
data sets and they are still not dependable as the errors in the epicenter locations are
far beyond the resolution of such studies.

7.2. Data evaluation

The maps produced include the information of a particular parameter

considering whole of the study area. The whole area is needed to be subdivided as the
decision rules should have to be created from the landslide related pixels. In order to
quantitatively discriminate the regions, a point mesh structure is laid over the region.
The landslide polygons are cropped out and the nodes are stored in “slided mass”
database with parameter map information.
A new approach is followed in the generation of decision rules of landsliding
mechanism, as it is believed that the best undisturbed morphological conditions
(conditions before landslide occurs) would be achieved by adding a buffer zone to the
crown and side areas. These buffer areas are then extracted from the point mesh and
the parameter map information is transformed. The resulted spreadsheet is converted
into a database, which is named as “seed cells”. The information in this database is
used to create the decision rules (Figure 7.2).

173
 The databases depending on “slided mass” are evaluated in Chapter 5. They
are used to characterize the landslides and to figure out the after-slide morphology.
Furthermore, the investigation of size and shape changes through 4 periods resulted in
quite large contributions for the reasons of failure. On the other hand the “seed cell”
database is used to create decision rules about sliding, as it reflects the most likely pre-
sliding conditions.
The data mining results of both the “slided mass” database and the “seed cells”
database are presented in detail in Chapter 5, hence only important issues will be raised here.
After the examination of 4 periods of polystats, fuzzystats and photo-
characteristics (photo checklist) database (Figure 7.3), it is found that by using these
three databases only very broad conclusions could have been achieved. Such as, the
removal or change of land cover and construction of significant engineering structures.
Among these databases the photo-characteristics database is found to be the most
useful one, unless there are different types of landslides in the project area. Unlike to
the literature (Carrara and Merenda, 1974; van Westen, 1993; Soeters and van Westen,
1996), it should have to be stressed that the single use of any of these databases would
neither result in any factual conclusion about the events going on in the
geomorphological history of the area nor would characterize the landslides.

Figure 7.2. Snapshot methodology for information transformation.

174
 The examination of the parameter distributions of “seed cell” database resulted
in some conclusions about the contributions of every single parameter to the landsliding
phenomena and /or the preferences of landslides on parameters. Other contributors in
literature such as Lessing et al. (1983); Turrini and Visintainer (1998); Gupta and Joshi
(1990), Carrara et.al (1991), Pachauri and Pant (1992); van Westen (1993), Chung et al.
(1995), Soeters and van Westen (1996), Gupta and Anbalagan, (1997) and Guzzetti et
al. (1999) tried to solve this parameter contributons by generalizing the whole event
such as, getting averages parameters in slope facets or creating unique condition areas
or creating dummy variables. However, none of these proposed methodologies are
applicable to individual pixel analyses; furthermore, no factual reason is valid to
generalize and to degrade the resolution of adequately detailed data. Although seed
cells let the user to upgrade the analyses to individual pixels, the interactions of
parameter maps could not still be explored efficiently.

Figure 7.3 Components of data evaluation stage

175
7.3. Hazard Analysis

The aggregation of information through the steps of data production and

evaluation yielded in a tailor-made hazard analysis. The analyses scheme started from
the very primitive one, ending with the complex statistical models in order to show the
limits of the proposed methodologies through the literature (Figure 7.4).
The most primitive analysis, Thematic Landslide Attribute Spatial Distribution
Analysis (TLASDA), is explained in Section 6.1., which is extensively used by van
westen and his colleaques (Wieczorek, 1984; van Westen, 1993; Soeters and van
Westen, 1996). However, it is found that this methodology could only be useful in
regional scale with very large number of landslides.
The second analysis, Landslide Activity Analysis (LACTA) is detailed in Section
6.2., which is found to be valuable, when used with polystats and fuzzystats databases,
as this methods validates the conclusions about the land cover and landuse change in
the study area. Single use of this analysis is of no use. However, the use of this
methodology is encouraged as being the only method to evaluate the hazard with a
historical perspective. Yet, it is still not advisable to base any hazard maps only on this
analysis. Further examples could be seen in works of Canuti et al. (1979) and van
Westen (1993).

Figure 7.4. Components of Hazard Analysis.

176
 The third analysis is Landslide Isopleth Analysis, which is explained in Section
6.3. This analysis reflects the current situation of the state, and to some extent shows
preferences of parameters that is responsible of landsliding. Parameter map crossing
could be exploited in order to show the relative preferences of landslides over parameter
maps. Although the analysis seems to be robust and free of subjective results, it can
turn into a very speculative map production method, as an experienced user can easily
manipulate the size of the counting grid and the threshold of landslide density in order to
show what he/she intends to emphasize. This analysis is also encouraged to use as an
entry for initiation of statistical analyses in regional scales.
By the fourth analysis the hazard analysis scheme enters to a new realm, the
statistical domain. Two main trends are observed in this realm as explained in Chapter
2, the bi-variate and multi-variate (Figure 7.5).

Figure 7.5. Components of Statistical analyses.

177
 In order to proceed with bi-variate analysis, the continuous parameter maps
should have to be converted into discrete maps, as the landslide densities in each class
of the parameter maps constitute the core of this analysis. However, this conversion
issue remains always unclear in literature as most of the authors use their expert
opinion for the boundaries of the classes (Carrara and Merenda, 1974; Meneraud and
Calvino, 1976; Kienholz, 1977, 1978; Stevenson, 1977; Malgot and Mahr, 1979;
Kienholz et al., 1983, 1988; Ives and Messerli, 1981; Rupke et al., 1988; Gupta and
Joshi, 1990; Pachauri and Pant,1992; van Westen, 1993, Soeters and van Westen,
1996; Gupta and Anbalagan, 1997). The use of expert opinion results in subjectivity and
removes the reproducibility of the proposed methods for different area in the globe.
Based on this issue, a data driven methodology is proposed, as the classes should be
selected according to the percentile divisions of seed cells. The data outside the seed
cell range is discarded as it is also been discarded by the nature by not having a single
landslide at these regions. In order to reduce complexity in the calculations stage, 10
classes are proposed to use, as even in 10 classes with 13 parameters, 130 different
classes should have to be maintained. The magic number of 10 as the class number, is
selected arbitrarily considering the complexity. However, it is believed that lower class
numbers would result in large generalizations in the final hazard map, where higher
class numbers would result in isolated pixels, that should have to be filtered out, where
filtering would alter the objectivity of the hazard map.
Furthermore, this data dependent division reduces the problem, in bi-variate
analyses, of what weight should be given to each parameter map, as each class acts
like a map and the ratio of the landslide density over the class area gives its natural
weight.
In bi-variate analysis, explained in Section 6.4.1., all continuous parameter
maps are converted to categorical variables, and corresponding weight values of each
class is calculated and added up to create the final hazard map. The complexity of this
analysis is moderate and it could easily be implemented to any scales. However, the
evaluation of parameter maps should be done in caution, as some parameters gets
quite large weights. Further divisions of parameter maps might be needed, as discussed
in Section 6.4.1, the distance to E-5 highway gets the maximum natural weight but it has
two distinct divisions. The mountain pass gets its desired weight, on contrary the
Kaynaşlı valley pass of this highway gets also the same weight without having a single
landslide nearby. Moreover, it should also be checked in detail that duplicate information
should not be present in the parameter weights, such as the distance to E-5 highway
and the road class of land cover in Table 6.7.
Further in the analysis, factor analysis is carried out. Although factor analysis
does not yield in a hazard map, the new factors could easily be used to validate the

178
natural weights of bi-variate analysis and could easily create a base for logical
regression analyses. Such that the largest component of factor analysis points out that
human activity is the most important factor, which is in concordance with the natural
weight scheme as the percentiles of distance to E-5 highway, land cover and distance to
settlement get quite large natural weights.
The last analysis performed in Asarsuyu Catchment is Logical Regression,
which is presented in section 6.4.2.2. Although this analysis is fairly new in Landslide
Hazard Assessment realm, a new approach is followed. The previous researches
(Atkinson and Massari, 1998; Dai et al. 2001; Lee and Min, 2001) use categorical
variables in the analysis, which creates doubts in the selection of classes of variables,
which was discussed in the former paragraphs. The new approach is to use the
continuous data as it is, in order not to alter the state and information present in the
parameter maps.
Although the understanding of this analysis requires a good statistical
background, the application is quite simple. It is encouraged to be used in medium and
large scale applications, if fed with adequate parameter maps. It is believed that the
analysis would end in more accurate results when the number of parameter maps
increases.
In both hazard maps 4 classes are used to define the degree of the hazard.
They are very low, low, high and very high. The moderate hazard class is not
implemented here, as the definition of moderate is obscure and the question of “being
moderate relative to what?” is still a great debate over hazard analysis. More willingly
than using 3 classed hazard scheme (low, moderate, high), a four class scheme is used
as dividing the moderate class into two (low and high), and stretching the end members
to very low and to very high.

7.4. Hazard Map Comparison

Upon the completion of the hazard maps, a quantitative comparison scheme

should have to be implemented. In order to do this, a two fold methodology is followed
(Figure 7.6). First one is to compare the areas of hazard classes and the corresponding
densities of landslides. This analysis is important because the idea in optimum zonation
refers to allocating minimum areas for high hazard zones, while covering most of the
landslides present in the area. Based on this optimum zonation concept, an index is
defined as Seed Cell Area Index (SCAI) and the two hazard maps are compared. The
logical regression hazard map is found to be more accurate and possesses acceptable
results relative to that of bivariate analyses.

179
 The second method is to compare the hazard classes of the two produced maps
via their spatial locations. In order to achieve this a re-coded matrix is prepared and
presented in Table 6.19. It is found that 80 % of the two maps are converging into
acceptable results. The remaining mismatched 20% of the area is reflected by the
deficiencies of bivariate analyses as discussed in previous sections. The overweight of
percentiles of distance to E-5 highway, distance to fault, fault density and the geological
map are the major sources of this mismatch.

Figure 7.6. Components of Hazard Map Comparison

After it has been decided to use the logical regression hazard map as the final
output, it is seen that, the northern slope of Asarsuyu catchment is definitely classified
as very low hazard. The attributes responsible for this classification is quite reasonable
as these areas are the least populated, the land cover is not disturbed and the cover is
dense forest, very distant to E-5 highway and the major active fault, the lithology is

180
resistant enough, although the drainage density and slope values are higher than the
rest of the area.
On the other hand the southeastern slopes are definitely on very high hazard
class. The reasons could be listed as: the removal of lateral supports by E-5 highway
cut slopes, fill areas of E-5 highway result in readily unstable unconsolidated material,
close location to active faults, high disturbance to land cover, high activity of highway
resulting in extra vibration and the presence of flyschoidal units.

181
 CHAPTER 8

CONCLUSIONS

The additional arguments to the existing literature, obtained from the landslide
hazard analysis of Asarsuyu catchment are grouped into two; conclusions about
landslide phenomena in Asarsuyu catchment and new approaches generated for data
driven landslide hazard assessment.

A. Landslide Phenomena in Asarsuyu Catchment:

1. The Asarsuyu catchment is investigated for four different periods. It is found

that except Bakacak and Bülbülderesi landslides there exists 33 individual
landslides in 1952, 45 in 1972, 39 in 1984 and 37 in 1994.
2. The major causes for generation and for reactivation of landslides is found
to be the change in land cover, and the increasing interaction of human
activity resulting in land use changes.
3. The northern slopes of Asarsuyu catchment is classified as very low hazard.
The attributes responsible for this classification is quite reasonable as these
areas are the least populated, the land cover is not disturbed, dense forest,
very distant to E-5 highway and to the major active fault and the lithology is
intact enough.
4. The Southeastern slopes, especially the Bolu Mountain Highway pass are
definitely on very high hazard class. The reasons could be listed as: the
removal of lateral supports by E-5 highway cut slopes, fill areas of E-5
highway resulting in readily unstable unconsolidated material, close location
to active faults, high disturbance to land cover, high activity of highway
resulting in extra vibration, and the presence of flyschoidal units.

182
B. New Approaches for Data Driven Landslide Hazard Assessment Procedures:

1. In the conversion of non attributed continuous data into categorical

variables, true distances are used rather than map distances.
2. In the generation of decision rules of landsliding mechanism, a buffer zone
is added to the crown and side areas, to reveal the best undisturbed
morphological conditions to be stored as attributes of a new database called
“Seed Cell Database”.
3. The decision rules are directly based on Seed Cell Database such as the
data outside the seed cell range is discarded as it is also been discarded by
the nature by not having a single landslide.
4. In diving the classes of parameter maps, a new method is proposed and
presented as percentile method, which is also dependent on “Seed Cell
Database”.
5. Factor analysis is used to figure out new obscure factors present in the
current data set.
6. Logical regression is implemented and forced to be used with both
categorical and continuous data sets.
7. For hazard map comparison a two fold methodology is proposed: 1).area
comparison in which a new index SCAI is introduced and 2). Spatial
comparison in which re-coding matrix is presented.

Following issues can be listed as recommendations for future research:

1. High resolution remote sensing products can be used to derive DEM’s of

any desired time period in order to check the system accuracy and to re-
establish the hazard assessment at any time
2. New data sets such as: rainfall, geotechnical, hydrological and soil map
related data, could be implemented in this hazard assessment procedures,
if problems in data production stages are solved.
3. Seismic data could be implemented regarding different lithologies, if
microtremor data, and P and S wave distribution maps are made available.

183
 REFERENCES

Abdülselamoğlu, S., 1959, Almacıkdağı ile Mudurnu ve Göynük civarının jeolojisi,

İstanbul Üniversitesi, Fen Fakültesi Monografileri, 14.

Afifi, A.A. and Clark, V, 1998, Computer aided multivariate analysis, Chapman-Hall,
London, 455p.

Agterberg, F.P., Bonham-Carter, G.F., and Wright, D.F., 1990, Statistical pattern
integration for mineral exploration, in: Proceedings COGEODATA Symposium, Espo
Finland, Gaal G. ed.

Aktimur, T., Algan, Ü., Ateş, Ş., Oral, A., Ünsal, Y., Karatosun, H., Öztürk, V. and
Sönmez, M., 1983, Bolu ve yakın çevresinin yerbilimleri sorunları ve muhtemel
çözümleri, MTA Rap No:7387 (unpublished)

Akyüz, H.S., Barka, A., Altunel, E., Hartleb, R. and Sunal, G., 2000, Field observations
and slip distributions of the November 12, 1999 Düzce earthquake (M=7.1), Bolu –
Turkey. In 1999 Izmit and Düzce Earthquakes: preliminary results, Barka, A., Kozacı,
Ö., Akyüz, S. and Altunel, E. Istanbul University Press. Istanbul. 63-70.

Ambraseys, N.N., 1970, Some characteristic features of the North Anatolian Fault Zone,
Tectonophysics, 9, 143-165.

Ambraseys, N.N. and Finkel, C., 1995, The Anatolian Earthquake of 17 August 1668,
Proc of the Symp.on Historical Seismograms and Earthquakes, 400-407.

Anderson, M.G. and Richards, K.S., 1987, Slope Stability. Geotechnical Engineering
and Geomorphology, Wiley and Sons, New York, 648 p.

Archibald, P.D., 1987, GIS and remote sensing data integration, Geocarto International,
Vol.3, pp. 67- 72.

Astaldi, 1990, Geological report: Preliminary design Anatolian Motorway Gümüşova-

Gerede stretch No:2, Report number 2014.

Astaldi, 1995, Updating geological research results in the Bolu and Düzce area, Barka,
A., Astaldi report no: 45,110/R/2097

Atkinson, P.M. and Massari, R., 1998, Generalized linear modelling of susceptibility to
landsliding in the central Apennines, Italy. Computers and Geosciences, 24, 4, 373-385.

Aydan, Ö., Ulusay, R., Kumsar, H. and Tuncay, E., 2000, Site investigation and
engineering evaluation of the Düzce-Bolu earthquake of November 12, 1999, Turkish
Earthquake Foundation, TDV/DR 09-51, 220p.

Aydan, Ö. and Dalgıç, S., 1998, Prediction of deformation behaviour of 3-lanes Bolu
Tunnels through squeezing rocks of North Anatolian Fault Zone (NAFZ). Regional Symp
on Sedimentary Rock Engineering. Taipei, Taiwan, 1998, 228-233

184
Aydın, M., Serdar., S., Şahintürk, Ö., Yazman, M., Çokuğraş, R.,Demir, O. and Özçelik,
Y., 1987, Çamdağ (Sakarya) – Sünnicedağ (Bolu) yöresinin jeolojisi, TJK bülteni, 30, 1-
14.

Barka, A., and Erdik, M., 1993, Active Faults of Gümüşova-Gerede highway, Bolu
Mountain, Astaldi report, 25p.

Batum, İ., 1968, 1/25,000 ölçekli Adapazarı G26-b3 paftasının Jeolojisi, MTA Rap
No:4778 (unpublished).

Blumenthal, M., 1948, Bolu civarı ile Aşağı Kızılırmak mecrası arasındaki Kuzey
Anadolu silsilelerinin jeolojisi, MTA Yayınları, Seri B, No:13, 265p.

Bonham-Carter, G.F., 1996, Geographic information systems for geoscientists,

modelling with GIS. Pergamon Press, Canada, 398 p.

Bonham-Carter, G.F., Agterberg, F.P., Wright, D.F., 1990, Weights of evidence

modelling: a new approach to mapping mineral potential. Geologcal Survey of Canada
Paper 8-9. Agterberg, F.P. and Bonham and Carter, G.F. (eds). Ottowa, Canada 171-
183 pp.

Brabb, E.E., 1984, Innovative approches to landslide hazard and risk mapping. In Proc.,
Fourth International Symposium on Landslides, Canadian Geotechnical Society,
Toronto, Canada, vol: 1, pp. 307-324.

Bromhead, E.N., 1986, The stability of slopes. Surrey University Press, Surrey, 333 p.

Brunsden, D., Doornkamp, J.C., Fookes, P.G., Jones, D.K.C. and Kelly, J.M.H, 1975,
Large scale geomorphological mapping and highway design. Quarterly Journal of
Engineering Geology, Vol 8, 227-253.

Burrough, P.A., 1986, Principals of geographical information systems and land

resources assessment, Clerandon Press, Oxford, England, 194 pp.

Canik, B., 1980, Bolu sıcak su kaynaklarının hidrojeolojisi, Selçuk Üniversitesi Fen
Fakültesi Yayınları, 1, 74 p.

Canuti P., Frascati, F., Garzonio, C.A., and Rodolfi, C., 1985, Stabilita dei versanti nell
area representative di Montespertoli (Firenze) S.E.L.C.A. Firenze. 2 map sheets.

Canuti P., Frascati, F., Garzonio, C.A., and Rodolfi, C., 1986, Slope stability mapping in
Tuscany, Italy. In International Geomorphology, Part 1, Gardiner, V. (ed). Wiley and
Sons, New York, 231-239.

Canuti, P., Frascati, F., Garzonio, C.A., and Rodolfi, C., 1979, Dinamica morfologica di
un ambiente sogetto a fenomeni franosi e ad Intensa attiva agricola. Publ. No: 142.
Consiglio Nazionalle delle Ricerche, Perugia, Italy, pp-81-102.

Carrara A., 1993, Uncertainity in evaluating landslide hazard and risk, in Prediction and
Percention of Natural Hazards ed. Nemec, J., Nigs, J.M. and Siccardi, F., Kluwer,
Dordrecht, The Netherlands, 101-111.

Carrara, A., 1983, Multivariate models for landslide hazard evaluation. Mathematical
Geology, vol.15, no.3, pp.403-427.

185
Carrara, A., 1988, Landslide hazard mapping by statistical methods: A “Black Box”
approach. In Workshop on Natural Disaster in European Medditerranean Countries,
Perugia, Italy, Consiglio Nazionale delle Ricerche, Perugia, pp. 205-224.

Carrara, A. and Merenda, L., 1974, Metodologia per un censimento degli eventio
franoso in Calabria. Geologia Applicata e Idrogeologia, Vol 10, 237-255.

Carrara, A., Cardinalli, M., Detti, R., Guzzetti F., Pasqui, V. and Reichenbach. P., 1990,
Geographical information systems and multivariate models in landslide hazard
evaluation. In ALPS 90 Alpine Landslide Practical Seminar, Sixth International
Conference and Field Workshop on Landslides, Aug 31 - Sept 12, Milan, Italy,
Universita degli Studi de Milano., pp.17-28.

Carrara, A., Cardinalli, M., Detti, R., Guzzetti F., Pasqui, V. and Reichenbach. P., 1991,
GIS techniques and statistical models in evaluating landslide hazard, Earth Surface
Processes and Landforms, vol.16, no. 5, pp. 427-445.

Carrara A., Cardinali M., and Guzzetti F., 1992, Uncertainity in assessing landslide
hazard and risk. ITC Journal, 2, 172-183.

Carrara, A., 1989, Landslide hazard mapping by statistical methods: A black box
approach. Proc. Int. Workshop Natural Disasters in Europ.-Mediterr. Countries, Perugia,
June 27 – July 1, 1988, CNR-USNSF, 205-224.

Cerit, O., 1983, Mengen (Bolu NE) yöresinin jeolojik incelenmesi, Yüksek lisans tezi,
Hacettepe Ü., 160 s.

Cerit, O., 1990, Bolu Masifinin jeolojik ve tektonik incelemesi, Doktora Tezi, Hacettepe
Ü., 217p.

Chandler J.H. and Brunsden D., 1995, Steady state behaviour of the Black Ven
Mudslide: the application of archival analytical photogrammetry to studies of landform
change. Earth Surface Processes and Landforms, 20, 255-275.

Chandler J.H. and Moore, R., 1989, Analytical photogrammetry: A method for
monitoring slope instability. Quarterly Journal of Engineering Geology, 22, 97-110.

Chung, C.J. and Fabbri, A.G., 1993, The representation of geoscience information for
data Integration. Nonrenewable Resources, 2-3, 122-139.

Chung, C.F., Fabbri, A.G., van Westen, C.J., 1995, Multivariate regression analyses for
landslide hazard zonation. In: Geographical Information Systems in Assessing Natural
Hazards, Carrara, A. and Guzzetti, F. (eds). Kluver Academic Publishers, Netherlands,
107-134.

Congalton, Russell G., 1991, A review of assessing the accuracy of classifications of

remotely sensed data, Remote Sensing Environment, Vol.37, pp. 35- 46.

Crippen, R.W. ,1990, Calculating the vegetation index faster. Remote Sensing of
Environment, vol 34., pp. 71-73.

Crist, E.P. and Cicone, R.C., 1984, Application of the Tasseled cap concept simulated to
Thematic mapper data. Photogrammetric engineering and remote sensing, 50, 3, 343-
352.

Crozier, M.J., 1973, Techniques for the morphometric analysis of landslips. In: Slope
Stability, Brunsden, D. and Prior, D.B. (eds), Wiley and Sons, New York, pp 103-142.

186
Crozier, M.J., 1986, Landslides: Causes, Consequences and Environment. Croom
Helm, England, 245 p.

Cruden, D.M., 1991, A simple definition of a landslide, Bulletin of International

Association of Engineering Geology, 43, 27-29.

Dai, F.C., Lee, C.F. and Xu, Z.W., 2001, Assessment of landslide susceptibility on the
natural terrain of Lantau island, Hong Kong. Environmental Geology, 40,3, 381-391.

Dalgıç, S., 1994a, Anadolu otoyolu Bolu dağı geçişinin mühendislik jeolojisi, Doktora
Tezi, İstanbul Üniversitesi, Fen bilimleri Enstitüsü, 213p. (unpublished)

Dalgıç, S., 1994b, Anadolu otoyolu Bolu dağı geçişinin mühendislik jeolojisi, Türkiye
Jeoloji Kurultayı Bülteni, 9, 393-397.

Dalgıç, S., 1997, Distribution of lithological and structural features along Bolu tunnel,
Anatolian Motorway, Geological Bulletin of Turkey, 40-2,39-48.

Dalgıç, S.,1998a, Slope stability problems of the weak rocks in the Asarsuyu pass of the
Anatolian Motorway. Bull. Eng. Geol. Env., 57, 199-206.

Dalgic S, 1998b, Selection of crushed rock quarries for the construction of the Anatolian
Motorway ,Environmental & Engineering Geoscience ,4: (4) 511-518

Dalgıç, S. and Gözübol, A.M., 1995, Bolu otoyol tünelinde stabilite problemleri.
Geosound, 27, 73-80.

Dalgic, S., Şimşek, O. and Gözübol, A.M., 1995 Anadolu otoyolu Bolu Yumrukaya
geçişinde heyelanların etkisi, İkinci Ulusal heyelan Sempozyumu, Adapazarı, 163-170.

Davis, J.C., 1986, Statistical and data analysis in geology, Wiley and Sons, New York,
546 pp.

Demirtaş., R. and Yılmaz, R., 1996, Türkiyenin Sismotektoniği, T.C. Bayındırlık ve İskan
Bakanlığı, Ankara., 91 p.

Emerson, J.D., and Hoaglin, D.C., 1983, Stem and leaf displays. In: Understanding
robust and exploratory data analysis. Hoaglin, D.C., Mossteiler, F. and Tukey, J.W.,
Wiley and Sons, New York.

Erendil, M., Aksay, S., Kuşçu, İ., Oral, A., Tunay, G. and Temren, A., 1991, Bolu masifi
ve çevresinin jeolojisi, MTA Rap No:9425 (unpublished)

Faust, N.L., Anderson, W.H., Star, J.L., 1991, Geographic information systems and
remote sensing future computing environment, Photogrammetric Engineering & Remote
Sensing, Vol.57, No. 6, pp. 655- 668.

Fookes, P.G., Dale, S.G., Land, J.M., 1991, Some observations on a comparative aerial
photography interpretation of a landslipped area. Quarterly Journal of Engineering
Geology, 24, 249-265.

Freedman, D., and Diaconis, P., 1981, On the histograms a density estimator L2 theory.
Zeitschrift für Wahrscheinlichkeustheorie und verwandte Gebette, 57, 453-476.

187
Gagon, H., 1975, Remote sensing of landslide hazards on quick clays of Eastern
Canada. Proceedings 10th International symposium on remote sensing of environment.
ERIM, Ann Arbor, Michigan, 803-810.

Gedik, A., ve Alkaş, İ, 1996, Bolu yöresinin Jeolojisi ve karbondioksid olanakları, MTA
rap No: 10147. (unpublished).

Gee, M.D., 1992, Classification of landslide hazard zonation methods and a test of
predictive capability. Proocedings 6th International symposium of landslides,
Christchurch, New Zealand, Vol 2, 947-952.

Gillespie, A.R., Kahle, A.B. and Walker, R.E., (1987), Color enhancement of highly
correlated images. I.Decorrelation and HSI contrast stretches. Remote Sensing of
Environment, 20, pp. 209-235.

Görmüş, S., 1980, Yığılca (Bolu NW) yöresinin jeolojik incelemesi, Doktora Tezi,
Hacettepe Üniversitesi. 210s.

Görmüş, S., 1982b, Yığılca (Bolu KB) yöresinin tektoniği ve jeolojik evrimi, H.Ü.
Yerbilimleri, 9, 133-140.

Görmüş, S., 1982a, Yığılca (Bolu KB) yöresinin stratigrafisi, H.Ü. Yerbilimleri, 9, 91-110.

Gözübol, A.M., 1978, Mudurnu-Dokurcun-Abant (Bolu ili) alanının jeoloji incelemesi ve

Kuzey Anadolu yarılımının yapısal özellikleri, Doktora tezi, İstanbul Üniversitesi Fen
Fakültesi Tatbiki Jeoloji Kürsüsü.

Graham, J., 1984, Methods of stability analysis. In: Slope Stability, Brunsden, D. and
Prior, D.B. (eds), Wiley and Sons, New York, pp 171-215.

Grunder, M., 1980, Beispiel einer anwendungsorientierten gefahrenkartierung 1:25.000

für forstliche sanierungsprojeckte im berner oberland (Schweiz). Proceedings
INTRAPRAEVENT, Bad Ischl,Austria, Band 4, p. 353-360.

Güler, B., 1999, Bolu Dağı Bakacak (Bolu) yöresi jipslerinin jeolojik incelemesi, Yüksek
Lisans Tezi, Hacettepe Ü., 46 s.

Gupta, P. and Anbalagan, R.,1997, Slope stability of Tehri dam reservoir area, India
using landslide hazard zonation (LHZ) mapping. Quarterly Journal of Engineering
Geology, 30, 27-36

Gupta, R.P. and Joshi, B.C., 1990, Landslide hazard zoning using the GIS approach –
A case study from the Ramganga catchment, Himalayas. Engineering Geology, 28,
119-131.

Guzzetti, F., Carrara, A., Cardinalli, M., Reichenbach, P.,1999, Landslide hazard
evaluation: A review of current techniques and their application In a multi-case study,
central Italy. Geomorphology, 31, 181-216

Hammond, C.J., Prellwitz, R.W., and Miller, S.M., 1992, Landslide hazard assessment
using monte carlo simulation. In Proc. Sixth International Symposium on Landslides
(Bell, D.H. ed), Christchurch, New Zealand, A.A. Balkema, Rotterdam, Netherlands,
vol.2, pp 959-964.

Hansen, A., 1984, Landslide hazard analysis, In: Brunsden D. and Prior, D.B., Slope
Instability, p. 523-602. Wiley, Newyork.

188
Hartlen, J. and Viberg, L., 1988, General report: evaluation of landslide hazard. In Proc.,
Fifth International Symposium in Landslides (C.Bonnard ed), Lausanne, A.A. Balkema,
Rotterdam, Netherlands, Vol.1, pp. 3-35.

Hearn, G.J., 1992, Terrain hazard mapping at Ok Tedi mine, Papua New Guinea.
Proocedings 6th International symposium of landslides, Christchurch, New Zealand, Vol
2, 971-976.

Hermelin, M., 1990, Bases fisicas para los planes de desarollo de los municipios de
Riseralda. AGID report no:13. Environmental geology and natural hazards of the
Andean region, 269-274.

Hermelin, M., 1992, Medio ambiente, planes de desarolloy toma de decisions.

Proceedings 1st Simposio Internacional sobre sensors remotos y sistemas de
informacion geografica (SIG) para el estudşo de riesgos naturals, Bogota, Colombia,
646-663.

Hoek, E. and Bray, J.W., 1981, Rockslope engineering, Instıtute of mining and
Metallurgy, London, 358 p.

Huang, S.L. and Chen, B.K., 1991, Intergration of Landsat and terrain information for
landslide study. Proceedings 8th Thematic Conference on Geological Remote Sensing
(ERIM), Denver, Colorado, Vol 2, 743-754.

International Association of Engineering Geology (IAEG), 1976, Engineering geological

maps: A guide to their preparation, UNESCO Press, Paris, 79pp.

Işın, Z.S., 1999, D-100 (E-5) Karayolu, Bolu Dağı geçişi, Kaynaşlı – Elmalık arasının
mühendislik jeolojisi ve Taşaltı Heyelan Bölgesinin jeoteknik değerlendirilmesi, Yüksek
Lisans Tezi. İ.T.Ü., 145s.

Ives, J.D. and Messerli, B., 1981, Mountain hazard mapping in Nepal. Introduction to
mountain research project. Mountain Research and Development, Vol 1, No: 3-4, 223-
230.

Izenman, A.J., 1991, Recent developments in non-parametric density estimation.

Journal of the American Statistical Association. 86, 413, 205-224.

Jenson, S.K. and Domingue, J.O. (1988). Extracting topographic structure from digital
elevation data for geographic information system analysis. Photogrammetric
Engineering and Remote Sensing, 54, 1593-1600.

Jones, F.O., Embody, D.R. and Peterson, W.C., 1961, Landslides along the Colombia
river valley, North-eastern Washintgon. U.S.G.S. prof. Paper 367: 98 p. Washington.

Kam, Tin-Seong , 1995, Integrating GIS & RS techniques for urban land cover and land
use analysis, Photogrammetric Engineering & Remote Sensing, Vol.57, 655- 668.

Kauth, R.J. and Thomas, G.S., 1976, The tasseled cap – A graphic description of
spectral – Temporal development of agricultural crops as seen by Landsat. Proceedings
2nd International Symposium on Machien Processing of Remotely Sensed data, Purdue
University, West Lafayette, IN. 41-51

Kaya, O., 1973, The Devonian and Lower Carboniferous stratigraphy of İstinye,
Bostancı and Büyükada subareas, Paleozoic of İstanbul, ed. Kaya,O., 40, 1-36.

189
Kaya, O., 1982, Ereğli, Yığılca, Bolu Kuzey, Mengen alanlarının stratigrafisi ve yapı
özellikleri. TPAO Rap No: 1639. (unpublished).

Kaya, O. and Dizer, A., 1981-1982 a, Mengen (Bolu) Eosen kömür havzasının
stratigrafisi. MTA dergisi 97/98, 123 – 133.

Kaya, O. and Dizer, A., 1981-1982 b, Bolu Kuzeyi Üst Kretase ve paleojen kayarının
stratigrafisi ve yapısı. MTA dergisi, 97/98, 57-77.

Kaya, O., Dizer, A., Tansel, İ. and Özer, S., 1986, Yığılca (Bolu) alanı üst Kretase ve
Paleojeninin stratigrafisi, MTA derg., 107, 13-31.

Ketin, İ, 1955, Akçakoca-Düzce bölgesinin jeolojik lövesi hakkında memuar, MTA Rap
No:2277 (unpublished)

Ketin, İ., 1966, Anadolunun Tektonik birlikleri. MTA Enstitüsü Dergisi, 66

Ketin, İ., 1967, Bolu – Gerede – Mengen ve Yığılca bölgesindeki Paleozoyik oluşuklara
ait jeolojik rapor, TPAO Arama Grubu Rap No: 379. (unpublished).

Ketin, İ., 1969, Kuzey Anadolu Fayı hakkında. MTA Enstitusü Dergisi, 72, 1-27.

Kienholz, H., 1977, Kombinierte geomorphologische gefahrenkarte 1:10,000 von

Grindelwald. Catena, vol.3, pp 265-294.

Kienholz, H., 1978, Maps of geomorphology and natural hazards of Grindelwald,

Switzerland, scale 1:10.000, Arctic and Alpine Research, 10, 169-184.

Kienholz, H., 1980, Zur Anwendung des luftbildes bei der mittelmasstaabingen
gefahrentkartierung für regionalplanereische zwecke in sclecht erschlossenen
gebirgsraumen anhand von erfahrungen aus kartierungen in den Colorado Rocky
Mountains. Proceedings INTRAPRAEVENT, Bad Ischl,Austria, Band 3, p. 155-172.

Kienholz, H., 1984, Hangstabilitatas un gefahrenbeurteilung im nepalischen

mittelgebirge. Proceedings INTRAPRAEVENT, Villach, Austria, Band 2, p. 331-342.

Kienholz, H., Bichsel, M., Grunder, M. and Mool, P., 1983, Kathmandu Kakani area,
Nepal: Mountain hazards and slope stability map. United Nations University, Mountain
Hazards mapping Project. Map No :4, Scale 1:10.000.

Kienholz, H., Mani, P and Klay, M., 1988, Rigi Nordlene, Beurteilung der naturgefahren
und waldbaliche prioritatenfestlegung. Proceedings INTRAPRAEVENT, Graz, Austria,
Band 1, p. 161-174.

Kleinbaum, D.G., 1991, Logistic regression a self learning text. Springer Verlag, New
York.

Kobashi, S., and Suzuki, 1988, Hazard index for the judgement of slope stability in the
Rokko Mountain region. In Proc., Interpraevent, 1988, Graz, Austria, vol 1., pp. 223-233.

Koral, H., Dalgıç, S. and Gözübol, 1994, Bolu masifi gabroyik kayalarında mikrofabrik
çalışma. Türkiye Jeoloji Kurultayı Bülteni, 9, 183-187.

Lambe, T.W. and Whitman, R.V., 1969, Soil Mechanics, Wiley and Sons, New York, 553 p.

Lee, S. and Min, K., 2001, Statistical analysis of landslide susceptibility at Yongin,
Korea. Environmental Geology, 40, 1095-1113.

190
Lessing, P., Messina, C.P. and Fonner, R.F., 1983, Landslide risk assessment.
Environmental Geology, Vol 5, No 2, 92-99.

Lillesand, T.M. and Kiefer, R.W., 1996, Remote sensing and image interpretation. Wiley
and Sons, New York, 750p.

Lunetta, Ross S., Congalton, and Russell G., 1991, Remote sensing and geographic
information systems data integration: error sources and research issues,
Photogrammetric Engineering & Remote Sensing, Vol.57, No. 6, pp. 677- 687.

Malgot, J. and Mahr, T., 1979, Engineering geological mapping of the west Carpathian
landslide areas. Bull.Int. Assoc. Eng. Geologists, 19, 116-121.

Mather, P.M., 1999, Computer processing of remotely sensed data: An Introduction,

Chichester, New York, Wiley and Sons, 292 p.

Mc Donald, H.C. and Grubbs, R.C., 1975, Landsat imagery analysis: an aid for
predicting landslide prone areas for highway construction. NASA Earth Resource
Survey Symposium, Houston, Texas, Vol 1-b, 769-778.

Meneraud, J.P. and Calvino, A., 1976, Carte ZERMOS, Zones exposees a des risques
lies aux mouvements du sol et du sous-sol a 1:25.000, region de la Moyenne Vesubie
(Alpes Maritimes). Bureau de Recherches Geologiques Minieres, Orleans, France, 11p.

Mollard, J.D., 1986, Early regional photointerpretattion and geological studies of

landslide terrain along the south Saskatchewan and Qu'appelle River valleys. Canadian
Geotechnical Journal, 23, 79-83.

Mulder, H.F.H.M, 1991, Assessment of landslide hazard, Ph.D. Thesis, University of

Utrecht, Netherlands, pp.150.

Mulder, H.F.H.M, and van Asch, T.W.J, 1988, A stochastic approach to landslide hazard
determination in a forested area. In Proc., Fifth International Symposium in Landslides
(C.Bonnard ed), Lausanne, A.A. Balkema, Rotterdam, Netherlands, Vol.2, pp. 1207-
1210.

Neugebauer, J., 1994, Closing-up Structures, alternatives to pull-apart basins; the effect
of bends in the North Anatolian Fault, Turkey, Terra Nova, 6, 359-365.

Neugebauer, J., 1995, Structures and kinematics of the North Anatolian Fault zone,
Adapazarı-Bolu region, northwest Turkey. Tectonophysics, 243, 119-134.

Neugebauer, J., Löflerr, M., Berckhemer, H., Yatman, A., 1997, Seismic observations at
an overstep of the western North Anatolian Fault (Abant-Spanca region, Turkey). Geol
Rundsch, 86, 93-102.

Neuland, H., 1976, A Prediction model for landslips. Catena, 3, 215-230.

Nurlu, M., 1993, Kuzey Anadolu Fay Zonunda, (Bolu-Sapanca gölü arası) etken olan
gerilimlerin fay analizleri ve uydu görüntüleri yardımı ile saptanması. 46. Türkiye Jeoloji
Kurultayı Bildiri özleri. p. 121.

Okimura, T. and Kawatani, T., 1986, Mapping of potential surface-failure sites on granite
mountain slopes. International Geomorphology, Wiley, New York, 121-138.

191
Orkan, N.İ., Aktimur, H., Sungur, G. and Işıklar, İ.S., 1977, Ankara-İstanbul otoyolu
güzergahı Bolu dağı jeoloji incelemesi, MTA Rap No:2277 (unpublished)

Öztürk A., İnan, S. and Tutkun, Z., 1984, Abant-Yeniçağa (Bolu) yöresinin tektoniği,
Cumhuriyet Universitesi, Yerbilimleri, 1, 1-18.

Pachauri, A.K. and Pant, M., 1992, Landslide hazard mapping based on geological
attributes. Engineering Geology, 32, 81-100.

Perrot, A., 1988, Cartographie des risques de glissement en Lorainne. Fifth International
Symposium in Landslides (C.Bonnard ed), Lausanne, A.A. Balkema, Rotterdam,
Netherlands, Vol.2, pp. 1217-1222.

Rengers, N., 1986, Use of (stereo-) orthophotography prepared from aerial and
terrestrial photographs for engineering geological maps and plans. Proc. 7 Th Int Symp
on Remote Sensing for Resources Development and Environ. 665-668.

Rengers, N., Soeters, R. and van Westen, C.J., 1992, Remote Sensing and GIS applied
to mountain hazard mapping. Episodes, 15, No:1, 36-45.

Rib, H.T. and Liang, T., 1978, "Recognition and identification", In: Schuster, R.L. and
Krizek, R.J.,(eds.). Landslides analysis and control. National Academy of Science,
Trans. Resear. Board Special Report, 176, p.34-80, Washington D.C.

Rupke, J., de Graaf, L.W.S., de Jong, M.G.G. and Verhofstad, J., 1987, A
geomorphological Mapping system at scale 1:10.000 for mountainous areas. Zeitschrift
für Geolomorfologie N.F., 31, No: 2, 229-242.

Rupke, J., Cammeraat, E., Seijmonsbergen, A.C., van Westen, C.J., 1988, Engineering
geomorphology of the Widentobel catchment, Apenzell and Sankt Gallen, Switzerland.
Ageomorphological inventory system applied to geotechnical appraisal of slope stability.
Engineering Geology, 26, 33-68.

Sabto, M., 1991, Probabilistic modeling applied to landslides in Central Colombia using
GIS procedures. Unpublished Msc thesis, ITC, Enschede, Netherlands, 26 p.

Saner, S., Taner, İ., Aksoy, Z., Siyako, M. and Burkan, K.A., 1979, Karabük-Safranbolu
yöresinin jeolojisi, TPAO rap no: 1322 (unbpublished).

Sauchyn, D.J. and Trench, N.R., 1978, Landsat applied to landslide mapping.
Photogrammetric Engineering and Remote Sensing, 44, No: 6, 735-741.

Scott, D.W., 1979, On optimal and data-based histograms. Biometrika, 66, 3, 605-610.

Scott, D.W., 1992, Multivariate density estimation. John Wiley, New York.

Seijmonsbergen, A.C., 1992, Geomorphological evolution of an Alpine area and its

application to geotechnical and natural hazard appraisal. Ph.D thesis, University of
Amsterdam, 109p.

Serdar, H. and Demir, C., 1983, Mengen-Bolu-Abant dolayı jeolojisi ve petrol olanakları.
TPAO raporu. (unpublished).

Sissakian, Kh. V, 1986,The study of mass movement from aerial photographs. Proc. 7
Th Int Symp on Remote Sensing for Resources Development and Environ. 675-678

192
Soeters, R and van Westen, C.J., 1996, Slope instability recognition analysis and
zonation, In Landslides: Investigation and Mitigation (Turner, K.T, and Schuster, R.L.
eds), Transportation Research Board National Research Council, Special Report no:
247, Washington D.C., pp 129-177.

Sözen, A., Erbayar, M., Çamaşırcıoğlu, A. ve Çeltek, N., 1996, Düzce (Bolu) - Devrek
(Zonguldak) yöresinin genel jeokimya raporu, MTA Rap No: 9922. (unpublished).

Spiegelhalter, D.J., 1986, Uncertainity in expert systems, in artificial intelligence and

statistics (Gale, W.A., ed), Addison Wessey, Reading, Mass., pp. 17-55.

Stephens, P.R., 1988, Use of satellite data to map landslides. Proceedings 9th Asian
Conference on Remote Sensing, Bangkok, 1-7.

Stevenson, P.C., 1977, An emprical method for the evaluation of relative landslide risk.
Bulletin of Int.Assoc. of Engineering Geologists, 16, p.69-72

Strandberg, C.A., 1967, Aerial Discovery manual. Wiley and Sons, New York, 249 p.

Sturges, H.A., 1926, The choice of a class interval. Journal of the American Statistical
Association, 21, 65-66.

Sucuoğlu, H.,Bakır, S., Tankut, T., Erberik, A., Gülkan, P., Özcebe, G., Gür, T., Yılmaz,
Ç., Ersoy, U., Yılmaz, T. and Akkar, S., 2000, Marmara ve Düzce depremleri
mühendislik raporu, Deprem Mühendisliği Araştırma Merkezi, 174 p.

Şaroglu, F., Boray, A. and Emre, O., 1987, Active Faults of Turkey, MTA Rap No: 8643
(Unpublished).

Şaroglu, F., Boray, A. and Emre, O., 1992, Active Faults of Turkey, MTA Publication

Şaroğlu, F., Herece, E., Sarıaslan, M., Emre, Ö., 1995, Yeniçağa-Gerede-Eskipazar
arasının jeolojisi ve Kuzey Anadolu fayıın genel özellikler. MTA Rap No: 9873
(Unpublished)

Şengör, A.M.C. and Canıtez, N., 1982, The North Anatolian Fault. In: Alpine
Mediterranean Geodynamics, Berekhemer, H. and Hsu, K. (eds). Deuticke, Vienna,
205-216.

Şengör, A.M.C., Görür, N., and Şaroğlu, F., 1985, Strike-slip faulting and related basin
formation in zones of tectonic escape; Turkey as a case study, In Strike slip
deformation, basin formation and sedimentation, edited by Biddle, K.T. and Blick, N.C.,
SEPM Spec.Pub., 37.

Şimşek, O. and Dalgıç, S., 1997, Consolidation properties of the clays at Düzce
plainand their relationship with geological evaluation, Geological Buletin of Turkey,
40(2), 29-38.

Taymaz, T., 2000, Seismotectonics of Marmara Region: Source characteristics of 1999

Gölcük – Sapanca – Düzce Earthquakes. In 1999 Izmit and Düzce Earthquakes:
preliminary results, Barka, A., Kozacı, Ö., Akyüz, S. and Altunel, E. Istanbul University
Press. Istanbul. 79-97

Terrill, W.R., 1994, A FAQ on Vegetation in Remote Sensing,

http://asio.jde.aca.mmu.ac.uk/new_gis/data/rsvegfaq.htm.

TNT Mips Manual, 2000, Combining Rasters, pp. 4-10.

193
Tokay, M., 1952, Karadeniz Ereğlisi, Alaplı, Kızıltepe, Alacaağzı bölgesi jeolojisi, MTA
derg, 42-43, 1-35.

Tokay, M., 1964, 1/500.000 ölçekli Türkiye Jeoloji Haritası Bolu Paftası, MTA Yayınları

Tokay, M., 1973, Kuzey Anadolu Fay Zonunun Gerede ile Ilgaz arasındaki kısmında
jeolojik gözlemler. Kuzey Anadolu Fayı ve deprem kuşağı sempozyumu. 12-29.

Turrini, M.C. and Visintainer, P., 1988, Proposal of a method to define areas of landslide
hazard and an application to an area of the dolomites, Italy. Engineering Geology, 50, 255-265

Unterberger, W. and Brandl, J., 2000, The effect of recent earthquake in Turkey on the
Bolu Tunnels, Felsbau, 18-2, 50-54.

Ustaömer, P.A. and Rogers, G., 1999, The Bolu Massif: remnant of a pre-Early
Ordovicien active margin in the west Pontides, northern Turkey. Geological Magazine,
136, 5, 579-592.

Uysallı, H., 1959, Bolu - Merkeşler bölgesinin jeolojisi ve linyit imkanları. MTA Enstitusü
dergisi sayı 52, 107 p.

Unesco, Working Party on World Landslide Inventory, 1993, A suggested method for
describing the activity of a landslide. Bull. of the Int Eng Geol, 47, 53-57.

Van Westen, C.J.,1992, Medium scale landslide hazard analysis using a PC based GIS:
A case study from chinchina. In Proc., 1er Simposio Internacional sobre Sensores
Remotes y Sistemas de Informacion Geografica (SIG) para el Estudo de Riesgos
Naturales, Bogota, Colombia (Alzate, J.B. ed), Instituto Geografico Agustin Codazzi,
Bogota, vol 2, pp20.

Van Westen, C.J., 1993, Application of geographic information systems to landslide

hazard zonation. ITC Publication no: 15. International Institute for Aerospace and Earth
Resources Survey, Enschede, The Netherlands, 245 pp.

Vargas, G.C., 1992, Methodologia pour l’etablissement de cartes de sensibilite aux

movements de terrain fonde sur l’utulisation d’un couple stereographique SPOT XS/TM.
Application a la region de Palz del Rio (Colomibie). Proceedings 1st Simposio
Internacional sobre sensors remotos y sistemas de informacion geografica (SIG) para el
estudio de riesgos naturals, Bogota, Colombia, 201-220.

Varnes, V., 1978, Slope movement types and processes, Landslides Analysis and
Control: Natiopnal Academy of Sciences Transportation Research Board Special Report
176, p. 12-33.

Varnes, D.J.,1984, Landslide Hazard Zonation: a review of principles and practice,

UNESCO, Natural Hazards, No:3, 61pp

Ward, T.J., Ruh-Ming Li and Simons, D.B., 1982, Mapping landslide hazards in forest
watershed. Journal of Geotechnical Engineering Division, Proceedings of The American
Society of Civil Engineers, vol 108, No: GT2, 319-324.

Wieczorek, G.F., 1984, Preparing a detailed landslide inventory map for hazard
evaluation and reduction, Bulletin of the Association of the Engineering Geologists,
vol.21, no:3, pp.337-342.

194
Wright, R.H., Campbell, R.H., and Nilsen, T.H., 1974, Preparation and use of isopleth
maps of landslide deposits, Geology, vol.2 pp 483-485.

Wrigley, N., 1984, Categorical data analysis for geographers and environmental
scientist. Longman. Harlow.

Yalçın, H. and Cerit, O., 1991, Bolu masifi örtü kayaçlarında diyajenetik ve çok düşük
dereceli metamorfik kil minerallerinin mineralojisi ve jeokimyası. C.Ü. Müh. Fak. Der.
Seri A. Yerbilimler C.8,1.

Yılmaz, Y., Gözübol A.M., Tüysüz, O. and Yiğitbaş., 1981, Abant (Bolu)-Dokurcun
(Sakarya) arasında Kuzey Anadolu Fay Zonunun kuzey ve güneyinde kalan tektonik
birliklerin evrimi, İ.Ü.Yerbilimleri der., 2, 231-261.

Yin, K.L., and Yan, T.Z., 1988, Statistical prediction model for slope instability of
metamorphosed rocks. In Proc., Fifth International Symposium in Landslides
(C.Bonnard ed), Lausanne, A.A. Balkema, Rotterdam, Netherlands, Vol.2, 1269-1272.

195
 VITA

MEHMET LÜTFİ SÜZEN

[email protected]
www.metu.edu.tr/~suzen

1972 Born in Ankara, TURKEY

1989 - 1994 B.Sc. in Geological Engineering, Department of Geological

Engineering, Middle East Technical University
(M.E.T.U.), Ankara, Turkey

1994 - 1996 M.Sc. in Geological Engineering, Department of Geological

Engineering, Middle East Technical University
(M.E.T.U.), Ankara, Turkey

1996 - 1997 Post Graduate Diploma Course at International Institute for

Aerospace and Earth Sciences (I.T.C.), Enschede,
The Netherlands.

1997 - 2002 Ph.D in Geological Engineering, Department of Geological

Engineering, Middle East Technical University
(M.E.T.U.), Ankara, Turkey

Areas of Interest
Geographical Information Systems, Remote Sensing, Natural Hazard Assessment

Others
Stock Markets, Portfolio Management, Scuba Diving, Yatchting, Motorsports

196

View publication stats