Original Paper

Landslides
DOI 10.1007/s10346-019-01174-y Yang Yang I Jintao Yang I Chengdong Xu I Chong Xu I Chao Song
Received: 1 May 2018
Accepted: 21 March 2019
© Springer-Verlag GmbH Germany Local-scale landslide susceptibility mapping using
part of Springer Nature 2019
the B-GeoSVC model

Abstract Local-scale landslide susceptibility mapping (LSM) pro- et al. 2016), frequency ratio (Li et al. 2016; Nicu 2018), weights of
vides detailed information for decision making and planning. evidence (Hong et al. 2017; Kayastha et al. 2012), information value
Most published landslide susceptibility maps lack reliable infor- (Che et al. 2012; Sarkar et al. 2013), neural network (Pham et al.
mation at the local scale due to the spatial heterogeneity being 2017), support vector machine (Hong et al. 2016), naive Bayes
ignored. To enrich the local-scale information of LSM, multiple (Tsangaratos and Ilia 2016a), and decision tree (Tsangaratos and
information fusion methods for the local spatial heterogeneity and Ilia 2016b). These methods assume that the contribution of the
regional trends of control factors are needed. However, no infor- control factors to landslides is consistent at different locations,
mation fusion method has been proposed for LSM yet. In this which is not necessarily the case (Chalkias et al. 2014; Erener and
paper, we developed a new integrated statistical method, named B- Düzgün 2010; Van Westen et al. 1999). However, in reality, the
GeoSVC, under the hierarchical Bayesian framework for LSM. contribution of each control factor for LSM is more likely to be
Specifically, this model applied the GeoDetector method to fit the spatially heterogeneous, i.e., the associations vary among different
regional trends of control factors and employed spatially varying geospatial units (Song et al. 2019). Therefore, traditional LSM
coefficients (SVC) model to fit the local spatial heterogeneity of based on methods above can be quite unreliable at local scales in
each control factor. Then, the regional trends and local spatial cases where the spatial variabilities of landslides and control
heterogeneity information were fused within the hierarchical factors have not been fully accounted for.
Bayesian framework. The B-GeoSVC model was verified using data To obtain detailed local information, methods based on phys-
from the Duwen basin of China, which was in the central region ical mechanistic processes are usually selected. For example, many
affected by the MS 8.0 Wenchuan earthquake that occurred on studies have applied limit equilibrium analysis (Saade et al. 2016),
May 12, 2008. Under a cross-validation experiment, the prediction finite element analysis (Ishii et al. 2012; Zhang et al. 2015), gray
accuracy rate of the B-GeoSVC model was 86.09%, and the area correlation analysis (Dai et al. 2016), and catastrophe progression
under the curve was 0.93, which suggested that the B-GeoSVC method (Wen et al. 2015) for stability analysis of landslides at the
model was able to achieve relatively accurate local-scale LSM and local scale. The advantage of these methods is that slope stability
provide richer local information than traditional regional scale can be analyzed at the local scale. Nevertheless, methods based on
LSM. More importantly, not only the B-GeoSVC model could be physical landslide mechanisms require a large number of detailed
employed as a general solution to fuse both regional and local- geotechnical parameters, which require considerable time and
scale information for landslide mapping, but also offer new in- labor input to carry out simulations, experiments, and field inves-
sights into the broader earth science and spatial statistics. tigations. Thus, it is almost impossible in application for a large
study area to analyze each local-scale landside considering phys-
Keywords Landslide susceptibility mapping . Spatial ical mechanistic processes.
heterogeneity . Regional and local information For the sake of obtaining more local-scale information in a
fusion . GeoDetector . SVC . Hierarchical Bayesian methodYang large area, some studies have used the geographically weighted
Yang and Jintao Yang contributed equally to this work. regression (GWR) model, that is, a local regression model, to
perform LSM (Chalkias et al. 2014; Erener and Düzgün 2010),
Introduction which takes the coefficient of the control factor as a variable
Landslide is one of the most hazardous natural disasters and related to spatial location. Although the GWR model fully fits
causes great loss of life and property (Petley 2012). In 2016, 9710 the spatial heterogeneity of the control factors, it requires a suffi-
geological disasters including landslides were reported in China, cient number of samples for the local model and this condition is
causing approximately 370 deaths and direct economic losses of usually unable to be met because landslides may not be locally
$488 million (http://data.mlr.gov.cn). Landslide susceptibility present. Therefore, most of the current studies are based on
mapping (LSM) can predict the potential occurrence rate of land- regional models with sufficient samples and relatively stable re-
slides and is an important tool in careful planning and mitigation gional trends for LSM. However, the regional model cannot reflect
strategies around such natural hazards (Dai et al. 2002; Fell et al. the spatial variability of the control factors at the local scale. As
2008). However, in actual landslide risk management, it is not only yet, the requirement of fusing both regional trends and local
necessary to regionalize landslide-prone areas but also to provide spatial variability of the control factors is not satisfied simulta-
more detailed information at local scales, which has increased the neously by current LSM models.
demand for local-scale LSM and poses new challenges for LSM Model integration is usually an effective solution when a single
modeling. model cannot satisfy multiple needs at the same time. For exam-
Determining and rationalizing the weighting of landslide con- ple, some studies integrated the factor selection process with LSM
trol factors are required when performing LSM and modeling. modeling to select important factors to avoid reducing the LSM
Many methods have been applied to determine these factors and reliability due to redundancy factors (Dou et al. 2015; Jebur et al.
their weighting, such as logistic regression (Akgun 2012; Pollett 2014). Some other studies integrated the geomorphon method (a

Landslides
 Original Paper
terrain analysis method in GRASS GIS) with LSM to produce performed under the hierarchical Bayesian framework. The B-
proper mapping units, which made the division of mapping units GeoSVC model is verified using the landslide data from the Duwen
in LSM more reasonable (Luo and Liu 2017). Although these basin in Sichuan, China.
integrations reduced the uncertainty of LSM, they were all simply
based on the data manipulation process, not based on a true Methods
integration method by using different models. Under this condi- In this paper, LSM is divided into two main steps: (1) data prepro-
tion, the information from the previous data process was not fully cessing, which includes the division of the grid cells and determi-
integrated with the latter data process. Therefore, an integrated nation of the regional trends of the contribution rate of each
model that can fuse both regional trends and local spatial vari- control factor fitted by GeoDetector; (2) integration of the model,
ability of the control factors at the same time is needed for local- which involves building a SVC model based on the hierarchical
scale LSM. Bayesian framework to fit the spatial heterogeneity and fuse the
In previous LSM studies, local scale and regional scale were regional trends from GeoDetector. Figure 1 shows the graphical
usually defined as 1:25,000–1:5000 and 1:250,000–1:25,000, respec- representation of the statistical methods employed in developing
tively (Corominas et al. 2014; Van Westen et al. 2006). Moreover, the integrated B-GeoSVC model, mainly including fusion of the
these definitions have also been widely utilized in landslide risk regional trends from GeoDetector and the local heterogeneity
management works (Ciurean et al. 2017; Ghosh et al. 2009). Local from SVC.
scale and regional scale are relative concepts, and in this study
local scale refers to each geographical grid and regional scale refers
to the entire study area. When it comes to information fusion, we GeoDetector
use regional trends to represent the weight of the control factors at GeoDetector is a spatial statistical method for detecting spatially
the regional scale. stratified heterogeneity and revealing the control factors behind
To address the problems discussed above, we propose a new spatial patterns (www.geodetector.org). It was first used to detect
integrated statistical method under the hierarchical Bayesian potential control factors for diseases (Wang et al. 2010).
framework, named the B-GeoSVC model, of which the full name GeoDetector has been widely used in the field of spatial data
is Bhierarchical Bayesian integration of GeoDetector model and analysis because it contains no linear assumption and has a clear
SVC model.^ Specifically, this integrated method employs the physical meaning (Luo and Liu 2017; Wang and Hu 2012). The core
GeoDetector to fit the regional weighting trend of the control idea of GeoDetector is that, if the control factor X has a potential
factors and applies the spatially varying coefficients (SVC) model influence on a spatial phenomenon Y, then the spatial phenome-
to fit the spatial variability of the local weighting heterogeneity. non Y represents a spatially stratified heterogeneity that is similar
The fusion of regional trends and local spatial heterogeneity is to the control factor (Wang et al. 2016). Thus, the stronger the

Fig. 1 Research flow of model development showing the two major phases of (1) data preparation and (2) information fusion

Landslides
spatially stratified heterogeneity of the phenomena, the stronger of the ith mapping unit. The spatially varying coefficients
the influence of the control factor, the degree of which can be regression model is as follows:
measured by the q value, as follows:
Y ¼ β 0 ðμi ; ϑi Þ þ β1 ðμi ; ϑi ÞX i1 þ … þ β j ðμi ; ϑi ÞX ip þ σi ð5Þ
1 m
q ¼ 1− ∑ N w σ2w ; q∈½0; 1 ð1Þ
Nσ2 w¼1
where σi is the error term and (μi,ϑi) is the spatial location. The
value of β(μi,ϑi) is solved using data only from the vicinity of
where m is the number of zones of the control factor, N is the position (μi,ϑi); therefore, βj(μi,ϑi) can fit the spatial heterogeneity
number of mapping units (grids) of the study area, and Nw is the of the control factor in different spatial positions. The spatial
number of mapping units in the wth zone. σ2w is the variance of Y heterogeneity of the contribution rate of slope instability is de-
in the wth zone, and σ2 is the regional variance of Y in the whole scribed below in Eq. (6).
area. The larger the value of q, the stronger the control of X on the
−1
spatially stratified heterogeneity of Y. β^^iðμi ; ϑi Þ ¼ X T ðμi ; ϑi ÞW ðμi ; ϑi ÞX ðμi ; ϑi Þ X T W ðμi ; ϑi ÞY ð6Þ
To perform LSM, the study area is divided into regular grids
(mapping units) and the number of landslide in the grid is
taken as the Y variable in GeoDetector. Assuming that X is a where XT is the design matrix of the neighborhoods of the ith
potential control factor (such as lithology or slope), the X layer mapping unit, and Y is the vector of observations. The term
is divided into three zones: a, b, and c according to their W(μi,ϑi) is a spatial weight matrix that controls the local sample
properties. If the landslide rate appears obvious as a spatially range in the local regression model. If the sample range includes
stratified heterogeneity influenced by X, then the q value is all samples of study area, then the values of βj(μi,ϑi) are equal in
large. Under perfect conditions, the spatially stratified hetero- different spatial positions. In this case, βj(μi,ϑi) is equivalent to βj
geneity of the landslide rate is completely controlled by X, and in Eq. (2), and Eq. (5) is consistent with the regional regression
q = 1. In other cases, there is no significant difference among model. After solving Eq. (6), the value of each coefficient βj is
the zones in terms of X. In such conditions, the sum of the obtained and then Eq. (3) is combined with Eq. (4) to obtain P.
local variances in each zone is close to the regional variance
and q = 0. The q value of the control factor is constant in the Hierarchical Bayesian framework
whole study area, and so can be taken as a regional trend. The local model fits the spatial heterogeneity of the control factor
well, but there may be an insufficient number of samples in the
SVC logistic regression model local region, which in this case reduces the reliability of the local
Logistic regression is a standard model for processing binary model. As described in the BGeoDetector^ section, GeoDetector
variables, the main purpose of which is to predict the probability can calculate the regional trend of each control factor (red line in
of an event based on the control factors (Zhang et al. 2016). Fig. 2) but cannot fully fit the spatial heterogeneity of control
Landslide events obey a binomial distribution, and logistic regres- factors at the local scale (green line in Fig. 2). Fusion of the two
sion in the LSM is applied as follows: trends may be a good choice for reducing the uncertainty of the
local-scale LSM (black line in Fig. 2).
Bayesian statistics assume that the unknown parameter θ is a
Y ¼ β 0 þ β1 X 1 þ β 2 X 2 þ … þ β j X j þ ε ði ¼ 1; 2; …; nÞ ð2Þ random variable that obeys a certain probability distribution
(Damien et al. 2013). Therefore, in the hierarchical Bayesian frame-
work, once prior knowledge of the control factor is provided, then
prior information of the control factor can be corrected based on
P the local sample information (Bivand et al. 2014). Bayesian prior
Y ¼ logitðpÞ ¼ ln ð3Þ
1−P knowledge compensates for the lack of information caused by the
low number of samples to a certain extent, and it has been widely
used in various disciplines (Bivand et al. 2014; Grassi and Proietti
eY 2014).
Pðz ¼ 1Þ ¼ ð4Þ
1 þ eY In the hierarchical Bayesian framework, the term β^i ðμi ; ϑi Þ is a
random variable and its value can be inferred by a combined
likelihood function with prior distribution information:
where Y is the transition variable, P is the probability of the

landslide occurring, X1, X2,…Xj are the control factors, and β0,
πðβjX; Y Þ∝pðβ Þ∙ f ðX; Y j β ð7Þ
β1,…βj are the coefficients.
In large-area LSM, the weight of each control factor varies
between locations (Zhang et al. 2016) in terms of spatial het-
erogeneity (Brunsdon et al. 1996). In this case, the regression where f (X, Y | β) is the likelihood function and p(β) is the prior
coefficient βj is no longer a constant but a spatially dependent knowledge.
variable. The spatially varying coefficients model is designed A random variable is generally described by gamma Г(λ, η) as
with a spatial function that is used to calculate the coefficient the prior knowledge in a hierarchical Bayesian framework, where

Landslides
 Original Paper

Fig. 2 Illustration of the effects of fusing the regional trend and local spatial heterogeneity information

λ is the shape parameter and η is the scale parameter, and param- control factor is corrected based on the SVC logistic regression
eters λ and η are greater than 0. The shape parameter λ represents model under the hierarchical Bayesian framework. Hence, a fusion
the contribution of a covariate in the regression model. The larger of regional trend information at a regional scale and spatial het-
the λ, the more prior contribution of a covariate. However, most of erogeneity at the local scale is carried out in this way.
the former studies applied non-informative (default) priors to
assume that each covariate contributes the same in a regression Study area and data
model (Song et al. 2018a, b).
Non-informative priors may weaken the model interpretation Study area
ability. Hence, we define informative priors by leveraging the The study area is along the Duwen Highway, in mountainous areas
strength from GeoDetector in our integrated method. The regional of southwest China. The tectonic structure of the area is related to
trend value (q-statistics) calculated by GeoDetector is also greater the transitional belt (Fig. 3) on the southeastern edge of the
than 0. Thus, it is assumed to follow the parameter λ in a prior Qinghai-Tibet Plateau and the western Sichuan Basin. An earth-
gamma distribution. Moreover, indicator q-statistics also qualifies quake of magnitude (MS) 8.0 occurred in the study area on May 12,
the contribution of a covariate. Under this condition, we incorpo- 2008, referred to as the Wenchuan earthquake. The secondary
rate q-statistics as a quantitative indicator to define shape param- geological disasters resulting from the earthquake included land-
eter λ, which is the larger the q-statistics, the larger the λ. In this slides and slope collapse (Tang et al. 2016), and it is expected that
way, the regional trend is taken as the prior knowledge, and it is these will be active for a period of 10–20 years (Cui et al. 2008).
fused with the local spatial heterogeneity. Lithologies in the study area consist mainly of granitoids, with
The scale parameter η represents the uncertainty (confidence some coal beds, pyroclastic rocks, and other sedimentary rocks.
interval) of the estimated parameter. The smaller the η, the lower Tectonic movement in the study area is frequent and is influenced
the uncertainty of the estimated parameter. Following previous by three major fault zones, collectively known as the
Bayesian studies (Blangiardo et al. 2013; Ugarte et al. 2014), we Longmenshan fault zone. This fault zone strikes northeast-
set η to 0.00005 in this study, which is also a default setting in R- southwest and crosses the Duwen Highway. The three major zones
INLA, to make sure that all the parameters are estimated with are the Guanxian-Jiangyou fault (also known as the Qianshan
relatively high accuracies. The expressions of the prior and likeli- fault), the Yingxiu-Beichuan fault (also known as the Central
hood functions are as follows: fault), and the Mao-wen fault (also known as the Houshan fault).
Additionally, the Yingxiu fault is the seismogenic fault zone related
  
π β j ðμi ; ϑi ÞjX ðμi ; ϑi Þ; Y ðμi ; ϑi Þ ∝Гðλjq; ηÞ∙ f ðX ðμi ; ϑi Þ; Y ðμi ; ϑi Þ β j ðμi ; ϑi Þ
to the Wenchuan earthquake (Zhuang et al. 2010; Yang et al. 2019).
The study area is approximately 935 km2 in extent, and the
ð8Þ climatic conditions and geographical environment vary greatly
between the north and south. The terrain in the study area is also
In this paper, the regional trend q of each control factor ob- strongly intersected, with relative elevation changes from 300 to
tained from the GeoDetector is used as a prior knowledge of the 2500 m, providing favorable conditions for natural disasters such
hierarchical Bayesian framework, and the prior knowledge of each as landslides and mudslides (Gan et al. 2011). Thus, the local

Landslides
Fig. 3 Landslide inventory map and digital elevation model (DEM) of the study area along the Duwen Highway (roads marked in red) in southwest China

spatial heterogeneity in terms of slope, geology, and climate can- (Fig. 4). In these areas, the landslides mainly include two types:
not be ignored, and forms an integral part of the local-scale rainfall-induced and earthquake-induced. Therefore, consider-
considerations in the model. ing the diversity of trigger conditions of landslides, we mainly
selected the common influence factors of rainfall landslides
Data and earthquake landslides as evaluation indicators. These fac-
A total of 4841 landslide records were acquired from 0.5-m tors include surface roughness of the terrain (Fig. 4a), slope
multi-band remote sensing imagery from the Pleiades satellite angle (Fig. 4b), distance from inhabited areas to the mapping
in 2014. Results of remote sensing interpretation were then unit (Fig. 4c), distance from the mapping unit to the nearest
combined with results of field investigation of geological di- road (Fig. 4d), digital elevation model (DEM) (Fig. 3), rock
sasters along the Duwen Highway, and seven control factors group (Fig. 4e), and land use (Fig. 4f). The factor Bsurface
were then selected from the geographic national survey roughness^ is the standard deviation of the terrain elevation,

Landslides
 Original Paper

Fig. 4 Thematic maps of control factors used in this study. a Surface roughness. b Slope angle. c Distance from inhabited area to mapping unit. d Distance from mapping
unit to the nearest road. e Rock mass. f Land use

the elevation unit is meter, so the Bsurface roughness^ is in Regional trend and spatial heterogeneity information fusion
meters; the larger the value, the rougher the surface. The factor To balance computational efficiency and spatial resolution, a
Brock group^ has five categories: the harder rock group, hard 200 × 200 m spatial resolution is chosen and the study area is
rock group, soft rock group, softer rock group, and loose rock divided into grids. If a landslide occurs in the grid, the y variable
group, denoted respectively as E, D, C, B, and A. The factor of the grid is assigned a value of 1, and if no landslide occurs, the
Bland use^ has six categories: cultivated land, woodland, grass- y variable is assigned a value of 0. The attribute values of the five
land, water body, construction land, and bare land, denoted control factors are then overlain into the grid as an x variable,
respectively as 1, 2, 3, 4, 5, and 6. and the attribute table of the grid is exported as the input for the

Results
The q-statistics value
Regional trend calculated by GeoDetector
As described in the BGeoDetector^ section, GeoDetector spec-
ifies that the x variable must be a zonal layer, and thus the road
continuous control factors such as slope, DEM, surface rough-
ness, habitat buffer, and road buffer are reclassified into five
zones. To construct the y variable of GeoDetector, the study
inhabited area
area is divided into regular grids as the basic mapping unit.
The number of landslide points in the grids is counted as roughness
landslide rate y. The attribute values of each zone of the
control factor are mapped into grids by the overlay layers,
and then the attribute table of the grid is exported as the input DEM
data for GeoDetector. The regional trend of the weighting of
each control factor is calculated from Eq. (1) (Fig. 5).
A larger q value indicates a higher control power of the slope
factor in the spatially stratified heterogeneity of the landslide
rate. In other words, such a factor provides a greater contri-
bution to the occurrence of the landslide. In this paper, the q-
statistic values of the control factors are used as the trend of
the weight of the control factor. Fig. 5 Regional trend weighting indices of the control factors used in this study

Landslides
Fig. 6 Receiver operating characteristic (ROC) curve of the B-GeoSVC model

LSM model. A SVC model based on hierarchical Bayesian frame In this study, cross-validation was run to evaluate the
was built using the INLA software package in the freely available predictive performance of the proposed method. Specifically,
statistical package R. The regional trend of each control factor we randomly sampled 30% from the existing data to create
obtained from the GeoDetector was used as a priori knowledge to the test set and used the rest of the data as the training set.
infer the model coefficient, which is the manner the fusion of We obtained AUC and PAR using the 30% test set. The AUC
spatial heterogeneity and regional trends is carried out. value of the B-GeoSVC model is 0.93 (Fig. 6), which proves
Prediction accuracy can reflect the reliability of the land- that the B-GeoSVC model has a good performance. The PAR
slide susceptibility map. Receiver operating characteristic values of 1 and 0 are 84.60% and 86.24%, respectively, and
(ROC) curves reflect the sensitivity and specificity of continu- the total PAR of the model is 86.09% (Table 1).
ous variables (Akgün and Bulut 2007). The area under the Regarding categorical control factors, it is difficult for the
curve (AUC) of the ROC curve reflects the performance of SVC model to calculate the spatial variability of land use and
the model, and the closer the AUC value is to 1, the better rock group because these factors remain unchanged in the
the performance. Logistic regression model divided the result adjacent grids at the local scale. Thus, we take these factors as
into two categories: 0 and 1, and the prediction accuracy rate the regional variables, and their regression coefficients are
(PAR) for values 0 and 1 can be calculated by the confusion shown in Table 2.
matrix using the test data (Song et al. 2019). Specifically, the
a d Local-scale LSM
PAR of 1 and 0 can be calculated by formulas aþb and cþd ,
respectively; and the total PAR was calculated by formulas The regression coefficients of control factors in different spa-
aþd tial positions are shown in Fig. 7a–e, and each cell value on
aþbþcþd. WhereBa^ denotes the number of samples whose true
the map represents the value of the regression coefficient.
values are 1 (e.g., landslide occurred) and the model-predicted
From Fig. 7, it is evident that the regression coefficient shows
results are also 1; Bb^ denotes the number of samples whose
a greater spatial heterogeneity in the entire study area. The
true values are 1, but the model-predicted results are 0 (e.g., no
reasons for this are discussed in the BDiscussion^ section of
landslide occurred); Bc^ denotes the number of samples whose
the paper.
true values are 0, but the model-predicted results are 1; and
As shown in Fig. 8, the landslide susceptibility map is
Bd^ denotes the number of samples whose true values are 0,
reclassified into three levels according to the landslide
and the model-predicted results are also 0.

Table 1 Confusion matrix and model prediction accuracy rate (PAR) of 30% test data
Model Landslide Prediction Percent Accurate
Yes No
B-GeoSVC Yes 610 111 84.60% 86.09%
No 854 5361 86.26%

Landslides
 Original Paper
Table 2 Discrete control factors posterior coefficients and their 95% confidence intervals (CIs)
Factors β 95% CI
0.025 quant 0.975 quant
Loose rock group 1.32 − 0.18 2.75
Softer rock group − 0.56 − 2.20 1.05
Soft rock group − 2.97 − 3.25 − 2.69
Hard rock group − 2.43 − 2.67 − 2.20
Harder rock group − 2.43 − 2.59 − 2.27
Cultivated land − 3.5 × 10−4 − 1.9 × 10−2 1.8 × 10−2
Woodland − 4.3 × 10−4 − 1.9 × 10−2 1.7 × 10−2
Grassland 2.1 × 10−4 − 1.8 × 10−2 1.9 × 10−2
−4 −2
Water body 3.6 × 10 − 1.8 × 10 1.9 × 10−2
Construction land 3.7 × 10−5 − 1.8 × 10−2 1.9 × 10−2
Bare land − 9.2 × 10−4 − 1.9 × 10−2 1.8 × 10−2

percentage. We find that 85.7% of the landslides occur in the were selected to show the different contribution of control
high-prone area, and the occurrence of a landslide in this area factors to the landslide and to analyze the susceptibility of
is higher than 4.9 per square kilometer; 12.4% of the landslides the grid at a local scale. The three grids are labeled a, b, and
occur in the moderate-prone area, and the occurrence of a c, as shown in Fig. 8, and each grid corresponds to a boxplot
landslide is higher than 0.76 per square kilometer. Landslides on the right.
in low-prone area accounts for 1.1% of the total landslides, and Each box in the boxplot represents a control factor (e.g.,
the occurrence of a landslide is lower than 0.54 per square DEM, habitat), and the corresponding coordinates of the box
kilometer in this area. The landslide susceptibility map on the y-axis represent the coefficient of the control factor in
matches well with the actual landslide spatial distribution the corresponding grid on the left. The larger the y-axis value,
(Fig. 3). Three grids with different geographical environments the greater the contribution of the control factor to landslide

Fig. 7 Coefficient maps (CM) of each control factor in the study area. a CM of road buffer. b CM of slope degree. c CM of habitat buffer. d CM of surface roughness. e CM
of DEM

Landslides
Fig. 8 Landslide susceptibility map (left panel) and box plots (a–c) of local-scale landslide susceptibility analysis in three locations

occurrence. The height of the box represents the variance of associations between landslide occurrence and control factors,
the posterior distribution of the regression coefficient, which which is very useful for slope prevention and controlling.
reflects the degree of concentration of the posterior distribu- However, the spatial variability of these control factors is not
tion of coefficients. The smaller the variance, the better the considered in a regional model; thus, it will weaken the con-
posterior fitting. tribution of the control factors.
With regard to landslide-road associations, we found that
Discussion the southern part of the study area is flatter and has a higher
In the Duwen basin of China, we found that the contribution population density and road density than the north. Therefore,
of control factors to landslide varied across locations. Mean- the contribution of the road buffer factor to landslide occur-
while, there were also some regularities on a regional scale that rence is quite different between the northern and southern
were detected, such as slope degree, DEM, and roughness that parts of the study area. The results indicated that the road
contributed the most to the landslide and decreased in turn, buffer contributes little to the slope instability within the
which means terrain elements were the main cause of land- regional trend in this paper (Fig. 4). However, after correcting
slides in this area. Factors, distance to the residential area and the regional trend of the control factors using the local SVC
the road, revealed the extent of human damage to the slope, model, the control factors show a strong spatial heterogeneity
and we found these factors contributed relatively little to a between the southern and northern parts of the study area
landslide in the Duwen basin of China. Despite this, the con- (Fig. 7a).
tribution of these factors in local scale is of great significance The model also detected considerable spatial variability for
for guiding human activities in mountainous areas. the other landslide-environment associations. For example, the
Going beyond the regional-scale control factors aforemen- areas with elevations > 3000 m are mainly concentrated in the
tioned, a more important contribution of this study is that we north, and the spatial distributions of rainfall intensity and
have detected the local-scale spatial variability in the human activity are mainly concentrated in the south. All of

Landslides
 Original Paper
these three control factors show significant differences between and successfully applied it in the Duwen basin of China. The
the northern and southern parts of the study area. This may be integrated model fuses regional trends and local spatial hetero-
due to that the higher coefficient values of habitat were mainly geneity of control factors and has shown relatively good perfor-
distributed in the north relative to the south. Similarly, road is mance on both regional and local scales. Our study expands the
more densely associated, with stronger human activity in the limited knowledge of the complex local-scale associations be-
south compared with the north, because maybe the influence tween various environmental factors and landslide occurrence,
of habitat on landslides is weaker in the south and stronger in revealing their spatially heterogeneous relationships in the
the north. Duwen basin of China. Analysis of grid at the local scale is also
Moreover, the areas with the smallest surface roughness are realized in this paper, which provides useful information for
distributed in the south with lower altitudes. In these areas, the slope prevention and controlling. More importantly, the hierar-
coefficients of surface roughness are relatively small, so the risk of chical Bayesian framework can take into account the regional
landslide occurrence is limited. The slope is also a key factor trends at a regional scale (GeoDetector) and the spatial hetero-
influencing landslide occurrences. Compared with sparsely popu- geneity at the local scale (SVC), providing a new general solution
lated areas, river valleys with road buffer areas containing slopes for fusing different models. The B-GeoSVC model not only can
are more prone to trigger landslides occurrence. Thus, the slope be reliably used as a universal solution for local-scale landslide
coefficients are higher in these areas than those in other areas. susceptibility mapping, but also offers new insights into the
The main contribution of this study is to propose the B- broader earth science and spatial statistics.
GeoSVC model for local-scale LSM. Compared with other similar
studies that applied the GWR model (Chalkias et al. 2014; Erener Acknowledgments
and Düzgün 2010), we summarized the advantages of the B- The authors are grateful to Jinfeng Wang (LREIS) for his valuable
GeoSVC model in two aspects as follows. First, a fusion of the suggestions to improve our work and supporting freeware
spatial heterogeneity obtained from the small, local sample num- GeoDetector. We appreciate Yan Zhen and the colleagues in the
bers with stable regional trends obtained from large, regional Spatial Information Technology and Big Data Mining Research
sample numbers provides the advantages of both the Center in Southwest Petroleum University for their assistance in
GeoDetector and SVC models. Second, the hierarchical Bayesian this study. We would also like to thank the editors and five
framework is proved useful to integrate different models, which anonymous reviewers for their constructive comments and valu-
is an effective solution where a single model cannot actually meet able suggestions in improving this manuscript.
multiple needs.
Although this study has a certain contribution to local-scale FundingThe work was jointly supported by the
LSM, limitations still exist. First, this model has a premised National Natural Science Foundation of China
assumption for control factors, that is, the control factors should (no. 41701448), a grant from State Key Laboratory
have strong spatial variability at the local scale, while the factors of Resources and Environmental Information Sys-
such as land use, rock group, and other category factors are tem (no. 201811), and the Young Scholars Devel-
usually with little variability in adjacent units. Under this condi- opment Fund of Southwest Petroleum University
tion, the model may be unable to fit the local-scale spatial (no. 201699010064), the Open Fund of the State
variabilities for such category factors properly. Second, one geo- Key Laboratory of Geoscience Spatial Information
graphic grid cell may contain several landslides or one landslide Technology, Ministry of Land and Resource (no.
may be shared by several adjacent grids; thus, the grid cannot
KLGSIT2016-03), the Technology Project of the
represent for a specific landslide. What caused this issue is called
Bspatial scale effect^ (e.g., different mapping units) that exists
Sichuan Bureau of Surveying, Mapping and
not only in LSM but also in the broader geographical research.
Geoinformation (no. J2017ZC05), and the Science
Though some studies have compared different mapping units to
and Technology Strategy School Cooperation Pro-
test the spatial scale effect on mapping (Erener and Duzgun jects of the Nanchong City Science and Technology
2012), a general approach to obtain an optimal mapping unit is Bureau (no. NC17SY4016, 18SXHZ0025).
not yet proposed, especially accounting for various landslide
types and complex geographical conditions. Encouragingly, the
B-GeoSVC model is proposed not restricted to be applied in
specific one mapping unit, and the evaluation results of this References
study indicate an acceptable prediction accuracy under the pres-
Akgun A (2012) A comparison of landslide susceptibility maps produced by logistic
ent mapping unit. Nevertheless, the spatial scale effect is an regression, multi-criteria decision, and likelihood ratio methods: a case study at İzmir,
important issue to be considered in future LSM studies. Third, Turkey. Landslides 9:93–106. https://doi.org/10.1007/s10346-011-0283-7
applying multi-temporal landslide inventories to verify model Akgün A, Bulut F (2007) GIS-based landslide susceptibility for Arsin-Yomra (Trabzon,
other than using the same period of data is more objective, which North Turkey) region. Environ Geol 51:1377–1387. https://doi.org/10.1007/s00254-
should also be further studied. 006-0435-6
Bivand RS, Gómez-Rubio V, Rue H (2014) Approximate Bayesian inference for spatial
econometrics models. Spatial Stat 9:146–165. https://doi.org/10.1016/
Conclusion j.spasta.2014.01.002
In this study, we integrated GeoDetector and SVC models based Blangiardo M, Cameletti M, Baio G, Rue H (2013) Spatial and spatio-temporal models
on the hierarchical Bayesian framework into a new method, with R-INLA. Spat Spatiotemporal Epidemiol 7:39–55
named B-GeoSVC, for producing more reliable local-scale LSM,

Landslides
Brunsdon C, Fotheringham AS, Charlton ME (1996) Geographically weighted regression: Luo W, Liu CC (2017) Innovative landslide susceptibility mapping supported by
a method for exploring spatial nonstationarity. Geogr Anal 28:281–298. https:// geomorphon and geographical detector methods. Landslides 15:465–474. https://
doi.org/10.1111/j.1538-4632.1996.tb00936.x doi.org/10.1007/s10346-017-0893-9
Chalkias C, Kalogirou S, Ferentinou M (2014) Landslide susceptibility, peloponnese Nicu IC (2018) Application of analytic hierarchy process, frequency ratio, and statistical
peninsula in South Greece. J Maps 10:211–222. https://doi.org/10.1080/ index to landslide susceptibility: an approach to endangered cultural heritage. Environ
17445647.2014.884022 Earth Sci 77:79. https://doi.org/10.1007/s12665-018-7261-5
Che VB, Kervyn M, Suh CE, Fontijn K, Ernst GGJ, Marmol MAD, Trefois P, Jacobs P (2012) Petley D (2012) Global patterns of loss of life from landslides. Geology 40:927–930.
Landslide susceptibility assessment in Limbe (SW Cameroon): a field calibrated seed https://doi.org/10.1130/G33217.1
cell and information value method. Catena 92:83–98. https://doi.org/10.1016/ Pham BT, Bui DT, Prakash I, Dholakia MB (2017) Hybrid integration of multilayer
j.catena.2011.11.014 perceptron neural networks and machine learning ensembles for landslide suscepti-
Ciurean RL, Hussin H, Van Westen CJ, Jaboyedoff M, Nicolet P, Chen L, Frigerio S, Glade T bility assessment at Himalayan area (India) using GIS. Catena 149:52–63. https://
(2017) Multi-scale debris flow vulnerability assessment and direct loss estimation of doi.org/10.1016/j.catena.2016.09.007
buildings in the eastern Italian Alps. Nat Hazards 85:929–957. https://doi.org/ Pollett WG, Gibbs P, Mclaughlin S, Eteuati J, Harold M, Marion K, Patel S, Jones I (2016)
10.1007/s11069-016-2612-6 Spatial prediction models for shallow landslide hazards: a comparative assessment of
Corominas J, Van Westen CJ, Frattini P, Cascini L, Malet JP, Fotopoulou S, Catani F, the efficacy of support vector machines, artificial neural networks, kernel logistic
Eeckhaut MVD, Mavrouli O, Agliardi F (2014) Recommendations for the quantitative regression, and logistic model tree. Landslides 13:361–378. https://doi.org/10.1007/
analysis of landslide risk. Bull Eng Geol Environ 73:209–263. https://doi.org/10.1007/ s10346-015-0557-6
s10064-013-0538-8 Saade A, Abou-Jaoude G, Wartman J (2016) Regional-scale co-seismic landslide assess-
Cui P, Wei F, He S (2008) Mountain disasters induced by the earthquake of May 12 in ment using limit equilibrium analysis. Eng Geol 204:53–64. https://doi.org/10.1016/
Wenchuan and the disasters mitigation. J Mt Sci 26:280–282 (in Chinese) j.enggeo.2016.02.004
Dai FC, Lee CF, Ngai YY (2002) Landslide risk assessment and management: an overview. Sarkar S, Roy AK, Martha TR (2013) Landslide susceptibility assessment using information
Eng Geol 64:65–87. https://doi.org/10.1016/S0013-7952(01)00093-X value method in parts of the Darjeeling Himalayas. J Geol Soc India 82:351–362.
Dai Z, Wei Y, Lv T, Luo J, Yao W (2016) Deformation influence factors of a landslide in https://doi.org/10.1007/s12594-013-0162-z
Three Gorges Reservoir area based on grey correlation analysis. The Chinese Journal Song C, He Y, Bo Y, Wang J, Ren Z, Yang H (2018a) Risk assessment and mapping of
of Geological Hazard and Control 27:32-37. https://doi.org/10.16031/j.cnki.issn.1003- hand, foot, and mouth disease at the county level in mainland China using spatio-
8035.2016.01.06 temporal zero-inflated bayesian hierarchical models. Int J Environ Res Public Health
Damien P, Dellaportas P, Polson NG, Stephens DA (2013) Bayesian theory and applica- 15:1476. https://doi.org/10.3390/ijerph15071476
tions. Oxford University Press, London Song C, Yang X, Shi X, Bo Y, Wang J (2018b) Estimating missing values in china’s official
Dou J, Tien BD, Yunus AP, Jia K, Song X, Revhaug I, Xia H, Zhu Z (2015) Optimization of socioeconomic statistics using progressive spatiotemporal Bayesian hierarchical
causative factors for landslide susceptibility evaluation using remote sensing and GIS modeling. Sci Rep 8:10055. https://doi.org/10.1038/s41598-018-28322-z
data in parts of Niigata, Japan. PLoS One 10:e0133262. https://doi.org/10.1371/ Song C, Shi X, Bo YC, Wang JF, Wang Y, Huang DC (2019) Exploring spatiotemporal
journal.pone.0133262 nonstationary effects of climate factors on hand, foot, and mouth disease using
Erener A, Düzgün HSB (2010) Improvement of statistical landslide susceptibility mapping Bayesian spatiotemporally varying coefficients (STVC) model in Sichuan, China. Sci
by using spatial and global regression methods in the case of more and Romsdal Total Environ 648:550–560. https://doi.org/10.1016/j.scitotenv.2018.08.114
(Norway). Landslides 7:55–68. https://doi.org/10.1007/s10346-009-0188-x Tang C, Van Westen CJ, Tanyas H, Jetten VG (2016) Analysing post-earthquake landslide
Erener A, Duzgun HSB (2012) Landslide susceptibility assessment: what are the effects of activity using multi-temporal landslide inventories near the epicentral area of the
mapping unit and mapping method? Environ Earth Sci 66:859–877. https://doi.org/ 2008 wenchuan earthquake. Nat Hazards Earth Syst Sci 16:1–26. https://doi.org/
10.1007/s12665-011-1297-0 10.5194/nhess-16-2641-2016
Fell R, Corominas J, Bonnard C, Cascini L, Leroi E, Savage WZ (2008) Guidelines for Tsangaratos P, Ilia I (2016a) Comparison of a logistic regression and naïve bayes classifier
landslide susceptibility, hazard and risk zoning for land use planning. Eng Geol in landslide susceptibility assessments: the influence of models complexity and
102:85–98. https://doi.org/10.1016/j.enggeo.2008.03.022 training dataset size. Catena 145:164–179. https://doi.org/10.1016/
Gan JJ, Huang RQ, Fan CR, Qian-Yin LI, Xiao-Hua YE (2011) A study of the slope failure j.catena.2016.06.004
along the Dujiangyan to Wenchuan highway after the Wenchuan earthquake. Tsangaratos P, Ilia I (2016b) Landslide susceptibility mapping using a modified decision
Hydrogeol Eng Geol 38:59–65 (in Chinese) tree classifier in the Xanthi perfection, Greece. Landslides 13:305–320. https://doi.org/
Ghosh S, Van Westen CJ, Carranza EJM, Ghoshal TB, Sarkar NK, Surendranath M (2009) A 10.1007/s10346-015-0565-6
quantitative approach for improving the bis (Indian) method of medium-scale Ugarte MD, Adin A, Goicoa T, Fernandez Militino A (2014) On fitting spatio-temporal
landslide susceptibility. J Geol Soc India 74:625–638. https://doi.org/10.1007/ disease mapping models using approximate Bayesian inference. Stat Methods Med
s12594-009-0167-9 Res 23:507–530. https://doi.org/10.1177/0962280214527528
Grassi S, Proietti T (2014) Characterising economic trends by bayesian stochastic model Van Westen CJ, Seijmonsbergen AC, Mantovani F (1999) Comparing landslide hazard
specification search. Comput Stat Data Anal 71:359–374. https://doi.org/10.1016/ maps. Nat Hazards 20:137–158. https://doi.org/10.1023/A:1008036810401
j.csda.2013.02.024 Van Westen CJ, Asch TWJV, Soeters R (2006) Landslide hazard and risk zonation-why is it
Hong H, Ilia I, Tsangaratos P, Chen W, Xu C (2017) A hybrid fuzzy weight of evidence still so difficult? Bull Eng Geol Environ 65:167–184. https://doi.org/10.1007/s10064-
method in landslide susceptibility analysis on the wuyuan area, China. Geomorphol- 005-0023-0
ogy 290:1–16. https://doi.org/10.1016/j.geomorph.2017.04.002 Wang JF, Hu Y (2012) Software, data and modelling news. In: Environmental health risk
Hong H, Pradhan B, Bui DT, Xu C, Youssef AM, Chen W (2016) Comparison of four kernel detection with geogdetector, vol 33. Elsevier Science Publishers B. V, pp 114–115.
functions used in support vector machines for landslide susceptibility mapping: a case https://doi.org/10.1016/j.envsoft.2012.01.015
study at Suichuan area (China). Geomat Nat Haz Risk 8:544–569. https://doi.org/ Wang JF, Li XH, Christakos G, Gu X, Gu X, Gu X, Zheng XY (2010) Geographical detectors-
10.1080/19475705.2016.1250112 based health risk assessment and its application in the neural tube defects study of
Ishii Y, Ota K, Kuraoka S, Tsunaki R (2012) Evaluation of slope stability by finite element the Heshun region, China. Int J Geogr Inf Sci 24:107–127. https://doi.org/10.1080/
method using observed displacement of landslide. Landslides 9:335–348. https:// 13658810802443457
doi.org/10.1007/s10346-011-0303-7 Wang JF, Zhang TL, Fu BJ (2016) A measure of spatial stratified heterogeneity. Ecol Indic
Jebur MN, Pradhan B, Tehrany MS (2014) Optimization of landslide conditioning factors 67:250–256. https://doi.org/10.1016/j.ecolind.2016.02.052
using very high-resolution airborne laser scanning (lidar) data at catchment scale. Wen C, Xiao H, Zeng J (2015) Evaluation of landslide stability based on catastrophe
Remote Sens Environ 152:150–165. https://doi.org/10.1016/j.rse.2014.05.013 progression method. J Nat Disast Sci 24:68–73 (in Chinese)
Kayastha P, Dhital MR, Smedt FD (2012) Landslide susceptibility mapping using the Yang JT, Song C, Yang Y, Xu CD, Guo F, Xie L (2019) New method for landslide
weight of evidence method in the Tinau watershed, Nepal. Nat Hazards 63:479–498. susceptibility mapping supported by spatial logistic regression and GeoDetector: a
https://doi.org/10.1007/s11069-012-0163-z case study of Duwen Highway Basin, Sichuan Province, China. Geomorphology
Li L, Lan H, Guo C, Zhang Y, Li Q, Wu Y (2016) A modified frequency ratio method for 324:62–71. https://doi.org/10.1016/j.geomorph.2018.09.019
landslide susceptibility assessment. Landslides 14:727–741. https://doi.org/10.1007/
s10346-016-0771-x

Landslides
 Original Paper
Zhang X, Krabbenhoft K, Sheng D, Li W (2015) Numerical simulation of a flow-like Chengdu, 610500, China
landslide using the particle finite element method. Comput Mech 55:167–177. Email: [email protected]
https://doi.org/10.1007/s00466-014-1088-z
Zhang M, Cao X, Peng L, Niu R (2016) Landslide susceptibility mapping based on global J. Yang : C. Xu : C. Song
and local logistic regression models in three Gorges reservoir area, China. Environ State Key Laboratory of Resources and Environmental Information System (LREIS),
Earth Sci 75:1–11. https://doi.org/10.1007/s12665-016-5764-5 Institute of Geographic Sciences and Natural Resources Research,
Zhuang J, Peng C, Ge Y, Zhu Y, Liu Y, Pei L (2010) Risk assessment of collapses and Chinese Academy of Sciences,
landslides caused by 5.12 wenchuan earthquake—a case study of Dujiangyan- Beijing, 100101, China
Wenchuan highway. Chin J Rock Mech Eng 29:3735–3742 (in Chinese)
C. Xu
Institute of Geology, China Earthquake Administration,
Beijing, 100029, China
Yang Yang and Jintao Yang contributed equally to this work.

Y. Yang : J. Yang : C. Song ())

Landslides