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

net/publication/279737812

Crop classification of upland fields using Random forest of time-series

Landsat 7 ETM+ data

Article  in  Computers and Electronics in Agriculture · July 2015

DOI: 10.1016/j.compag.2015.05.001

CITATIONS READS

81 802

4 authors:

Kenichi Tatsumi Yosuke Alexandre Yamashiki

Tokyo University of Agriculture and Technology Kyoto University
17 PUBLICATIONS   211 CITATIONS    162 PUBLICATIONS   1,398 CITATIONS   

SEE PROFILE SEE PROFILE

Miguel Angel Canales Torres Cayo Ramos

Universidad Nacional Agraria La Molina Universidad Nacional Agraria La Molina
1 PUBLICATION   81 CITATIONS    6 PUBLICATIONS   151 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Interdisciplinary Study on Environmental Transfer of Radionuclides from the Fukushima Daiichi NPP Accident View project

Numerical simulation of the flood wave caused by breaking of the earth dam "Poechos" View project

All content following this page was uploaded by Yosuke Alexandre Yamashiki on 03 September 2015.

The user has requested enhancement of the downloaded file.

 Computers and Electronics in Agriculture 115 (2015) 171–179

Contents lists available at ScienceDirect

Computers and Electronics in Agriculture

journal homepage: www.elsevier.com/locate/compag

Crop classification of upland fields using Random forest of time-series

Landsat 7 ETM+ data
Kenichi Tatsumi a,⇑, Yosuke Yamashiki b, Miguel Angel Canales Torres c, Cayo Leonidas Ramos Taipe c
a
Department of Environmental and Agricultural Engineering, Tokyo University of Agriculture and Technology, Tokyo, Japan
b
Graduate School of Advanced Integrated Studies in Human Survivability, Kyoto University, Kyoto, Japan
c
Departamento de Recursos Hídricos, Universidad Nacional Agraria La Molina, La Molina, Peru

a r t i c l e i n f o a b s t r a c t

Article history: Crop classification of homogeneous landscapes and phenology is a common requirement to estimate land
Received 21 January 2014 cover mapping, monitoring, and land use categories accurately. In recent missions, classification methods
Received in revised form 16 February 2015 using medium or high spatial resolution data, which are multi-temporal with multiple frequencies, have
Accepted 2 May 2015
become more attractive. A new mode of incorporating spatial and temporal dependence in a homoge-
Available online 14 June 2015
neous region was tried using the Random Forest (RF) classifier for crop classification. A time-series of
medium spatial resolution enhanced vegetation index (EVI) and its summary statistics obtained from
Keywords:
Landsat 7 Enhanced Thematic Mapper Plus (Landsat 7 ETM+) were used to develop a new technique
Crop classification
Upland field
for crop type classification. Eight classes were studied: alfalfa, asparagus, avocado, cotton, grape, maize,
Random forest mango, and tomato. Evaluation was based on several criteria: sensitivity to training dataset size, the
Landsat 7 ETM+ number of variables, and mapping accuracy. Results showed that the training dataset size strongly affects
Enhanced vegetation index the classifier accuracy, but if the training data increase, the rate of improvement decreases. The RF algo-
rithm yielded overall accuracy of 81% and a Kappa statistic of 0.70, indicating high model performance.
Additionally, the variable importance measures demonstrated that the mode and sum of EVI had extre-
mely important variables for crop class separability. RF had computationally good performance. They can
be enhanced by choosing an appropriate classifier for multiple statistics and time-series of Landsat ima-
gery. It might be more economical to use no-cost imaging for crop classification using open-source
software.
Ó 2015 Elsevier B.V. All rights reserved.

1. Introduction profiles of crop phenology and classification (Ryerson et al.,

1985; Panigrahy and Sharma, 1997; Oetter et al., 2000;
Earth observation satellites are indispensable for the estimation Murakami et al., 2001; Simonneaux et al., 2008; Yang et al.,
of crop classification and land-cover monitoring (Asner et al., 2002; 2011; Mellor et al., 2013). Multi-date, high-resolution, and
Jakubauskas et al., 2002; Rodriguez-Galiano et al., 2012a). multi-spectral bands generally yield better classification for crops
Large-scale information related to the spatial distribution of crop than single and few spectral signatures, especially similar spectral
types is critical to many applications, from crop modeling and signatures.
management and estimation of cultivated areas (Loveland, 1991; In Peru, GDP grew 42% during 2002–2011. Agriculture
Mkhabela et al., 2011). Satellite image data have been used more accounted for 7% of total GDP in 2011. Peru is also a leading expor-
widely than airborne imagery for large-scale crop classification ter of organic products such as asparagus, coffee, and bananas.
because of their synoptic scale. As described above, classification However, environmental degradation related to El Niño/La Niña,
and area mapping of crop types using image data from satellite heavy rain, drought, decrease of groundwater, salt damage, acidifi-
sensors have been performed. Several researchers have also cation of soil and desertification hinders stable productivity. In
applied single-date and multi-date multispectral imagery for considering the national export policy, management and under-
standing of crop items are important in terms of the crop price,
supply, and quality. Furthermore, a direct and strong correlation
⇑ Corresponding author at: Department of Environmental and Agricultural exists between the change in the cultivated area, crop varieties
Engineering, Tokyo University of Agriculture and Technology, 3-5-8 Saiwai-cho, and crop supply and demand, cultivation technique, and the num-
Fuchu, Tokyo 183-8509, Japan. Tel./fax: +81 42 367 5679. ber of producers in Peru. For this reason, crop classification
E-mail address: [email protected] (K. Tatsumi).

http://dx.doi.org/10.1016/j.compag.2015.05.001
0168-1699/Ó 2015 Elsevier B.V. All rights reserved.
172 K. Tatsumi et al. / Computers and Electronics in Agriculture 115 (2015) 171–179

mapping development is an urgent necessity for labor-saving of a objectives of this study were (1) to evaluate the functional perfor-
field survey related to land use information and grasping cropping mance and availability of RF for classifying eight upland field crops
information easily and widely. However, several limitations exist in medium homogeneous study areas of about 23,000 ha using pre-
in relation to crop classification that must be resolved. First, sepa- dictor variables obtained solely from Landsat 7 ETM+ across the Ica
rating one crop from another is difficult because of variations in region in Peru; and (2) to assess the effectiveness and efficiency of
moisture, elevation, temperature, soil properties, fertilization, irri- the RF classifier for crop classification using open-source software,
gation, planting dates, and tillage practices. Second, many limita- which is Geographic Resources Analysis Support System (GRASS)
tions of crop classification are related to the similarity of (GRASS Development Team, 2012), R 3.1.0 (R Development Core
reflectance of upland field crops, and field-to-field spatial and Team, 2011) and freely available data, thereby avoiding the use
spectral variations, which are different from the patterns of indi- of ancillary and high-cost data.
vidual crop phenology. Third, crop classification requires methods
that can be interpreted readily and which can be simplified and 2. Materials and methods
operated easily with user-defined parameters that are automated
to adjust in practice. 2.1. Study site
An examination of crop classification methods revealed that
many available methods have used remotely sensed image data. The Ica region, located in southern Peru at approximately 14°S,
Traditional unsupervised methods such as ISODATA or K-means, 75.7°W (Fig. 1), is the target area of this study. Agriculture is con-
parametric supervised algorithms such as maximum likelihood, ducted on the flat Ica landscape, which dominates the southwest-
machine learning algorithms such as artificial neural networks, ern parts of this region, with circumjacent barren areas and
support vector machines, decision trees and ensembles of classi- mountains. Agriculture in the region relies on an aquifer fed by gla-
fiers have been applied to land cover using remote sensing datasets cial melt water. The aquifer is being rapidly depleted, leading to
(Foody, 2004; Lippitt et al., 2008; Mathur and Foody, 2008; Rogan calls for more efficient irrigation, or addition of reservoirs.
et al., 2008; Guerrero et al., 2012; Rodriguez-Galiano et al., Cropland areas occupy about 23,000 ha, with elevations extending
2012a,b). These algorithms were evaluated using multiple remote from sea level to 500 m. The eastern part of this region has topog-
sensing data and ancillary data from many crop-growing environ- raphy that is too rugged for agriculture. The Ica region tends to a
ments. They are effective because they are independent of data dis- desert climate. Temperatures are hot during the summer months
tribution assumptions, which can improve classification accuracy. and warm during winter months. The main crop classes are cotton,
Machine learning techniques, which use ensembles of classifica- grapes, asparagus, maize, and tomato. The following eight classes
tions, have been estimated for many applications recently (Wang were included in this study: alfalfa, asparagus, avocado, cotton,
et al., 2004; Yang et al., 2011; Zhang et al., 2013; Cracknell and grape, maize, mango and tomato.
Reading, 2014). Based on the background presented above,
Random Forest (RF) classifiers increasingly provide a new means 2.2. Landsat time-series data
to predict land-cover classification maps that are robust to varia-
tions in class reflectance caused by land use or disturbances in The database used for this study consists of a four-year time
regional scale mapping. The RF classifiers are ensemble decision series of optical data from Landsat 7 ETM+. Multi-time-series
trees developed in the field of machine learning. The classifiers Landsat 7 ETM+ are used to characterize phenological variation
combine bootstrap sampling to construct a large set of in the state of farming crops. Satellite sensor images of Ica, cap-
classification of individual decision trees (Breiman, 2001; tured by Landsat-7 of worldwide reference system (WRS) during
Rodriguez-Galiano et al., 2012a,b; Mellor et al., 2013). In fact, RF 2008–2011 (path 6, row 70, total 53 scenes), with WGS84 datum,
has high accuracy for land-cover classification across heteroge- projection UTM18S, have been used for crop classification analyses.
neous landscapes. It is more sensitive to noise than the other clas- Level processing of these data is the Landsat Terrain Corrected
sification methods (Pal, 2005; Waske and Braun, 2009; Oliveira Product (L1T), which uses ground control and relief models to
et al., 2012; Rodriguez-Galiano et al., 2012a,b). Moreover, RF clas- attain absolute geodetic accuracy. These data are co-registered
sifier runs efficiently with huge datasets. Recent studies showed and orthorectified. Several studies have demonstrated that a com-
that RF can incorporate multiple variables of remote sensing with bination of multi-time series images can increase the distinction
categorical land use data to improve classification performance between spectrally similar covers representing the phenological
and to discriminate between forests and other ground cover vegetation condition (Lunetta and Balogh, 1999; Yuan et al.,
(Lawrence et al., 2006; Martinuzzi et al., 2009; Ghimire et al., 2005). For these data, we (1) transformed the calibrated digital
2010; Latifi et al., 2010; Guo et al., 2011; Oliveira et al., 2012; number (DN) of Landsat 7 ETM+ products to at-surface reflectance
Rodriguez-Galiano et al., 2012a,b). In addition, topographic (e.g., with simplified atmospheric correction (Tizado, 2013) and (2)
elevation, slope and aspect) and bioclimatic (e.g., temperature, pre- removed cloud effects.
cipitation) variables used in combination with spectral image data Radiometric calibration converts DN to at-sensor radiance
have been demonstrated to enhance forest/non-forest, habitat and (RAD), RAD is defined as
vegetation classification (Franklin, 1995; Fahsi et al., 2000; Joy
ðLmax  Lmin Þ
et al., 2003; Gislason et al., 2006; Sesnie et al., 2008). RAD ¼ ðDN  DNmin Þ  þ Lmin ð1Þ
Most reports of previous work related to crop identification pri- ðDNmax  DNminÞ
marily describe land-cover categories (e.g., urban, grass, water, for- where RAD stands for the at-sensor radiance (W=ðm2  sr  lm)), Lmax
est, and crops). Additionally, multi-source remote sensing and and Lmin are the calibration constants which is described the Landsat
ancillary (e.g., topographic, bioclimate) data, which are big data, metadata file, and DNmax and DN min respectively denote the highest
could also be used to discriminate forest/non-forest using the RF and the lowest points of the range of rescaled radiance in a calibrated
classifier. Few studies have examined the use of multi-date ima- digital number. The at-surface reflectance REF surf is defined as:
gery for upland field crops. In this study, to conduct highly accurate h i " h i #
and simple analysis related to crop classification, we specifically cosðsatt t
sinðeÞ 2
examined the temporal profiles of crop phenology as attested in Srad ¼ e zenith Þ
 ESUN  sinðeÞ  e þ p  RADdark =ðp  d Þ
the Enhanced Vegetation index (EVI) and its summary statistic
obtainable from Landsat 7 ETM+ 30 m resolution data. The ð2Þ
 K. Tatsumi et al. / Computers and Electronics in Agriculture 115 (2015) 171–179 173

Study site

Fig. 1. Study area location.

REF surf ¼ ðRAD  RADdark þ 0:01  Srad Þ=Srad ð3Þ within the time-series data (diversity). These are calculated for
one-month periods during the calendar year. The ‘‘r.series’’ module
where Srad represents the sun radiance, satzenith denotes the satellite in the GRASS is used to produce these raster datasets, and obtained
zenith angle, and Esun signifies the mean solar exoatmospheric irra- datasets are used as input variables in RF classification.
diance (W=ðm2  lm)). Furthermore, d is the earth sun distance in
astronomical units, e stands for the solar elevation angle, and 2.3. Reference data
RADdark signifies the at-sensor radiance calculated from the darkest
object that shows DN with at least 1000 pixels for the entire image. Land cover maps were obtained from a combination of digital
Second, cloud-cover effects are removed from images using an aerial photographic interpretations of color aerial photographs
automated cloud-cover assessment algorithm from Irish (2000) and a detailed field survey of the ground across the study area
with constant values for the pass filter algorithm from Irish et al. acquired during 2008–2011 (Fig. 5a). The field survey was con-
(2006). The present study does not use cloud pixels. ducted by La Molina National Agrarian University. The created
The corrected Landsat 7 ETM+ showed a failure of the Scan Line polygon data were converted to raster format to align with the
Corrector (SLC) on May 31, 2003. Therefore, without an operating 30 m  30 m pixels of Landsat satellite imagery. These reference
SLC, the ETM+ line of sight now traces a zigzag pattern along the data were used for classification and validation procedures.
satellite ground track. Consequently, the imaged area is duplicated,
with width that increases toward the scene edge. Since that time, 2.4. Random forest approach
images have had wedge-shaped gaps on both sides of each scene,
resulting in an estimated 22% of any given scene being lost An ensemble classification algorithm, RF, consists of a group of
(USGS, 2013). Therefore, the portion including the data loss is not tree-based classifiers {h(x, Hk, k = 1,. . .)}, where x is the input vec-
used in the analysis. tor and Hk are independent and identically distributed random
To produce variables for the RF classifier model, EVI is first pro- vectors (Breiman, 2001; Hastie et al., 2009). RF uses bootstrapping
duced using the Landsat 7 ETM+ surface reflectance band 1 (blue), with replacement to enhance the diversity of classification trees,
3 (red), and 4 (near infrared). The enhanced vegetation index (EVI) which allocate each pixel to a class in accordance with the maxi-
is commonly used in studies using remote sensing data because it mum number of votes from the collection of trees. This method,
is optimized to enhance the vegetation signal, yielding improved although it has shown high accuracy and ability to model complex
vegetation monitoring and improved sensitivity in high-biomass interactions among variables, is a ‘‘black-box’’ because the individ-
regions (deFries et al., 1995; Liu and Huete, 1995). Moreover, EVI ual trees cannot be estimated separately (Prasad et al., 2006).
can reduce the influences of canopy background and atmospheric To run the RF model, it was necessary to define several impor-
variation compared to NDVI (Huete et al., 2002). We proceed by tant adjustable parameters. The primary parameters are the num-
determining the EVI values for each Landsat-7 ETM+ scene with ber of predictors at each decision tree node split (mtry) and the
the following: number of decision trees to run (ntree). Liaw and Wiener (2002)
report that mtry = 1 can give good performance. Rodriguez-
NIR  R
EVI ¼ 2:5  ð4Þ Galiano et al. (2012a) showed that reducing mtry weakens each
NIR þ 6:0  R  7:5  B þ 1 tree of the model, but it also reduces the correlation among
Therein, NIR, R, and B respectively denote atmospherically cor- individual trees, which increases the model accuracy. Oliveira
rected surface reflectance in near-infrared, red, and blue bands. et al. (2012) reported that an increase in values of mtry would
The images are enhanced spectrally using multi-date EVI values. result in a higher predictive performance of the model and attribu-
In this study, we make each output cell value a function of the val- tion of higher importance to fewer variables. In consideration of
ues assigned to the corresponding cells in the input raster map lay- these points, it is necessary to optimize the parameters mtry and
ers for 2008–2011. This procedure produced 7 statistic features: ntree to maximize the model accuracy.
average value (average), most frequently occurring value (mode), First, to evaluate the model performance, all data were divided
lowest/highest value range of values (max/min), sum of values with stratified random sampling ranging from 10% (11,781 pixels)
(sum), statistical variance (variance), number of different values to 90% (105,994 pixels) in increments of 10% for test data, left out
174 K. Tatsumi et al. / Computers and Electronics in Agriculture 115 (2015) 171–179

of the training data. The set of test data, which is an independent coefficient in the classification of the training data as altering the
validation set, was used only for the model evaluation. Moreover, training sets in increments of 10%, from 10% to 90%.
the remaining training dataset was divided to 75% (training data- Table 1 shows the relation accuracy and mtry equal to the 2, 7
set) and 25% (validation dataset) for the sake of a repeated (sqrt(p)), 25, 49 (maximum p), respectively. It showed that the
leave-group out cross-validation (LGOCV) strategy. This procedure model accuracy depends on mtry to some extent. Accuracy results
is repeated 10 times to estimate robust prediction performance. across tuning mtry parameters from LGOCV reached a maximum
Each datum of the validation data and test data is used to compute level when mtry equals 25. However, these results showed classifi-
accuracies and error rates averaged over all predictions and to esti- cation accuracy could not be significantly affected by a change in
mate each variable’s importance in the classification. mtry
To reduce data redundancy and to assist the model interpreta- Fig. 2 shows accuracy/kappa when sum of the training and val-
tion and the absolute values of pairwise correlation coefficients idation data were 11,781 (10% of all dataset) to 105,994 (90% of all
were considered. Predictors with near zero-variance values were dataset). For LGOCV, the training percentage was set to 75% to con-
removed. If two variables are highly correlated (>0.75), then, the struct the RF model, with intervals of about 11,700 and mtry = 25.
variable with the largest mean absolute correlation is automati- As a boundary about 30,000 training sample sizes, the rate of
cally removed from the model. improvement of accuracy/kappa is different. From the 30% thresh-
The kappa coefficient (Cohen, 1960), and the producer and user old, accuracy decreases more abruptly to attain kappa equal to 0.45
accuracy are calculated to evaluate the crop identification perfor- (Fig. 2). The ranking of the mtry variables in terms of importance
mance. The kappa coefficient, a statistical measure of inter-rater did not change significantly with different mtry. These results are
reliability, is calculated as follows: consistent with those presented by Cutler et al. (2007) and
Oliveira et al. (2012), and the final number of variables using RF
ðPðaÞ  PðeÞÞ model may be not important in an improvement performance
kappa ¼ ð5Þ
1  PðeÞ compared to the number of training dataset in this case. After all,
larger numbers of training datasets will increase the accu-
Therein, P(a) denotes the overall percent agreement, which rep- racy/kappa, but the rate of improvement is not constant.
resents the relative observed agreement fraction. P(e) is the hypo- Based on the evaluation presented above, the final models were
thetical probability of change agreement, which stands for the bolstered with the 49 predictor variables to estimate many vari-
expected probability fraction between observed data and the RF able importance, and the training data are 79,499 pixel points.
predictions. Therefore, kappa = 1 shows that the raters are in com- ntree initial parameter was set to 100 to produce stable results
plete agreement; kappa = 0 indicates no agreement among raters. because a tree number of 30 or less strongly affects the classifier
Producer accuracy is the proportion of a crop class on the reference accuracy. The steady state is reached at 100 trees or more. Since,
ground that is classified correctly in the field. The user accuracy is the number of ntree eventually converge in model, it is not need
the proportion of a predicted class on a map, which matches the to be a parameter study.
corresponding class on the reference ground. Producer accuracy
stands for the share of ground data that are consistent with classi-
fication results obtained using RF prediction, whereas user accu- 3.2. Variable importance
racy measures the percentage of classification results that are
classified correctly (see Table 2 footnote). Moreover, the selection RF algorithms accommodate two features of the importance of
of the most relevant variables to include the final RF model is done the variables by the mean decrease in accuracy (MDA) and the
by ranking the variables according to their importance in all sam- mean decrease in Gini index (MDG). Fig. 3 shows the respective
ples. The Random Forest package is included in a statistics package contributions of predictor variables to the RF final classification
(R 3.1.0; R Development Core Team, 2011) that is an open source model generated with respect to the monthly time-series EVI vari-
language and software environment. It is used for statistical ables in terms of MDA and MDG in accuracy, respectively.
computing. According to MDA and MDG, the average, mode, sum, and variance
of monthly EVI series were shown to be more important variables,
with values greater than 0.02 and 2000, respectively. On the other
3. Results and discussion hand, the contributions of max, min, and diversity in classifying the
data are less than the others. Especially, averages of February and
3.1. Effects of the number of predictor variables, and training data on April, mode and sum of May showed a greater contribution to the
classification accuracy RF model accuracy. Higher values of MDA and MDG indicate vari-
ables that are more important to the classification the data.
The collection of ground truth information in the target area to Though, it is an important to reduce the initial number of variables
train the model classifier is a difficult and time-consuming task with respects to computational costs and efficiency, it should be
that is also expensive, especially when processing numerous and careful not to have either too few variables or too many variables.
similar phenologies of categories, some of which have high Fig. 4 shows the MDA variable importance measures of each
intra-class variation. The largest number of training areas possible upland crops for the RF classification. It showed that the overall
must be used to represent the entire variation in a category (Pal trend related to variable importance is the same as the overall con-
and Mather, 2003). Nevertheless, it is possible to design an opti- tribution (Figs. 3 and 4). As the overall trend, when the mode and
mum sampling scheme that can operate rapidly and economically sum have not been used for RF classifier, classification accuracy
with an acceptable classification accuracy level (Lippitt et al., 2008; significantly decreases. Therefore, for each crop, the mode and
Rogan et al., 2008; Rodriguez-Galiano et al., 2012a,b). This study sum obtained from the EVI time-series have great importance in
examines eight crop species that have high variation and similar the class classification of field crops. Especially, the mode and
phenology. Consequently, it is necessary to use numerous training sum of May and July for all crops have a greater importance for
datasets to construct the RF classifier (Rodriguez-Galiano et al., the classification. In this period, all crops are under development
2012a,b). The effects of the number of predictor variables (p) stage, mid-season stage, and late season stage, and the mode and
(mtry), and training data on classification accuracy were evaluated sum of EVI could have high sensitivity to vegetation signal due to
using the overall accuracy (see Table 2 footnote) and kappa dense vegetation cover.
 K. Tatsumi et al. / Computers and Electronics in Agriculture 115 (2015) 171–179 175

Table 1
Effects of training set size and the number of variables on classification accuracy.

Training sample size

mtry 8839 17,671 26,503 35,336 44,167 53,000 61,833 70,667 79,499
2 0.656 0.681 0.702 0.715 0.726 0.735 0.744 0.753 0.757
7 0.666 0.692 0.715 0.727 0.739 0.748 0.757 0.766 0.771
25 0.678 0.705 0.730 0.743 0.755 0.765 0.774 0.784 0.789
49 0.673 0.700 0.725 0.740 0.750 0.761 0.770 0.778 0.783

Fig. 2. Variation of accuracy and kappa index with increasing number of training data sizes.

Fig. 3. Average of all crops variable importance with respect to summary statistics of EVI in terms of (a) the mean decrease in accuracy and (b) mean decrease in Gini index.

Peru is the second largest asparagus production country in the by this independence. A random selection of input variables fea-
world, and the cultivated area in this region is about 11,605 ha. tures seems well suited for multi-temporal approaches. Moreover,
Though asparagus grows year-round in this region, it generally the ensemble performance is increased further by combining this
begin to grow from early May after dormant stage (January to feature selection with random selection (Waske and Braun, 2009).
March). Therefore, the explanatory variable during winter season Error matrices were used to assess accuracy assessment for crop
may relatively contribute to increased identification of asparagus type classification using the RF model. Table 2 shows the accuracy
cultivated area. However, since the Landsat data in the study area assessment for crop type mapping obtained through the RF final
were missing through January to March, the importance of contri- model using an independent sample (test data) of 11,773 pixels.
bution to the classification in these months will be future tasks. The results of the accuracy assessment through the RF model con-
Avocado, a crop of great economic importance, is being produced firm the generally good performance of classifier ensembles con-
increasingly in recent years. The yield in the study area was sidering the large number of farming crops. The overall accuracy
9.7 t/ha in 2008, 9.8 t/ha in 2009, 14.2 t/ha in 2010, and 13.2 t/ha is 81%, with a kappa coefficient of 0.70 for test data. These results
in 2011 as reported by La Molina National Agrarian University. showed an almost complete match according to Landis (1977).
Harvest of avocado is mainly performed in winter season (Agro User and producer accuracy values of individual classes were,
Peruano, 2012), therefore, dense crop cover provided an important respectively, 89% and 60% on average. However, producer accuracy
predictor variables. However, why the lower values of mean was lower for alfalfa (44%), avocado (39%), and tomato (39%)
decrease in accuracy in June was unclear. Even though we cannot classes. In other words, the classification map missed 56%, 61%,
conclude that important variable obtained from this study is also and 61% (omission error) of the alfalfa, avocado, and tomato areas
applicable under other region, it is important to take into account on the ground, indicating a tendency for the model to misclassify
as many variables, which may be the agronomic important. alfalfa, avocado, and tomato as cotton (11/34, 231/467, 98/222).
Therefore, this high omission error was attributable mainly to
3.3. Classification accuracy spectral similarities between alfalfa, avocado, and cotton in fields.
Therefore, we think that alfalfa and avocado were typically mis-
Fig. 5 presents classification results of the ground truth (a) and classified within cotton. However, the misclassification between
those obtained using RF (b). Visual assessment of the classification alfalfa and avocado is unclear. Asparagus, grape, maize, and mango
results shows good performance by RF classifier ensembles, and were also misclassified within cotton. In this study, the cultivated
shows the general structures of the study area. However, classifica- area of cotton accounts for 50% of the total cultivated area. The
tion results obtained using RF include noise even in homogeneous mode properties of cotton show similar characteristics to those
areas. The main reason for the higher accuracy is certainly the of other crops. Therefore, crops with a small cultivated area com-
underlying assumption of classifier ensembles. The generation of pared to cotton are likely to be misclassified as cotton in pixels.
independent classifiers and the performance is influenced directly Waske and Braun (2009) show that the different crop types
176 K. Tatsumi et al. / Computers and Electronics in Agriculture 115 (2015) 171–179

Fig. 4. Variable importance with respect to summary statistics of EVI in terms of the mean decrease in accuracy (a) alfalfa, (b) asparagus, (c) avocado, (d) cotton, (e) grape, (f)
maize, (g) mango and (f) tomato.

(cereals, orchards, rapeseed and root crop) and complexity of the In this study, the highest producer accuracy was accomplished
classes (e.g., orchards, urban) still yield accurate and stable results, for cotton (98%), which in the Ica region occupies the largest culti-
which is extremely promising. Results of our study show that a cer- vated area. The lowest producer accuracy was attained for avocado
tain problem remains in accuracy for upland field crops, which has and tomato. It occurred in a small cultivated area, which suggests
similar phenology and summary statistics. that crop classification accuracy increases as the validation data
However, the user accuracy is a rigorous independent validation become numerous. The categories most difficult to classify were
for all crops (everywhere > 76%). User accuracies were excellent for those with the same variable importance ranks with cotton in top-
alfalfa, asparagus, avocado, mango, and tomato (94%, 91%, 95%, side, such as alfalfa and avocado. Our results show something
90%, and 98%, respectively) and very good for grape and maize about the findings of other studies using RF, that this ensemble
(86% and 83%, respectively). The producer accuracy for cotton RF classifier is useful to learn multiple crop cover types. In this
was excellent (98%), but the user accuracy was only 76%. In fact, study, the spatial resolution of the crop fields is the same as the
135/5854 cotton pixels on ground truth were misclassified as other image sensor resolution. Alfalfa, avocado, and tomato have lower
crops (omission error), whereas 1796/7515 cotton pixels stratified producer’s classification accuracy and occur in a small cultivated
on the classification map belonged to the other seven classes (com- area, but a completely mixel-free classification is unlikely at the
mission error). However, results show that the classification result field boundary. Therefore, methods applying much finer spatial
of the models is correct in general, and that the reliability of the resolution (e.g., IKONOS, GeoEye, WorldView and SAR data) might
image classification is high. Apart from a few exceptions, results greatly enhance the accuracy. Finally, the RF classifier model of EVI
indicate that application of the RF model for agricultural classes time-series data shows considerable promise as a tool for
as their areal quantification are advancing and might demonstrate crop-type identification monitoring. Especially, changes in the
the classifier’s utility in crop type mapping management actions mode and sum of EVI time-series are indicative of significant land
and yield forecasts. surface changes.
 K. Tatsumi et al. / Computers and Electronics in Agriculture 115 (2015) 171–179 177

Fig. 5. Classification results maps generated from the Landsat 7 ETM+ image dataset and the corresponding ground truth (a), using RF (b) (diagonal lines of (b) is missed by
SLC).

Table 2
Error matrix and accuracy measures based on the RF final model.

Predicted class Reference class

Alfalfa Asparagus Avocado Cotton Grape Maize Mango Tomato Total User accuracy%
Alfalfa 15 0 0 0 0 1 0 0 16 94a
Asparagus 1 1733 11 57 51 24 14 4 1895 91
Avocado 0 1 183 4 1 2 0 1 192 95
Cotton 11 494 231 5719 417 402 143 98 7515 76
Grape 1 41 20 29 800 15 4 17 927 86
Maize 2 31 19 37 30 726 12 14 871 83
Mango 4 1 3 7 4 6 242 1 268 90
Tomato 0 0 0 1 0 1 0 87 89 98
Total 34 2301 467 5854 1303 1177 415 222 11,773
Producer accuracy % 44a 75 39 98 61 62 58 39 81b
a
Producer’s accuracy of alfalfa = 15/34 = 44%, and User’s accuracy of alfalfa = 15/16 = 94%.
b
Overall accuracy = (15 + 1773 + 183 + 5719 + 800 + 726 + 242 + 87)/11773 = 81%.

4. Conclusions 1. RF requires only three user-defined parameters to be set:

the number of trees, the number of random split variables,
This study was undertaken to investigate the applicability of the and the training dataset sizes. The training dataset sizes
RF classifier model for crop cover identification in a geomorpholog- have a strong effect on the classifier accuracy, but if the
ical homogeneous crop area (23,441 ha) in Ica, Peru. We incorpo- number of training data increases, the rate of improvement
rated a suite of multi-temporal Landsat 7 ETM+ satellite imagery decreases. In addition, the number of variables only slightly
variables into an RF model. The RF performed well in the classifica- influences the classifier accuracy.
tions with eight upland field crops. Moreover, we examined the RF 2. The RF model can evaluate the variable importance for the
accuracy with the number of variables, training dataset sizes. The overall classification of the crop classification and for the
number of variables has little effect on the classifier accuracy. classification of each field crop category by the mean
The more the training data are increased, the higher the classifier decrease accuracy and mean decrease Gini index. Variable
accuracy becomes. Therefore, there is less need for tuning the importance measures demonstrate that the mode and
model parameters. The following are summaries of the findings. sum of EVI have extremely important variables for crop
178 K. Tatsumi et al. / Computers and Electronics in Agriculture 115 (2015) 171–179

class separability. The variance and diversity of EVI also Huete, A., Didan, K., Miura, T., Rodriguez, E.P., Gao, X., Ferreira, L.G., 2002. Overview
of the radiometric and biophysical performance of the MODIS vegetation
have importance for classification in crop categories. It
indices. Remote Sens. Environ. 83 (1–2), 195–213.
might be more economical to use no-cost imagery for crop Irish, R.R., 2000. Landsat 7 Automatic Cloud Cover Assessment. In Shen, S.S.,
classification. However, we have difficulty understanding Descour, M.R. (Eds.): Algorithms for Multispectral, Hyperspectral, and
the precise rules used to generate the conclusive crop clas- Ultraspectral Imagery VI, Proceedings of SPIE, 4049, pp. 348–355.
Irish, R.R., Barker, J.L., Goward, S.N., Arvidson, T., 2006. Characterization of the
sifications because of the many multiple classification trees Landsat-7 ETM+ Automated Cloud-Cover Assessment (ACCA) Algorithm.
are created from resampling the same dataset. Photogramm. Eng. Remote Sens. 72 (10), 1179–1188.
3. Results of this study demonstrate that the application of RF Jakubauskas, M.E., Legates, D.R., Kastens, J.H., 2002. Crop identification using
harmonic analysis of time-series AVHRR NDVI data. Comput. Electron. Agr. 37
classification approaches is useful for crop classification and (1–3), 127–139.
land management. They can be useful to detect mislabeled Joy, S.M., Reich, R.M., Reynolds, R.T., 2003. A non-parametric, supervised
areas, even multiple crop species in this area that have sim- classification of vegetation types on the Kaibab National Forest using decision
trees. Int. J. Remote Sens. 24 (9), 1835–1852.
ilar phenology. The overall accuracy and kappa coefficient Landis, J.R., Koch, G.G., 1977. The measurement of observer agreement for
for crops were 81% and 0.70, respectively. Therefore, techni- categorical data. Biometrics 33 (1), 159–174.
cal challenges exist for crop classification using the RF Latifi, H., Nothdurft, A., Koch, B., 2010. Non-parametric prediction and mapping of
standing timber volume and biomass in a temperate forest: application of
model in the case of the small training dataset and similarly multiple optical/LiDAR-derived predictors. Forestry 83 (4), 395–407.
variable importance rank for crop species (e.g., inclusion of Lawrence, R.L., Wood, S.D., Sheley, R.L., 2006. Mapping invasive plants using
elevation and climatic variables in the model). hyperspectral imagery and Breiman Cutler classifications (randomForest).
Remote Sens. Environ. 100 (3), 356–362.
Liaw, A., Wiener, M., 2002. Classification and regression by random forest. R News 2
The approaches and methods presented in this study use only (3), 18–22.
the no-cost dataset. They can be useful for different crops in med- Lippitt, C.D., Rogan, J., Li, Z., Eastman, J.R., Jones, T.G., 2008. Mapping Selective
ium areas. We must evaluate and compare these approaches and Logging in Mixed Deciduous Forest: A Comparison of Machine Learning
Algorithms. Photogrammetric Engineering & Remote Sensing 74 (10), 1201–
methods using other types of remote sensing data and additional 1211.
variables. Liu, H., Huete, A., 1995. A feedback based modification of the NDVI to minimize
canopy background and atmospheric noise. IEEE Trans. Geosci. Remote Sens. 33
(2), 457–465.
Acknowledgements Loveland, T.R., Merchant, J.W., Ohlen, D.O., Brown, J.F., 1991. Development of a land-
cover characteristics database for the conterminous US. Photogrammetric
Engineering & Remote Sensing 57 (11), 1453–1463.
We thank JICA–JSPS for providing funding for this study. We
Lunetta, R.S., Balogh, M.E., 1999. Application of Multi-temporal Landsat 5 TM
also appreciate La Molina National Agrarian University for field imagery for Wetland Identification. Photogrammetric Engineering and Remote
data collection and ground verification, which supported this Sensing 65 (11), 1303–1310.
study. Martinuzzi, S., Vierling, L.A., Gould, W.A., Falkowski, M.J., Evans, J.S., Hudak, A.T.,
Vierling, K.T., 2009. Mapping snags and understory shrubs for a LiDAR-based
assessment of wildlife habitat suitability. Remote Sens. Environ. 113 (12),
References 2533–2546.
Mathur, A., Foody, G.M., 2008. Crop classification by support vector machine with
intelligently selected training data for an operational application. Int. J. Remote
Agro Peruano: Abriendo un nuevo horizonte junto con el Japón, 2012. <http://www.
Sens. 29 (8), 2227–2240.
ccipj.org.pe/LINKSforWEB/no11dec2012/LINKS11_especial.pdf> (accessed
Mellor, A., Haywood, A., Stone, C., Jones, S., 2013. The performance of Random
20.11.2014).
Forests in an Operational Setting for Large Area Sclerophyll Forest Classification.
Asner, G.P., Keller, M., Pereira, R., Zweede, J.C., 2002. Remote sensing of selective
Remote Sensing 5 (6), 2838–2856.
logging in Amazonia Assessing limitations based on detailed field observations,
Mkhabela, M.S., Bullock, P., Raj, S., Wang, S., Yang, Y., 2011. Crop yield forecasting on
Landsat ETM+, and textural analysis. Remote Sens. Environ. 80 (3), 483–496.
the Canadian Prairies using MODIS NDVI data. Agric. For. Meteorol. 151 (3),
Breiman, L., 2001. Random forests. Mach. Learn. 45 (1), 5–32.
385–393.
Cohen, J., 1960. A coefficient of agreement for nominal scales. Educ. Psychol.
Murakami, T., Ogawa, S., Ishitsuka, N., Kumagai, K., Saito, G., 2001. Crop
Measur. 20 (1), 37–46.
discrimination with multitemporal SPOT/HRV data in the Saga Plains. Japan.
Cracknell, M.J., Reading, A.M., 2014. Geological mapping using remote sensing data:
International Journal of Remote Sensing 22 (7), 1335–1348.
a comparison of five machine learning algorithms, their response to variations
Oetter, D.R., Cohen, W.B., Berterretche, M., Maiersperger, T.K., Kennedy, R.E., 2000.
in the spatial distribution of training data and the use of explicit spatial
Land cover mapping in an agricultural setting using multiseasonal Thematic
information. Comput. Geosci. 63, 22–33.
Mapper data. Remote Sens. Environ. 76 (2), 139–155.
Cutler, D.R., Edwards Jr., T.C., Beard, K.H., Cutler, A., Hess, K.T., Gibson, J., Lawler, J.J.,
Oliveira, S., Oehler, F., San-Miguel-Ayanz, J., Camia, A., Pereira, J.M.C., 2012.
2007. Random forests for classification in ecology. Ecology 88 (11), 2783–2792.
Modeling spatial patterns of fire occurrence in Mediterranean Europe using
deFries, R., Hansen, M., Townshend, J., 1995. Global discrimination of land cover
Multi Regression and Random Forest. For. Ecol. Manage. 275, 117–129.
from metrices derived from AVHRR data sets. Remote Sens. Environ. 54 (3),
Pal, M., 2005. Random forest classifier for remote sensing classification. Int. J.
209–222.
Remote Sens. 26 (1), 217–222.
Fahsi, A., Tsegaye, T., Tadesse, W., Coleman, T., 2000. Incorporation of digital
Pal, M., Mather, P.M., 2003. An assessment of the effectiveness of decision
elevation models with Landsat–TM data to improve land cover classification
tree methods for land cover classification. Remote Sens. Environ. 86 (4),
accuracy. For. Ecol. Manage. 128, 57–64.
554–565.
Foody, G.M., 2004. Supervised image classification by MLP and RBF neural networks
Panigrahy, S., Sharma, S.A., 1997. Mapping of crop rotation using multidate Indian
with and without an exhaustively defined set of classes. Int. J. Remote Sens. 25
Remote Sensing Satellite digital data. ISPRS J. Photogramm. Remote Sens. 52 (2),
(15), 3091–3104.
85–91.
Franklin, J., 1995. Predictive vegetation mapping: geographic modelling of
Prasad, A.M., Iverson, L.R., Liaw, A., 2006. Newer classification and regression tree
biospatial patterns in relation to environmental gradients. Prog. Phys. Geogr.
techniques: bagging and random forests for ecological prediction. Ecosystems 9
19 (4), 474–499.
(2), 181–199.
Ghimire, B., Rogan, J., Miller, J., 2010. Contextual land-cover classification:
R Development Core Team, 2011. R: A Language and Environment for Statistical
incorporating spatial dependence in land-cover classification models using
Computing; R Foundation for Statistical Computing: Vienna, Australia. <http://
random forests and the Getis statistic. Remote Sens. Lett. 1 (1), 45–54.
www.r-project.org> (accessed 30.11.2013).
Gislason, P.O., Benediktsson, J.A., Sveinsson, J.R., 2006. Random Forests for land
Rodriguez-Galiano, V.F., Chica-Olmo, M., Abarca-Hernandez, F., Atkinson, P.M.,
cover classification. Pattern Recogn. Lett. 27 (4), 294–300.
Jeganathan, C., 2012a. Random Forest classification of Mediterranean land cover
GRASS Development Team. Geographic Resources Analysis Support System (GRASS)
using multi-seasonal imagery and multi-seasonal texture. Remote Sens.
Software; Version 6.4; Open Source Geospatial Foundation Project; 2012.
Environ. 121, 93–107.
<http://grass.osgeo.org/> (accessed 30.11.2013).
Rodriguez-Galiano, V.F., Ghimire, B., Rogan, J., Chica-Olmo, M., Rigol-Sanchez, J.P.,
Guerrero, J.M., Pajares, G., Montalvo, M., Romeo, J., Guijarro, R.M., 2012. Support
2012b. An assessment of the effectiveness of a random forest classifier for land-
vector machines for crop/weeds identification in maize fields. Expert Syst. Appl.
cover classification. ISPRS J. Photogramm. Remote Sens. 67, 93–104.
39 (12), 11149–11155.
Rogan, J., Franklin, J., Stow, D., Miller, J., Woodcock, C., Roberts, D., 2008. Mapping
Guo, L., Chehata, N., Mallet, C., Boukir, S., 2011. Relevance of airborne lidar and
land-cover modifications over large areas: a comparison of machine learning
multispectral image data for urban scene classification using Random Forests.
algorithms. Remote Sens. Environ. 112 (5), 2272–2283.
ISPRS J. Photogramm. Remote Sens. 66 (1), 56–66.
Ryerson, R.A., Dobbins, R.N., Thibault, C., 1985. Timely crop area estimates from
Hastie, T., Tibshirani, R., Friedman, J., 2009. Random forests. The Elements of
Landsat. Photogramm. Eng. Remote Sens. 51 (11), 1735–1743.
Statistical Learning. Springer, New York, pp. 587–604.
 K. Tatsumi et al. / Computers and Electronics in Agriculture 115 (2015) 171–179 179

Sesnie, S.E., Gessler, P.E., Finegan, B., Thessler, S., 2008. Integrating Landsat TM and Wang, L., Sousa, W.P., Gong, P., Biging, G.S., 2004. Comparison of IKONOS and Quick-
SRTM-DEM derived variables with decision trees for habitat classification and Bird images for mapping mangrove species on the Caribbean coast of Panama.
change detection in complex neotropical environments. Remote Sens. Environ. Remote Sens. Environ. 91 (3–4), 432–440.
112 (5), 2145–2159. Waske, B., Braun, M., 2009. Classifier ensembles for land cover mapping using
Simonneaux, V., Duchemin, B., Helson, D., Er-Raki, S., Olioso, A., Chehbouni, A.G., multitemporal SAR imagery. ISPRS J. Photogramm. Remote Sens. 64 (5), 450–457.
2008. The use of high-resolution image time series for crop classification and Yang, C., Everitt, J.H., Murden, D., 2011. Evaluating high resolution SPOT 5 satellite
evapotranspiration estimate over an irrigated area in central Morocco. Int. J. imagery for crop identification. Comput. Electron. Agr. 75 (2), 347–354.
Remote Sens. 29 (1), 95–116. Yuan, F., Bauer, M.E., Heinert, N.J., Holden, G.R., 2005. Multi-level land cover
Tizado, E.J., 2013. GRASS GIS manual. <http://grass.osgeo.org/grass64/manuals/ mapping of the Twin Cities (Minnesota) metropolitan area with multi-seasonal
i.landsat.toar.html> (accessed 28.11.2013). Landsat TM/ETM+ Data. Geocarto Int. 20 (2), 5–13.
USGS, 2013. SLC-off Products: Background. <http://landsat.usgs.gov/products_ Zhang, W., Yang, Y., Wang, Q., 2013. A study on software effort prediction using
slcoffbackground.php> (accessed 29.11.2013). machine learning techniques. Commun. Comput. Inform. Sci. 275, 1–15.

View publication stats