Importance of spatially distributed hydrologic variables for land use change modeling

Environmental Modelling & Software 83 (2016) 245e254

Contents lists available at ScienceDirect

Environmental Modelling & Software journal homepage: www.elsevier.com/locate/envsoft

Importance of spatially distributed hydrologic variables for land use change modeling € rn Waske Paul D. Wagner*, Bjo €t Berlin, D-12249 Berlin, Germany Remote Sensing and Geoinformatics, Institute of Geographical Sciences, Freie Universita

a r t i c l e i n f o

a b s t r a c t

Article history: Received 17 February 2016 Received in revised form 4 May 2016 Accepted 4 June 2016

Land use changes have a pronounced impact on hydrology. Vice versa, hydrologic changes affect land use patterns. The objective of this study is to test whether hydrologic variables can explain land use change. We employ a set of spatially distributed hydrologic variables and compare it against a set of commonly used explanatory variables for land use change. The explanatory power of these variables is assessed by using a logistic regression approach to model the spatial distribution of land use changes in a meso-scale Indian catchment. When hydrologic variables are additionally included, the accuracies of the logistic regression models improve, which is indicated by a change in the relative operating characteristic statistic (ROC) by up to 11%. This is mostly due to the complementarity of the two datasets that is reflected in the use of 44% commonly used variables and 56% hydrologic variables in the best models for land use change. © 2016 Elsevier Ltd. All rights reserved.

Keywords: Land use modeling Land-Use/Cover Change (LUCC) Logistic regression Hydrologic modeling India

1. Introduction Anthropogenic land cover transformation and land-use activities, particularly due to urbanization, farmland expansion and agricultural intensification, are among the most important components of global environmental change (Lambin et al., 2001; Turner et al., 2007). Regional land use and land cover change affects water resources (Foley et al., 2005) and vice versa, land use and land cover change is affected by changes in water resources (Lambin et al., 1999). Consequently, land use is a key component of hydrologic models (e.g., MIKE SHE, Refsgaard and Storm, 1995; SWAT, Arnold et al., 1998) and a key research topic in hydrology (DeFries and Eshleman, 2004). However, in studies and models of land use change, the potential explanatory power of hydrologic patterns for land use change is rarely exploited. Several land use change models use spatially distributed explanatory variables to predict the spatial distribution of land use changes (e.g., CLUE, Verburg and Overmars, 2009; GEOMOD, Pontius et al., 2001). These explanatory variables for land use change can broadly be classified in biophysical variables (e.g., topography, soil characteristics, climatic variables) and socioeconomic variables (e.g., population density, distance to roads or

* Corresponding author. E-mail address: [email protected] (P.D. Wagner). http://dx.doi.org/10.1016/j.envsoft.2016.06.005 1364-8152/© 2016 Elsevier Ltd. All rights reserved.

cities; Veldkamp and Lambin, 2001). Established variables with proven explanatory power include topographic variables, soil characteristics, population variables, and distance variables (e.g., ~ ate-Valdivieso and Baumann et al., 2011; Mas et al., 2004; On Bosque Sendra, 2010; Schneider and Pontius, 2001; Verburg et al., 2002, 2004). Sometimes also hydrologic variables like precipitation, available water capacity (Sohl et al., 2007), evapotranspiration (Prishchepov et al., 2013), a yearly moisture index (Rutherford et al., 2007), or a wetness index (Kim et al., 2014) are used to explain land use change. However, the distance to water bodies or streams often serves as the only proxy variable for water availability (e.g., Huang and Cai, 2007; Verburg et al., 2004; Xu et al., 2013), possibly due to the ease of computation. One reason that could have led to the rare consideration of hydrologic variables might be the limited availability of such spatially distributed variables. Moreover, a considerable amount of preprocessing is required e.g., to derive soil moisture patterns from satellite data (Koyama et al., 2010), evapotranspiration estimates using SEBAL (Bastiaanssen et al., 1998), or interpolated precipitation patterns (Buytaert et al., 2006). In summary, hydrologic variables are not among the wellestablished and commonly used variables in land use change modeling, even though they may affect land use change. Particularly in data scarce regions or in areas where few explanatory variables are available as spatially explicit data (Veldkamp and Lambin, 2001), a combination of different datasets to optimize information content seems promising. Moreover, due to the

246

P.D. Wagner, B. Waske / Environmental Modelling & Software 83 (2016) 245e254

interdependence of hydrology and land use, the inclusion of hydrologic information in land use change models could be valuable. In this study, we compare a set of commonly used biophysical and socio-economic variables to a set of hydrologic variables with regard to their potential to explain land use change. The study is carried out in an Indian catchment that experiences seasonally limited rainfall, where hydrologic variables have a particular importance with regard to land use changes (Lambin et al., 1999). Moreover, data availability for the study area is limited, so that additional spatially distributed information is valuable. To evaluate the potential of including hydrologic variables in land use change models we focus on three main objectives. We test i) if spatially distributed hydrologic variables can explain patterns of land use change in the study area, and ii) if they provide complementary information as compared to commonly used biophysical and socioeconomic variables. Moreover, we iii) analyze and evaluate the explanatory variables and the derived probability maps for land use changes.

2. Materials and methods 2.1. Study area The catchment of the Mula and Mutha Rivers is situated in western India upstream of the city of Pune. Major parts of the meso-scale catchment (2036 km2) are located in the Western Ghats (Fig. 1). Elevation ranges from 550 m in Pune up to 1300 m a.s.l. on the top ridges in the Western Ghats. The climate is tropical wet and dry with a dependence of rainfall on the summer monsoon (June to October). Annual rainfall amounts decrease from approximately 3500 mm a
1 in the western part of the catchment to 750 mm a
1 in the eastern part of the catchment (Gadgil, 2002; Gunnell, 1997). The majority of the study area (92.5%) is covered by a sandy clay loam while the rest (7.5%) is covered by clay (Food and Agriculture Organization of the United Nations (FAO), 2003). Land use is

dominated by semi-natural vegetation (70.7%), with forests mainly on the higher elevations in the west, whereas shrubland and grassland cover lower elevations (Fig. 1). Cropland (13.5%) is mainly found in proximity to water sources and settlements and is dominated by small fields (<1 ha) with rainfed agriculture during the monsoon season and irrigation during the dry season (rice-wheat rotation, sugarcane, mixed cropland). Typically, two crops per year are grown. Urban area (10.1%) is predominantly found in the eastern part of the catchment where the city of Pune and its surrounding settlements are located (Fig. 1). Reservoirs and rivers account for 5.7% of the study area (Wagner et al., 2013). 2.2. Land use change We use three land use maps for the cropping years 1989/90, 2000/01, and 2009/10 from Wagner et al. (2013) to determine land use changes between 1989/90 and 2000/01 and between 2000/01 and 2009/10, respectively. The maps are based on the classification of multispectral remote sensing data from different cropping seasons, using a stratified knowledge-based approach and a maximum likelihood classifier. The classification is challenging due to strong seasonal variations that result in two main growing seasons, various cropping practices and crop types, small-scale agriculture, and the limited availability of cloud-free satellite data. Therefore, the averaged overall accuracy per cropping year varies between 62% and up to more than 90%. While some classes are mapped with relative high accuracies (e.g., as sugarcane, rice, forest), classification accuracies for grassland, shrubland and mixed cropland are lower, compared to the accuracies achieved for other classes (Wagner et al., 2013). One reason for this might be that the differentiation between such land cover classes is often challenging, due to spectral ambiguity within the multispectral data (Song et al., 2002; Stefanski et al., 2014). A post-classification change detection between the land use maps of 1989/90 and 2009/10 shows that the catchment has been affected by a pronounced urbanization from 5.1% to 10.1% of the study area as well as an expansion of farmland from 9.7% to 13.5%. Consequently, semi-natural vegetation has decreased from 79.8% to 70.7%. Due to lower class specific classification accuracies, changes within the semi-natural classes can partly be interpreted as pseudo change, e.g., 59.8% of changed shrubland areas and 79.5% of changed grassland areas in 2009/10 have been converted from the respective other class when compared to the land use of 1989/90 (Wagner et al., 2013). 2.3. Explanatory variables

Fig. 1. Land use of the Mula and Mutha Rivers catchment in 2009/10.

Three sets of variables are tested to explain land use changes: i) commonly used biophysical and socio-economic variables in land use science, hereafter referred to as common variables, ii) modeled hydrologic variables, and iii) all variables together (variable sets i and ii). The first variable set consists of population density, distance to roads, distance to rivers, elevation, slope, aspect, and soil type (Table 1). These variables are widely used to explain land use ~ ate-Valdivieso and Bosque Sendra, 2010; Sohl changes (e.g., On et al., 2007; Verburg et al., 2002). We have used OpenStreetMap data from which the distance to roads variable has been calculated. The roads polygon shapefile (OpenStreetMap, 2015) captures the general characteristics of the road network quite well, even though some smaller roads are missing when compared to satellite data. Unfortunately, road networks for 1990 or 2000 have not been available to this study, so that changes in this variable are not represented. Moreover, we have employed a 30 m digital elevation model (DEM) based on ASTER satellite data which was processed and evaluated by Wagner et al. (2011). ASTER data has been proven

P.D. Wagner, B. Waske / Environmental Modelling & Software 83 (2016) 245e254

247

Table 1 Explanatory variables. Variable

Unit

Source

Spatial resolution

Population density

Persons per km2

Distance to roads Distance to rivers Elevation Slope Aspect Soil type

m m m Percent Degree e

Gridded Population of the World Version 3 CIESIN (2005) available for 1990 and 2000 Calculated from roads layer from OpenStreetMap (2015) Calculated from SWAT model streams (Wagner et al., 2013) ASTER DEM, processing Wagner et al. (2011) Calculated from ASTER DEM Calculated from ASTER DEM Digital Soil Map of the World (FAO, 2003)

2.5 arc-minutes resolution, resampled to 30 m 30 m 30 m 30 m 30 m 30 m Scale 1:5,000,000, resampled to 30 m

Water yield

mm per month

WYLD from hydrologic

HRU level, resampled to 30 m

Surface runoff

mm per month

SURQ_GEN from hydrologic

HRU level, resampled to 30 m

Interflow

mm per month

LATQ from hydrologic

HRU level, resampled to 30 m

Baseflow

mm per month

GW_Q from hydrologic

HRU level, resampled to 30 m

Transmission losses

mm per month

TLOSS from hydrologic

HRU level, resampled to 30 m

Evapotranspiration

mm per month

ET from hydrologic

HRU level, resampled to 30 m

Potential evapotranspiration

mm per month

PET from hydrologic

HRU level, resampled to 30 m

Soil water content

mm per month

SW_END from hydrologic

HRU level, resampled to 30 m

Groundwater recharge

mm per month

GW_RCHG from hydrologic

HRU level, resampled to 30 m

Water in the shallow aquifer returning to the root zone Irrigation amount

mm per month

Mean model output for variable model (Wagner et al., 2013) Mean model output for variable model (Wagner et al., 2013) Mean model output for variable model (Wagner et al., 2013) Mean model output for variable model (Wagner et al., 2013) Mean model output for variable model (Wagner et al., 2013) Mean model output for variable model (Wagner et al., 2013) Mean model output for variable model (Wagner et al., 2013) Mean model output for variable model (Wagner et al., 2013) Mean model output for variable model (Wagner et al., 2013) Mean model output for variable model (Wagner et al., 2013)

REVAP from hydrologic

HRU level, resampled to 30 m

mm per month

Water stress days

d per month

Mean model output for variable IRR from hydrologic model (Wagner et al., 2013) Mean model output for variable W_STRS from hydrologic model (Wagner et al., 2013)

suitable for the derivation of digital elevation models and derivatives such as slope and aspect even in mountainous regions (Eckert et al., 2005). The Digital Soil Map of the World (FAO, 2003) represents only two soils in the study area, so that we include the soil type as categorical variable. The second dataset was derived from a hydrologic model application of the Soil and Water Assessment Tool (SWAT, Arnold et al., 1998) for the studied catchment (Wagner et al., 2013). The model shows a reliable performance in the study area, which is supported by Nash-Sutcliff efficiencies of 0.69 and 0.67 for daily rainy season discharge at the two gauges in the study area that are less affected by dam management (G1, G4, Fig. 1). A more detailed description of model setup and model evaluation is available in Wagner et al. (2011, 2012). The SWAT model simulates the water balance based on hydrologic response units (HRUs), which are areas consisting of the same land use, soil and slope class. While the HRU distribution is often simplified by merging small HRUs that do not meet a user defined threshold (Her et al., 2015); we use a fully distributed HRU model setup (no threshold applied). This setup provides the opportunity to link the HRU output from the SWAT model to the respective spatially identified HRU, to produce a spatially distributed output. The derived values represent the spatial extent of the HRU. We use the spatial distribution of the following hydrologic variables (Table 1): The water balance components water yield and evapotranspiration as well as the hydrologic variables that are used in SWAT to calculate water yield (sum of surface runoff, interflow and baseflow, subtracting transmission losses) and evapotranspiration (potential evapotranspiration). Additionally, other modeled variables that influence water availability are included such as soil water content, groundwater recharge, water in the shallow aquifer returning to the root zone,

HRU level, resampled to 30 m HRU level, resampled to 30 m

irrigation amount, and water stress days (Table 1). For each variable we calculate the mean of the model output. Two model runs are used: i) a model run with the land use of 1989/90 from 1989 to 1999 to derive the explanatory variables for land use modeling, and ii) a model run with the land use of 2000/01 from 2000 to 2008 to update the explanatory variables for a temporal validation of the land use change models.

2.4. Logistic regression approach We analyze the land-based land use change classes that are present in all land use maps, i.e., forest, shrubland, grassland, urban, mixed cropland, sugarcane, and rice. Logistic regression models are set up based on the land use changes between 1989/90 and 2000/ 01, employing the different sets of explanatory variables. For each class C a binary coding of the study area is applied, where 1 indicates a change from any other class to class C and 0 indicates no change to class C. As we are interested in land use changes, areas with persisting land use C are excluded. The analysis focuses on land-based change, so that water areas are masked and are not considered in the logistic regression approach. For each land use class C the probability Pi for each pixel i for a change to this class is described as a function of the explanatory variables Xn;i , as follows:

 Log

 Pi ¼ b0 þ b1 X1;i þ b2 X2;i þ b3 X3;i þ b4 X4;i þ b5 X5;i 1
Pi

(1)

where bn are the regression coefficients for the variables Xn;i . Aiming at a comparable approach that yields the main explanatory variables we restrict the number of explanatory variables n to five. This number is regarded as a compromise, as it allows on the one

248

P.D. Wagner, B. Waske / Environmental Modelling & Software 83 (2016) 245e254

hand for a combination of variables from different datasets. On the other hand preliminary results have shown that often a larger number of variables has not led to a pronounced improvement of the logistic regression models. This restriction of variables is also supported by other studies that yield reasonable results by employing a maximum number of five explanatory variables ~ ate-Valdivieso and Bosque Sendra, 2010). (Baumann et al., 2011; On All possible combinations of five explanatory variables are used to set up regression models. Non-significant variables that do not lead to an improvement of model performance are removed from these models. The best model in terms of the highest relative operating characteristic (ROC, Pontius and Schneider, 2001) is chosen. The ROC is based on a curve of the rate of true positives versus the rate of false positives for a range of threshold values to classify the probabilities into two classes. The area under this curve is the ROC statistic, which ranges from 0.5 (random separation) to 1 (perfect discrimination). This methodology results in 21 logistic regression models one for each of the seven land use classes and the three different sets of variables. To avoid spatial autocorrelation a stratified random sample of 20% of the number of changed pixels is taken from the binary change raster. Additionally, adjacent pixels of the same class are not included in this sample. The 20% sample is randomly split in two equal parts, a 50% sample for training the logistic regression model and another 50% test sample for validating the logistic regression model. The training sample is used to extract the values of the explanatory variables at the respective locations. To exclude collinear variables we assess Pearson's correlation coefficient (r) between all variable pairs. If the absolute value of r is greater than 0.7, only the variable that better explains the land use changes is kept while the other one is removed from the variable set (Baumann et al., 2011). We use a logistic regression with only one variable and evaluate the regression result with the ROC to assess which variable is more suitable to explain the class specific land use change. Moreover, we analyze the regression models that are based on both variable sets in detail. To assess the importance of the individual variables, we count how often a variable is included in the 20 best models for each class (Baumann et al., 2011). The percentage of the inclusion of a variable in the best regression models is used as an indicator for variable importance. All calculations and analyses are carried out in R (R Core Team, 2015) and with the help of the Rpackages raster (Hijmans, 2015) and ROCR (Sing et al., 2005).

as these classes are not considered in the logistic regression model. The updated modeled hydrologic variables and the other common explanatory variables are used with the class specific logistic regression models to derive a probability map for each land use change class. The results are compared to the changes between 2000/01 and 2009/10 as derived from the land use maps and are evaluated with the help of the ROC statistic. 3. Results 3.1. Comparison of variable sets Fig. 2 shows the spatial validation values of the best logistic regression models for each dataset and land use change class. The assessment indicates that the common variable set is more suitable for mixed cropland and rice, whereas the modeled hydrologic variables are more suitable for semi-natural areas (forest, shrubland, grassland) and sugarcane. Due to the hydrologic properties of areas changed to urban areas, the hydrologic variable set results in a slightly higher ROC for urban areas as compared to the common variables. In summary, the information content of the hydrologic variable set is at least on the same level as the information content of the common variable set. For all land use change classes the use of both variable sets results in an increase of the ROC by 0.01e0.11 as compared to the use of the common variable set (Fig. 2). The probability maps that result from the different variable sets are shown in Fig. 3. The spatial patterns of the explanatory variables are visible in the probability maps that derive from the set of common variables (first column in Fig. 3): In the eastern part of the catchment the coarse population density grid shows up clearly in the predictions for changes in shrubland, urban area,

2.5. Validation We evaluate the logistic regression models in two steps, from now on referred to as i) spatial and ii) temporal validation (Pontius et al., 2004). First, we use the test sample for validating changes between 1989/90 and 2000/01. Therefore, the performance of the logistic regression models which are based on changes between 1989/90 and 2000/01 are spatially evaluated. This validation step is applied to the derived models based on all three variable sets. Second, we assess the temporal performance of the models that used both variable sets. Therefore, we apply the logistic regression models to model changes between 2000/01 and 2009/10. The hydrologic variables are updated by extracting their average values from a SWAT model run with the land use of 2000/01 from 2000 to 2008. Furthermore, the variable population density is updated by using data for the year 2000 instead of 1990. A binary change raster for each land use class is generated based on the changes between 2000/01 and 2009/10. For each land use class a new stratified random sample without adjacent pixels is taken that consists of 10% of the number of changed pixels. Water areas as well as changes to the new 2009/10 land use class “bare soil” are masked,

Fig. 2. Comparison of the relative operating characteristic (ROC) of the best logistic regression models for all land use change classes using common variables, hydrologic variables, and all variables as derived from the spatial validation.

P.D. Wagner, B. Waske / Environmental Modelling & Software 83 (2016) 245e254

249

Fig. 3. Probability maps for land use changes for the best logistic regression models using common variables (first column), hydrologic variables (second column) and all variables (third column). Probabilities for a change to each land use class are shown (rows). The relative operating characteristic statistic for the spatial validation in 2000 is provided for each probability map. Water areas are masked and depicted in white.

250

P.D. Wagner, B. Waske / Environmental Modelling & Software 83 (2016) 245e254

and sugarcane; the soil type is part of the model for changes in forest, shrubland, mixed cropland, and rice and the outline of one of the two soil types can be seen in the north east; elevation is part of all models and leads to the clear topographic pattern in all probability maps. The probability maps that have been calculated based on the hydrologic variables (column 2 in Fig. 3) show more detailed patterns as all variables are available on the HRU level. Some variables such as water yield (WYLD) and potential evapotranspiration (PET) depend to a large degree on the sub-basin specific climate data, so that the sub-basin structure shows up in the predicted probability maps of e.g., shrubland (PET), grassland (PET), mixed cropland (PET and WYLD), and sugarcane (WYLD). Irrigation patterns near the rivers are visible in the probability map for changes to sugarcane (indicating a higher probability for higher irrigation amounts) and grassland (indicating a lower probability for higher irrigation amounts). The main defining variable for changes to forest is the pattern of interflow, which is low in the flat north eastern part of the catchment and indicates low probabilities there. The use of both variable sets is reflected in the mixture of patterns in the probability maps (column 3, Fig. 3). Mostly, the patterns that are defining very high or very low probabilities in the single dataset assessments are also used in the logistic regression models that use both variable sets e.g., the interflow pattern for changes to forest, the population density pattern for changes to shrubland, and the irrigation pattern for changes to sugarcane. When compared to the spatial distribution of the classes in the 2009/10 land use map (Fig. 1) the logistic regression model that uses both variable sets results in more reasonable patterns as compared to the use of only one variable set. For example, probabilities for changes to forest are low in the north eastern part of the catchment and forest is mainly located in the western part; high probabilities for agricultural uses are found in the valleys near the water sources and are more accurately defined when including hydrologic variables. However, the highest probabilities for urban growth in the eastern part of the catchment are present in all three logistic regression models. The analysis of the probability maps is therefore in agreement with the assessment of the ROC values for the different variable sets (Fig. 2).

3.2. Model evaluation and variable impact In the following, we analyze the models that use both variable sets in more detail. All included explanatory variables are significant (p < 0.01, Table 2). The evaluation of the models shows that the ROC for calibration (ROC_cal) is on the same level as the ROC for the spatial validation that derives from independent samples (ROC_val_2000/01; Table 2). Hence the logistic regression model is robust in space. The ROC values indicate that all models can be used to distinguish accurately between two classes and that the prediction is better than random (ROC > 0.5). However the ROC is relatively low for shrubland and grassland (0.67, 0.66), but indicates a reasonable performance for all other land use change classes with values between 0.79 and 0.92, which are within a range of reliable precision (0.7e0.9, Wu et al., 2009). Applying the model to the changes between 2000/01 and 2009/10 results in a decrease of the ROC by about 0.1 for most classes (Table 2). Only changes to urban and to sugarcane show the same level of accuracy for this temporal validation. Nevertheless, ROC values between 0.75 and 0.87 indicate that the logistic regression models provide suitable predictions of spatial land use change for most classes. Only the changes to shrubland and grassland are harder to predict which might result from a higher percentage of pseudo change between these classes (see section 2.2). Table 2 shows that all logistic regression models use explanatory variables from both variable sets, underlining the difference as well as the complementarity of the two variable sets for land use change analysis. Most variables are used at least once for one of the logistic regression models. From the set of explanatory variables (Table 1) only slope, aspect, water in the shallow aquifer returning to the root zone, and water stress are not included. As a threshold for agricultural areas has been used in the classification approach (Wagner et al., 2013), the variable slope has not been used for the agricultural classes. The two hydrologic variables are not included due to collinearity (see section 2.4) with other hydrologic variables such as groundwater recharge and baseflow (positive correlation with water in the shallow aquifer returning to the root zone, r ¼ 0.89 and r ¼ 0.88, respectively) and evapotranspiration (negative correlation with water stress, r ¼
0.75) that better explain the patterns of land use change.

Table 2 Best logistic regression models for each land use change class: Odds ratios for the explanatory variables and relative operating characteristic statistic (ROC) for calibration, spatial validation, and temporal validation. Change to forest Population density Distance to roads Distance to rivers Elevation Soil type Water yield Surface runoff Interflow Baseflow Transmission losses Evapotranspiration Potential evapotranspiration Soil water content Groundwater recharge Irrigation amount

Change to shrubland

Change to grassland

Change to urban

0.9998

0.9999 1.0001

0.9997

0.9998 1.0055

0.9831 2.1534

5.3331

1.6580 1.4935 8.5  10
18

1.0035 0.7333

0.9966

5.3  10
6

8.4  10
6 0.9766

Change to mixed crops

Change to sugarcane

Change to rice

0.9846

0.9814 0.2341

0.9847 0.0529

0.9956

0.9791

3.5  102 1.0339 0.9205

2.0  104

1.0189 0.7322 0.9768

1.0486 0.8237

1.0604

Significance of all included variables

p < 0.001

p < 0.001

p < 0.001

p < 0.001

p < 0.001

p < 0.01

p < 0.001

ROC_cal ROC_val_2000/01 ROC_val_2009/10

0.90 0.90 0.77

0.67 0.67 0.57

0.66 0.66 0.56

0.88 0.88 0.87

0.80 0.79 0.78

0.92 0.92 0.84

0.84 0.83 0.75

P.D. Wagner, B. Waske / Environmental Modelling & Software 83 (2016) 245e254

251

benefit of using both datasets. Although more hydrologic variables are available, these are often collinear and therefore not all variables can be used at the same time. For the land use changes in the study area the explanatory variables distance to roads, distance to rivers, elevation, aspect, soil type, and transmission losses are very important as they show at least some importance for all different land use changes (Table 3). Water yield, baseflow, surface runoff, evapotranspiration, and groundwater recharge seem particularly important for changes to sugarcane, forest, grassland, mixed cropland, and urban area, respectively (100% inclusion). Moreover, elevation is very important for changes to urban, mixed cropland, and rice, and transmission losses are important for changes to shrubland and grassland (100% inclusion, Table 3). Only one variable, the water in the shallow aquifer returning to the root zone, is never included in one of the best 20 models for land use changes as the information content is already present in the patterns of other collinear hydrologic variables such as groundwater recharge and baseflow (r ¼ 0.89 and r ¼ 0.88, respectively) that have been given preference to during variable selection. Similarly, the collinearity of slope and interflow (r ¼ 0.77) results in the choice of either the one or the other variable (i.e., slope for changes to grassland and interflow for changes to forest and shrubland, Table 3).

The odds ratio for the n-th explanatory variable in a logistic regression model is defined as expðbn Þ and is given in Table 2. For metric variables, the odds ratios can be interpreted as follows. Odds ratios below 1 indicate that the probability of land use change to a specific class will decrease with an increase in the explanatory variable. Values above 1 indicate that an increase in the explanatory variable will lead to an increase of the probability of land use change. The magnitude of the odds ratio is scaled to the magnitude of the explanatory variable, i.e. small values of the variable (e.g., transmission losses) result in very large or very small odds ratios. Our results show that with increasing elevation, change to forest becomes more probable (0.55% per meter), whereas changes to urban and all agricultural classes are less probable (about 2% per meter, Table 2). With an increase of the mean monthly irrigation amount by 1 mm the probability for a change to grassland decreases by 2.32% (odds ratio ¼ 0.9768) whereas the probability for a change to sugarcane increases by 6.04% (odds ratio ¼ 1.0604). The odds ratios are in agreement with general assumptions e.g., with increasing population density the probability for a change to shrubland (0.9998) or grassland (0.9999) decreases, and with an increasing distance to roads changes to grassland (1.0001) are more likely whereas the probability for changes to urban areas (0.9997) decreases. Also indirect influences become obvious that result from the hydrologic properties of changed areas, e.g., an increase in evapotranspiration increases the probability for a change to mixed crops (1.0339) and decreases the probability for a change to urban area (0.9766).

3.4. Evaluation of the derived probability maps for land use change modeling The probability map for each land use change class using the best logistic regression model and variables from both variable sets is shown in column three of Fig. 3. An advantage of the derived probability maps is that they can be used to assess competition between different land use classes. A simple model of three competing land uses can be illustrated using an R-G-B color composite. Fig. 4A shows the probability maps of agricultural land uses as an R-G-B color composite image, where rice is depicted as red, sugarcane as green, and mixed cropland as blue. White areas have high probabilities of change to all agricultural classes and are found close to the rivers. With an increasing distance from the rivers violet colors dominate indicating suitability for rice and mixed cropland. At the largest distance to the river and particularly in the southern part of the catchment the high probability for mixed

3.3. Variable importance In addition to the analysis of the best regression models we also evaluate the 20 best models for each land use change. To assess the importance of the variables we count how often a variable is included in one of the 20 best models. Mostly, those variables that have been chosen for the best model (Table 2) also have the largest importance with regard to the 20 best models (bold percentages in Table 3). In terms of the ROC statistic that is used to select the best model the 20 best models only differ by 0.01. Therefore, the 20 different variable combinations yield similar accuracies. Overall, 44% of the variables have been taken from the common variable set and 56% from the hydrologic variable set. This underlines the

Table 3 Percentage of inclusion of each explanatory variable in the best 20 logistic regression models for each change class. The best variable combination is depicted in bold. Change to forest Population density Distance to roads Distance to rivers Elevation Slope Aspect Soil type Water yield Surface runoff Interflow Baseflow Transmission losses Evapotranspiration Potential evapotranspiration Soil water content Groundwater recharge Irrigation amount Water stress days Min ROC_cal Max ROC_cal

Change to shrubland

30% 55% 65%

70% 5% 5% 45%

20% 25%

20% 50%

80% 100% 80%

Change to grassland 50% 35% 5% 50% 15% 20% 5%

60% 65%

100%

100%

100%

55%

15%

Change to urban

Change to mixed crops

15% 60% 30% 100%

15% 45% 15% 100%

30% 25% 15% 85%

5% 40% 15% 100%

10% 60%

15% 20%

5% 35%

30% 30%

60%

100%

65% 100% 50%

35%

75% 85%

50%

35% 45%

20% 65% 30% 100%

85% 35% 0.89 0.90

Change to sugarcane

Change to rice

80%

25% 0.66 0.67

0.65 0.66

0.88 0.88

0.80 0.80

0.92 0.92

0.83 0.84

252

P.D. Wagner, B. Waske / Environmental Modelling & Software 83 (2016) 245e254

Fig. 4. R-G-B-Composites of the probability maps indicating competing land use changes. Water areas are masked and depicted in blue. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

cropland is shown by a larger percentage of blue in the darker violet color. In summary, this assessment represents the reasonable sequence of sugarcane, rice, and mixed cropland that are ordered with regard to their irrigation water demand and proximity to the water source. Similarly, the competition of urban area (R), forest (G), and mixed cropland (B) is illustrated in Fig. 4B. The pattern clearly shows that a change to forest would be dominating at higher elevations in the west. The competition of urban area and cropland becomes clear from the pink to violet colors of the flat areas in the eastern part of the catchment. Unlike the competition of agricultural classes, no areas exist that are equally suitable for all three classes. Moreover, no yellow or cyan colors are depicted, indicating that forests are not competing with urban or mixed cropland areas. Even though shrubland and grassland are predicted with the lowest accuracies a competition map of the three semi-natural classes (grassland (R), forest (G), shrubland (B)) provides interesting insights (Fig. 4C). Bright areas that indicate competition of these classes are mainly located at the ridges of the Western Ghats. Forest and shrubland (green and blue colors) dominate in the western part, whereas grassland and shrubland (red and blue colors) dominate in the eastern part of the catchment and in the valley bottoms. The black color in the east indicates that changes to semi-natural classes are highly unlikely in and around Pune. 4. Discussion While the importance of land use change for hydrologic studies has been emphasized (Stonestrom et al., 2009), our results underline the importance of hydrologic patterns for studies of land use change. Hence, our findings contribute to the previously highlighted research priority on water issues in regional land use and land cover change (Lambin et al., 1999). The dependence of land use on water availability is evident both in theory and practice (e.g., Guerschman et al., 2003) and the suitability of single hydrologic variables has been shown in other studies (e.g., Kim et al., 2014; Rutherford et al., 2007; Sohl et al., 2007). However, our study is the first that evaluates a set of hydrologic variables against a set of biophysical and socio-economic variables. Our results indicate that the additional hydrologic variables lead to an improvement in the explanation of land use change patterns, which is supported by the improvement of the ROC by up to 0.11 as well as the visual interpretation of Fig. 3. Moreover, their importance is clearly shown as 56% of the variables in the 20 best regression models for each land use change come from the hydrologic dataset. This clearly underlines that hydrologic and biophysical/socio-economic variables provide different, complementary information on land use change. In the absence of measured data, we used the output from a hydrologic model. However, also variables derived from satellite imagery or measured data may be suitable (Sohl et al., 2007;

Prishchepov et al., 2013). While twelve variables have been tested in this study, their patterns have often been similar and variables have been excluded due to collinearity (e.g., positive correlation of groundwater recharge and water in the shallow aquifer returning to the root zone, r ¼ 0.89). In the best logistic regression models, between two and four of five variables came from the hydrologic dataset. Therefore, also fewer additional hydrologic variables may result in an improvement of model performance, particularly in situations of limited data availability. The quantitative (Table 2) as well as the visual evaluation of the probability maps (Fig. 3) indicates that the derived models can contribute towards explaining land use changes in the study area. The patterns are consistent with the land use map for 2009/10 (Fig. 1). Moreover, all ROC values derived by the spatial (ROC_val_2000/01: 0.66e0.92) and temporal validation (ROC_val_2009/10: 0.56e0.87) are greater than 0.55 indicating that the predictions are at least better than random. The range of ROC values is on a similar level as compared to other studies of land use change that used logistic regression models (e.g., 0.65 for deforestation, Pontius and Schneider, 2001; 0.7 for land abandonment, Prishchepov et al., 2013; 0.75e0.81 for new residential, industrial and recreational areas, Verburg et al., 2004). It should be noted that we restricted the models to five explanatory variables and that slightly better results can be achieved if all significant explanatory variables are included. Therefore, the achieved results compare well with other studies. The possibility to update (hydrologic) variables and the temporal validation indicate the potential of the simple binary modeling approach to contribute towards an accurate prediction of the spatial distribution of possible future changes. Even though the ROC values for the logistic regression models that use only hydrologic variables are comparatively high (Fig. 2), the processes behind the influence of these variables must carefully be reviewed, before applying the variables in another environment. For urban expansion hydrologic properties are not the defining drivers of land use change. Our results show that the probability for urban area increases with decreasing groundwater recharge and decreasing evapotranspiration (Table 2). This is mainly due to the hydrologic properties of areas converted to urban areas, which are primarily situated around Pune in the eastern part of the catchment. Groundwater recharge is lower in these parts as it depends on the amount of rainfall which declines from west to east, and evapotranspiration is lower there due to a different soil type. Moreover, grassland areas that have comparatively low groundwater recharge and evapotranspiration values are more easily converted to urban areas than e.g., agricultural areas. While these variables can contribute towards explaining the location of changes to urban areas through these indirect relationships, hydrologic variables should primarily be used to explain changes in land use classes that are dependent on water availability, like semi-natural and agricultural areas.

P.D. Wagner, B. Waske / Environmental Modelling & Software 83 (2016) 245e254

Our assessment has shown that competing land use changes yield reasonable land use patterns (Fig. 4). An ordered sequence of agricultural land uses with regard to their irrigation water demand and proximity to the water source seems reasonable (Fig. 4A). The global increase of agricultural and urban areas (Foley et al., 2005) results in a competition for land (Müller and Munroe, 2014). Previous research in the region has shown that cropland and urban area compete for similar areas (Samal and Gedam, 2015), resulting in a relocation of cropland due to urban growth in the study area (Wagner et al., 2013). This competition is well reflected in the R-G-B composite in Fig. 4B. It is also shown that forest areas do not compete with agricultural or urban areas, which may result from the fact that parts of the forest areas are reserved forests. Also the land use maps indicate little forest loss to these classes between 1990 and 2010 (Wagner et al., 2013). Newer developments such like the construction of a new city (Lavasa) in the forested Western Ghats area (Wagner et al., 2016) have started after the training period and can therefore not be represented by the probability maps. Forest, shrubland, and grassland are a continuum and therefore the competition map shows large areas where high probabilities for at least two of these classes overlap. It becomes very clear, that a change to one of these classes in and near Pune is highly unlikely (Fig. 4C). In summary, the R-G-B composite maps show that the probability maps are suitable to model the competition of different land use changes. Hence, the probability maps can be used by land use change models (e.g., CLUE, Verburg and Overmars, 2009) that weigh different probability maps to derive the spatial land use distribution for future scenarios of aggregated land use changes. This study shows that it is useful to integrate hydrologic modeling results in land use change models. In a second step land use information in hydrologic models can be updated using a land use change model. This bi-directional coupling will allow for the integration of (hydrologic) feedbacks, which is one of the main challenges in land use modeling (Verburg et al., 2006; Veldkamp and Lambin, 2001). Particularly with regard to climate change induced droughts (Verburg et al., 2011), this coupling through hydrologic variables will mean a promising novel development. Moreover, such integrated environmental modeling approaches are more and more in need to provide a means to analyze and solve increasingly complex real-world problems (Kelly et al., 2013; Laniak et al., 2013). 5. Conclusions Spatially distributed hydrologic variables can contribute towards explaining land use changes. For all land use change classes the model accuracy increases if hydrologic variables are used together with commonly used biophysical and socio-economic explanatory variables for land use change. In all of the best logistic regression models, at least one variable from each dataset is used and in the 20 best models for all land use change classes, 44% come from the common set of biophysical and socio-economic variables while 56% come from the hydrologic variable set. This underlines the difference as well as the complementarity of the two variable sets. It is therefore strongly recommended to include hydrologic variables in studies of land use change. The derived logistic regression models are useful for the prediction of the spatial distribution of possible future land use changes. Probability maps and ROC values between 0.75 and 0.87 underline this for the most classes. Only changes to grassland and shrubland are harder to predict. Our assessment shows that competition between the different classes can be reasonably modeled with the derived probability maps. This is also promising with regard to the possibility of coupling land use and hydrologic

253

models through hydrologic model outputs and predicted land use changes, yielding an integrated modeling system that can take feedbacks into account. Software and data availability We use the freely available open source hydrologic model SWAT (Arnold et al., 1998, available from http://swat.tamu.edu/) to produce spatially distributed hydrologic variables for the studied catchment (Wagner et al., 2013). Data sources for all utilized explanatory variables for land use change are provided in Table 1. All calculations and analyses are carried out in R (R Core Team, 2015) and with the help of the R-packages raster (Hijmans, 2015) and ROCR (Sing et al., 2005), which are available from https://www. r-project.org/. R and SWAT run on Microsoft Windows and linux computers with no special hardware requirements. Please contact the corresponding author for any further information at paul. [email protected]. Acknowledgements We are thankful to Erach Bharucha and Shamita Kumar from the Institute of Environment Education & Research at Bharati Vidyapeeth University Pune, who supported our research in the study area in many ways and from the very beginning. The authors thank the editor and two anonymous reviewers for their constructive comments. References Arnold, J.G., Srinivasan, R., Muttiah, R.S., Williams, J.R., 1998. Large area hydrologic modeling and assessment d part 1: model development. J. Am. Water Resour. Assoc. 34, 73e89. Baumann, M., Kuemmerle, T., Elbakidze, M., Ozdogan, M., Radeloff, V.C., Keuler, N.S., Prishchepov, A.V., Kruhlov, I., Hostert, P., 2011. Patterns and drivers of postsocialist farmland abandonment in Western Ukraine. Land Use Policy 28 (3), 552e562. http://dx.doi.org/10.1016/j.landusepol.2010.11.003. Bastiaanssen, W.G.M., Pelgrum, H., Wang, J., Ma, Y., Moreno, J.F., Roerink, G.J., van der Wal, T., 1998. A remote sensing surface energy balance algorithm for land (SEBAL). Part 2: Validation. J. Hydrol. 212e213, 213e229. http://dx.doi.org/ 10.1016/S0022-1694(98)00254-6. vre, B., Wyseure, G., 2006. Spatial and Buytaert, W., Celleri, R., Willems, P., De Bie temporal rainfall variability in mountainous areas: a case study from the south Ecuadorian Andes. J. Hydrol. 329, 413e421. CIESIN (Center for International Earth Science Information Network), Columbia University, and Centro Internacional de Agricultura Tropical (CIAT), 2005. Gridded Population of the World. Version 3 (GPWv3): Population Density Grid. NASA Socioeconomic Data and Applications Center (SEDAC), Palisades, NY. http://dx.doi.org/10.7927/H4XK8CG2 (accessed 11.08.15.). DeFries, R., Eshleman, K.N., 2004. Land-use change and hydrologic processes: a major focus for the future. Hydrol. Process 18, 2183e2186. Eckert, S., Kellenberger, T., Itten, K., 2005. Accuracy assessment of automatically derived digital elevation models from aster data in mountainous terrain. Int. J. Remote Sens. 26, 1943e1957. Foley, J.A., DeFries, R., Asner, G.P., Barford, C., Bonan, G., Carpenter, S.R., Chapin, F.S., Coe, M.T., Daily, G.C., Gibbs, H.K., Helkowski, J.H., Holloway, T., Howard, E.A., Kucharik, C.J., Monfreda, C., Patz, J.A., Prentice, I.C., Ramankutty, N., Snyder, P.K., 2005. Global consequences of land use. Science 309 (5734), 570e574. Food and Agriculture Organization of the United Nations (FAO), 2003. Digital Soil Map of the World and Derived Soil Properties. FAO, Rome. Gadgil, A., 2002. Rainfall characteristics of Maharashtra. In: Diddee, J., Jog, S.R., Kale, V.S., Datye, V.S. (Eds.), Geography of Maharashtra. Rawat Publications, Jaipur, pp. 89e102. Guerschman, J.P., Paruelo, J.M., Burke, I.C., 2003. Land use impacts on the normalized difference vegetation index in temperate Argentina. Ecol. Appl. 13, 616e628 http://dx.doi.org/10.1890/1051-0761(2003)013[0616:LUIOTN] 2.0.CO;2. Gunnell, Y., 1997. Relief and climate in South Asia: the influence of the Western Ghats on the current climate pattern of peninsular India. Int. J. Climatol. 17, 1169e1182. Her, Y., Frankenberger, J., Chaubey, I., Srinivasan, R., 2015. Threshold effects in HRU definition of the soil and water assessment tool. Trans. ASABE 58 (2), 367e378. Hijmans, R.J., 2015. Raster: Geographic Data Analysis and Modeling. R package version 2.4-20. Huang, Q., Cai, Y., 2007. Simulation of land use change using GIS-based stochastic

254

P.D. Wagner, B. Waske / Environmental Modelling & Software 83 (2016) 245e254

model: the case study of Shiqian County, Southwestern China. Stoch. Environ. Res. Risk Assess. 21 (4), 419e426. Kelly (Letcher), R.A., Jakeman, A.J., Barreteau, O., Borsuk, M.E., ElSawah, S., Hamilton, S.H., Henriksen, H.J., Kuikka, S., Maier, H.R., Rizzoli, A.E., van Delden, H., Voinov, A.A., 2013. Selecting among five common modelling approaches for integrated environmental assessment and management. Environ. Model. Softw. 47, 159e181. Kim, I., Le, Q.B., Park, S.J., Tenhunen, J., Koellner, T., 2014. Driving forces in archetypical land-use changes in a mountainous watershed in east Asia. Land 3, 957e980. Koyama, C.N., Korres, W., Fiener, P., Schneider, K., 2010. Variability of surface soil moisture observed from multi-temporal C-band SAR and field data. Vadose Zone J. 9, 1014e1024. Lambin, E.F., Baulies, X., Bockstael, N., Fischer, G., Krug, T., Leemans, R., Moran, E.F., Rindfuss, R.R., Sato, Y., Skole, D., Turner II, B.L., Vogel, C., 1999. Land-use and Land-cover Change: Implementation Strategy. IGBP Report No. 48 and IHDP Report No. 10, IGBP, Stockholm, 125 pp. Lambin, E.F., Turner, B.L., Geist, H.J., Agbola, S.B., Angelsen, A., Bruce, J.W., Coomes, O.T., Dirzo, R., Fischer, G., Folke, C., George, P.S., Homewood, K., Imbernon, J., Leemans, R., Li, X., Moran, E.F., Mortimore, M., Ramakrishnan, P.S., Richards, J.F., Skånes, H., Steffen, W., Stone, G.D., Svedin, U., Veldkamp, T.A., Vogel, C., Xu, J., 2001. The causes of land-use and land-cover change: moving beyond the myths. Glob. Environ. Change 11 (4), 261e269. http://dx.doi.org/ 10.1016/S0959-3780(01)00007-3. Laniak, G.F., Olchin, G., Goodall, J., Voinov, A., Hill, M., Glynn, P., Whelan, G., Geller, G., Quinn, N., Blind, M., Peckham, S., Reaney, S., Gaber, N., Kennedy, R., Hughes, A., 2013. Integrated environmental modeling: a vision and roadmap for the future. Environ. Model. Softw. 39, 3e23.  pez, A., 2004. Modelling deforestation using Mas, J.F., Puig, H., Palacio, J.L., Sosa-Lo GIS and artificial neural networks. Environ. Model. Softw. 19 (5), 461e471. http://dx.doi.org/10.1016/S1364-8152(03)00161-0. Müller, D., Munroe, D.K., 2014. Current and future challenges in land-use science. J. Land Use Sci. 9 (2), 133e142. http://dx.doi.org/10.1080/1747423X.2014.883731. ~ ate-Valdivieso, F., Bosque Sendra, J., 2010. Application of GIS and remote sensing On techniques in generation of land use scenarios for hydrological modeling. J. Hydrol. 395 (3e4), 256e263. http://dx.doi.org/10.1016/j.jhydrol.2010.10.033. OpenStreetMap, 2015. Roads Polygon Shapefile India. Available at: http://download. geofabrik.de/asia/india.html (accessed 08.09.15.). Pontius Jr., R.G., Schneider, L.C., 2001. Land-cover change model validation by an ROC method for the Ipswich watershed, Massachusetts, USA. Agric. Ecosyst. Environ. 85, 239e248. Pontius Jr., R.G., Cornell, J.D., Hall, C.A.S., 2001. Modeling the spatial pattern of landuse change with GEOMOD2: application and validation for Costa Rica. Agric. Ecosyst. Environ. 85, 191e203. Pontius Jr., R.G., Huffaker, D., Denman, K., 2004. Useful techniques of validation for spatially explicit land-change models. Ecol. Model. 179 (4), 445e461. http:// dx.doi.org/10.1016/j.ecolmodel.2004.05.010. Prishchepov, A.V., Müller, D., Dubinin, M., Baumann, M., Radeloff, V.C., 2013. Determinants of agricultural land abandonment in post-Soviet European Russia. Land Use Policy 30 (1), 873e884. R Core Team, 2015. R: a Language and Environment for Statistical Computing. R Foundation for Statistical Computing, Vienna. Refsgaard, J.C., Storm, B., 1995. MIKE SHE. In: Singh, V.P. (Ed.), Computer Models of Watershed Hydrology. Water Resources Publications, Highlands Ranch, Colorado, pp. 809e846. Rutherford, G.N., Guisan, A., Zimmermann, N.E., 2007. Evaluating sampling strategies and logistic regression methods for modelling complex land cover changes. J. Appl. Ecol. 44, 414e424. http://dx.doi.org/10.1111/j.1365-2664.2007.01281.x. Samal, D.R., Gedam, S.S., 2015. Monitoring land use changes associated with urbanization: an object based image analysis approach. Eur. J. Remote Sens. 48, 85e99.

Schneider, L.C., Pontius Jr., R.G., 2001. Modeling land-use change in the Ipswich watershed, Massachusetts, USA. Agric. Ecosyst. Environ. 85 (1e3), 83e94. http://dx.doi.org/10.1016/S0167-8809(01)00189-X. Sing, T., Sander, O., Beerenwinkel, N., Lengauer, T., 2005. ROCR: visualizing classifier performance in R. Bioinformatics 21 (20), 3940e3941. Sohl, T.L., Sayler, K.L., Drummond, M.A., Loveland, T.R., 2007. The FORE-SCE model: a practical approach for projecting land cover change using scenario-based modeling. J. Land Use Sci. 2 (2), 103e126. http://dx.doi.org/10.1080/ 17474230701218202. Song, C., Woodcock, C.E., Li, X., 2002. The spectral/temporal manifestation of forest succession in optical imagery: the potential of multitemporal imagery. Remote Sens. Environ. 82, 285e302. Stefanski, J., Kuemmerle, T., Chaskovskyy, O., Griffiths, P., Havryluk, V., Knorn, J., Korol, N., Sieber, A., Waske, B., 2014. Mapping land management regimes in western Ukraine using optical and SAR data. Remote Sens. 6 (6), 5279e5305. http://dx.doi.org/10.3390/Rs6065279. Stonestrom, D.A., Scanlon, B.R., Zhang, L., 2009. Introduction to special section on impacts of land use change on water resources. Water Resour. Res. 45 http:// dx.doi.org/10.1029/2009WR007937. W00A00. Turner II, B.L., Lambin, E.F., Reenberg, A., 2007. Land change science special feature: the emergence of land change science for global environmental change and sustainability. Proc. Natl. Acad. Sci. U. S. A. 104, 20666e20671. Veldkamp, A., Lambin, E.F., 2001. Predicting land-use change. Agric. Ecosyst. Environ. 85, 1e6. Verburg, P.H., Kok, K., Pontius Jr., R.G., Veldkamp, A., 2006. Modeling land-use and land-cover change. In: Lambin, E.F., Geist, H.J. (Eds.), Land-use and Land-cover Change e Local Processes and Global Impacts. Springer, Berlin, pp. 117e135. Verburg, P.H., Neumann, K., Nol, L., 2011. Challenges in using land use and land cover data for global change studies. Glob. Change Biol. 17, 974e989. http:// dx.doi.org/10.1111/j.1365-2486.2010.02307.x. Verburg, P.H., Overmars, K.P., 2009. Combining top-down and bottom-up dynamics in land use modeling: exploring the future of abandoned farmlands in Europe with the Dyna-CLUE model. Landsc. Ecol. 24 (9), 1167e1181. http://dx.doi.org/ 10.1007/s10980-009-9355-7. Verburg, P.H., Ritsema van Eck, J., de Nijs, T.C., Dijst, M.J., Schot, P., 2004. Determinants of land-use change patterns in the Netherlands. Environ. Plan. B Plan. Des. 31, 125e150. http://dx.doi.org/10.1068/b307. Verburg, P.H., Soepboer, W., Veldkamp, A., Limpiada, R., Espaldon, V., Mastura, S.S.A., 2002. Modeling the spatial dynamics of regional land use: the CLUE-S model. Environ. Manag. 30, 391e405. Wagner, P.D., Fiener, P., Wilken, F., Kumar, S., Schneider, K., 2012. Comparison and evaluation of spatial interpolation schemes for daily rainfall in data scarce regions. J. Hydrol. 464e465, 388e400. http://dx.doi.org/10.1016/ j.jhydrol.2012.07.026. Wagner, P.D., Kumar, S., Fiener, P., Schneider, K., 2011. Hydrological modeling with SWAT in a monsoon-driven environment: experience from the Western Ghats, India. Trans. ASABE 54 (5), 1783e1790. http://dx.doi.org/10.13031/2013.39846. Wagner, P.D., Kumar, S., Schneider, K., 2013. An assessment of land use change impacts on the water resources of the Mula and Mutha Rivers catchment upstream of Pune, India. Hydrol. Earth Syst. Sci. 17, 2233e2246. http://dx.doi.org/ 10.5194/hess-17-2233-2013. Wagner, P.D., Bhallamudi, S.M., Narasimhan, B., Kantakumar, L.N., Sudheer, K.P., Kumar, S., Schneider, K., Fiener, P., 2016. Dynamic integration of land use changes in a hydrologic assessment of a rapidly developing Indian catchment. Sci. Total Environ. 539, 153e164. http://dx.doi.org/10.1016/ j.scitotenv.2015.08.148. Wu, X., Hu, Y., He, H.S., Bu, R., Onsted, J., Xi, F., 2009. Performance evaluation of the SLEUTH model in the Shenyang metropolitan area of northeastern China. Environ. Model. Assess. 14 (2), 221e230. Xu, L., Li, Z., Song, H., Yin, H., 2013. Land-use planning for urban Sprawl based on the CLUE-S model: a case study of Guangzhou, China. Entropy 15, 3490e3506.