Introduction
The roe deer as a meso-mammals of Palearctic distribution can be classified into two distinct species: the European roe deer (Capreolus capreolus) and the Siberian roe deer (Capreolus pygargus). The habitats of Capreolus pygargus is continental Asia and some regions of Eastern Europe, from the Khoper River and Don River bend to the Ural Mountains and across southern Siberia. It is distributed throughout northern Mongolia and east to the coastlines of the East Sea, and the Yellow Sea, including the Korean Peninsula especially Jeju Island (Koh & Randi, 2001; Lee et al., 2015; 2016; Sokolov et al., 1982). The European roe deer travel from Europe to Western Russia and the Siberian roe deer migrate from Russia to Korea (Park et al., 2014). The differences between these species are their body sizes and morphometric characteristics (Danilkin & Hewison, 1996). Abundance of roe deer species as an inevitable food source is one of the main factor to control many raptors population (Jiang et al., 2015). Although International Union for the Conservation of Nature declared that Siberian roe deer is in the Red List as the least concern specie (Lovari et al., 2016), their population have considerably declined owing to trapping and overhunting. Therefore, habitat potential map can play a prominent role in developing conservation plans and maintaining biodiversity of roe deer as a game species.
In recent years, because of rapid increasing of the quantity and quality of geospatial data, attempts have been made to generate more accurate habitat potential maps for a variety of species. To achieve this goal, geographic information system (GIS) as a powerful instrument can be applied to assess the spatial correlations between species occurrences and spatial variables. Based on the literature review, various types of GIS-based models have been employed to develop habitat potential maps for different plant and animal species. These methods can be categorized into four groups, namely bivariate, multivariate, multi-criteria decision-making (MCDM), and artificial intelligence (AI) algorithms (Kadirhodjaev et al., 2020). The first two methods are statistical modeling which can be implemented easily and the outcomes can be interpreted simply. In spite of the uncomplicated performance of these data-driven methods, the size and precision of input data can affect the predictive accuracy of employed models. Additionally, selection effective parameters among multicollinearity variables can cause challenges in statistical models (Farrell et al., 2019). The most representative of bivariate and multivariate approaches are the frequency ratio (FR) (Choi et al., 2011a), weights-of-evidence model (Choi et al., 2011b), logistic regression (Imam & Kushwaha, 2013; Pereira & Itami, 1991), and evidential belief function (Lee et al., 2019).
Multi-criteria decision-making (MCDM) algorithms are a type of knowledge-driven methods applied to assign weights to conditioning factors based on evidence of varying quality, guideline, and expert opinions (Al-Abadi et al., 2016). Analytical hierarchy process is the most widely used MCDM method for habitat mapping (Aini et al., 2015; Imam & Tesfamichael, 2013; Mesfin & Berhan, 2016; Sanare et al., 2015).
Nowadays, machine learning methods received growing attention of researchers in a wide range of scientific fields due to their predictive ability to identify patterns in historical data and find the nonlinear connection between variables. The most prevalent machine learning approaches implemented for species geographic distributions modeling includes artificial neural network (Lee et al., 2013; 2017), random forest (Ng et al., 2018; Robert et al., 2016), maximum entropy model (Park et al., 2018; Zhang et al., 2020a), support vector machine (Cui et al., 2020), multivariate adaptive regression spline (Leathwick et al., 2005), generalized additive models (Kosicki, 2020; Sanchez et al., 2008; Schmiing et al., 2013) and classification and regression tree (Garzón et al., 2006). In addition, Oh et al. (2019) showed the effectiveness of support vector machine and naïve bayes techniques to produce potential ruditapes philippinarum habitat map in the Geunso Bay, South Korea. Rahimian Boogar et al. (2019) used support vector machine and maximum entropy methods to predict habitat suitability of Juniperus spp. in the Southern part of Zagros Mountains, Iran. Farrell et al. (2019) developed habitat suitability maps of wild turkeys in Mississippi using three machine learning algorithms, namely maximum entropy, random forests, and support vector machines.
With the advent of big data and in order to overcome the drawbacks of mentioned techniques, deep learning methods were introduced as a subfield of machine learning techniques powered by AI which outperformed the prediction precision. The most popular deep learning algorithms include convolutional neural network (CNN), recurrent neural network (RNN), long short-term memory (LSTM) network, and deep belief networks. As emphasized by a great number of studies, these algorithms are able to find relationship between training sample data and apply it to new sets. In facts, deep learning algorithms have promising capacity of processing unstructured data which contains multiple features. Moreover, these approaches can enhance the robustness and accuracy of analyzing.
The main objective of the present study is to investigate applicability of CNN and LSTM algorithms to produce Siberian roe deer habitat distributions map in South Korea for the first time in the field of species distribution model. Furthermore, the accuracy of predictions are compared with FR method as a bivariate statistical model. Finally, the spatial relationships between the several habitat distributions and various environmental influencing factors will be elucidated which can be used for defining long-term programs to conserve this specie.
Materials|Methods
Materials
Study area
The Korean Peninsula is situated in Northeast Asia. The country and all its islands with an approximately area of 100,302 km2 lies between latitudes 33° and 39° N, and longitudes 124° and 130° E (Fig. 1). The country reaches to Democratic People’s Republic of Korea in the north and it is surrounded by the East Sea to the east, the Korea Strait and the East China Sea to the south, and the Yellow Sea to the west. There are large coastal plains in the southern and western regions and mountains covered majority parts of South Korea (70%) are located in the eastern and northern parts. Owing to these topographical structures, only 30% of the total land in South Korea, which is located in the west and southeast, are suitable for agricultural purposes.
South Korea has more than three thousand islands which most of them are small and uninhabited located in the western and southern coasts of South Korea. Jeju, the largest Island in this country, is about 1,825 km2 and 100 km off the southern coast of South Korea. Jeju is warmer than the rest of South Korea because of its humid and subtropical climate. Winters are cool and dry while summers are hot, humid, and even sometimes rainy. The Halla-san Mountain as an extinct shield volcano is one of the three main mountains of South Korea; with elevation 1,950 m above sea level is in Jeju which makse this Island to be the highest point in South Korea. The two other high mountains are Jirisan and Seoraksan.
The longest river in South Korea is Nakdong (521 kilometers) and Han and Geum rivers comes after. These major rivers respectively flow from the north to south and the east to west, and they ended up the Yellow Sea or the Korea Strait. Their water flow fluctuates seasonally and they are broad and shallow.
South Korea climate can be classified as the humid–temperate zone with both continental and oceanic characters. The maximum and minimum annual temperature during 2010-2020 took the values between 19.1 to 17.6°C and 7.8 to 9°C, respectively. The sum of annual precipitation were in the interval 95.8 to 126.2 mm/yr, approximately 37% of which falls in the summer season, between June and Aguste (Korea meteorological administration: KMA, 2020).
Dataset preparation for spatial modeling
A full-scale investigation into the species distributions have been launched since 1997 by the Korean National Institute for Environmental Studies to monitor species populations particularly for endangered ones. The survey was carried out based on extensive field observations to identify special spreading of wildlife habitats. The observed or detected locations were documented using hand-held global positioning system. As a result, roe deer was discerned at 741 points in the study area according to experts report (Fig. 1). For the purpose of building model and validating its performance, both habitat and non-habitat locations are needed. Therefore, the same number of habitat locations (741 points) were identified as non-habitat ones. Afterwards, the yield data as the habitat and non-habitat locations were randomly divided into two groups including training and validation in the ratio of 70:30, respectively (Chen et al., 2019; Jaafari et al., 2019; Kamali Maskooni et al., 2020; Oh et al., 2019; Razavi Termeh et al., 2018). These points (training: validation) were selected utilizing ‘Create Random Points’ function in ArcGIS 10.8 software (Esri, Redlands, CA, USA). Then, 70% of the habitat and non-habitat locations were merged to build training dataset (1,038 points) which was applied for constructing model. The remaining 30% of points were amalgamated to make the testing dataset employed for model validation (444 Points) (Dodangeh et al., 2020; Panahi et al., 2020a). Finally, the attributes of each point was extracted by overlaying both training and validation datasets with habitat influential factors.
Habitat-related factors
Habitat selection is an optimization procedure which interaction between diverse environmental variables can affect habitat distributions of wild animals. Indeed, determination of contributing factors exerts undue influence on the accuracy of habitat potential mapping and there are no comprehensive guidelines for selecting appropriate affecting factors. In the current study, 15 factors were opted depending upon the literature review and availability of data. Afterwards, FR model was applied to assess the correlation between roe deer habitats and the affecting factors. Finally, According to the results of FR method, the potential maps were developed based on the integration of 10 evaluated thematic maps with FR values greater than 0.50. Results implied that the highest FR value belongs to slope height (FR=0.94), followed by drainage density (0.86), altitude (0.85), road density (0.76), topographic wetness index (TWI; 0.73), valley depth (0.69), morphological feature (0.68), normalized difference water index (NDWI; 0.65), radar intensity (0.59) and topographic position index (TPI, 0.52) (Fig. 2). The mentioned maps were prepared and analyzed using ArcGIS 10.8 software. First, a digital elevation model (DEM) with a pixel size of 30×30 m was prepared using topographical maps published by the National Geographic Information Institute. Afterwards, DEM applied to derive topographic factors, namely valley depth, slope height, TPI, TWI and morphological features.
Field observations show that the roe deer population increase at higher elevations distant from urban areas and transport infrastructure. TPI involves the analyzing of topographic landscape terrain units into the upper, middle and lower parts. By declining elevation of each cell in a DEM, the TPI values decrease. The TPI value for flat ground surface located in the foot of hills or mid sloppy region is near zero (Arya et al., 2020). TWI is another key parameter which greatly affect animal distribution and their abundance. IWI indicates the soil moisture which impacts on vegetation patterns of an area. The habitat preferences is determined fundamentally by the existence of sufficient food in a territory.
Water availability is another decisive factor in the distribution of animal and plant species. The higher availability promotes plant growth and, thus, supports more food sources for roe deer (Ng et al., 2018). Water body has low radiation and strong absorption in the range from visible to infrared wavelengths; therefore, NDWI map provides information about water body in a region. NDWI map at the spatial resolution of 30 m was produced using the green and Near Infrared bands of Sentinel-2 images. The value of NDWI for water bodies is positive while vegetative cover or soil mass gain negative values (Biswas et al., 2020). The land cover map was prepared and published by the Ministry of Environment was employed to develop drainage and road density maps. Several studies clarified the effects of barriers such as road infrastructure and water surfaces on roe deer displacements and dispersal movements (Duarte et al., 2010; Rosell et al., 1996).
Landforms can be considered as an important parameter influences population distribution and patterns of habitat selection by animals. Morphometric features, which characterize the landform of a terrain, was extracted from DEM using SAGA–GIS software ( http://saga-gis.org). The inventory map was categorized into plane, pit, ridge, channel, peak, and pass. The concentration of roe deer habitats is in areas with moderate altitudes and gentle slopes, such as river valleys and flat areas. They prefer to reduce their movements in areas with steep topography (ridges) which confirms the studies carried out by Acevedo et al. (2011), López-Martín et al. (2009), Loro et al. (2016), and Pays et al. (2012). In fact, intermediate and concave reliefs lead to decrease of habitat quality. Convex and convex/concave structure have positive impact on it in both summer and winter (Reimoser et al., 2009).
Satellite radar intensity image produced by processing Sentinel-1 (ESA operating) images of the Korean Peninsula in 2020. Each pixel in an intensity image indicates the proportion of microwave backscattering strength from that area on the ground. The radar intensity map was utilized to determine the surface roughness and the moisture level of the area. These parameters can play a vital role in the distribution of wild animal populations (Evcin et al., 2019).
Methods
Description of models
To identify roe deer potential habitats, three big data analysis algorithms including FR, CNN, and LSTM were employed in the current study. Additionally, the receiver operating characteristic (ROC) curve was applied to determine the accuracy rate of prediction. The methodologies of study were clarified in the following sections and the flowchart is shown the different processing steps of adopted approaches in present study (Fig. 3).
Frequency ratio model: The FR is a type of bivariate statistical methods with the capability of deducing the potential probabilistic correlation between an incident and each related variable. This ability leads to prophesy an event rest on an assumption that the conditions will not be altered (Lee & Talib, 2005). FR of each effective factors in mapping roe deer potential habitats can be expressed as follows:
where FRij denotes the FR of a i-th class for the j-th factor, Hij is the number of pixels contains habitat location in the i-th class of the j-th factor and TH shows the total number of habitats. Fij represents the number of pixels in the i-th class of the j-th factor and Tf indicates the total number of pixels of each factor. Finally, the habitat potential maps is prepared by summation of FR model result for each contributing factor. The value of FR method signifies the amount of correlation between the class and habitat potentiality. In other words, the higher value of FR specifies the higher probability of consideration a location as a possible habitat, and vice versa (Altafi Dadgar et al., 2017).
Convolutional neural network algorithm: Deep learning algorithms are one of the artificial neural networks (ANN) model which are structurally and numerically different from ANN. CNN is a class of deep learning neural network method in which at least one of its layers is allocated to the convolution operation (Tekerek, 2021). The main structure of CNN model composed of input layer, convolution layer, pooling or subsampling layer, fully connected layer, and output layers (Panahi et al., 2020b). The input layer is designed for converting the input data to numerical format which had been stored in a matrix. Based on the input properties, matrix property may be vary. Convolutional layer is the major feature which causes the CNN to be dissimilar to other standard neural network approaches. The sum of matrix multiplication process between kernel and input will be applied in order to extract new features in convolutional layer, and then this process will be followed by mathematical operations over the input that is called activation functions. Activation functions are necessary for introducing non-linearity into neural networks because they help to catch up non-linear features from the input data (Zhang et al., 2019).
Reducing the numerous parameters as well as the feature selection is being done by pooling layers. There are two common types of pooling function that are called maximum pooling and average pooling. Maximum or average pooling functions are employed to calculate maximum or average value for each patch of the feature map, respectively (Zhang et al., 2020b).
Fully connected layer normally comes after all convolutional and pooling layers when high-level features were extracted. This layer has the function to manage the classification settings likes prevailing feed-forward ANN (Kumar & Hati, 2020). Fully connected layer consists of all neurons that are connected to the past layer and mathematically can be defined by the following equation:
where y shows the output of the fully connected layer. Xi denotes the input vector, a is the weight, u stands for bias of the fully connected layer, and the activation function is f(.) (Tang et al., 2020). Softmax function is the activation function of fully connected layer and computes the probabilities of i-th target class (σ(Xi)). Activation function can be expressed as follows:
where zi is the i-th output neuron and N indicates the number of possible target classes for the specific inputs (Zare & Ayati, 2020). The last fully connected layer is interpreted as an output layer in which every neuron is responsible for representing final probabilities of each class.
Long short-term memory algorithm: RNNs has various classes that one of them are LSTM which has been developed to process time-dependent variables presented in sequential data (Vu et al., 2020). In comparison with ordinary RNNs, in the hidden layer of LSTM there is the exception of the summation units that will be replaced by memory blocks (Wei, 2020). Each LSTM unit comprises of three gates including forget gate (ft), input gate (it), and output gate (ot) which all of them remove or add information to each cell state. Since LSTM has the ability to learn hidden long-term sequential dependencies, it becomes popular among scholars (Yang et al., 2017). For a defined time point t, the input of a LSTM cell is xt as well as its previous output (ht-1). C̃t and Ct represent the cell input and output state, and Ct-1 is its previous state. Based on the LSTM architecture, Ct and ht will be passed to the next cell in the network. Ct and ht can be determined by the following equations (Tin et al., 2019):
where W, b, σ and tanh are weighted matrices, bias vector, sigmoid function, and hyperbolic tangent function. The forget gate has the role of reserving or forgetting information. New information will be added with proper scale by input gate. In Eqs.5 and 6, the values updates by sigmoid activation function and new candidate values generates by tanh function. The updated new candidate can be computed according to Eq.7 (Supreetha et al., 2020). A tanh layer, a sigmoid layer, and a pointwise multiplication operation (⨀) make the output gate (ot). Finally, the output gate takes a decision which information will be output according to the cell state Ct as well as the input xt and ht–1, by using the following equations (Shi et al., 2021):
Evaluation of models
ROC curve analysis is the most common statistical method employed to assess quantitatively the prediction capability of developed model for both training and testing phase. The ROC curve plots false positive rate on the X-axis (1-specificity) versus true positive rate on the Y-axis (sensitivity) (Al-Abadi et al., 2017). In current research, the 1-specificity displays the portion of pixels inaccurately predicted by the presence or absence of habitat potential and the sensitivity is the portion of pixels classified accurately (Pradhan et al., 2020). In fact, the area under the ROC curve (AUC) demonstrates the ability of the method to predict where an event happens or not. The AUC value can vary between 0 and 1. The AUC equals to 0.5 specifies random prediction and a higher value, which is closer to one, is as an indicator of better model predictability (Chen et al., 2019; Nguyen et al., 2020). Therefore, values between 0.5-0.6 indicates low accuracy, 0.6-0.7 moderate, 0.7-0.8 high, 0.8 to 0.9 very high accuracy (Arabameri et al., 2019). If AUC takes a value bigger than 0.9, the model prediction accuracy is perfect. AUC is calculated by training and validation dataset separately as called success and prediction rate curves, respectively. The success rate curve illustrates the model ability to fit the observed events and prediction rate curve displays the model prediction quality.
Results
Impact of influential factor subclasses on habitat distributions
The FR method was applied to determine the effect of influential factors on roe deer habitat distribution. With regard to the results presents in Table 1, the higher elevations raised the possibility of roe deer concentration due to providing a safe habitat away from human activities, especially from hunting. In the valley depth parameter, the class of 0-16.02 and 16.03-35.24 had the most impact on choosing a place as a habitat with the FR values 2.38 and 1.17, respectively. Observation showed that the population of roe dear increased with the increase of slope height. Highest value was found in case of 60.24-1,097.06 m. The FR values had been computed for the TPI factor as a proxy for topographic landscape terrain units, showed that higher values were related to the class of 0.42-3.73 (FR=1.41) followed by 3.74-6.22 and 6.23-9.13 classes (FR=1.21 and 1.27, respectively). For TWI, the estimated FR values decreased with increase this index. The class of 1.65-5.06 had the highest FR value (1.79), followed by the class of 5.07-5.96 (1.11). The spatial analysis revealed that NDWI class of 0.63-0.67 had the higher FR value (FR=1.92) and significantly affected roe deer habitat distribution. The highest weighted class was observed in drainage density, the second class with a FR value 1.50, followed by the road density class of 0-2.68 (FR=1.45) and the third class of the variable radar intensity (FR=1.26). Greater road density correlated with low habitat suitability for roe deer species. In case of morphological feature, roe deer tend to live in regions located in a convex/concave land formation.
Habitat potential maps
Roe deer habitat potential map based on FR model was generated by the summation of FR values of each influential factor and their subclasses using ArcGIS 10.8 through the following equation:
where FRi is the FR value of each factor’s class and N denotes the total number of influential factors. As shown in Fig. 4, the resultant map was classified into five classes using quantile method including very low, low, moderate, high and very high habitat potential zones, covering about 19.71%, 19.83%, 20.17%, 20.11%, and 20.19% of the study area, respectively. The two other maps were prepared by applying CNN and LSTM models. The CNN model predicted that approximately 19.87%, 19.85%, 19.25%, 19.47%, and 21.57% areas were designated to the very low, low, moderate, high and very high habitat potential zones whereas theses values for the LSTM model were 19.70%, 19.81%, 19.31%, 19.86%, and 21.31%, respectively. The models prediction accuracy were concluded by using AUC analysis. During the training phase the AUC of FR, CNN, and LSTM models were 0.74, 0.73 and 0.76, respectively (Fig. 5). Moreover, The FR, CNN, and LSTM models had AUC of 0.73, 0.70 and 0.73 during the testing phase.
Discussion
The present study was executed to assess the ability of deep learning algorithms (CNN and LSTM) to identify the spatial relationships between the possible roe deer habitat and related environmental factors for the first time. In fact, the prediction ability of deep learning approaches directly depends on input data accuracy and imprecise habitat location data can lead to inaccurate prediction. Concerning the performance evaluation, LSTM model had relatively higher prediction ability in comparison to FR and CNN models according to its AUC values during training and validation process (AUC=0.76% and 0.73%). Percentage of roe deer habitat locations within each class of habitat potential map (Fig. 6) display that the most numbers of locations fell into very high class; therefore, models have shown acceptable performances.
The results of variable importance analysis of influential factors denoted that TPI, TWI, altitude and valley depth had the greatest contribution in the modelling procedure with a mean effective weight of 63.3%, 58.1%, 57.8%, and 54.8%, respectively. On the other contrary, morphological feature, drainage density and radar intensity had lowest impact on habitat potential mapping in descending order. Moreover, the final maps reveal that there are slight differences between stability and robustness of developed models and two algorithms are a class of same family. Additionally, not only the values of AUC were close but also the distribution of the potential classes had remarkable similar pattern. The northeastern and central parts of the inland area of South Korea and the central part of Jeju Island have very high potential. However, LSTM model performance was superior to CNN and FR model. It should be mentioned that although FR model is easy and the result can be interpreted simply but has limitation in assessing relationships between an event occurrence and effective factors because of defining statistical assumptions prior to any study in comparison with machine learning methods. Finally, the resulting habitat potential map can make an outstanding contribution to define effective conservation plans and manage the roe deer habitats.
Figures and Tables
Fig. 2
Influential factors: (A) altitude, (B) valley depth, (C) slope height, (D) TPI, (E) TWI, (F) NDWI, (G) drainage density, (H) road density, (I) radar intensity, and (J) morphological feature. TPI, topographic position index; TWI, topographic wetness index; NDWI, normalized difference water index.
Fig. 3
Overview of applied methodology for delineation of roe deer potential habitats. GIS, geographic information system; ROC, receiver operating characteristic; CNN, convolutional neural network; LSTM, long short-term memory; TPI, topographic position index; TWI, topographic wetness index; NDWI, normalized difference water index.
Fig. 4
Habitat potential maps developed using (A) FR, (B) CNN, (C) LSTM models. FR, frequency ratio; CNN, convolutional neural network; LSTM, long short-term memory.
Fig. 5
Comparison of models prediction power using AUC (A) training phase, (B) testing phase. AUC, area under the receiver operating characteristic curve; FR, frequency ratio; CNN, convolutional neural network; LSTM, long short-term memory.
Fig. 6
Percentage of roe deer habitat locations within each class of habitat potential map. FR, frequency ratio; CNN, convolutional neural network; LSTM, long short-term memory.
Table 1
The effect of each influential factor subclasses on habitat distributions using FR model
| Influential factors | Classes | Number of pixels | Number of habitat | FR |
|---|---|---|---|---|
| Altitude (m) | 0-45.65 | 22,173,780 | 32 | 0.31 |
| 45.66-129.33 | 22,198,200 | 34 | 0.33 | |
| 129.34-243.45 | 22,214,775 | 69 | 0.66 | |
| 243.46-433.65 | 22,038,609 | 151 | 1.46 | |
| 433.66-1,940 | 22,014,492 | 233 | 2.26 | |
| Valley depth (m) | 0-16.02 | 20,315,890 | 227 | 2.38 |
| 16.03-35.24 | 23,898,669 | 131 | 1.17 | |
| 35.25-64.07 | 22,805,822 | 80 | 0.75 | |
| 64.08-121.73 | 22,041,141 | 49 | 0.47 | |
| 121.74-816.88 | 21,578,334 | 32 | 0.32 | |
| Slope height (m) | 0-8.60 | 16,874,116 | 31 | 0.39 |
| 8.61-12.91 | 19,399,497 | 42 | 0.46 | |
| 12.92-30.12 | 38,234,399 | 123 | 0.69 | |
| 30.13-60.23 | 18,843,481 | 142 | 1.61 | |
| 60.24-1,097.06 | 17,288,363 | 181 | 2.23 | |
| TPI | 0-0.41 | 23,970,287 | 37 | 0.33 |
| 0.42-3.73 | 21,751,996 | 144 | 1.41 | |
| 3.74-6.22 | 21,640,415 | 123 | 1.21 | |
| 6.23-9.13 | 21,647,975 | 129 | 1.27 | |
| 9.14-105.78 | 21,629,183 | 86 | 0.85 | |
| TWI | 1.65-5.06 | 18,901,824 | 159 | 1.79 |
| 5.07-5.96 | 25,030,432 | 130 | 1.11 | |
| 5.97-7.47 | 22,989,581 | 99 | 0.92 | |
| 7.48-12.38 | 22,105,070 | 96 | 0.93 | |
| 12.39-27.22 | 21,612,949 | 35 | 0.35 | |
| NDWI | –1-0.25 | 21,374,407 | 19 | 0.19 |
| 0.26-0.52 | 21,839,899 | 80 | 0.78 | |
| 0.53-0.62 | 19,426,220 | 117 | 1.28 | |
| 0.63-0.67 | 21,359,487 | 154 | 1.53 | |
| 0.68-1 | 26,219,650 | 149 | 1.21 | |
| Drainage density | 0 | 25,265,819 | 176 | 1.48 |
| 0.01-3.18 | 22,500,741 | 159 | 1.50 | |
| 3.19-5.57 | 22,665,901 | 82 | 0.77 | |
| 5.58-8.75 | 20,907,668 | 63 | 0.64 | |
| 8.76-101.45 | 18,878,276 | 39 | 0.44 | |
| Road density | 0-2.68 | 33,123,325 | 230 | 1.45 |
| 2.69-7.74 | 19,301,104 | 85 | 0.92 | |
| 7.75-13.4 | 18,128,869 | 100 | 1.15 | |
| 13.41-21.44 | 18,395,607 | 68 | 0.77 | |
| 21.45-75.93 | 18,936,390 | 33 | 0.36 | |
| Radar intensity | –125.56-–25.54 | 18,047,153 | 68 | 0.79 |
| –25.53-–21.7 | 22,669,814 | 123 | 1.14 | |
| –21.69-–17.85 | 24,776,910 | 149 | 1.26 | |
| –17.84-–13.23 | 21,357,501 | 107 | 1.05 | |
| –13.22-–70.63 | 21,898,743 | 72 | 0.69 | |
| Morphological features | Peak | 25,575,559 | 81 | 0.68 |
| Ridge | 150,186 | 0 | 0.00 | |
| Pass | 39,090,576 | 135 | 0.74 | |
| Plane | 444,837 | 3 | 1.44 | |
| Channel | 45,271,226 | 299 | 1.41 | |
| Pit | 107,472 | 1 | 1.98 |