remote sensing An Efﬁcient User-Friendly Integration Tool for Landslide Susceptibility Mapping Based on Support Vector Machines: SVM-LSM Toolbox

: Landslide susceptibility mapping (LSM) is an important element of landslide risk assessment, but the process often needs to span multiple platforms and the operation process is complex. This paper develops an efﬁcient user-friendly toolbox including the whole process of LSM, known as the SVM-LSM toolbox. The toolbox realizes landslide susceptibility mapping based on a support vector machine (SVM), which can be integrated into the ArcGIS or ArcGIS Pro platform. The toolbox includes three sub-toolboxes, namely: (1) inﬂuence factor production, (2) factor selection and dataset production, and (3) model training and prediction. Inﬂuence factor production provides automatic calculation of DEM-related topographic factors, converts line vector data to continuous raster factors, and performs rainfall data processing. Factor selection uses the Pearson correlation coefﬁcient (PCC) to calculate the correlations between factors, and the information gain ratio (IGR) to calculate the contributions of different factors to landslide occurrence. Dataset sample production includes the automatic generation of non-landslide data, data sample production and dataset split. The accuracy, precision, recall, F1 value, receiver operating characteristic (ROC) and area under curve (AUC) are used to evaluate the prediction ability of the model. In addition, two methods—single processing and multiprocessing—are used to generate LSM. The prediction efﬁciency of multiprocessing is much higher than that of the single process. In order to verify the performance and accuracy of the toolbox, Wuqi County, Yan’an City, Shaanxi Province was selected as the test area to generate LSM. The results show that the AUC value of the model is 0.8107. At the same time, the multiprocessing prediction tool improves the efﬁciency of the susceptibility prediction process by about 60%. The experimental results conﬁrm the accuracy and practicability of the proposed toolbox in LSM.


Introduction
The occurrence of landslide disasters causes great losses to the economy and human life all over the world every year [1,2]. Natural events such as rainfall [3,4], earthquakes [5,6] and floods [7] often lead to a series of landslides. Landslide susceptibility mapping (LSM) is used to determine the probability of future landslides in the study area by comprehensively analyzing various topographic, geological and hydrological factors, as well as human activity, alongside historical landslide activity in the study area [8,9]. LSM is of great significance to landslide risk management, human life safety and urban future planning.
In recent years, LSM has attracted the attention of many scholars, and various related articles have been published. The methods of generating landslide susceptibility mapping mainly include empirical modeling based on expert experience [10,11], physically based models [12], data-driven statistical modeling [13][14][15] and machine learning models [16][17][18][19]. Compared with traditional methods, the machine learning models do not rely on expert experience, which reduce the subjectivity of evaluation and generally have high accuracy. With the development of geographic information system (GIS) software and open-source machine learning libraries, the machine learning methods are becoming increasingly popular.
Compared with other machine learning algorithms, the support vector machine (SVM) method has been widely used in calculating landslide susceptibility because of its advantages in solving small-sample, nonlinear and high-dimensional classification problems [5,8]. However, the process of landslide susceptibility assessment using SVM is complicated, consisting of multiple steps such as data preprocessing, influencing factor selection, dataset production, model training and prediction. Generally, when using SVM to generate LSM, researchers must work with a cross-platform. Terrain factors based on the Digital Elevation Model (DEM) (e.g., slope, aspect) rely on platforms such as ArcGIS or QGIS. Model training and parameter optimization usually adopt widely used programming languages such as Python, R or MATLAB. In addition, Excel, SPSS software or programming languages have been used for model accuracy evaluation and statistical analysis in most previous studies.
Tools related to landslide susceptibility mapping are usually available in the form of academic code, which requires users to have programming skills. Some studies have proposed and applied several tools to evaluate landslide susceptibility. Osna et al. [20] developed an independent application (GeoFIS) to generate landslide susceptibility maps using the Mamdani fuzzy inference system (FIS). Sezer et al. [11] developed an LSM module based on expert experience with NetCAD architecture software. Jebur et al. [21] created a landslide susceptibility mapping toolbox using bivariate statistical analysis (BSA) based on ArcGIS. Zhang et al. [15] provided a landslide susceptibility assessment tool based on the optimized frequency ratio method, which itself is based on the ArcGIS platform. Torizin et al. [22] provided an independent landslide susceptibility assessment application written in Python, Project Manager Suite (LSAT PM). Bragagnolo et al. [23] developed a free and open-source plug-in, namely r.landslide, based on the GRASS software of opensource GIS, to generate landslide susceptibility mapping based on an artificial neural network. Sahin et al. [24] integrated R and ArcGIS software and developed a landslide susceptibility mapping toolkit (LSM tool pack) based on logistic regression and random forest. Guo et al. [25] introduced a Python QGIS plugin [26] named FSLAM, which allows us to compute regional shallow landslide susceptibility based on the effective antecedent water recharge and the event rainfall.
Most of the above toolboxes are based on expert experience models or statistical models, such as the weight of evidence method, frequency ratio method and so on. These methods are simple in principle and easy to implement, but with limited accuracy. To date, only a limited number of previous studies have involved the development of landslide susceptibility mapping tools based on machine learning methods. At the same time, most tools only involve model training and prediction, instead of the whole process of LSM. In addition, most studies only use the single-factor pixel value corresponding to landslide point locations as samples for model training. However, landslides usually occur within a region and are affected by characteristics from the surrounding environment. Therefore, problems exist when constructing samples based on a single pixel [27,28]. The realization of regional-scale data construction is often complicated and time-consuming.
To solve the above-mentioned problems, this research develops an LSM toolbox based on the ArcGIS platform (SVM-LSM toolbox). The toolbox includes data preprocessing, factor selection, SVM model training and evaluation, and landslide susceptibility map prediction, involving the whole process of LSM. Moreover, this toolbox only uses the ArcGIS platform, which avoids cross-platform operation and reduces user input parameters as much as possible. The operation is simple, convenient and user-friendly. The susceptibility prediction process based on sliding windows is time-consuming. This tool Remote Sens. 2022, 14, 3408 3 of 22 provides a multiprocessing rapid prediction tool to sufficiently improve the production efficiency of landslide susceptibility mapping. In addition, a tool for the rapid production of multi-channel block datasets is constructed to improve the efficiency of dataset making. It is worth noting that this toolbox is not limited to the mapping of landslide susceptibility based on SVM and can also be used for other binary classification problems based on SVM. Section 2 of this paper introduces the basic functions of the toolbox and a description of each module; Section 3 discusses the experimental research on the landslide susceptibility mapping of the toolbox in Wuqi County, Shaanxi Province, China, and provides an analysis of the relevant results; and Section 4 presents the conclusion.

LSM Workflow
An overall flow chart of LSM based on SVM is shown in Figure 1. The process of generating LSM based on SVM consists of data collection, data preprocessing, dataset making, feature selection, model training and susceptibility map prediction. The data collection includes historical landslide data, the coverage of the study area and landslide influencing factors, such as roads, rivers, faults, Normalized Difference Vegetation Index (NDVI), DEM, lithology and rainfall. Among them, landslide points, the coverage of the study area, roads, rivers and faults are vector data, NDVI, DEM and lithology are grid data, and rainfall is the NetCDF-4 (nc4) format. Data preprocessing includes calculating topographic factors (such as slope, aspect, etc.) based on DEM, converting line vector data to continuous raster factors, and nc4 data processing. For raster data, it is also necessary to clip them to the same study area range. Subsequently, based on landslide points and the range of the study area, the same number of non-landslide points are randomly selected to construct negative samples. Then, the dataset is randomly divided into training samples and test samples in the ratio of 7:3. In addition, the Pearson correlation coefficient (PCC) and information gain ratio (IGR) are calculated for all the samples. Influencing factors are selected based on the calculation results; factors with high correlations or with less importance to landslide occurrence are removed. Then, the training and test sets are reconstructed according to the results of the feature selection. Finally, the training set is used to train the model, and an optimal SVM model is obtained through the comprehensive analysis of parameters and evaluation indicators such as accuracy, precision, recall, F1 value, receiver operating characteristics (ROC) and area under the curve (AUC). The optimal model is finally used to predict the susceptibility index of the study area and generate a susceptibility map of the study area for subsequent analysis.
In this paper, a toolbox is presented to generate landslide susceptibility maps according to the above-mentioned workflow. The LSM toolbox includes three sub-toolboxes: "1 influence factor production", "2 factor selection and dataset production" and "3 model training and prediction", as shown in Figure 2. This toolbox is developed based on ArcPy and Python language and can be directly integrated into ArcGIS 10.1 (or higher) or ArcGIS Pro software. It is efficient and user-friendly.
x FOR PEER REVIEW 4 of 23 In this paper, a toolbox is presented to generate landslide susceptibility maps according to the above-mentioned workflow. The LSM toolbox includes three sub-toolboxes: "1 influence factor production", "2 factor selection and dataset production" and "3 model training and prediction", as shown in Figure 2. This toolbox is developed based on ArcPy and Python language and can be directly integrated into ArcGIS 10.1 (or higher) or ArcGIS Pro software. It is efficient and user-friendly.  In this paper, a toolbox is presented to generate landslide susceptibility maps ing to the above-mentioned workflow. The LSM toolbox includes three sub-toolb influence factor production", "2 factor selection and dataset production" and " training and prediction", as shown in Figure 2. This toolbox is developed based o and Python language and can be directly integrated into ArcGIS 10.1 (or higher) or Pro software. It is efficient and user-friendly.

Influencing Factor Production
Landslide influencing factors are various factors that affect the occurrence slides through the study of the occurrence mechanism of landslides in the study a

Influencing Factor Production
Landslide influencing factors are various factors that affect the occurrence of landslides through the study of the occurrence mechanism of landslides in the study area. The occurrence of landslides is affected by various influencing factors. At present, there is no unified standard for the selection of influencing factors. Pourghasemi et al. [29] conducted a statistical analysis on the influencing factors used in the study and found that topographic factors, geological factors and human activities are the most commonly used factors for landslide occurrence. This toolbox provides a tool for generating relevant topographic factors based on DEM, a tool for converting roads, faults and rivers into continuous raster data, and a rainfall processing tool.

Topographic Factor Calculation
This tool integrates other factors calculated by DEM, and automatically calculates other topographic factors such as slope, aspect, curvature, plane curvature, profile curvature, relief amplitude, surface roughness, topographic wetness index (TWI) and other topographic factors based on DEM data in the study area. As shown in Figure 3a, it is necessary to only input DEM data and select the factors that need to be calculated. These factors can be calculated optionally according to the needs of users by checking the box in front of the factors to be calculated, but aspect must be calculated when calculating plane curvature, and slope must be calculated when calculating profile curvature, surface roughness or TWI.

Convert Line Vector Data to Continuous Raster Factor
This tool automatically converts the line vector data of the study area into continuous raster data, such as distance to roads, distance to faults and distance to rivers. The conversion principle adopts Euclidean distance. As shown in Figure 3b, the user only needs to input the line vector data to be converted and the result save path.

Convert Line Vector Data to Continuous Raster Factor
This tool automatically converts the line vector data of the study area into continuous raster data, such as distance to roads, distance to faults and distance to rivers. The conversion principle adopts Euclidean distance. As shown in Figure 3b, the user only needs to input the line vector data to be converted and the result save path.

Rainfall Data Processing
The National Aeronautics and Space Administration (NASA, https://gpm.nasa.gov/, accessed on 24 December 2020) provides a Global Precipitation Measurement Mission (GPM). These are high-precision precipitation data obtained using multi-sensors, multisatellites and multi-algorithms combined with the satellite network and rainfall gauge inversion, with a spatial and temporal resolution up to 0.5 h, 0.1 • × 0.1 • [30]. The monthly or daily rainfall data downloaded from NASA are in the .nc4 format, which is time-consuming and laborious to convert into raster data one by one. Therefore, this tool provides a rainfall batch conversion tool to convert the .nc4 format data to the .tif format raster data. As shown in Figure 3c, the user only needs to input the rainfall data and specify the raster data output coordinate system.

Batch Clipping of Each Factor Layer
After the production of the factor layer data, the row and column numbers and coverage of each factor layer data are usually inconsistent. This tool is used to batch clip the raster data of each factor layer according to the vector data of the study area in order to obtain the factor layer data of the study area. As shown in Figure 3d, this tool only needs Remote Sens. 2022, 14, 3408 7 of 22 the user to set the folder where the raster factors are located and the vector data of the study area; it can automatically iteratively select the .tif format files for clipping. All the raster data resolutions should be consistent.

Non-Landslide Data Generation
This tool is used to generate non-landslide point data within the study area vector data layer. As shown in Figure 4a, the user inputs landslide points and the study area vector file and specifies the number of non-landslide points to be selected outside of a buffer and how many meters from the landslide point. First, the tool generates a buffer zone at a specified distance from the landslide point and erases the buffer zone layer on the study area layer to obtain the selectable range of non-landslide sample points. It then uses random points to generate the same number of non-landslide points within the optional range. The non-landslide points should be selected as far from landslide points as possible.
PEER REVIEW 8 of 23

Data Sample Production
This tool is used to generate multi-channel block sample raster data from vector point data. As shown in Figure 4b, the user inputs the vector point elements and the multichannel factor layer data and specifies the buffer distance, which is half of the actual distance represented by the cropped raster size. The tool uses vector point data to create a buffer and iteratively selects the buffer range corresponding to each point vector in order to cut the multi-channel raster data one by one, resulting in a single multi-channel block

Data Sample Production
This tool is used to generate multi-channel block sample raster data from vector point data. As shown in Figure 4b, the user inputs the vector point elements and the multi-channel factor layer data and specifies the buffer distance, which is half of the actual distance represented by the cropped raster size. The tool uses vector point data to create a buffer and iteratively selects the buffer range corresponding to each point vector in order to cut the multi-channel raster data one by one, resulting in a single multi-channel block dataset of each vector point named after the "FID" attribute value. When the buffer distance is less than the resolution of the raster data, the obtained sample has reached the point at which the landslide point is located.

Dataset Split
When using the machine learning methods for model training, it is common to split the samples into a training set and a test set in a certain ratio. The training set is used to train the model and the test set is used to test the generalization of the model and prevent overfitting. As shown in Figure 4c, users can specify the ratio of the training and test sets by themselves. Generally, the ratio of the training and test sets is 7:3. Finally, the sample paths and labels of the training and test sets will be given, respectively (0 for non-landslide and 1 for landslide), and the results are saved in a txt file.

PCC and IGR Calculation
Determining the most effective combination of the influencing factors for landslide susceptibility mapping is of great importance. If the influencing factors are not evaluated, Remote Sens. 2022, 14, 3408 9 of 22 this will not only cause data redundancy but will also affect the execution efficiency and prediction accuracy of the model [31]. At present, there is no optimal solution for the selection of influencing factors, but they typically consist of two parts: correlation analysis and importance evaluation. This toolbox provides two of the most commonly used influencing factor selection methods: PCC and IGR. The PCC is an index used to measure the correlation between the influencing factors. The closer its absolute value is to 1, the stronger the correlation between the two factors. The information gain ratio is an index used to evaluate the importance of each factor layer on landslide occurrence. The higher the IGR value, the greater the impact of this factor on landslide occurrence. Any factor with zero IGR does not influence landslide occurrence. As shown in Figure 4d, this tool calculates PCC and IGR based on the generated data samples and saves them in a txt file. Upon comprehensively considering the calculation results, factors with strong correlation and little influence on landslide occurrence are eliminated based on the principle that the lower the correlation is, the greater the importance is.

Image Generation to Be Predicted
The different factor layers are stacked in a certain order to form multi-channel raster data, which is the image to be predicted. It is used for sample production and susceptibility map prediction. As shown in Figure 5a, this tool only requires the input of the path and stacking order of each factor layer. Here, the stacking order of the factor layers used for the image to be predicted should be consistent with the order of the factor layers in the model training samples.
x FOR PEER REVIEW 10 of 23

Model Training and Performance Evaluation of SVM
This tool is used to generate SVM models with given parameters and provide evaluation results of the accuracy of each model. As shown in Figure 5b, the user enters the directory in which the dataset sample is located along with the number of rows, columns and channels of the dataset. At the same time, the optional values of parameter gamma

Model Training and Performance Evaluation of SVM
This tool is used to generate SVM models with given parameters and provide evaluation results of the accuracy of each model. As shown in Figure 5b, the user enters the directory in which the dataset sample is located along with the number of rows, columns and channels of the dataset. At the same time, the optional values of parameter gamma and penalty factor C to be adjusted also should be given. The parameter adjustment method used in this tool is the grid search algorithm.
SVM has certain advantages in solving the problem of small-sample classification [32]. The kernel function and slack variable are used to deal with the linear indivisibility of the sample data. At the same time, because the classifier is only determined by the support vector, SVM can effectively avoid overfitting. SVM attempts to classify samples by introducing kernel functions that map landslide influencing factors to a high-dimensional feature space, from which it attempts to locate the optimal hyperplane with the maximum spacing between landslides and non-landslides from the feature space [33]. Xu et al. [5] discussed the influence of different kernel functions of SVM on landslide susceptibility mapping. The results show that the prediction effect of the radial basis function (RBF) in SVM is optimum. Therefore, the kernel function of this tool defaults to RBF.
The susceptibility map is equivalent to a binary classification problem. Landslides are marked as "1" and non-landslides marked as "0". Thus, the confusion matrix can be constructed according to different combinations of real value and predicted value, and the model accuracy evaluation index can be constructed based on the confusion matrix. In this tool, accuracy, precision, recall, F1 value, receiver operating characteristic (ROC) and area under curve (AUC) were used to evaluate the prediction ability of the model. The calculation formula [6] is as follows.

Landslide Susceptibility Map Prediction
This tool is used to predict landslide susceptibility in the study area, based on the optimal model, and obtain the landslide susceptibility map in the study area. In this tool, a sliding window with the same row and column numbers as the dataset is constructed to select the data to be predicted for input into the optimal model to obtain the susceptibility index until all rows and columns are sliding. The tool provides two options: single process ( Figure 5c) and multiprocessing (Figure 5d). Single-process and multiprocessing tools can be used under ArcGIS and ArcGIS Pro, but the single-process tool speed is slow and the multiprocessing tool is fast. In a single process, the user must only give the image to be predicted, the optimal model and the number of rows and columns of the dataset. In multiprocessing, the user must also specify "pythonw.exe" location.

Results
Taking Wuqi County, Shaanxi Province, China as an example, the developed toolbox was applied to carry out a landslide susceptibility assessment.

Study Area
The study area is located in Wuqi County, Yan'an City, Shaanxi Province (107 • 38 57 E1 08 • 32 49 E, 36 • 33 33 N~37 • 24 27 N). It covers a total area of 3791.5 km 2 , encompasses a total population of 145,700 and has an altitude of 1233~1809 m. The study area has a warm, temperate, continental, semi-arid climate. It is dry and windy in spring, sees alternating drought and flood conditions in summer, is cool and wet in autumn and is cold and dry in winter, and the annual average temperature is 7.8 • C. The average annual rainfall is 483.4 mm, and the total coverage of forest and grass is 49.6%. The Wuding and Beiluo River systems lie within the study area. The landform belongs to the hilly and gully area of the Loess Plateau. The terrain fluctuates greatly, the gully is long and the slope is steep [34]. The landslide type in the study area mainly belongs to Loess landslides. During the flood season, rainstorms or continuous rainfall will often induce landslides, collapses and debris flow of different scales, seriously threatening the lives and property safety of local people. Therefore, it is of great practical significance to carry out landslide susceptibility evaluation in Wuqi County. The location of the study area is shown in Figure 6.
nual rainfall is 483.4 mm, and the total coverage of forest and grass is 49.6%. The Wuding and Beiluo River systems lie within the study area. The landform belongs to the hilly and gully area of the Loess Plateau. The terrain fluctuates greatly, the gully is long and the slope is steep [34]. The landslide type in the study area mainly belongs to Loess landslides. During the flood season, rainstorms or continuous rainfall will often induce landslides, collapses and debris flow of different scales, seriously threatening the lives and property safety of local people. Therefore, it is of great practical significance to carry out landslide susceptibility evaluation in Wuqi County. The location of the study area is shown in Figure 6.

Preprocessing of Influencing Factors
The influence factor data sources used in this example include DEM, roads, rivers, lithology, NDVI and rainfall. Lithology and NDVI were pre-processed into 30 m resolution raster data. For the acquired DEM data, the "topographic factor calculation" tool is used to generate slope, aspect, curvature, plane curvature, profile curvature, relief amplitude, surface roughness and a topographic wetness index (TWI). At the same time, the

Preprocessing of Influencing Factors
The influence factor data sources used in this example include DEM, roads, rivers, lithology, NDVI and rainfall. Lithology and NDVI were pre-processed into 30 m resolution raster data. For the acquired DEM data, the "topographic factor calculation" tool is used to generate slope, aspect, curvature, plane curvature, profile curvature, relief amplitude, surface roughness and a topographic wetness index (TWI). At the same time, the "convert line vector data to continuous raster factor" tool is used to produce the distance to rivers and distance to roads. Since there is no active fault in the study area and it is not affected by active faults, the distance to the fault is not considered. For the rainfall data (.nc4), the "rainfall data processing" tool is used to convert the monthly rainfall data obtained by NASA into the corresponding raster data in batches, and the raster calculator is used to accumulate monthly rainfall data in order to obtain annual rainfall data. Finally, the "batch clipping of each factor layer" tool is used to batch cut the generated influencing factor data according to the vector data of the study area. Finally, a total of 14 landslide influencing factors are generated (Figure 7), and the spatial resolutions of all the factor data are 30 m.

Factor Selection and Sample Generation
There are 789 historical landslides in the study area, which can be divided into 175 large landslides, 417 medium landslides and 197 small landslides. In this study, all the landslide locations are used to construct the landslide dataset. Based on the landslide point data, the "non-landslide data generation" tool was used to randomly generate the same number of non-landslide points, each of which should be at least 1 km away from all of the landslide points in the study area.
Since the calculation of IGR and PCC must be based on all the sample data, the dataset needs to be created before the selection of influencing factors. Firstly, the "image generation to be predicted" tool is used to stack the generated data of 14 influencing factors in the study area in multiple channels. Then, the "data sample production" tool is used to make landslide and non-landslide block datasets based on the superimposed multi-channel images. In addition, the "dataset split" tool is used to divide the training samples and test samples in the ratio of 7:3, before saving the path and labels of the samples to the corresponding txt file, respectively. Finally, all the block datasets have fourteen channels, eight rows and eight columns. There are 1104 images in the training set and 474 images in the test set, in which the landslide dataset is marked as 1 and the non-landslide dataset is marked as 0.
After using the "PCC and IGR calculation" tool to calculate the PCC and information gain ratio of each factor layer based on the data samples, Figure 8 shows the results of the PCC calculation. It can be seen that the correlation coefficients between plane curvature and slope, TWI and slope, and relief amplitude and surface roughness are greater than 0.5. The study area is located in the hinterland of the Loess Plateau which is a typical hilly and gully landscape with high topographic fragmentation and loose soils. The reason for such strong correlations is that the study area often suffers from severe rainfall erosion and river erosion. On the one hand, the greater the slope, the more severe the soil erosion. Therefore, the more complex the surface morphology, the greater roughness and relief amplitude of the surface. On the other hand, the steep slopes with low water retention capacity lead to low soil water content (TWI), and vice versa. Figure 9 presents the calculation results of the information gain ratio. The IGR values of 14 landslide influencing factors are greater than 0, indicating that these factors have an impact on the occurrence of landslides in the corresponding areas. In this study area, lithology has the greatest impact on landslide occurrence, followed by NDVI, plane curvature, profile curvature and TWI, while curvature and relief amplitude have the least impact. Upon a comprehensive analysis of PCC and IGR, the two influencing factors of slope and relief amplitude were removed for Wuqi County, and the remaining 12 influencing factors were used for subsequent research.
According to the evaluation results, the steps of "image generation to be predicted", "data sample production" and "dataset split" should be repeated in decreasing order of information gain ratio (i.e., lithology, plane curvature, profile curvature, NDVI, TWI, aspect, surface roughness, distance to rivers, DEM, distance to roads, rainfall and curvature) to obtain the final image and sample data for further prediction. The number of channels of all the block datasets used is 12, and their row and column numbers are both eight in the subsequent analysis. Figure 8. Pearson correlation coefficient matrix for the Wuqi County case study. Note that "slp" represents slope, "asp" represents aspect, "cur" represents curvature, "plancur" represents plane curvature, "profilecur" represents profile curvature, "rivers" represents distance to rivers, "roads" represents distance to roads, "lithology" represents lithology, "SroughnessC" represents surface roughness, "relief" represents relief amplitude, and "rainfall" represents rainfall. Figure 9. Information gain ratio for the Wuqi County case study. Note that "slp" represents slope, "asp" represents aspect, "cur" represents curvature, "plancur" represents plane curvature, "profilecur" represents profile curvature, "rivers" represents distance to rivers, "roads" represents distance to roads, "lithology" represents lithology, "SroughnessC" represents surface roughness, "relief" represents relief amplitude, and "rainfall" represents rainfall. Figure 8. Pearson correlation coefficient matrix for the Wuqi County case study. Note that "slp" represents slope, "asp" represents aspect, "cur" represents curvature, "plancur" represents plane curvature, "profilecur" represents profile curvature, "rivers" represents distance to rivers, "roads" represents distance to roads, "lithology" represents lithology, "SroughnessC" represents surface roughness, "relief" represents relief amplitude, and "rainfall" represents rainfall.

Model Training and Performance Evaluation
The "model training and performance evaluation of SVM" tool is used to train the model based on the generated training data, evaluate the performance with the test set and plot the ROC curve. Of these, the SVM model uses the RBF kernel function. The model has two parameters: gamma and penalty factor C. The grid search algorithm is used to optimize the parameters, find the optimal set of model parameters and generate the optimal model. The values of parameters gamma and C are selected from 0.01, 0.02, 0.05, 0.08, 0.1, 0.2, 0.5, 0.8, 1, 2 and 5. Figure 10 shows the AUC values and the difference in accuracy between the training and test sets for different gamma and C values, which used gamma values as horizontal coordinates and C values as vertical coordinates. In the figure, the size of the circle represents the AUC value. The larger the circle, the greater the AUC value and the better the model performance. The color of the circle represents the accuracy difference between the training and test sets. If it exceeds 0.5, it is represented by 0.5. The greater the accuracy difference, the higher the degree of overfitting of the model and the worse the generalization performance. Consequently, comprehensive analysis shows that when gamma is 0.02 and C is 2, the AUC value is high, the accuracy difference is small, and the model is optimal.

Model Training and Performance Evaluation
The "model training and performance evaluation of SVM" tool is used to train the model based on the generated training data, evaluate the performance with the test set and plot the ROC curve. Of these, the SVM model uses the RBF kernel function. The model has two parameters: gamma and penalty factor C. The grid search algorithm is used to optimize the parameters, find the optimal set of model parameters and generate the optimal model. The values of parameters gamma and C are selected from 0.01, 0.02, 0.05, 0.08, 0.1, 0.2, 0.5, 0.8, 1, 2 and 5. Figure 10 shows the AUC values and the difference in accuracy between the training and test sets for different gamma and C values, which used gamma values as horizontal coordinates and C values as vertical coordinates. In the figure, the size of the circle represents the AUC value. The larger the circle, the greater the AUC value and the better the model performance. The color of the circle represents the accuracy difference between the training and test sets. If it exceeds 0.5, it is represented by 0.5. The greater the accuracy difference, the higher the degree of overfitting of the model and the worse the generalization performance. Consequently, comprehensive analysis shows that when gamma is 0.02 and C is 2, the AUC value is high, the accuracy difference is small, and the model is optimal.  Table 1 shows the performance of the optimal model with the testing dataset, and Figure 11 shows its corresponding ROC curve. Among the 474 testing datasets, 169 landslides and 171 non-landslides were correctly predicted, while 68 landslides and 66 non-landslides were incorrectly predicted. The correct samples predicted by the model accounted for 71.73% of the total samples, with a precision of 71.55% and a recall rate of 72.15%. At the same time, the AUC value of the model is 0.8029, indicating that the model has good prediction performance and the result of the landslide susceptibility map is reliable.  Table 1 shows the performance of the optimal model with the testing dataset, and Figure 11 shows its corresponding ROC curve. Among the 474 testing datasets, 169 landslides and 171 non-landslides were correctly predicted, while 68 landslides and 66 nonlandslides were incorrectly predicted. The correct samples predicted by the model accounted for 71.73% of the total samples, with a precision of 71.55% and a recall rate of 72.15%. At the same time, the AUC value of the model is 0.8029, indicating that the model has good prediction performance and the result of the landslide susceptibility map is reliable.

Confusion matrix
Landslide  Figure 11. The ROC curve of the optimal model.

Landslide Susceptibility Map Generation and Analysis
With the trained optimal model, the "landslide susceptibility map prediction" tool is used to predict the generated image unit by unit according to the optimal model. The probability of each evaluation unit being predicted as a landslide is obtained to generate a landslide susceptibility map for the study area. The predicted susceptibility indexes lie between 0 and 1. The larger the susceptibility index is, the more susceptible the area is to landslides. The generated susceptibility map is divided into five levels-very low, low, moderate, high and very high-using the natural break method in ArcGIS. The landslide susceptibility map of Wuqi County after classification is obtained by SVM, as shown in Figure 12.
It is clear in Figure 12 that the areas in Wuqi County with high and very high susceptibility to landslides are mainly concentrated on both sides of rivers severely affected by soil erosion. Low-and very-low-susceptibility areas are mainly distributed in high-altitude areas with limited human activity. The locations of historical landslides are well fitted with the predicted results. The areas where landslides are relatively concentrated are predicted as high and very high susceptibility areas, which is in line with the actual situation. Table 2 shows the proportion of each graded area and the density of landslide points within each grade. It can be seen that the proportion of high-and very-high-susceptibility areas is 29.97%, and the proportion of low-and very-low-susceptibility areas is 49.18%. With increased susceptibility grade, the density of landslide points increases continuously, which is in line with the actual situation of the susceptibility grade. The density of Figure 11. The ROC curve of the optimal model.

Landslide Susceptibility Map Generation and Analysis
With the trained optimal model, the "landslide susceptibility map prediction" tool is used to predict the generated image unit by unit according to the optimal model. The probability of each evaluation unit being predicted as a landslide is obtained to generate a landslide susceptibility map for the study area. The predicted susceptibility indexes lie between 0 and 1. The larger the susceptibility index is, the more susceptible the area is to landslides. The generated susceptibility map is divided into five levels-very low, low, moderate, high and very high-using the natural break method in ArcGIS. The landslide susceptibility map of Wuqi County after classification is obtained by SVM, as shown in Figure 12.
It is clear in Figure 12 that the areas in Wuqi County with high and very high susceptibility to landslides are mainly concentrated on both sides of rivers severely affected by soil erosion. Low-and very-low-susceptibility areas are mainly distributed in high-altitude areas with limited human activity. The locations of historical landslides are well fitted with the predicted results. The areas where landslides are relatively concentrated are predicted as high and very high susceptibility areas, which is in line with the actual situation. Table 2 shows the proportion of each graded area and the density of landslide points within each grade. It can be seen that the proportion of high-and very-high-susceptibility areas is 29.97%, and the proportion of low-and very-low-susceptibility areas is 49.18%. With increased susceptibility grade, the density of landslide points increases continuously, which is in line with the actual situation of the susceptibility grade. The density of landslide points in very-high-susceptibility areas is 0.77 and that in very-low-susceptibility areas is 0.04. landslide points in very-high-susceptibility areas is 0.77 and that in very-low-susceptibility areas is 0.04.

Toolbox Operation Efficiency Evaluation
Although the "landslide susceptibility map prediction (single process)" and "landslide susceptibility map prediction (multiprocessing)" tools can be used under ArcGIS and ArcGIS Pro, it is recommended that they be used with ArcGIS Pro. Since Python 2.7 installed in ArcGIS is generally 32-bit, it has extremely limited use of memory resources and can only use a maximum of 2G of memory when processing massive data. If it exceeds 2G, a "Memory Error" will appear. Meanwhile, the Python 3 environment used by ArcGIS Pro is 64-bit, which can use more memory than the 32-bit Python, and therefore the "Memory Error" rarely occurs.

Toolbox Operation Efficiency Evaluation
Although the "landslide susceptibility map prediction (single process)" and "landslide susceptibility map prediction (multiprocessing)" tools can be used under ArcGIS and ArcGIS Pro, it is recommended that they be used with ArcGIS Pro. Since Python 2.7 installed in ArcGIS is generally 32-bit, it has extremely limited use of memory resources and can only use a maximum of 2G of memory when processing massive data. If it exceeds 2G, a "Memory Error" will appear. Meanwhile, the Python 3 environment used by ArcGIS Pro is 64-bit, which can use more memory than the 32-bit Python, and therefore the "Memory Error" rarely occurs.   Notes: "Data sample production", "dataset split" and "image generation to be predicted" tools must be run twice. * indicates that the first run time and the second run time, and † shows the total single process running time and the total multiprocessing running time.
As shown in Table 3, the total time of the SVM-LSM toolbox for the ArcGIS single process is 5 h 19 min 27 s and that for the ArcGIS Pro single process is 2 h 58 min 39 s, which improves running efficiency by 44.08%. The main gap in running time is concentrated in the operation of the "susceptibility map prediction" tool. At the same time, the total time of the SVM-LSM toolbox in ArcGIS multiprocessing is 2 h 48 min 3 s and the total time in ArcGIS Pro multiprocessing is 1 h 52 min 4 s, which improves running efficiency by 33.31%. The main difference in the running time is concentrated in the step of the "model training and performance evaluation of SVM". The abovementioned two differences are mainly due to their difference in the number of bits. Therefore, it is recommended that the toolbox in ArcGIS Pro is run with 64-bit Python. In addition, under the ArcGIS platform, the running time of the "landslide susceptibility map prediction (multiprocessing)" tool is 2 h 48 min 3 s and the running time of the "landslide susceptibility map prediction (single process)" tool is 5 h 19 min 27 s, which shortens running time by nearly 2 h 31 min 24 s and improves running efficiency by 47.39%. Under the ArcGIS Pro platform, the running time of the "landslide susceptibility map prediction (multiprocessing)" tool is 20 min 12 s and the running time of the "landslide susceptibility map prediction (single process)" tool is 1 h 26 min 47 s, which shortens running time by nearly 1 h 6 min 35 s and improves running efficiency by 76.72%. This shows that the multiprocessing prediction tool for the sliding window in this tool can greatly improve the efficiency of susceptibility mapping.

Model Selection: SVM
As mentioned earlier, SVM is used in the toolbox. To assess whether it is optimal to employ SVM, comparisons with two other commonly used models, namely, decision tree (DT) and random forest (RF), are performed. Table 4 shows the operation efficiency and AUC values of different models. The DT model requires two parameters to be adjusted: max_depth and min_samples_leaf ; the RF model requires five parameters to be adjusted: max_depth, max_features, n_estimators, min_samples_leaf and min_samples_split; and the SVM model requires two parameters to be adjusted: gamma and C. For the grid search method, the greater the number of model parameters, the higher the model training time complexity, and the more time-consuming the model tuning is. In terms of model accuracy, for the same training and testing datasets in Wuqi County, the AUC of the optimal RF model is 0.8372, the AUC of the optimal SVM model is 0.8029, and the AUC of the optimal DT model is 0.7774. The AUC values of SVM and RF model are both higher than 0.8, indicating that these two models can better reflect the landslide susceptibility in this area. Therefore, compared with the three models, the SVM model is friendlier to beginners, with fewer parameters to be adjusted, short running time and high accuracy. Therefore, we choose the SVM model to build the LSM toolbox.

Conclusions
This paper develops a tool known as the SVM-LSM toolbox, which integrates the whole process of landslide susceptibility mapping. The toolbox consists of three sub-toolboxes: (1) influence factor production, (2) factor selection and dataset production, and (3) model training and prediction. The tool can be integrated into ArcGIS 10.1 (or higher) as well as ArcGIS Pro. The interface is user-friendly, easy to implement and provides multiprocessing prediction, which greatly improves prediction efficiency. In order to assess the performance of the toolbox, Wuqi County (an area highly prone to Loess landslides) is selected as the study area. Six basic factors are selected and a total of fourteen landslide influencing factors are obtained based on the influencing factor production tool. In the selection of influencing factors, the slope and relief amplitude factors are eliminated according to the results of PCC and IGR. Finally, the model training tool is used to obtain the optimal model according to various evaluation indexes and generate a susceptibility map of the study area.
The results show that the model has good prediction performance and high prediction accuracy. The susceptibility areas of Wuqi County are mainly concentrated along rivers severely affected by soil erosion. In short, the SVM-LSM toolbox optimizes the complex susceptibility mapping process, avoids the cross-platform operation of traditional workflow and greatly shortens the prediction time of the susceptibility map. At present, the toolbox has only been tested with ArcGIS and ArcGIS Pro software on the Windows system. In the future, it will be integrated into other commonly used GIS processing software, such as QGIS, for expansion. Furthermore, more machine learning models can be incorporated, and automatic parameter tuning function can be developed to further improve the userfriendliness and universality of the toolbox.