A landslide susceptibility assessment method based on multi-source remote sensing data and machine learning

Abstract

The application discloses a landslide susceptibility assessment method based on multi-source remote sensing data and machine learning, and belongs to the technical field of geological disaster risk assessment. The method obtains ASTER GDEM digital elevation data, Sentinel-2 vegetation index data, land use classification data, road and water system data and geological structure data and the like, extracts nine evaluation factors including elevation, slope, slope direction, terrain relief, normalized vegetation index, land use type, distance from road, distance from water system and distance from fault, collects historical landslide points and non-landslide point samples to construct a training data set, trains four machine learning models including logistic regression, random forest, support vector machine and multilayer perception, selects an optimal model through cross validation and multi-index evaluation, generates a landslide susceptibility spatial distribution map and performs danger classification. The application solves the problems of strong subjectivity and low precision of a traditional assessment method, and the accuracy reaches more than 97% and the ROC-AUC reaches 0.973 in application in Tengchong County, thereby providing a scientific basis for geological disaster prevention.

Description

Technical Field

[0001] This invention belongs to the field of geological disaster risk assessment technology, specifically relating to a landslide susceptibility assessment method based on multi-source remote sensing data and machine learning, and particularly relating to a method for extracting environmental factors from multi-source remote sensing data, training and comparing models using various machine learning algorithms, and achieving high-precision assessment of landslide susceptibility. Background Technology

[0002] Landslides are one of the most common geological hazards worldwide, causing enormous casualties and property losses every year. Landslide susceptibility assessment is a crucial foundation for geological hazard risk management, and is of great significance for disaster early warning, land planning, and engineering construction. Traditional landslide susceptibility assessment methods mainly rely on expert experience and simple statistical analysis, which have the following shortcomings: First, traditional methods such as the Analytic Hierarchy Process (AHP) rely heavily on expert subjective judgment, and different experts may yield significantly different assessment results for the same area, resulting in poor objectivity and repeatability. Second, traditional methods struggle to handle multi-dimensional, large-scale remote sensing data and fail to fully utilize the rich information provided by modern remote sensing technology. Third, traditional statistical methods (such as the information content method and the coefficient of determination method) assume independence between factors, failing to effectively capture the complex nonlinear relationships and interactions between environmental factors. Fourth, traditional methods lack systematic model validation and accuracy assessment mechanisms, making it difficult to quantify the reliability of the assessment results.

[0003] In recent years, with the rapid development of remote sensing technology and machine learning methods, landslide susceptibility assessment methods based on remote sensing data and machine learning have gradually emerged. However, existing methods still have the following problems: limitations of single data sources, with most studies using only DEM data or single satellite data, failing to fully integrate the advantages of multi-source remote sensing data; blind algorithm selection, lacking systematic comparison of multiple algorithms, making it difficult to determine the most suitable algorithm for a specific region; empirical parameter settings, with model parameter settings lacking theoretical basis and experimental verification; and an imperfect evaluation system, lacking a comprehensive performance evaluation index system.

[0004] Therefore, there is an urgent need to develop a new method for assessing landslide susceptibility that integrates multi-source remote sensing data, employs multiple algorithms for comparative optimization, and has a comprehensive evaluation system. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a landslide susceptibility assessment method based on multi-source remote sensing data and machine learning. By integrating multi-source remote sensing data, constructing a scientific evaluation factor system, using various machine learning algorithms for training and comparison, and establishing a comprehensive model evaluation mechanism, a high-precision assessment of landslide susceptibility can be achieved. The technical solution of the present invention is as follows: A landslide susceptibility assessment method based on multi-source remote sensing data and machine learning includes the following steps: (1) Acquisition of multi-source remote sensing data: ASTER GDEM digital elevation model data (30-meter resolution), Sentinel-2 multispectral remote sensing image data, China Land Cover Dataset (CLCD), road network vector data, water system network vector data, and geological structure data of the study area were acquired. These data come from different remote sensing platforms and databases, are complementary, and can comprehensively reflect the topography, vegetation cover, land use, infrastructure, and geological structure characteristics of the study area.

[0006] (2) Evaluation Factor Extraction: Four topographic factors—elevation, slope, aspect, and topographic relief—were extracted from the digital elevation model (DEM) data. Elevation was obtained directly from the DEM; slope was calculated by determining the elevation change rate of adjacent pixels; aspect was calculated by determining the direction of maximum slope; and topographic relief was calculated by the ratio of the elevation range to the mean within a 300-meter radius moving window. Near-infrared and red bands were extracted from Sentinel-2 image data, and the Normalized Difference Vegetation Index (NDVI) was calculated. Land use type factors were extracted from land use classification data, categorized into seven types: forest, shrubland, construction land, cultivated land, grassland, unused land, and water area. Using ArcGIS spatial analysis tools, the Euclidean distances from each raster cell to the nearest road, water system, and fault were calculated, yielding three triggering factors: distance to road, distance to water system, and distance to fault.

[0007] (3) Sample Data Construction: Historical landslide data of the study area were obtained from the Resource and Environmental Science Data Platform of the Chinese Academy of Sciences as positive samples. After determining the number of positive samples, negative samples were collected in non-landslide areas at a ratio of 1.2-1.8. The collection of negative samples followed the following principles: the distance from any historical landslide point was greater than 500 meters to avoid overlap with the landslide-affected area; the slope of the location was greater than 5 degrees to exclude flat terrain; and the samples were evenly distributed in the value space of each evaluation factor to ensure the representativeness of the samples. The values ​​of the positive and negative samples on the nine evaluation factors were combined to form a complete training dataset.

[0008] (4) Data Preprocessing: The training dataset was randomly divided into a training set and a test set in a 7:3 ratio. Stratified sampling was used to ensure that the ratio of positive to negative samples in the training and test sets remained consistent with the overall population. Z-score standardization was performed on the nine evaluation factors, i.e., subtracting the mean and dividing by the standard deviation to unify the numerical range of each factor to a standard normal distribution with a mean of 0 and a standard deviation of 1. The standardization parameters (mean and standard deviation) were calculated from the training set, and the same standardization parameters were used for both the training and test sets to avoid data leakage.

[0009] (5) Multi-model training: Four models were trained using the Scikit-learn machine learning library. The logistic regression model used L2 regularization with a regularization coefficient of 0.01-0.1 and a maximum of 1000 iterations. The random forest model had 100 decision trees (range 80-120) and a maximum tree depth of 20 (range 15-25), using information gain (Gini coefficient) as the splitting criterion. The support vector machine model used radial basis function (RBF) kernel, with a penalty coefficient C of 1-10 and kernel parameter γ of 0.01-0.1, and enabled probability prediction. The multilayer perceptron model had two hidden layers with 64 and 32 neurons, respectively, using ReLU as the activation function, Adam as the optimization algorithm, a learning rate of 0.001, a maximum of 500 iterations, and enabled early stopping to prevent overfitting. After training on the training set, each model was used to make predictions on the test set.

[0010] (6) Model Evaluation and Selection: Six metrics were used to comprehensively evaluate model performance. Accuracy measures the proportion of correctly classified samples. Precision measures the proportion of actual landslides among samples predicted as landslides. Recall measures the proportion of actual landslide samples correctly identified. The F1 score is the harmonic mean of precision and recall, comprehensively reflecting model performance. The area under the ROC curve (ROC-AUC) measures the model's classification ability at different thresholds. The area under the PR curve (PR-AUC) is more sensitive to imbalanced data, reflecting the trade-off between precision and recall. Based on the six metrics, the random forest model was selected as the final prediction model when its F1 score was greater than 0.90 and its ROC-AUC was greater than 0.95. Simultaneously, the importance scores of each evaluation factor were calculated using the random forest model to identify the controlling factors.

[0011] (7) Susceptibility Assessment: The final prediction model is applied to all grid cells in the study area. The values ​​of the nine evaluation factors for each cell are input, and the landslide susceptibility probability value (between 0 and 1) is output. Based on the probability distribution characteristics and the natural breakpoint method, the study area is divided into four hazard levels: a susceptibility probability less than 0.3 indicates a low susceptibility zone, meaning the probability of landslides is very small; a susceptibility probability of 0.3-0.5 indicates a medium susceptibility zone, meaning the probability of landslides is relatively low; a susceptibility probability of 0.5-0.7 indicates a high susceptibility zone, meaning the probability of landslides is relatively high; and a susceptibility probability greater than 0.7 indicates an extremely high susceptibility zone, meaning the probability of landslides is very high. A susceptibility classification map is generated to provide a basis for geological disaster prevention and control.

[0012] The beneficial effects of this invention are: First, multi-source data integration. This invention integrates multi-source remote sensing data such as ASTER GDEM, Sentinel-2, CLCD, road network, water system network, and geological structure, comprehensively characterizing the features of the study area from multiple dimensions such as topography, vegetation cover, land use, infrastructure, and geological structure. Compared with single-source data source assessment methods, it provides richer information and a more comprehensive assessment.

[0013] Second, a scientific factor system. This invention constructs a scientific evaluation system containing nine evaluation factors, covering topographic factors (elevation, slope, aspect, topographic relief), environmental factors (NDVI, land use type), and triggering factors (distance from roads, water systems, and faults). The selection of factors is based on theoretical principles, and the calculation methods are standardized.

[0014] Third, multiple algorithms are compared and optimized. This invention simultaneously trains four algorithms: logistic regression, random forest, support vector machine, and multilayer perceptron. The optimal algorithm is selected through systematic comparison, avoiding the limitations of a single algorithm. Experiments show that random forest performs best in landslide susceptibility assessment, with an accuracy of 97% and an ROC-AUC of 0.973.

[0015] Fourth, a comprehensive evaluation system. This invention uses six metrics—accuracy, precision, recall, F1 score, ROC-AUC, and PR-AUC—to comprehensively evaluate model performance, which is more scientific and reliable than traditional methods that only use accuracy as a single evaluation metric.

[0016] Fifth, the application results are significant. The invention has achieved excellent results in Tengchong County, Yunnan Province. The random forest model F1 score reached 0.964, accurately identifying 78% of the landslide points concentrated in the eastern region. Feature importance analysis showed that the distance from the fault (0.713) and the distance from the road (0.486) were the main controlling factors, providing a scientific basis for geological disaster prevention and control in this region. Attached Figure Description

[0017] Figure 1 This is a technical flowchart of the landslide susceptibility assessment method in an embodiment of the present invention; Figure 2 This is a schematic diagram showing the location of the study area in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the extraction of digital elevation model data and derived factors in an embodiment of the present invention; Figure 4 This is a schematic diagram of land use classification data in an embodiment of the present invention; Figure 5 This is a schematic diagram of the normalized vegetation index distribution in an embodiment of the present invention; Figure 6 This is a schematic diagram of infrastructure data (road network and water system network) in an embodiment of the present invention; Figure 7 This is a schematic diagram of geological structural data (fault distribution) in an embodiment of the present invention; Figure 8 This is a schematic diagram of the spatial distribution of historical landslide points in an embodiment of the present invention; Figure 9 This is a diagram showing the relationship between evaluation factors and landslides (slope) in an embodiment of the present invention. Figure 10 This is an analysis diagram (elevation) showing the relationship between evaluation factors and landslides in an embodiment of the present invention. Figure 11 This is a diagram showing the relationship between evaluation factors and landslides (slope aspect) in an embodiment of the present invention. Figure 12 This is a diagram showing the relationship between evaluation factors and landslides (NDVI) in an embodiment of the present invention. Figure 13 This is a diagram showing the relationship between evaluation factors and landslides (topographic relief) in an embodiment of the present invention. Figure 14 This is a diagram showing the relationship between evaluation factors and landslides (land use) in an embodiment of the present invention. Figure 15 This is a diagram showing the relationship between evaluation factors and landslides (distance from the road) in an embodiment of the present invention. Figure 16 This is a diagram showing the relationship between evaluation factors and landslides (distance from water system) in an embodiment of the present invention. Figure 17 This is a diagram showing the relationship between evaluation factors and landslides (distance from fault) in an embodiment of the present invention. Figure 18 This is a performance comparison chart of four machine learning models in the embodiments of the present invention; Figure 19 This is a comparison chart of the ROC curves and PR curves of four machine learning models in this embodiment of the invention; Figure 20 This is a feature importance analysis diagram of the random forest model in an embodiment of the present invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0019] Example: Landslide Susceptibility Assessment in Tengchong County, Yunnan Province Tengchong County, located in southwestern Yunnan Province, lies within the collision zone of the Eurasian and Indian plates. Its complex geological structure, well-developed faults, and frequent earthquakes make it a high-risk area for landslides. The region ranges in altitude from 937 to 3700 meters, with a terrain that slopes from northwest to southeast. It has a low-latitude plateau monsoon climate, with annual precipitation of 1400-1600 mm, concentrated in the rainy season from May to October. This embodiment applies the method of the present invention to assess the landslide susceptibility of Tengchong County. The specific steps are as follows: Step 1: Acquisition of multi-source remote sensing data like Figure 2 As shown, Tengchong County is located between 24°38' and 25°52' north latitude and 98°05' and 98°46' east longitude, with a total area of ​​5,845 square kilometers. The following data were obtained in this embodiment: (1) ASTER GDEM V2 digital elevation model data, 30-meter resolution, obtained from the geospatial data cloud platform (http: / / www.gscloud.cn). The data has been processed by hole filling and smoothing.

[0020] (2) Sentinel-2 multispectral remote sensing image data, containing 13 bands with a resolution of 10-60 meters, obtained from the Copernicus Open Access Hub of the European Space Agency in 2022 cloudless imagery.

[0021] (3) The Wuhan University 30-meter resolution China Land Cover Dataset (CLCD 2022) was generated using random forest classification Landsat images and includes seven categories: forest, shrub, construction land, cultivated land, grassland, unused land and water.

[0022] (4) Road network vector data, obtained from the Rivermap website (Rivermap.cn), includes major highways, secondary highways and local roads.

[0023] (5) Vector data of the water system network, obtained from OpenStreetMap, including Longchuan River, Daying River and their tributaries.

[0024] (6) Geological structure data: 1:200,000 geological maps were obtained from the China Geological Survey Geological Cloud Platform (https: / / geocloud.cgs.gov.cn), and fault lines were digitally extracted.

[0025] Step 2: Evaluation Factor Extraction (1) Extraction of terrain factors. For example... Figure 3As shown, four topographic factors were extracted from ASTER GDEM data using ArcGIS 10.8.2 software. The elevation factor was directly obtained from the DEM, with a value range of 937-3700 meters and an average of 1381.5 meters. The slope factor was calculated using the Slope tool, with a value range of 0-70 degrees and an average of 30 degrees; 85.58% of the area has a slope within the 0-30 degree range. The aspect factor was calculated using the Aspect tool, with a value range of 0-360 degrees, categorized into nine types: North, Northeast, East, Southeast, South, Southwest, West, Northwest, and Flat. The topographic relief factor was calculated using the Focal Statistics tool, with a 300-meter radius circular window. The ratio of the elevation range (maximum minus minimum) to the mean within the window was calculated, with a value range of 0-550 meters.

[0026] (2) Extraction of normalized vegetation index. For example... Figure 5 As shown, Band 8 (near-infrared, 842nm) and Band 4 (red, 665nm) were extracted from Sentinel-2 imagery, and NDVI was calculated as (B8-B4) / (B8+B4). The NDVI value ranges from -0.5 to 1.0, with an average of 0.7. High NDVI values ​​(>0.7) correspond to forest areas, while low NDVI values ​​(<0.3) correspond to built-up areas and bare land.

[0027] (3) Extraction of land use type. For example... Figure 4 As shown, land use types were extracted from CLCD 2022 data. Land use in Tengchong County is primarily forest (82.8%), followed by cultivated land, construction land, and shrubs.

[0028] (4) Distance factor calculation. For example... Figure 6 and Figure 7 As shown, the Euclidean Distance tool in ArcGIS was used to calculate the straight-line distance from each 30-meter raster cell to the nearest road, river, and fault. The distance to the road ranges from 0 to 5000 meters, the distance to the river ranges from 0 to 3000 meters, and the distance to the fault ranges from 0 to 10000 meters.

[0029] Step 3: Sample Data Construction like Figure 8 As shown, historical landslide data for Tengchong County from 2010 to 2020 were obtained from the Resource and Environmental Science Data Platform of the Chinese Academy of Sciences. After ground verification and remote sensing image verification, 133 valid landslide points were confirmed as positive samples. The landslide points showed an asymmetrical distribution, with more in the east and fewer in the west, with 104 in the east (78%) and 29 in the west (22%), mainly distributed in the low-to-medium altitude area of ​​937-2200 meters.

[0030] 199 negative sample points were collected in non-landslide areas at a ratio of 1:1.5. The negative sample collection followed the following principles: (1) the distance from any historical landslide point was greater than 500 meters; (2) the slope of the location was greater than 5 degrees, excluding valleys and plains; (3) stratified random sampling was used in the value space of the 9 evaluation factors to ensure that the sample covered the typical value range of each factor.

[0031] The values ​​of 332 sample points (133 positive samples and 199 negative samples) on 9 evaluation factors were extracted to form a data table with 332 rows and 10 columns. The first 9 columns are the evaluation factor values, and the 10th column is the label (1 indicates landslide and 0 indicates non-landslide).

[0032] Step 4: Data Preprocessing The 332 samples were divided into a training set (232 samples) and a test set (100 samples) in a 7:3 ratio. A stratified sampling strategy was used for the split, employing the `train_test_split` function from Scikit-learn. The `stratify` parameter was set to the label column, and `random_state` was set to 42 to ensure reproducibility. After the split, the training set contained 93 positive samples (40.09%) and 139 negative samples (59.91%), while the test set contained 40 positive samples (40.00%) and 60 negative samples (60.00%), maintaining a consistent ratio of positive to negative samples.

[0033] Z-score standardization was performed on the nine evaluation factors. First, the mean μ and standard deviation σ of each factor in the training set were calculated. Then, the transformation formula z=(x-μ) / σ was applied to both the training and test sets. After standardization, the numerical distribution of each factor is a standard normal distribution with a mean of 0 and a standard deviation of 1, eliminating the influence of different factor dimensions and numerical ranges.

[0034] Step 5: Multi-model training Four machine learning models were trained on a standardized training set using the Python Scikit-learn library (version 1.0.2): (1) Logistic Regression: Parameters were set as max_iter=1000, solver='lbfgs', penalty='l2', C=0.1, random_state=42. After training, the accuracy on the test set was 54.00%, the F1 score was 0.512, and the ROC-AUC was 0.543. The performance was close to random guessing, indicating that the linear model could not effectively capture the complex nonlinear relationship between landslides and environmental factors.

[0035] (2) Random Forest: Parameters were set as follows: n_estimators=100, max_depth=20, min_samples_split=2, min_samples_leaf=1, random_state=42. After training, the accuracy on the test set was 97.00%, precision was 93.02%, recall was 100.00%, F1 score was 0.964, ROC-AUC was 0.973, and PR-AUC was 0.968, achieving optimal performance.

[0036] (3) Support Vector Machine: Parameters were set as kernel='rbf', C=1.0, gamma=0.01, probability=True, random_state=42. After training, the accuracy on the test set was 91.00%, precision was 56.41%, recall was 94.87%, F1 score was 0.696, and ROC-AUC was 0.956. Although the ROC-AUC was high, the F1 score was significantly lower than that of the random forest, indicating an imbalance between precision and recall.

[0037] (4) Multi-layer Perceptron: Parameters set: hidden_layer_sizes=(64,32), activation='relu', solver='adam', learning_rate_init=0.001, max_iter=500, early_stopping=True, random_state=42. After training, the accuracy on the test set was 82.00%, precision was 9.80%, recall was 97.44%, F1 score was 0.178, and ROC-AUC was 0.918. Although the recall was high, the precision was extremely low, and the F1 score was very poor, indicating that the sample size (332 samples) was insufficient for training the deep learning model.

[0038] Step 6: Model Evaluation and Selection like Figure 18 As shown, the comparison results of the four models across six performance metrics are as follows: Accuracy: RF (97.00%) > SVM (91.00%) > MLP (82.00%) > LR (54.00%) Precision: RF (93.02%) > LR (52.34%) > SVM (56.41%) > MLP (9.80%) Recall rate: RF (100.00%) > MLP (97.44%) > SVM (94.87%) > LR (51.23%) F1 score: RF (0.964) > SVM (0.696) > LR (0.512) > MLP (0.178) ROC-AUC: RF (0.973) > SVM (0.956) > MLP (0.918) > LR (0.543) PR-AUC: RF (0.968) > SVM (0.842) > MLP (0.723) > LR (0.534) like Figure 19 As shown, the ROC curve of Random Forest is closest to the top left corner, with an ROC-AUC of 0.973; the PR curve is closest to the top right corner, with a PR-AUC of 0.968. Considering all six metrics, Random Forest performs best or second best in all metrics, with an F1 score of 0.964 (>0.90) and an ROC-AUC of 0.973 (>0.95). Therefore, Random Forest is chosen as the final prediction model.

[0039] like Figure 20 As shown, the importance scores of each evaluation factor were calculated using a random forest model. Distance from the fault had the highest importance score (0.713), followed by distance from the road (0.486) and aspect (0.467), indicating that geological structure (faults) and human activities (road construction) are the main factors controlling landslide occurrence in Tengchong County. Topographic relief (0.239), slope (0.229), and NDVI (0.177) had moderate importance. Distance from the water system (0.163), elevation (0.099), and land use (0.089) had relatively low importance.

[0040] Step 7: Susceptibility Assessment The trained random forest model was applied to every 30-meter grid cell across the entire Tengchong County. The standardized values ​​of the nine evaluation factors for each cell were input, and the model output the landslide susceptibility probability value (between 0 and 1) for that cell. A raster map of the landslide susceptibility probability distribution in Tengchong County was generated.

[0041] Based on the distribution characteristics of susceptibility probability values ​​and the natural breakpoint method, Tengchong County is divided into four risk levels: Low-risk areas: The probability of landslides is <0.3. They are mainly distributed in the western high-altitude forest area and the eastern river valley plain area, accounting for 45.2% of the total area. There are 3 historical landslide sites (2.3%).

[0042] The most prone area has a landslide probability of 0.3-0.5, mainly distributed in the central hilly area, accounting for 28.7% of the total area, with 15 historical landslide points (11.3%).

[0043] High-risk areas: 0.5-0.7 probability of occurrence, mainly distributed near the eastern fault zone and along major roads, accounting for 18.6% of the total area, with 43 historical landslide points (32.3%).

[0044] Extremely high risk area: Risk probability > 0.7, mainly distributed in the eastern fault-dense area and road construction concentrated area, accounting for 7.5% of the total area, with 72 historical landslide points (54.1%).

[0045] The assessment results show that the high-risk and extremely high-risk areas, accounting for 26.1% of the total area, contain 86.4% of the historical landslide points, indicating that the model predictions are highly consistent with the actual situation. The spatial distribution of the high-risk and extremely high-risk areas highly overlaps with the distribution of fault zones and road networks, verifying the conclusions of the feature importance analysis.

[0046] The application of this invention in Tengchong County demonstrates that the landslide susceptibility assessment method based on multi-source remote sensing data and machine learning has high accuracy and reliability, providing a scientific basis for geological disaster prevention and control, land planning, and engineering construction. By identifying high-risk and extremely high-risk areas, it can guide disaster prevention and mitigation efforts such as optimizing the layout of geological disaster monitoring networks, relocating residents in dangerous areas, and implementing engineering measures in key areas.

[0047] The above description is only a preferred embodiment of the present invention. All equivalent changes and modifications made within the scope of the claims of the present invention should be included in the scope of the present invention.

Claims

1. A landslide susceptibility assessment method based on multi-source remote sensing data and machine learning, characterized in that, Includes the following steps: (1) Acquisition of multi-source remote sensing data: acquire digital elevation model data, multispectral remote sensing image data, land use classification data, road network data, water system data and geological structure data of the study area; (2) Evaluation factor extraction: Elevation, slope, aspect and topographic relief factors are extracted from the digital elevation model data; Normalized vegetation index is calculated from the multispectral remote sensing image data; Land use type factors are extracted from the land use classification data; distance factors from roads, distance from water systems, and distance from faults are calculated from the road network data, water system data, and geological structure data, respectively. (3) Sample data construction: Collect historical landslide points as positive samples, and collect negative samples in non-landslide areas according to the spatial distribution characteristics of landslide points. The number of negative samples is 1.2-1.8 times the number of positive samples. Combine the evaluation factor data of positive and negative samples to form a training dataset. (4) Data preprocessing: The training dataset is divided into a training set and a test set in a 7:3 ratio. The partitioning process adopts a stratified sampling strategy to maintain a consistent ratio of positive and negative samples. The evaluation factor data of the training set and the test set are standardized. (5) Multi-model training: Train the logistic regression model, random forest model, support vector machine model and multilayer perceptron model respectively. The random forest model has 80-120 decision trees, the support vector machine model uses radial basis kernel function, and the multilayer perceptron model contains 2 hidden layers. (6) Model evaluation and selection: The performance of each model is evaluated using six indicators: accuracy, precision, recall, F1 score, area under the ROC curve, and area under the PR curve. The model with the best overall performance is selected as the final prediction model. (7) Susceptibility assessment: The landslide susceptibility of each grid cell in the study area is predicted using the final prediction model, and a susceptibility probability distribution map is generated; Based on the probability of occurrence, the study area is divided into four levels: low-risk area, medium-risk area, high-risk area, and extremely high-risk area.

2. The landslide susceptibility assessment method according to claim 1, characterized in that, The digital elevation model data has a resolution of 30 meters and is sourced from ASTER GDEM V2; the multispectral remote sensing image data is sourced from the European Space Agency's Sentinel-2 satellite; and the land use classification data is sourced from a 30-meter resolution Chinese land cover dataset.

3. The landslide susceptibility assessment method according to claim 1, characterized in that, The method for calculating the terrain relief factor is as follows: a moving window with a radius of 300 meters is set, and the ratio of the range to the average elevation within the window is calculated.

4. The landslide susceptibility assessment method according to claim 1, characterized in that, The formula for calculating the Normalized Difference Vegetation Index (NDVI) is: NDVI = (NIR - R) / (NIR + R), where NIR is the reflectance in the near-infrared band and R is the reflectance in the red band.

5. The landslide susceptibility assessment method according to claim 1, characterized in that, The collection of negative samples follows these principles: the distance between the negative sample point and any historical landslide point is greater than 500 meters; the slope of the location of the negative sample point is greater than 5 degrees; and the negative sample points are evenly distributed in the value space of each evaluation factor.

6. The landslide susceptibility assessment method according to claim 1, characterized in that, The standardization process uses the Z-score standardization method, calculated as: z = (x - μ) / σ, where x is the original data, μ is the mean of the training set, and σ is the standard deviation of the training set; the test set is standardized using the mean and standard deviation of the training set.

7. The landslide susceptibility assessment method according to claim 1, characterized in that, The logistic regression model uses L2 regularization with a regularization coefficient of 0.01-0.1; the maximum tree depth of the random forest model is 15-25; the penalty coefficient C of the support vector machine model is 1-10, and the kernel function parameter γ is 0.01-0.1; the number of neurons in the two hidden layers of the multilayer perceptron model is 64 and 32, respectively.

8. The landslide susceptibility assessment method according to claim 1, characterized in that, During the model evaluation process, when the F1 score of the random forest model is greater than 0.90 and the area under the ROC curve is greater than 0.95, the random forest model is selected as the final prediction model.

9. The landslide susceptibility assessment method according to claim 1, characterized in that, The risk classification criteria are as follows: a probability of occurrence less than 0.3 is a low-risk area, a probability of occurrence between 0.3 and 0.5 is a medium-risk area, a probability of occurrence between 0.5 and 0.7 is a high-risk area, and a probability of occurrence greater than 0.7 is an extremely high-risk area.

10. The landslide susceptibility assessment method according to any one of claims 1-9, characterized in that, The method also includes a feature importance analysis step: using a random forest model to calculate the importance score of each evaluation factor, identifying the main controlling factors that have the most significant impact on landslide occurrence, and providing a basis for geological disaster prevention and control strategies.

Images