Landslide susceptibility evaluation method based on deep neural network and considering time sequence InSAR and time sequence rainfall

Abstract

The invention provides a landslide susceptibility evaluation method based on a deep neural network and considering time sequence InSAR and time sequence rainfall, belongs to the technical field of geological disasters and geographic information, and particularly relates to a landslide susceptibility prediction method based on the deep neural network. The method comprises the following steps: establishing a buffer area through landslide points, selecting a research area to randomly generate non-landslide points, screening related static characteristic factors by using a Pearson's correlation coefficient matrix and a VIF method, obtaining GCP points by using PS-InSAR, introducing the generated points as SBAS-InSAR processing parameters, generating dynamic characteristic factors of surface deformation, and calculating the deformation of the surface deformation. The time sequence average rainfall of the region is obtained through spatial interpolation, the rainfall before earth surface deformation is obtained through data processing codes to serve as another dynamic characteristic factor, finally evaluation is conducted through the constructed ResNetconvLSTMUnet, and the susceptibility index is divided into five grades; according to the method, the dual-time-sequence dynamic factors are creatively fused, the spatial-temporal feature extraction capability is enhanced, the data redundancy is reduced, the sample selection is reasonable, the evaluation precision is high, and scientific support can be provided for landslide disaster prevention and reduction.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of geological disasters and geographic information technology, specifically involving a landslide susceptibility prediction method, and more particularly a landslide susceptibility assessment method based on deep neural networks that integrates time-series InSAR surface deformation data, time-series rainfall data and static disaster-causing factors. It is applicable to landslide risk assessment, land spatial planning and disaster prevention and mitigation decision-making in mountainous areas, geologically fragile areas and earthquake-prone areas. Background Technology

[0002] Landslides, as one of the major natural disasters worldwide, pose a serious threat to human life and property, regional ecological environment stability, and sustainable socio-economic development. In the field of landslide disaster prevention and control research, landslide susceptibility mapping (LSM) is a core technical means for disaster early warning and risk management. Its core objective is to quantitatively depict the instability tendency of slopes in a specific area, providing a scientific and reliable basis for decision-making in land spatial planning, disaster emergency response plan formulation, and engineering protection measure deployment, which is of great significance for improving regional landslide disaster prevention and mitigation capabilities.

[0003] Currently, landslide susceptibility assessment methods can be mainly divided into two categories: qualitative assessment models and quantitative assessment models. Qualitative assessment models, typically represented by expert experience and knowledge-driven methods, were widely used in the early stages when technological conditions were limited. These models rely on the subjective judgment of experts in the field regarding factors such as geological conditions, topography, and hydrological characteristics of the study area to classify susceptibility levels. However, they have significant limitations: on the one hand, the assessment results are highly dependent on the experts' personal experience, exhibiting strong subjectivity and poor repeatability; on the other hand, large-scale regional assessments involve extremely high labor costs, and the assessment conclusions are difficult to verify with objective data, thus failing to meet the needs of modern high-precision, large-scale landslide susceptibility assessment.

[0004] With the development of mathematical statistics theory and computer technology, quantitative evaluation models have gradually become the mainstream in landslide shaving (LSM) research. Their core is to quantify the correlation between hazard-causing factors and landslide occurrence through mathematical models, mainly including physical-driven models and data-driven models. In recent years, the rapid development of artificial intelligence technology has promoted the widespread application of data-driven models in LSM. Machine learning (ML) techniques, represented by random forests, support vector machines, and Bayesian networks, have significantly improved evaluation accuracy due to their ability to fit nonlinear relationships. However, ML models exhibit regional adaptability differences; different models often show biases in their predictions for the same region. Therefore, researchers need to compare multiple models to select the optimal one, increasing the complexity and cost of the method's application.

[0005] Meanwhile, deep learning (DL) models, with their stronger feature learning capabilities, are gradually being applied to landslide identification and LSM (Landslide Slope Management). Algorithms such as convolutional neural networks (CNNs), deep neural networks (DNNs), and gated recurrent unit (GRU) networks have become research hotspots. However, existing DL models still have technical bottlenecks: CNNs have high computational complexity and long training time when processing large-scale data, and convolution and pooling operations are prone to losing local feature information, affecting the identification accuracy of small-scale landslide areas; DNN models are prone to overfitting during training and have poor adaptability to small sample datasets, making it difficult to meet the evaluation needs of data-scarce regions; although GRU networks can process time-series data, their ability to capture complex spatiotemporal coupling relationships is limited, and they cannot fully explore the dynamic correlations between multiple factors. Therefore, developing DL models that balance accuracy, efficiency, and adaptability has become a key direction for overcoming the technical bottlenecks of existing LSM.

[0006] Furthermore, existing LSM studies have significant shortcomings in the selection of causative factors: most studies only use static environmental factors as model inputs, such as geological factors (lithology, fault distribution), topographic factors (slope, aspect), hydrological factors (distance from rivers), and human engineering activities (distance from roads, land use type), neglecting the triggering effect of dynamic factors in the time dimension on landslide occurrence. Although satellite-based temporal interferometry (InSAR) technology has been widely used for surface deformation monitoring, and both rainfall dynamics and landslide deformation characteristics are key dynamic factors determining slope stability, existing technologies mostly analyze them as independent factors, failing to achieve the organic integration of the two types of dynamic information. This results in models that cannot comprehensively reflect the spatiotemporal evolution process of landslide incubation, development, and occurrence, making it difficult to further improve the scientific rigor and reliability of LSM.

[0007] In summary, most methods rely solely on static factors, failing to integrate temporal InSAR surface deformation and rainfall data, thus failing to reflect the disaster-causing logic of both static background and dynamic evolution. Traditional machine learning has limited fitting capabilities for nonlinear spatiotemporal relationships; single deep learning models are prone to losing local features, and GRU struggles to capture complex spatiotemporal coupling relationships, resulting in low accuracy in small-scale landslide identification and poor efficiency in large-area evaluation. Developing a deep neural network framework that can effectively integrate static factors, temporal rainfall factors, and InSAR deformation factors, while possessing high accuracy and efficiency, has become a pressing technical requirement in the current LSM field. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a landslide susceptibility assessment method based on deep neural networks and taking into account temporal InSAR and temporal rainfall. By integrating static disaster-causing factors, temporal InSAR surface deformation dynamic factors, and temporal rainfall dynamic factors, a ResNet_convLSTM_Unet model with both spatial feature extraction and temporal dependency capture capabilities is constructed to achieve high-precision assessment of landslide susceptibility and provide more scientific and reliable technical support for landslide prevention and mitigation work.

[0009] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: a landslide susceptibility assessment method based on deep neural networks and taking into account temporal InSAR and temporal rainfall, comprising the following steps:

[0010] 1. A landslide susceptibility assessment method based on deep neural networks and taking into account temporal InSAR and temporal rainfall, characterized by comprising the following steps:

[0011] S1. Data preparation: Obtain satellite imagery data and rainfall data for a specific time period within the study area. Select 15 static disaster-causing factors based on the conditions for landslide occurrence and collect historical landslide data within the study area.

[0012] S2. Negative sample selection in the study area: Use the selected positive samples to generate the corresponding landslide buffer zone, and then generate random negative samples in the study area.

[0013] S3. Selection of static characteristic factors of the study area: The Pearson correlation coefficient matrix is ​​used to initially select and eliminate static factors with excessively high absolute values, i.e., relatively redundant static factors. Then, the VIF method is used to perform multicollinearity test to further eliminate static factors with excessively high values. Finally, 9 static factors are selected and special factors are processed separately.

[0014] S4. Dynamic characteristic factors of the study area are processed and combined with static characteristic factors for fusion evaluation, and then the data is integrated to obtain the dataset;

[0015] S5. Construct the ResNet_convLSTM_Unet model, extract static spatial features and dynamic spatiotemporal features through a parallel encoder, and obtain the landslide susceptibility evaluation model through feature fusion and decoding. Then, select the optimal parameters to obtain the ResNet_convLSTM_Unet landslide susceptibility evaluation model.

[0016] S6. Landslide susceptibility assessment and verification: The landslide susceptibility assessment model is used to predict the landslide susceptibility of the study area. The natural discontinuity method is used to divide the landslide susceptibility probability value into five susceptibility levels, and finally the spatial distribution of earthquake landslide disasters in the disaster area is obtained.

[0017] Further, in step S1, the satellite image data selected is SAR image data from the Sentinel-1 satellite, with 22 scenes of the study area selected, VV / VH dual polarization, and a spatial resolution of 10m; the rainfall data selected is the hourly average rainfall observation data from meteorological stations in and around the study area, with a time span consistent with the satellite image data, and a data accuracy of not less than 0.1mm; the 15 static disaster-causing factors include slope aspect, elevation, distance from fault, land use rate, plane curvature profile curvature, distance from river, distance from road, slope, surface roughness, lithological distribution, normalized building index (NDBI), normalized vegetation index (NDVI), normalized water index (NDWI), and surface relief; the historical landslide point data includes accurate geographical coordinates, landslide occurrence time, and landslide scale information.

[0018] Furthermore, in step S2, the radius of the landslide buffer zone is 100m; negative samples must not be within any landslide buffer zone and the straight-line distance to the nearest positive sample must not be less than the buffer zone radius; the ratio of the number of negative samples to the number of positive samples is controlled at 1:1.

[0019] Furthermore, in step S3, the threshold for Pearson correlation coefficient matrix analysis is set to |r|≥0.9, and redundant factors with absolute correlation coefficient values ​​exceeding this threshold are eliminated; the threshold for the VIF method is set to 10, and factors with VIF values ​​greater than 10 are eliminated; the special factors include categorical variable factors such as lithology and land use type, and continuous variable factors such as slope and altitude. Categorical variable factors are processed using one-hot encoding, and continuous variable factors are processed using normalization.

[0020] Further, in step S4, the dynamic characteristic factors include temporal InSAR surface deformation factors and temporal rainfall factors; the temporal InSAR surface deformation factors are obtained by selecting GCP points through PS-InSAR and then processing them through SBAS-InSAR to obtain the annual average surface deformation rate and the maximum cumulative deformation; the temporal rainfall factors are spatially interpolated using the Kriging interpolation method, and then the cumulative rainfall during the flood season and the cumulative rainfall corresponding to the maximum number of consecutive rainy days are statistically analyzed.

[0021] Furthermore, step S5 specifically includes the following steps:

[0022] At the bottleneck layer of the network, S51 fuses the high-level static spatial features output by the ResNet encoder with the final dynamic spatiotemporal memory output by the ConvLSTM encoder using a concatenate cascade structure.

[0023] The feature map fused by S52 is used as the starting input of the UNet decoder;

[0024] In each upsampling layer of the decoder, S53 fuses the decoded features of the current layer with the corresponding scale features of the ResNet encoder through skip connections; at the end of the decoder, a pixel-level landslide probability map is output through 1×1 convolution and sigmoid activation function.

[0025] Further, S51 includes: the encoder has two parallel branches, B is the batch size, T is the time step, P is the spatial size, C is the number of channels, i is the number of levels, and E is the encoding feature; the static branch input tensor (B, C) static (B×T, C) is formed after being copied T times in the time dimension. static The dynamic feature map {E} is prepared by combining features from the P, P) domains and outputting a multi-scale static feature map through pre-trained ResNet multi-layer encoding. S,i The shape is (B, C) i P / 2 i P / 2 i ); Dynamic branch input tensor (B, T, C) dynamic After adjusting the dimensions, the vectors (P, P) flow through a ConvLSTM encoding block to capture spatiotemporal features and output a dynamic feature map {E}. D,i The shape is (B, C) i P / 2 i P / 2 i ).

[0026] Furthermore, in step S6, the five susceptibility levels are extremely low susceptibility, low susceptibility, moderate susceptibility, high susceptibility, and extremely high susceptibility; the evaluation results are verified for accuracy using ROC curves and AUC values, with an AUC value ≥ 0.85.

[0027] Compared with the prior art, the present invention has the following significant advantages:

[0028] 1. Innovative integration of dual time-series dynamic factors to comprehensively reflect landslide disaster-causing mechanisms: This invention is the first to use time-series InSAR surface deformation data and time-series rainfall data as dynamic disaster-causing factors, organically integrating them with traditional static factors. This breaks through the limitations of existing methods that rely solely on static factors, and can more comprehensively reflect the interaction between static background conditions and dynamic triggering factors during landslide formation, thereby improving the scientificity and rationality of the evaluation results.

[0029] 2. Constructing a novel deep neural network model to enhance spatiotemporal feature extraction capabilities: The ResNet_convLSTM_Unet model designed in this invention integrates the residual connections of ResNet to solve the gradient vanishing problem in deep networks; the temporal feature capture of convLSTM to mine the temporal dependencies of dynamic factors; and the fine segmentation of Unet to improve spatial prediction accuracy. These three advantages enable it to simultaneously and efficiently extract the spatial features of static factors and the spatiotemporal coupling features of dynamic factors, significantly outperforming traditional machine learning models and single-structure neural network models.

[0030] 3. The static factor selection strategy is scientific, reducing data redundancy and multicollinearity: Through a two-step screening using the Pearson correlation coefficient matrix and the VIF method, redundant static factors are effectively eliminated, reducing data dimensionality and model training complexity. At the same time, the impact of multicollinearity among factors on the evaluation results is avoided, thus improving the stability and interpretability of the model.

[0031] 4. Reasonable sample selection ensures model training effect: The negative sample selection method with slope buffer zone constraint is adopted to ensure the spatial independence of positive and negative samples and the balance of the sample set. It avoids model overfitting or prediction bias caused by improper sample selection and provides data guarantee for high-precision model training.

[0032] 5. High accuracy of evaluation results and significant application value: The evaluation method of this invention can output high-precision spatial distribution results of landslide susceptibility, with clear classification of susceptibility levels. It can provide scientific and reliable technical support for landslide disaster risk early warning, land spatial planning, engineering construction site selection, and disaster prevention and mitigation plan formulation, and has important practical application value and social and economic benefits. Attached Figure Description

[0033] Figure 1 This is an overall flowchart of the landslide susceptibility assessment method of the present invention;

[0034] Figure 2 This is a graph showing the results of the Pearson correlation coefficient test on the characteristic factors;

[0035] Figure 3 This is a graph showing the results of a multiple linearity test on the remaining eigenfactors using VIF.

[0036] Figure 4 This is a schematic diagram of the ResNet_convLSTM_Unet model structure;

[0037] Figure 5 It is a graph showing the results of analysis of multiple test indicators;

[0038] Figure 6 This is a graph showing the results of the AUC index analysis;

[0039] Figure 7 This is a map showing the susceptibility partitioning results for each model in the study area. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0042] A landslide susceptibility assessment method based on deep neural networks and taking into account temporal InSAR and temporal rainfall, such as Figure 1 As shown, it includes:

[0043] S1. Data Preparation: Satellite imagery data for the region was obtained using SAR imagery from the Sentinel-1 satellite, with 22 scenes selected. The images were VV / VH dual polarization and a spatial resolution of 10m. Rainfall data was obtained from hourly average rainfall observations at meteorological stations within and around the study area, maintaining the same time span as the satellite imagery data, with a data accuracy of no less than 0.1mm. Fifteen static disaster-causing factors were included: slope aspect, elevation, distance from fault, land use rate, planar curvature profile curvature, distance from river, distance from road, slope, surface roughness, lithological distribution, normalized building index (NDBI), normalized vegetation index (NDVI), normalized water index (NDWI), and surface relief. Historical landslide point data included accurate geographical coordinates, landslide occurrence time, and landslide scale information.

[0044] S2. Negative sample selection in the study area: Use the selected positive samples to generate a landslide buffer zone with a radius of 100m; negative samples must not be within any landslide buffer zone and their straight-line distance from the nearest positive sample must not be less than the buffer zone radius; the ratio of the number of negative samples to the number of positive samples must be controlled at 1:1.

[0045] S3. Selection of static characteristic factors of the study area: The Pearson correlation coefficient matrix is ​​used to initially select and eliminate static factors with excessively high absolute values, i.e., relatively redundant static factors. Then, the VIF method is used to perform multicollinearity test to further eliminate static factors with excessively high values. Finally, 9 static factors are selected and special factors are processed separately.

[0046] S4. Dynamic feature factors of the study area are processed and combined with static feature factors for fusion evaluation. Then, the data is integrated to obtain the training dataset.

[0047] S5. Construct the ResNet_convLSTM_Unet model, extract static spatial features and dynamic spatiotemporal features through a parallel encoder, and obtain the landslide susceptibility evaluation model through feature fusion and decoding. Then, select the optimal parameters to obtain the ResNet_convLSTM_Unet landslide susceptibility evaluation model.

[0048] S6. Landslide susceptibility assessment and verification: The landslide susceptibility assessment model is used to predict the landslide susceptibility of the study area. The natural discontinuity method is used to divide the landslide susceptibility probability value into five susceptibility levels, and finally the spatial distribution of earthquake landslide disasters in the disaster area is obtained.

[0049] Specifically, S1 includes the following steps:

[0050] (1) Based on the disaster-causing mechanism, disaster-causing pattern, and disaster-causing environment of co-seismic landslides in complex mountainous areas, disaster-causing factors are selected and then extracted. Specifically, topographic disaster-causing factors are extracted from DEM data using ArcGIS Pro software, including slope, aspect, plane curvature, profile curvature, surface relief, and surface roughness. Geological environmental data such as stratigraphy and faults are extracted from 1:500,000 geological maps. NDVI is calculated based on red and near-infrared bands in Landsat optical remote sensing images, NDWI is calculated based on green and short-wave infrared bands, and NDBI is calculated based on near-infrared and short-wave infrared bands. Relevant regional land cover data are extracted from global land cover data, relevant regional water systems are extracted from national water system data as environmental factors, and relevant regional road data are extracted from national road network data.

[0051] (2) Perform Euclidean distance processing on the fault, water system and road data to obtain the distance data from the fault, water system and road respectively. Convert the distance data from the fault, water system and road and other disaster-causing factors in S12 into rectangular grids to obtain disaster-causing factor grid data.

[0052] Specifically, S3 includes the following steps:

[0053] (1) The Pearson correlation coefficient is used to measure the strength of the linear relationship between pairwise feature factors. Based on the correlation coefficient matrix, highly correlated feature factor pairs are selected. Then, based on the previous importance ranking, one of the highly correlated factors is removed. The threshold is set to |r|≥0.9. Redundant factors with absolute correlation coefficient values ​​exceeding this threshold are removed to complete the initial screening. The formula is:

[0054]

[0055] In the formula, , The two factors are respectively in the th The attribute values ​​of each sample point , These are the means of the two factors, The total number of samples, Pearson correlation coefficient;

[0056] (2) Perform a global collinearity test on the remaining feature factors. Calculate the VIF value of each feature factor. A higher VIF value indicates stronger collinearity among factors. Set the threshold for the VIF method to 10 and remove factors with VIF values ​​greater than 10. The formula is as follows:

[0057] In the formula, For the first The coefficient of determination of one factor on all other factors by multiple linear regression.

[0058] (3) Resample all raster data to a uniform spatial resolution and coordinate system. Special factors include categorical variables such as lithology and land use type, and continuous variables such as slope and altitude. Categorical variables are processed using one-hot encoding, and continuous variables are processed using normalization.

[0059] Specifically, S4 includes the following steps:

[0060] (1) Using the SARScape plugin of ENVI software, the SAR image data of Sentinel-1 satellite is combined with DEM data to complete the PS-InSAR interferometric workflow. PS is used for the first inversion and the second inversion. The geocoding is used to obtain GCP points. Then the SBAS-InSAR interferometric workflow is completed. The orbit refinement and normalization are carried out using the GCP points obtained by PS-InSAR. The first inversion and the second inversion are used. The geocoding is used to obtain the annual average surface deformation rate and the maximum cumulative deformation.

[0061] (2) Spatial interpolation of time-series rainfall factors at multiple meteorological stations was performed using the spatial analysis tool of ArcGIS Pro software, and then the cumulative rainfall during the flood season and the cumulative rainfall corresponding to the maximum number of consecutive rainy days were statistically analyzed.

[0062] (3) The obtained dynamic features are integrated with the selected static feature factors as the dataset.

[0063] Specifically, S5 includes the following steps:

[0064] (1) In the input layer, the static factor X static Its characteristic does not change over time, and its shape is (B, C). static(,H,W), where B is the batch size, C static H represents the number of static feature channels, and H and W represent the height and width of the data block. Similarly, the dynamic factor X... dynamic Its characteristic does not change over time, and its shape is (B, C). dynamic (,H,W), where C dynamic The static feature channel number is denoted as t. To combine spatial and temporal information, the model iterates through T time steps. At each time step t, the static factor tensor X is... static The dynamic factor X at the current time step t dynamic The pieces are then joined together, ultimately forming a sequence (B, C). dynamic+static The H, W) shape input network model.

[0065] (2) In the temporal coding layer, a pre-trained ResNet-18 structure is used, which is executed cyclically over T time steps to extract the deep spatial features at each time step. The feature sequence S output by the deepest layer of the encoder over T time steps is S = (e 3,1 , e 3,2 ,…, e 3,T The data are fed into ConvLSTM cells to capture the spatiotemporal dependence of landslide evolution.

[0066] (3) At the bottleneck layer, the ConvLSTM unit uses a series of gating mechanisms (input gate i) t Forgotten Gate t Output gate o t To update its internal cell state C t and hidden state H t The final output H of the bottleneck layer T This is the hidden state at the last time step t=T, encapsulating all the spatiotemporal information from the T time steps. It is then processed through a 1×1 convolution and a Dropout layer for channel adjustment and regularization to obtain x. bottle .

[0067] (4) At the decoder layer, the U-Net architecture is adopted, and the spatial resolution is gradually restored through a series of upsampling and skip connections to achieve pixel-level prediction.

[0068] (5) Output layer: The highest resolution feature map d0 of the decoder is fed into a 1×1 convolutional layer to compress the feature channels to 1 and generate the final landslide susceptibility prediction map.

[0069] The following description, in conjunction with the accompanying drawings, further illustrates this embodiment:

[0070] Figure 2 shows the Pearson correlation coefficient test results for 15 static factors. Highly correlated factor pairs such as "surface relief - slope" (r=0.95) and "NDVI - NDWI" (r=-0.97) can be clearly identified, providing a basis for eliminating redundant factors. However, slope is the main factor and is retained.

[0071] As shown in Figure 3, the VIF test results show that the VIF values ​​of factors such as "distance from the river" and "NDWI" exceed 10 and need to be removed.

[0072] As shown in Figure 4, the ResNet_convLSTM_Unet model extracts static and dynamic features through a dual-branch encoder, and after fusion through a bottleneck layer, outputs the prediction results through a U-Net decoder.

[0073] As shown in Figure 6, the validation set AUC value gradually increased and stabilized above 0.85 during model training, indicating that the model performance was good.

[0074] As shown in Figure 7, the extremely high susceptibility areas in the study area are mainly concentrated along the fault line, in areas with high slope and high rainfall, which is highly consistent with the actual landslide distribution.

[0075] The embodiments described above merely illustrate specific implementation methods of this application, and while the descriptions are detailed and specific, they should not be construed as limiting the scope of protection of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the technical solution of this application, and these modifications and improvements all fall within the scope of protection of this application.

Claims

1. A landslide susceptibility evaluation method based on a deep neural network and taking into account time-series InSAR and time-series rainfall, characterized in that, The method comprises the following steps: S1, data preparation, obtaining satellite image data in a specific time period in the study area and rainfall data in the same period, selecting 15 static disaster-causing factors according to the occurrence conditions of landslide disasters, and collecting historical landslide point data in the study area; S2, selection of negative samples in the study area, using the selected positive samples (historical landslide points) to generate corresponding landslide buffer zones, and then generating random negative samples in the study area; S3, optimization of static characteristic factors in the study area, using the Pearson correlation coefficient matrix to initially optimize and eliminate static factors with absolute values that are too high and are relatively redundant, then using the VIF method to perform multiple collinearity test, further eliminating static factors with high values, and finally selecting 9 static factors and processing the special factors separately; S4, processing of dynamic characteristic factors in the study area, combining static characteristic factors for fusion evaluation, and then integrating data to obtain a data set; S5, constructing a ResNet_convLSTM_Unet model, extracting static spatial features and dynamic spatio-temporal features through parallel encoders, and obtaining a landslide susceptibility evaluation model through feature fusion and decoding, and then selecting the best parameters to obtain a ResNet_convLSTM_Unet landslide susceptibility evaluation model; S6, landslide susceptibility evaluation and verification, using the landslide susceptibility evaluation model to predict the landslide susceptibility of the study area, using the natural breakpoint method to divide the landslide susceptibility probability value into five susceptibility grades, and finally obtaining the overall spatial distribution of the disaster area of the earthquake landslide disaster.

2. The method of claim 1, wherein: In step S1, the satellite image data is selected from SAR image data of Sentinel-1 satellite, 22 scenes of VV / VH dual polarization with a spatial resolution of 10m are selected; the rainfall data is selected from the time-averaged rainfall observation data of the meteorological stations in the study area and the surrounding area, the time span is consistent with the satellite image data, and the data accuracy is not less than 0.1mm; the 15 static disaster-causing factors include slope direction, elevation, distance from fault, land use rate, plane curvature, profile curvature, distance from river, distance from road, slope, ground roughness, lithology distribution, normalized building index (NDBI), normalized vegetation index (NDVI), normalized water body index (NDWI), and ground undulation; the historical landslide point data includes accurate geographic position coordinates, landslide occurrence time, and landslide scale information.

3. The method of claim 1, wherein: In step S2, the radius of the landslide buffer zone is 100m; the negative samples need to meet the conditions of not being in any landslide buffer zone and the straight-line distance from the nearest positive sample being not less than the buffer zone radius; the ratio of the number of negative samples to the number of positive samples is controlled at 1:

1.

4. The method of claim 1, wherein: In step S3, the threshold of Pearson correlation coefficient matrix analysis is set to |r|≥0.9, and the redundant factors with the absolute value of the correlation coefficient exceeding the threshold are removed; the threshold of VIF method is set to 10, and the factors with the VIF value greater than 10 are removed; the special factors include classified variable type factors such as lithology and land use type, and continuous variable type factors such as slope and altitude, the classified variable type factors are processed by one-hot encoding, and the continuous variable type factors are processed by normalization.

5. The method of claim 1, wherein: In step S4, the dynamic characteristic factors include time-series InSAR ground deformation factors and time-series rainfall factors; the time-series InSAR ground deformation factors are obtained by selecting GCP points through PS-InSAR and then processing the annual average ground deformation rate and the maximum cumulative deformation amount through SBAS-InSAR; The time-series rainfall factors are spatially interpolated through Kriging interpolation, and then the cumulative rainfall in the flood season and the cumulative rainfall corresponding to the maximum continuous rainfall days are calculated.

6. The method of claim 1, wherein step S5 specifically comprises the following steps: S51, at the bottleneck layer of the network, the high-level static spatial features output by the ResNet encoder are fused with the final dynamic spatio-temporal memory output by the ConvLSTM encoder using a concatenate cascade structure; S52, the fused feature map is used as the starting input of the UNet decoder; S53, at each upsampling layer of the decoder, the decoding features of the current layer are fused with the features of the corresponding scale of the ResNet encoder through skip connections; The decoder outputs a pixel-level landslide occurrence probability map through 1x1 convolution and Sigmoid activation function.

7. The method of claim 6, wherein, The S51 includes: the encoder contains two parallel branches, B is batch size, T is time step, P is spatial size, C is channel number, i is level number, E is encoding feature; the static branch input tensor (B, C static , P, P) is copied T times in time dimension to form (BxT, C static , P, P) to prepare for dynamic feature combination, through pre-training ResNet multi-layer encoding, output multi-scale static feature map {E S,i} shape is (B, C i , P / 2 i , P / 2 i ); the dynamic branch input tensor (B, T, C dynamic , P, P) is adjusted in dimension to flow through the ConvLSTM encoding block, captures the space-time feature, and outputs the dynamic feature map {E D,i} shape is (B, C i , P / 2 i , P / 2 i ).

8. The method of claim 1, wherein: In step S6, the five susceptibility levels are extremely low susceptibility, low susceptibility, medium susceptibility, high susceptibility, and extremely high susceptibility; the evaluation results are verified by the ROC curve and the AUC value, and the AUC value is greater than or equal to 0.85.

Images