Landslide susceptibility prediction method based on environmental factor importance dynamic evaluation

Abstract

The invention relates to a landslide susceptibility prediction method based on dynamic evaluation of environmental factor importance, and the method comprises the steps: collecting landslide and non-landslide samples and environmental factor data, and constructing a data set; and constructing a landslide susceptibility prediction model, inputting the data set into the training model to obtain a landslide susceptibility prediction result, and finally rating the landslide susceptibility of the region and performing visualization processing on the prediction result. The method comprises the following steps: constructing an internal environment factor model comprising a convolution unit, an SE layer and an LSTM unit, and mining the feature representation of an internal environment factor; an external driving environment factor model comprising a convolution unit, an SE layer, a Transform module and a SeqPool layer is constructed, and feature information of external driving environment factors is captured and analyzed; and the feature representation of the internal environment factor model and the feature information of the external driving environment factor model are subjected to weighted fusion through a self-attention mechanism, so that the main driving factors of the occurrence of the landslide disaster can be effectively identified, the prediction error is effectively reduced, and the prediction performance of the landslide prediction model is improved.

Description

Technical Field

[0001] The invention relates to the technical field of geological disaster risk analysis, and in particular to a landslide susceptibility prediction method based on dynamic evaluation of environmental factor importance. Background Art

[0002] Landslide is one of the most destructive natural phenomena in the world. It occurs when the rock and soil in the slope lose stability under the action of gravity and slide downward along a certain sliding surface. The occurrence of landslide is usually accompanied by the movement of a large amount of soil and rock, which has a significant impact on the natural environment and human society. It poses a serious threat to human life, property and regional sustainable development.

[0003] Landslide susceptibility prediction is a quantitative or qualitative assessment and prediction of the likelihood of landslides in a certain area under certain geological conditions. It is based on the assumption that landslides are more likely to occur in places with conditions similar to those that have occurred in the past. Therefore, susceptibility prediction can be used to predict the potential locations of future landslides. However, historical landslide inventories can only reveal the actual distribution of historical landslides; landslide susceptibility prediction results can provide indispensable information for regional landslide risk assessment and the design of disaster mitigation measures.

[0004] At present, the models for predicting the susceptibility of landslide disasters can be divided into physics-based models, statistical models and machine learning models. Physical models evaluate the susceptibility of landslides by using slope stability theory to conduct stability analysis; statistical models establish the relationship between landslide susceptibility and factors such as elevation, rainfall and human influence based on statistical analysis of landslide databases; machine learning-based models can usually more accurately construct the mapping between landslide causes and landslide occurrence, including support vector machines, decision trees, logistic regression and artificial neural networks; deep learning models are an extension of machine learning. In recent years, researchers have applied deep learning models to landslide susceptibility prediction, improving the accuracy of traditional statistical models and classic machine learning models.

[0005] However, most of these models consider geological and environmental conditions equally, and it is difficult to effectively identify the main driving factors of landslide disasters in different scenarios, which affects the prediction performance of the model and leads to errors in the prediction of landslide susceptibility. Summary of the invention

[0006] The purpose of the present invention is to provide a landslide susceptibility prediction method based on dynamic evaluation of environmental factor importance in view of the above problems, construct an intrinsic environmental factor model comprising a convolutional neural network, a squeeze and excitation module (Squeeze-and-Excitation, SE) and an LSTM network for mining the feature representation of the intrinsic environmental factor; construct an extrinsic driving environmental factor model comprising a convolutional neural network, an SE layer and a Transformer network for capturing and parsing the feature information of the extrinsic driving environmental factor; and construct an internal and external driving self-attention fusion model as a landslide susceptibility prediction model by weighted fusion of the feature representation of the intrinsic environmental factor model and the feature information of the extrinsic driving environmental factor model through a self-attention mechanism, so as to effectively identify the main driving factors of the occurrence of landslide disasters, improve the prediction performance of the landslide prediction model, and reduce the prediction error.

[0007] The technical solution of the present invention is: The landslide susceptibility prediction method based on dynamic evaluation of environmental factors importance includes the following steps: S1: Collect historical landslide data, select landslide samples and non-landslide samples, and obtain landslide sample and non-landslide sample layers; S2: Consider environmental factors, collect environmental factor data, and obtain landslide environmental factor layer; S3: The landslide and non-landslide sample layers are superimposed on the environmental factor layer respectively, and the environmental factor layer is extracted again to generate the data set for input model; S4: After randomly dividing the dataset into training set, validation set and test set according to the proportion, the dataset is normalized; S5: Construct a landslide susceptibility prediction model based on dynamic evaluation of environmental factors importance; S6: using the data set obtained in step S3 to train the landslide susceptibility prediction model, and verifying the accuracy and generalization error of the model, and using the trained landslide susceptibility prediction model to obtain a landslide susceptibility prediction result; S7: Classifying the landslide susceptibility of the region according to the landslide susceptibility prediction result of the landslide susceptibility prediction model in step S6.

[0008] Furthermore, in step S1, non-landslide samples are randomly selected outside the buffer zone and do not overlap with the landslide sample area.

[0009] Furthermore, in step S1, the random selection method of non-landslide samples outside the buffer zone is to filter the samples within 3 kilometers and then evenly select non-landslide samples far away from the landslide area in the target area until the number is consistent with the number of landslide samples.

[0010] Furthermore, in step S2, the multi-source data of the target area establish environmental influencing factors, and the multi-source data are imported into ArcGIS for processing to obtain a plane layer of landslide environmental factors; the environmental factors include intrinsic environmental factors and external driving environmental factors; the intrinsic environmental factors include elevation, soil, slope, slope direction, water flow intensity index and terrain moisture index; the external driving environmental factors include distance from roads, distance from water systems, normalized vegetation index, land use and peak rainfall.

[0011] Furthermore, in step S3, the landslide and non-landslide sample layers are respectively superimposed on the environmental factor layers, and the environmental factor layers are extracted again to generate a data set for the input model; the step is to extract the values ​​of all environmental factors at the same spatial position after the landslide and non-landslide samples are superimposed, which are used as the attribute values ​​of the corresponding landslide and non-landslide samples to generate the input data set of the model.

[0012] Furthermore, in step S4, the data set is randomly divided into a training set, a validation set, and a test set in proportion, and then normalized; the ratio is 6:2:2; the training set is used to learn the landslide characteristics and train the internal parameters of the model; the validation set is used to optimize the model hyperparameters; and the test set is used to verify the generalization error and evaluate the robustness of the model.

[0013] Furthermore, step S5 includes the following sub-steps: S51: Construct the internal environmental factor model and the external driving environmental factor model; S52: Use the intrinsic environmental factor model to mine the characteristic representation of the intrinsic environmental factors, and use the extrinsic driving environmental factor model to capture and analyze the characteristic information of the extrinsic driving environmental factors; S53: The features of the intrinsic environmental factor model and the extrinsic driving environmental factor model are weightedly fused through the self-attention mechanism to construct an internal and external driving self-attention fusion model.

[0014] Further, in step S5, the intrinsic environmental factor model takes the intrinsic environmental factor as input, and includes a first convolution unit, a first SE layer and an LSTM layer connected in sequence; the first convolution unit feature extraction includes a convolutional neural network, a ReLU activation function, a batch normalization layer and a maximum pooling layer; the first SE layer includes a global average pooling and a fully connected layer; the LSTM layer includes a double-layer LSTM network, and outputs the hidden state of the last time step as a feature representation; ; In the formula, represents the first convolutional neural network n The convolution kernel is in the output feature map i The first k The features of the output at each position; Represents the standardized input data, with a shape of (B, C, L), where B represents the batch size and C represents the number of channels. L Indicates the length of the sequence; Represents the input data i The sample j Channels in sequence length L Middle k + m -1 position value; W Represents the weight of the convolution kernel, the shape is (C, K), K is the size of the convolution kernel, represents the weight from the jth input channel to the nth output channel at the mth position of the convolution kernel, represents the bias term, and ReLU is the activation function; The first SE layer uses global average pooling and fully connected layers for feature extraction and mapping. The feature extraction formula is: ; In the formula, represents the recalibrated feature map, represents the value of the 𝑘th position of the 𝑐th channel of the input feature map, and are the weight matrices of the fully connected layers, represents the ReLU activation function, represents the Sigmoid activation function, Indicates the length of the feature map.

[0015] Further, in step S5, the external driving environment factor model takes the external driving environment factor as input, and the model includes a second convolution unit, a second SE layer, a Transformer module and a SeqPool layer connected in sequence; the second convolution unit includes a convolutional neural network, a ReLU activation function, a batch normalization layer and a maximum pooling layer; the second SE layer is implemented by global average pooling and a fully connected layer; the Transformer module includes a multi-head self-attention mechanism, LayerNorm and a feed-forward fully connected layer, and has multiple attention heads; the SeqPool layer aggregates sequence features through attention weights; The calculation formula of SeqPool network layer is: ; In the formula, represents the output features of the SeqPool layer, Represents the output feature map of the Transformer module Locations, is the weight of the attention layer, is the weight of the output layer, represents the length of the feature map, Represents the Softmax function.

[0016] Furthermore, in step S53, the features of the intrinsic environmental factor model and the extrinsic driving environmental factor model are weighted and fused through the self-attention mechanism. The formula of the internal and external driving self-attention fusion model is as follows: ; In the formula, and They are the characteristic representations of the internal environmental factor model and the external driving environmental factor model, represents the self-attention mechanism processing function, and are the weights and biases of the fully connected layer, is the Sigmoid activation function.

[0017] Furthermore, the visualization processing of the landslide susceptibility prediction results includes classifying the landslide susceptibility of the target area into five levels: very high, high, medium, low and very low, and generating a landslide susceptibility prediction map.

[0018] Compared with the prior art, the beneficial effects of the present invention include: 1) The present invention provides a landslide susceptibility prediction method based on dynamic evaluation of the importance of environmental factors. By constructing an intrinsic environmental factor model including an SE layer, the feature representation of the intrinsic environmental factors is effectively mined; an extrinsic driving environmental factor model including an SE layer and a Transformer network is constructed to effectively capture and analyze the feature information of the extrinsic driving environmental factors; and a landslide susceptibility prediction model is obtained by weightedly fusing the feature representation of the intrinsic environmental factor model and the feature information of the extrinsic driving environmental factor model through a self-attention mechanism. The method can effectively identify the main driving factors of landslide disasters, effectively reduce the prediction error, and improve the prediction performance of the landslide prediction model.

[0019] 2) A landslide susceptibility prediction method based on dynamic evaluation of the importance of environmental factors, through the combination of SE module, attention mechanism and convolutional neural network, LSTM, Transformer network, etc., can adaptively clarify the main driving factors causing landslide disasters in different scenarios, thereby helping to improve the accuracy of landslide susceptibility assessment.

[0020] 3) The landslide susceptibility prediction method based on the dynamic evaluation of the importance of environmental factors takes into account the internal conditions and external driving conditions respectively, and constructs an internal and external driving self-attention fusion model, which can not only improve the accuracy of landslide susceptibility prediction, but also effectively identify the main driving factors under different geological and environmental conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0022] Figure 1 A schematic diagram of a landslide susceptibility prediction method according to a specific embodiment of the present invention; Figure 2 A landslide point distribution diagram according to an embodiment of the present invention; Figure 3 Schematic diagram of landslide susceptibility classification according to an embodiment of the present invention. DETAILED DESCRIPTION

[0023] like Figure 1 As shown, a landslide susceptibility prediction method based on dynamic evaluation of environmental factor importance disclosed in an embodiment of the present invention comprises the following steps: S1: Collect historical landslide data of the selected target area, select landslide samples and non-landslide samples, and obtain plane layers of landslide samples and non-landslide samples; S2: Considering environmental factors, collect environmental factor data and import them into the geographic information system to obtain the plane layer of landslide environmental factors; S3: Overlay the landslide and non-landslide sample layers with the environmental factor layers respectively, extract the environmental factor layers again, and generate a data set; S4: The dataset is randomly divided into training set, validation set and test set in a ratio of 6:2:2, and then normalized; S5: Construct a landslide susceptibility prediction model based on dynamic evaluation of environmental factors importance; S6: using the data set obtained in step S3 to train the landslide susceptibility prediction model, and verifying the accuracy and generalization error of the model, and using the trained landslide susceptibility prediction model to obtain a landslide susceptibility prediction result; S7: Classifying the landslide susceptibility of the region according to the landslide susceptibility prediction result of the landslide susceptibility prediction model in step S6.

[0024] Further, in step S1, the target areas include Chongqing, Yichang and Enshi in Hubei Province, and Anhua County in Hunan Province. Chongqing is located in southwest China and is an important part of the Three Gorges Reservoir area. It has a complex geological structure and is prone to geological disasters such as landslides and collapses. Yichang and Enshi are located in the middle and upper reaches of the Yangtze River and are significantly affected by the water level changes in the Three Gorges Reservoir area, making landslides prone to occur. Anhua County is located in the central part of Hunan Province, with undulating terrain, concentrated precipitation, and frequent geological disasters. The target areas and landslide points are distributed as follows: Figure 2 shown.

[0025] Preferably, landslide sample data is obtained based on the historical list of landslides in the target area; non-landslide samples are randomly selected outside the buffer zone, and the number of non-landslide samples is the same as the number of landslide samples to ensure that there is no overlap between landslide samples and non-landslide samples. The number of non-landslide points initially generated is 3-4 times the number of landslide points in the target area. For these initial non-landslide samples, a 3-kilometer threshold is used to filter out points within 3 kilometers of the nearest landslide point. The minimum distance of the remaining non-landslide samples is 3 kilometers, and then the non-landslide samples in the target area are uniformly selected to obtain the final sample set.

[0026] Preferably, a total of 925 landslide samples and 925 non-landslide samples are collected, and the grid layer of landslide and non-landslide distribution contains 136,375 grid units.

[0027] Furthermore, in step S2, the multi-source data of the target area are analyzed to establish environmental factors, and the multi-source data are imported into ArcGIS for processing to obtain the plane layer of landslide environmental factors; the environmental factors include intrinsic environmental factors and external driving environmental factors. The intrinsic environmental factors include elevation, soil, slope, slope aspect, stream intensity index (SPI), and terrain wetness index (TWI); the external driving environmental factors include distance from roads, distance from water systems, normalized difference vegetation index (NDVI), land use, and peak rainfall. ArcGIS software is used to unify the coordinate system and spatial resolution of various environmental factors.

[0028] Preferably, the environmental factor values ​​of the landslide and non-landslide sample points extracted from each environmental factor layer using the spatial analysis tool in ArcGIS are used as the data set for model training.

[0029] Furthermore, before establishing the prediction model in step S3, multicollinearity analysis is performed to calculate the variance inflation factor (VIF), identify and deal with multicollinearity problems, and use Pearson correlation coefficient analysis to evaluate the linear correlation between environmental factors to obtain the final data set.

[0030] Furthermore, in step S4, the data set is randomly divided into a training set, a validation set, and a test set in a ratio of 6:2:2. The training set is used to learn the landslide characteristics and train the internal parameters of the prediction model, the validation set is used to optimize the hyperparameters of the prediction model, and the test set is used to check the final generalization error and evaluate the robustness of the model; the data set is normalized and input into the model for landslide susceptibility prediction.

[0031] Further, step S5 includes the following sub-steps: S51: Construct the internal environmental factor model and the external driving environmental factor model; S52: The intrinsic environmental factor model mines the characteristic representation of the intrinsic environmental factors, and the extrinsic driving environmental factor model captures and analyzes the characteristic information of the extrinsic driving environmental factors; S53: The features of the intrinsic environmental factor model and the extrinsic driving environmental factor model are weightedly fused through the self-attention mechanism to construct an internal and external driving self-attention fusion model.

[0032] Preferably, in step S5, the intrinsic environmental factor model takes the intrinsic environmental factor as input, and includes a first convolution unit, a first SE layer and an LSTM unit connected in sequence; the first convolution unit includes a first convolutional neural network, a batch normalization layer and a maximum pooling layer; the first SE layer includes a global average pooling and a fully connected layer; the LSTM unit includes a two-layer LSTM network, and outputs the hidden state of the last time step as a feature representation; The feature extraction formula of the first convolutional neural network is: ; In the formula, represents the first convolutional neural network n The convolution kernel is in the output feature map i The first k The features of the output at each position; Represents the standardized input data, with a shape of (B, C, L), where B represents the batch size and C represents the number of channels. L Indicates the length of the sequence; Represents the input data i The sample j Channels in sequence length L Middle k + m -1 position value; W Represents the weight of the convolution kernel, the shape is (C, K), K is the size of the convolution kernel, represents the weight from the jth input channel to the nth output channel at the mth position of the convolution kernel, represents the bias term, and ReLU is the activation function; The first SE layer uses global average pooling and fully connected layers for feature extraction and mapping. The feature extraction formula is: ; In the formula, represents the recalibrated feature map, represents the value of the 𝑘th position of the 𝑐th channel of the input feature map, and are the weight matrices of the fully connected layers, represents the ReLU activation function, represents the Sigmoid activation function, Indicates the length of the feature map.

[0033] Preferably, in step S5, the external driving environment factor model takes the external driving environment factor as input, and the model includes a second convolution unit, a second SE layer, a Transformer module and a SeqPool layer connected in sequence; the second convolution unit includes a second convolutional neural network, a batch normalization layer and a maximum pooling layer; the second SE layer includes a global average pooling and a fully connected layer; the Transformer module includes a multi-head self-attention mechanism, LayerNorm and a feed-forward fully connected layer, and has multiple attention heads; the SeqPool layer aggregates sequence features through attention weights; The calculation formula of SeqPool network layer is: ; In the formula, represents the output features of the SeqPool layer, Represents the output feature map of the Transformer module Locations, is the weight of the attention layer, is the weight of the output layer, represents the length of the feature map, Represents the Softmax function.

[0034] Preferably, in step S53: The characteristics of the intrinsic environmental factor model and the extrinsic driving environmental factor model are weightedly fused through the self-attention mechanism. The calculation formula of the internal and external driving self-attention fusion model is: ; In the formula, represents the characteristic representation of the intrinsic environmental factor model, Represents the characteristic information of the external driving environmental factor model, represents the self-attention mechanism processing function, and are the weights and biases of the fully connected layer, Represents the Sigmoid activation function.

[0035] Furthermore, in step S7, the visualization processing of the landslide susceptibility prediction result includes dividing the landslide susceptibility of the target area into five levels: very high, high, medium, low and very low.

[0036] Further, in step S8, visualization processing is performed to visualize the landslide susceptibility prediction results according to the landslide susceptibility level to generate a landslide susceptibility prediction map, such as Figure 3 shown.

[0037] Based on the above inventive concept, this embodiment further provides a system for implementing the above landslide susceptibility prediction, which includes the following functional modules: Data acquisition function module: used to collect landslide historical data, select landslide samples and non-landslide sample data, and generate landslide and non-landslide sample layers; Environmental factor module: used to obtain and process environmental factor data, generate internal environmental factor layer and external driving environmental factor layer, and overlay landslide and non-landslide sample layers to generate data sets.

[0038] Intrinsic environmental factor module: used to mine the feature representation of intrinsic environmental factors. The intrinsic environmental factor module takes the intrinsic environmental factors as input and includes the first convolution unit, the first SE layer and the LSTM unit connected in sequence; External driving environment factor module: used to capture and analyze the characteristic information of external driving environment factors; the external driving environment factor module takes the external driving environment factors as input, and includes the second convolution unit, the second SE layer, the Transformer module and the SeqPool layer connected in sequence; Internal and external driving self-attention fusion module: It is used to weight the features of the internal environmental factor module and the external driving environmental factor module through the self-attention mechanism, input the data set to train the model, and use the trained landslide susceptibility prediction model to obtain the landslide susceptibility prediction results. Landslide susceptibility visualization module: It is used to classify the landslide susceptibility of the target area based on the landslide susceptibility prediction results, divide the landslide susceptibility into five levels: very high, high, medium, low and very low, and generate a landslide susceptibility prediction map.

Claims

1. A landslide susceptibility prediction method based on dynamic evaluation of environmental factors, characterized in that: The following steps are involved: S1: Collect historical landslide data, select landslide samples and non-landslide samples, and obtain landslide sample and non-landslide sample layers; S2: Consider environmental factors, collect environmental factor data, and obtain landslide environmental factor layer; S3: Overlay the landslide and non-landslide sample layers with the environmental factor layers respectively, extract the environmental factor layers again, and generate a data set; S4: The dataset is randomly divided into training set, validation set and test set according to the proportion; S5: Construct a landslide susceptibility prediction model based on dynamic evaluation of environmental factors importance; S6: using the data set obtained in step S3 to train the landslide susceptibility prediction model, and verifying the accuracy and generalization error of the model, and using the trained landslide susceptibility prediction model to obtain a landslide susceptibility prediction result; S7: Classifying the landslide susceptibility of the region according to the landslide susceptibility prediction result of the landslide susceptibility prediction model in step S6.

2. The method for predicting landslide susceptibility based on dynamic evaluation of environmental factor importance according to claim 1, characterized in that: In step S1, the non-landslide samples are randomly selected outside the buffer zone and do not overlap with the landslide sample area.

3. The method for predicting landslide susceptibility based on dynamic evaluation of environmental factor importance according to claim 2, characterized in that: In the step S1, after filtering the samples within 3 kilometers, non-landslide samples far from the landslide area in the target area are uniformly selected until the number of non-landslide samples is consistent with that of landslide samples.

4. The method for predicting landslide susceptibility based on dynamic evaluation of environmental factor importance according to claim 3 is characterized in that: In step S2, the environmental factors include intrinsic environmental factors and external driving environmental factors; the intrinsic environmental factors include elevation, soil, slope, aspect, water flow intensity index and terrain moisture index; the external driving environmental factors include distance to roads, distance to water systems, normalized difference vegetation index, land use and peak rainfall.

5. The method for predicting landslide susceptibility based on dynamic evaluation of environmental factor importance according to claim 4, characterized in that: The step S5 specifically includes the following sub-steps: S51: Construct the internal environmental factor model and the external driving environmental factor model; S52: Use the intrinsic environmental factor model to mine the characteristic representation of the intrinsic environmental factors, and use the extrinsic driving environmental factor model to capture and analyze the characteristic information of the extrinsic driving environmental factors; S53: Through the self-attention mechanism, the feature representation of the intrinsic environmental factor model and the feature information of the extrinsic driving environmental factor model are weightedly fused to construct an internal and external driving self-attention fusion model.

6. The method for predicting landslide susceptibility based on dynamic evaluation of environmental factor importance according to claim 5, characterized in that: In step S5, the intrinsic environmental factor model takes the intrinsic environmental factor as input, and the intrinsic environmental factor model includes a first convolution unit, a first SE layer and an LSTM unit connected in sequence; the first convolution unit includes a first convolutional neural network, a batch normalization layer and a maximum pooling layer; the first SE layer includes a global average pooling and a fully connected layer; the LSTM unit includes a two-layer LSTM network to output the hidden state of the last time step as a feature representation.

7. The method for predicting landslide susceptibility based on dynamic evaluation of environmental factor importance according to claim 6, characterized in that: The feature extraction formula of the first convolutional neural network is: ; In the formula, represents the first convolutional neural network n The convolution kernel is in the output feature map i The first k The features of the output at each position; Represents the standardized input data, with a shape of (B, C, L), where B represents the batch size and C represents the number of channels. L Indicates the length of the sequence; Represents the input data i The sample j Channels in sequence length L Middle k + m -1 position value; W Represents the weight of the convolution kernel, the shape is (C, K), K is the size of the convolution kernel, represents the weight from the jth input channel to the nth output channel at the mth position of the convolution kernel, represents the bias term, and ReLU is the activation function; The first SE layer uses global average pooling and fully connected layers for feature extraction and mapping. The feature extraction formula is: ; In the formula, represents the recalibrated feature map, represents the value of the 𝑘th position of the 𝑐th channel of the input feature map, and are the weight matrices of the fully connected layers, represents the ReLU activation function, represents the Sigmoid activation function, Indicates the length of the feature map.

8. The method for predicting landslide susceptibility based on dynamic evaluation of environmental factor importance according to claim 7, characterized in that: In step S5, the external driving environment factor model takes the external driving environment factor as input, and the external driving environment factor model includes a second convolution unit, a second SE layer, a Transformer module and a SeqPool layer connected in sequence; the second convolution unit includes a second convolutional neural network, a batch normalization layer and a maximum pooling layer; the second SE layer includes a global average pooling and a fully connected layer; the Transformer module includes a multi-head self-attention mechanism, LayerNorm and a feedforward fully connected layer, and has multiple attention heads; the SeqPool layer aggregates sequence features through attention weights; The calculation formula of SeqPool network layer is: ; In the formula, represents the output features of the SeqPool layer, Represents the output feature map of the Transformer module Locations, is the weight of the attention layer, is the weight of the output layer, represents the length of the feature map, Represents the Softmax function.

9. The method for predicting landslide susceptibility based on dynamic evaluation of environmental factor importance according to claim 8, characterized in that: In step S53, the calculation formula of the internal and external driving self-attention fusion model is: ; In the formula, represents the characteristic representation of the intrinsic environmental factor model, Represents the characteristic information of the external driving environmental factor model, represents the self-attention mechanism processing function, and are the weights and biases of the fully connected layer, Represents the Sigmoid activation function.

10. The method for predicting landslide susceptibility based on dynamic evaluation of environmental factor importance according to any one of claims 1 to 9, characterized in that: In step S7, the landslide susceptibility of the target area is divided into five levels: very high, high, medium, low and very low, and the landslide susceptibility prediction results are visualized according to the landslide susceptibility levels to generate a landslide susceptibility prediction map.

Images