A landslide susceptibility analysis method based on a graph neural network

Abstract

The application discloses a landslide susceptibility analysis method based on a graph neural network, which comprises the following steps: acquiring image data sets of a target region at different time points based on a preset time period; dividing the target region into a plurality of geological units according to topography, and setting each geological unit as a node in a graph structure; extracting surface deformation feature vectors of each geological unit node at different time points; iteratively optimizing the surface deformation feature vectors of each geological unit node at different time points to obtain optimized surface deformation feature vectors; sending the optimized surface deformation feature vectors to a preset graph neural network model to obtain landslide susceptibility probability values of the corresponding geological units; and triggering an alarm information if the landslide susceptibility probability value of any one of the geological units is greater than a preset probability threshold. The application significantly improves the accuracy and reliability of landslide susceptibility reanalysis through dynamic calibration of multi-time sequence characteristics.

Description

Technical Field

[0001] This invention relates to the field of geological monitoring technology, and more specifically, to a landslide susceptibility analysis method based on graph neural networks. Background Technology

[0002] Landslides are highly destructive geological hazards that pose a serious threat to people's lives and property. Traditional methods for landslide susceptibility analysis mainly include deterministic models, statistical models, and machine learning models. Deterministic models are based on physical and mechanical principles, require a large number of geological parameters, and have limited applicability; statistical models, such as logistic regression and the weight of evidence method, rely on the distribution of historical landslide data and are difficult to handle complex nonlinear relationships.

[0003] In recent years, the application of deep learning technology in the field of geological disasters has gradually increased. However, most existing methods are based on convolutional neural networks to process regular grid data, which cannot effectively represent the irregular boundaries of geological units and the topographic relationships between different geological units. At the same time, existing studies mostly use current data for analysis, lacking a comprehensive consideration of the surface deformation feature vectors at different time points in geological units, ignoring the dynamic changes and continuous characteristics of landslide influencing factors, resulting in insufficient accuracy in long-term monitoring. Summary of the Invention

[0004] To address at least one of the aforementioned technical problems, the present invention aims to provide a landslide susceptibility analysis method based on graph neural networks, which can dynamically calibrate multiple temporal features, significantly improving the accuracy and reliability of complex landslide susceptibility analysis.

[0005] This invention provides a landslide susceptibility analysis method based on graph neural networks, comprising:

[0006] Based on a preset time period, acquire image data sets of the target area at different time points;

[0007] The target area is divided into multiple geological units according to the terrain, and each geological unit is set as a node in the graph structure;

[0008] For each geological unit node, extract its surface deformation feature vector at different time points;

[0009] The surface deformation feature vector of each geological unit node at different time points is iteratively optimized to obtain the optimized surface deformation feature vector.

[0010] The optimized surface deformation feature vector is sent to a preset graph neural network model to obtain the landslide susceptibility probability value of the corresponding geological unit.

[0011] If the landslide susceptibility probability value of any geological unit is greater than the preset probability threshold, an alert message will be triggered.

[0012] In this solution, the step of dividing the target area into multiple geological units according to the terrain specifically includes:

[0013] The target area is divided into multiple sub-target areas according to a preset grid.

[0014] Extract the elevation value of the center point in the sub-target region and set it as the elevation value of the sub-target region.

[0015] The terrain location index (TPI) of the sub-target area is set as follows:

[0016] ,in This represents the terrain location index of the sub-target region M. This represents the elevation value of the sub-target region M. This represents the elevation value of the m-th sub-target region with radius R centered on sub-target region M, where n is the total number of sub-target regions with radius R centered on sub-target region M.

[0017] Cluster analysis was performed on the TPI of the sub-target regions to extract adjacent sub-target regions of the same type, and then the adjacent sub-target regions of the same type were set as a geological unit.

[0018] In this scheme, the step of performing cluster analysis based on the TPI of the sub-target regions specifically includes:

[0019] S201, randomly extract a sub-target region and set it as a cluster node;

[0020] S202, calculate the difference between the TPI of the cluster node and the TPI of any sub-target region to obtain the terrain location index difference;

[0021] S203, If the difference in terrain location index is less than or equal to the preset index difference, then the corresponding two sub-target areas will be set to the same type.

[0022] S204. After traversing all sub-target regions, select a new sub-target region from the remaining sub-target regions as the clustering node, and repeat steps S202 to S203 until all sub-target regions are clustered, and record the number of geological units.

[0023] S205, After traversing all sub-target regions as cluster nodes, determine the number of geological units corresponding to different cluster analyses;

[0024] S206, cluster analysis is performed on the sub-target region using the cluster nodes corresponding to the minimum number of geological units.

[0025] This plan also includes:

[0026] Extract the area value of each geological unit;

[0027] If the area value of a geological unit is greater than the preset area threshold, the corresponding geological unit will be split into two geological units until the area value of all geological units is less than or equal to the preset area threshold.

[0028] Extract the distance value between any two points in each geological unit;

[0029] If the distance between any two points is greater than a preset distance threshold, the geological unit is split into two geological units by the perpendicular line connecting the two points, until the distance between any two points in all geological units is less than or equal to the preset distance threshold.

[0030] This plan also includes:

[0031] Based on the same type of geological unit, if the sum of the areas of two adjacent geological units is less than or equal to a preset area threshold and the distance between any two points in the two adjacent geological units is less than or equal to a preset distance threshold, then the two adjacent geological units will be merged to obtain the merged geological unit.

[0032] In this scheme, the step of iteratively optimizing the data in the set of surface deformation feature vectors at different time points for each geological unit node to obtain the optimized surface deformation feature vectors specifically includes:

[0033] For each geological unit node, the data in the set of surface deformation feature vectors at different time points are arranged in chronological order to obtain the set of deformation feature vectors.

[0034] The accumulated amount of feature change is obtained by weighting the values ​​in the deformation feature vector set based on a preset time decay factor. Its formula is:

[0035] ,in This represents the cumulative amount of characteristic changes of geological unit node i at the current time. The time decay factor represents the time point (Tt). , Let T and (t-1) represent the surface deformation feature vectors of geological unit node i at time points t and (t-1), respectively; T represents the total number of time points collected, and t represents the time point index.

[0036] The optimized surface deformation feature vector is set as Its formula is:

[0037] ,in This represents the timing reliability adjustment coefficient, where 'a' is the corresponding parameter.

[0038] In this scheme, after obtaining the set of deformation feature vectors, the method further includes:

[0039] Based on the same geological unit node, the deformation feature vector at the current time point is subtracted from the deformation feature vector at the previous time point to obtain the deformation feature vector difference at the current time point.

[0040] If the difference in deformation feature vectors at the current time point is greater than the preset threshold for the difference in deformation feature vectors, then extract the difference in deformation feature vectors of other geological unit nodes of the same type to which the geological unit node belongs at the current time point and set it as the comparison difference in deformation feature vectors.

[0041] The error index is determined by comparing and analyzing the difference in deformation eigenvectors at the current time point with the difference in deformation eigenvectors compared with those at the previous time point.

[0042] If the error index is less than the preset error index threshold, the difference in the deformation feature vector at the current time point is normal.

[0043] If the error index is greater than or equal to the preset error index threshold, the deformation feature vector difference at the current time point is considered abnormal. The deformation feature vector difference at the current time point is then revised based on the comparison of the deformation feature vector differences to obtain the revised deformation feature vector difference.

[0044] The deformation feature vector of the geological unit node at the current time point is revised based on the difference in the revised deformation feature vector.

[0045] In this scheme, after obtaining the set of deformation feature vectors, the method further includes:

[0046] If the deformation feature vector of a geological unit node is greater than the preset deformation feature vector threshold, mark the geological unit node and the corresponding time point.

[0047] Based on the marked time points, extract the number of different geological unit nodes marked within a preset range at a set time.

[0048] If the quantity value is greater than the preset quantity threshold, then the landslide susceptibility probability will be marked as 100% within the preset range.

[0049] In this scheme, the step of sending the optimized surface deformation feature vector to a preset graph neural network model to obtain the landslide susceptibility probability value of the corresponding geological unit specifically includes:

[0050] The optimized surface deformation features are updated using graph convolutional layers in a graph neural network model, resulting in the first updated geological unit node features. ;

[0051] The updated geological unit node features are sent to the graph attention layer to obtain the updated geological unit node features. ;

[0052] Based on the same geological unit, the second updated geological unit node features at all time points within a preset sliding time window are weighted and summed to obtain the temporally fused geological unit node features. ;

[0053] Let P be the probability value of landslide susceptibility for a geological unit, and its formula is:

[0054] ,in This represents the probability value of a landslide occurring in geological unit i. This represents the trainable weight vector of the final prediction layer. The trainable bias scalar of the final prediction layer, Represents the dot product of vectors;

[0055] The graph neural network model includes a graph convolutional layer, a graph attention layer, a temporal feature fusion layer, and a final prediction layer.

[0056] One or more technical solutions proposed in this application have at least the following technical effects:

[0057] 1. Existing technologies are mostly based on static or single-phase data for analysis, which cannot effectively capture the dynamic evolution of landslide influencing factors over time. This application highlights the cumulative effect of landslide deformation through time iteration characteristics, thereby improving the accuracy of landslide susceptibility analysis.

[0058] 2. This application divides the target area first according to the terrain, then further subdivides large-area geological units or narrow geological units, and merges geological units with the same terrain. This avoids fixed grid division, effectively reduces the amount of feature iteration calculation, and can effectively express the irregular boundaries of geological units. Attached Figure Description

[0059] Figure 1 The flowchart of a landslide susceptibility analysis method based on graph neural networks according to the present invention is shown. Detailed Implementation

[0060] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0061] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0062] Figure 1 The flowchart of a landslide susceptibility analysis method based on graph neural networks according to the present invention is shown.

[0063] like Figure 1 As shown, this invention discloses a landslide susceptibility analysis method based on graph neural networks, comprising:

[0064] S101, Based on a preset time period, acquire image data sets of the target area at different time points;

[0065] S102, the target area is divided into multiple geological units according to the terrain, and each geological unit is set as a node in the graph structure;

[0066] S103, for each geological unit node, extract its surface deformation feature vector at different time points;

[0067] S104, iteratively optimize the surface deformation feature vector of each geological unit node at different time points to obtain the optimized surface deformation feature vector;

[0068] S105, The optimized surface deformation feature vector is sent to the preset graph neural network model to obtain the landslide susceptibility probability value of the corresponding geological unit;

[0069] S106 If the landslide susceptibility probability value of any geological unit is greater than the preset probability threshold, an alarm message will be triggered.

[0070] According to an embodiment of the present invention, the image data set includes image data from at least three different time points. The image data can be three-dimensional scan image data, remote sensing image data, or a combination of both. The target area is divided into multiple geological units by topography, with each geological unit serving as a node in the graph structure. Edge connections are constructed based on the adjacency relationships and hydrological connectivity of the geological units to form a geological unit graph structure, or simply graph structure. The monitoring rigor of the target area is determined according to a preset probability threshold, wherein the smaller the preset probability threshold, the greater the monitoring rigor of the target area. For example, the preset probability threshold is set to 20%.

[0071] According to an embodiment of the present invention, the step of dividing the target area into multiple geological units according to the terrain specifically includes:

[0072] The target area is divided into multiple sub-target areas according to a preset grid.

[0073] Extract the elevation value of the center point in the sub-target region and set it as the elevation value of the sub-target region.

[0074] The terrain location index (TPI) of the sub-target area is set as follows:

[0075] ,in This represents the terrain location index of the sub-target region M. This represents the elevation value of the sub-target region M. This represents the elevation value of the m-th sub-target region with radius R centered on sub-target region M, where n is the total number of sub-target regions with radius R centered on sub-target region M.

[0076] Cluster analysis was performed on the TPI of the sub-target regions to extract adjacent sub-target regions of the same type, and then the adjacent sub-target regions of the same type were set as a geological unit.

[0077] It should be noted that if a portion of the area of ​​a sub-target region lies within a circle centered at sub-target region M with radius R, then the area corresponding to sub-target region M is calculated accordingly. At the same time, the topographic location index of the corresponding sub-target area is also included in the calculation; furthermore, when the locations of sub-target areas of the same type are not connected, the two sub-target areas of the same type but not connected are respectively set as two different geological units; the elevation value of the sub-target area M can be directly replaced by the elevation value of the center point of the corresponding sub-target area, and m belongs to n.

[0078] According to an embodiment of the present invention, the step of performing cluster analysis based on the TPI of the sub-target region specifically includes:

[0079] S201, randomly extract a sub-target region and set it as a cluster node;

[0080] S202, calculate the difference between the TPI of the cluster node and the TPI of any sub-target region to obtain the terrain location index difference;

[0081] S203, If the difference in terrain location index is less than or equal to the preset index difference, then the corresponding two sub-target areas will be set to the same type.

[0082] S204. After traversing all sub-target regions, select a new sub-target region from the remaining sub-target regions as the clustering node, and repeat steps S202 to S203 until all sub-target regions are clustered, and record the number of geological units.

[0083] S205, After traversing all sub-target regions as cluster nodes, determine the number of geological units corresponding to different cluster analyses;

[0084] S206, cluster analysis is performed on the sub-target region using the cluster nodes corresponding to the minimum number of geological units.

[0085] It should be noted that by using cluster analysis and selecting the minimum number of geological units, the computational load of subsequent features is reduced while ensuring data accuracy, thus improving the efficiency of landslide susceptibility analysis.

[0086] According to an embodiment of the present invention, it further includes:

[0087] Extract the area value of each geological unit;

[0088] If the area value of a geological unit is greater than the preset area threshold, the corresponding geological unit will be split into two geological units until the area value of all geological units is less than or equal to the preset area threshold.

[0089] Extract the distance value between any two points in each geological unit;

[0090] If the distance between any two points is greater than a preset distance threshold, the geological unit is split into two geological units by the perpendicular line connecting the two points, until the distance between any two points in all geological units is less than or equal to the preset distance threshold.

[0091] It should be noted that by reasonably controlling the area value of geological units and the distance between any two points, the difference in surface deformation feature vectors at the locations of each pixel in the same geological unit is reduced; the accuracy and representativeness of the surface deformation feature vectors of each geological unit node are improved; for example, the preset area threshold is 5 square meters and the preset distance threshold is 5 meters.

[0092] According to an embodiment of the present invention, based on the same type of geological unit, if the sum of the areas of two adjacent geological units is less than or equal to a preset area threshold and the distance between any two points in the two adjacent geological units is less than or equal to a preset distance threshold, then the corresponding two adjacent geological units are merged to obtain a merged geological unit.

[0093] It should be noted that, in order to prevent the geological unit from being divided too much, the area and the distance between any two points of the same type of adjacent geological unit can be checked to improve the rationality of the geological unit division.

[0094] According to an embodiment of the present invention, the step of iteratively optimizing the data in the set of surface deformation feature vectors at different time points for each geological unit node to obtain the optimized surface deformation feature vector specifically includes:

[0095] For each geological unit node, the data in the set of surface deformation feature vectors at different time points are arranged in chronological order to obtain the set of deformation feature vectors.

[0096] The accumulated amount of feature change is obtained by weighting the values ​​in the deformation feature vector set based on a preset time decay factor. Its formula is:

[0097] ,in This represents the cumulative amount of characteristic changes of geological unit node i at the current time. The time decay factor represents the time point (Tt). , Let T represent the surface deformation feature vectors of geological unit node i at time points t and (t-1), respectively; T represents the total number of time points collected, and t represents the time point index, t=1,2,...,T-1,T, where t=1 represents the earliest time point and t=T represents the current time point;

[0098] The optimized surface deformation feature vector is set as Its formula is:

[0099] ,in This represents the timing reliability adjustment coefficient, where 'a' is the corresponding parameter.

[0100] It should be noted that a pre-defined temporal attention mechanism is used to calculate the confidence score of each feature at each time point. Its formula is ,in It is the Sigmoid activation function. , Indicates learnable parameters, ; This represents the concatenation of two corresponding feature vectors; the time weight is calculated using the confidence score. Its formula is Then, based on the time weight, the corresponding time series reliability adjustment coefficient is determined. .

[0101] According to an embodiment of the present invention, after obtaining the set of deformation feature vectors, the method further includes:

[0102] Based on the same geological unit node, the deformation feature vector at the current time point is subtracted from the deformation feature vector at the previous time point to obtain the deformation feature vector difference at the current time point.

[0103] If the difference in deformation feature vectors at the current time point is greater than the preset threshold for the difference in deformation feature vectors, then extract the difference in deformation feature vectors of other geological unit nodes of the same type to which the geological unit node belongs at the current time point and set it as the comparison difference in deformation feature vectors.

[0104] The error index is determined by comparing and analyzing the difference in deformation eigenvectors at the current time point with the difference in deformation eigenvectors compared with those at the previous time point.

[0105] If the error index is less than the preset error index threshold, the difference in the deformation feature vector at the current time point is normal.

[0106] If the error index is greater than or equal to the preset error index threshold, the deformation feature vector difference at the current time point is considered abnormal. The deformation feature vector difference at the current time point is then revised based on the comparison of the deformation feature vector differences to obtain the revised deformation feature vector difference.

[0107] The deformation feature vector of the geological unit node at the current time point is revised based on the difference in the revised deformation feature vector.

[0108] It should be noted that if the difference in deformation feature vectors at the current time point is greater than the preset threshold for deformation feature vector difference, it indicates that the surface deformation feature vector value at the current time point is large. Therefore, the accuracy of the deformation feature vector at the current time point is determined by comprehensive analysis with other comparative deformation feature vector differences. The error index represents the comprehensive error index of the difference between two deformation feature vectors in terms of direction and value. For example, it can be obtained by determining the directional similarity value and the numerical similarity value of the difference between two deformation feature vectors, multiplying the directional similarity value and the numerical similarity value by the corresponding weighting coefficient, and then summing them to obtain the similarity index of the difference between the two deformation feature vectors. The preset value is then subtracted from the corresponding similarity index to obtain the error index of the difference between the two deformation feature vectors. The preset value can be set to 1. When there are multiple comparative deformation feature vector differences, the average value of the multiple comparative deformation feature vector differences is compared with the difference in deformation feature vectors at the current time point for analysis. If the error index is greater than or equal to the preset error index threshold, the difference between the comparison deformation feature vectors is set as the difference between the deformation feature vectors after the revision at the current time point. Then, the difference between the deformation feature vectors after the revision at the current time point is added to the deformation feature vector of the previous time point to obtain the deformation feature vector of the geological unit node at the current time point.

[0109] According to an embodiment of the present invention, after obtaining the set of deformation feature vectors, the method further includes:

[0110] If the deformation feature vector of a geological unit node is greater than the preset deformation feature vector threshold, mark the geological unit node and the corresponding time point.

[0111] Based on the marked time points, extract the number of different geological unit nodes marked within a preset range at a set time.

[0112] If the quantity value is greater than the preset quantity threshold, then the landslide susceptibility probability will be marked as 100% within the preset range.

[0113] It should be noted that, for example, the preset range is a region with a radius of R centered on the marked geological unit node, and the midpoint of the set time is the marked time point. For example, if the set time is 3 minutes, then the corresponding marked time point should be 1 minute and 30 seconds.

[0114] According to an embodiment of the present invention, the step of sending the optimized surface deformation feature vector to a preset graph neural network model to obtain the landslide susceptibility probability value of the corresponding geological unit specifically includes:

[0115] The optimized surface deformation features are updated using graph convolutional layers in a graph neural network model, resulting in the first updated geological unit node features. ;

[0116] The updated geological unit node features are sent to the graph attention layer to obtain the updated geological unit node features. ;

[0117] Based on the same geological unit, the second updated geological unit node features at all time points within a preset sliding time window are weighted and summed to obtain the temporally fused geological unit node features. ;

[0118] Let P be the probability value of landslide susceptibility for a geological unit, and its formula is:

[0119] ,in This represents the probability value of a landslide occurring in geological unit i. This represents the trainable weight vector of the final prediction layer. The trainable bias scalar of the final prediction layer, Represents the dot product of vectors;

[0120] The graph neural network model includes a graph convolutional layer, a graph attention layer, a temporal feature fusion layer, and a final prediction layer.

[0121] It should be noted that the calculation formula included in the graph convolutional layer is as follows: ,in This represents activation functions such as ReLU. This represents the set of neighboring nodes of a geological unit node in the graph structure. , Let i and j represent the degrees of node i and node j, respectively, which are the number of edges connected to that node. This represents the symmetric scaling factor used for normalization. This represents the trainable weight matrix of the first layer graph convolution; This represents the optimized surface deformation feature vector of neighbor node j (or itself j=i); the graph attention layer assigns differentiated attention weights to different neighbors, first calculating the original attention score. Its formula is Then, the attention weights are obtained by normalization. Its formula is Then, weighted aggregation is obtained. Its formula is ,in and These are learnable parameters; This represents the standardized attention weight assigned to node i by its neighbor node j. The temporal feature fusion layer integrates the geological unit node features updated a second time at different time points, and its formula is: ,in These are the weighting coefficients for the corresponding time points. This represents the feature of the i-th node after passing through the second graph attention layer at time t.

[0122] This invention discloses a landslide susceptibility analysis method based on graph neural networks, comprising: acquiring image data sets of a target area at different time points based on a preset time period; dividing the target area into multiple geological units according to topography, and setting each geological unit as a node in a graph structure; extracting surface deformation feature vectors of each geological unit node at different time points; iteratively optimizing the surface deformation feature vectors of each geological unit node at different time points to obtain optimized surface deformation feature vectors; sending the optimized surface deformation feature vectors to a preset graph neural network model to obtain the landslide susceptibility probability value of the corresponding geological unit; if the landslide susceptibility probability value of any geological unit is greater than a preset probability threshold, an alert is triggered. This invention significantly improves the accuracy and reliability of landslide susceptibility analysis through dynamic calibration of multiple temporal features.

[0123] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components can be combined, or integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0124] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0125] In addition, in the various embodiments of the present invention, each functional unit can be integrated into one processing unit, or each unit can be a separate unit, or two or more units can be integrated into one unit; the integrated unit can be implemented in hardware or in the form of hardware plus software functional units.

[0126] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0127] Alternatively, if the integrated units of this invention are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

Claims

1. A landslide susceptibility analysis method based on graph neural networks, characterized in that, include: Based on a preset time period, acquire image data sets of the target area at different time points; The target area is divided into multiple geological units according to the terrain, and each geological unit is set as a node in the graph structure; For each geological unit node, extract its surface deformation feature vector at different time points; The surface deformation feature vector of each geological unit node at different time points is iteratively optimized to obtain the optimized surface deformation feature vector. The optimized surface deformation feature vector is sent to a preset graph neural network model to obtain the landslide susceptibility probability value of the corresponding geological unit. If the landslide susceptibility probability value of any geological unit is greater than the preset probability threshold, an alert message will be triggered. The step of iteratively optimizing the data in the set of surface deformation feature vectors at different time points for each geological unit node to obtain the optimized surface deformation feature vectors specifically includes: For each geological unit node, the data in the set of surface deformation feature vectors at different time points are arranged in chronological order to obtain the set of deformation feature vectors. The accumulated amount of feature change is obtained by weighting the values ​​in the deformation feature vector set based on a preset time decay factor. Its formula is: ,in This represents the cumulative amount of characteristic changes of geological unit node i at the current time. The time decay factor represents the time point (Tt). , Let T and (t-1) represent the surface deformation feature vectors of geological unit node i at time points t and (t-1), respectively; T represents the total number of time points collected, and t represents the time point index. The optimized surface deformation feature vector is set as Its formula is: ,in This represents the timing reliability adjustment coefficient, where 'a' is the corresponding parameter. This represents the surface deformation feature vector of geological unit node i at the current time point; The step of sending the optimized surface deformation feature vector to a preset graph neural network model to obtain the landslide susceptibility probability value of the corresponding geological unit specifically includes: The optimized surface deformation features are updated using graph convolutional layers in a graph neural network model, resulting in the first updated geological unit node features. ; The updated geological unit node features are sent to the graph attention layer to obtain the updated geological unit node features. ; Based on the same geological unit, the second updated geological unit node features at all time points within a preset sliding time window are weighted and summed to obtain the temporally fused geological unit node features. ; Let P be the probability value of landslide susceptibility for a geological unit, and its formula is: ,in This represents the probability value of a landslide occurring in geological unit i. This represents the trainable weight vector of the final prediction layer. The trainable bias scalar of the final prediction layer, Represents the dot product of vectors; The graph neural network model includes a graph convolutional layer, a graph attention layer, a temporal feature fusion layer, and a final prediction layer.

2. The landslide susceptibility analysis method based on graph neural networks according to claim 1, characterized in that, The step of dividing the target area into multiple geological units according to the terrain specifically includes: The target area is divided into multiple sub-target areas according to a preset grid. Extract the elevation value of the center point in the sub-target region and set it as the elevation value of the sub-target region. The terrain location index (TPI) of the sub-target area is set as follows: ,in This represents the terrain location index of the sub-target region M. This represents the elevation value of the sub-target region M. This represents the elevation value of the m-th sub-target region with radius R centered on sub-target region M, where n is the total number of sub-target regions with radius R centered on sub-target region M. Cluster analysis was performed on the TPI of the sub-target regions to extract adjacent sub-target regions of the same type, and then the adjacent sub-target regions of the same type were set as a geological unit.

3. The landslide susceptibility analysis method based on graph neural networks according to claim 2, characterized in that, The steps for clustering analysis based on the TPI of the sub-target regions specifically include: S201, randomly extract a sub-target region and set it as a cluster node; S202, calculate the difference between the TPI of the cluster node and the TPI of any sub-target region to obtain the terrain location index difference; S203, If the difference in terrain location index is less than or equal to the preset index difference, then the corresponding two sub-target areas will be set to the same type. S204. After traversing all sub-target regions, select a new sub-target region from the remaining sub-target regions as the clustering node, and repeat steps S202 to S203 until all sub-target regions are clustered, and record the number of geological units. S205, After traversing all sub-target regions as cluster nodes, determine the number of geological units corresponding to different cluster analyses; S206, cluster analysis is performed on the sub-target region using the cluster nodes corresponding to the minimum number of geological units.

4. The landslide susceptibility analysis method based on graph neural networks according to claim 1, characterized in that, Also includes: Extract the area value of each geological unit; If the area value of a geological unit is greater than the preset area threshold, the corresponding geological unit will be split into two geological units until the area value of all geological units is less than or equal to the preset area threshold. Extract the distance value between any two points in each geological unit; If the distance between any two points is greater than a preset distance threshold, the geological unit is split into two geological units by the perpendicular line connecting the two points, until the distance between any two points in all geological units is less than or equal to the preset distance threshold.

5. The landslide susceptibility analysis method based on graph neural networks according to claim 4, characterized in that, Also includes: Based on the same type of geological unit, if the sum of the areas of two adjacent geological units is less than or equal to a preset area threshold and the distance between any two points in the two adjacent geological units is less than or equal to a preset distance threshold, then the two adjacent geological units will be merged to obtain the merged geological unit.

6. The landslide susceptibility analysis method based on graph neural networks according to claim 1, characterized in that, After obtaining the set of deformation feature vectors, the process further includes: Based on the same geological unit node, the deformation feature vector at the current time point is subtracted from the deformation feature vector at the previous time point to obtain the deformation feature vector difference at the current time point. If the difference in deformation feature vectors at the current time point is greater than the preset threshold for the difference in deformation feature vectors, then extract the difference in deformation feature vectors of other geological unit nodes of the same type to which the geological unit node belongs at the current time point and set it as the comparison difference in deformation feature vectors. The error index is determined by comparing and analyzing the difference in deformation eigenvectors at the current time point with the difference in deformation eigenvectors compared with those at the previous time point. If the error index is less than the preset error index threshold, the difference in the deformation feature vector at the current time point is normal. If the error index is greater than or equal to the preset error index threshold, the deformation feature vector difference at the current time point is considered abnormal. The deformation feature vector difference at the current time point is then revised based on the comparison of the deformation feature vector differences to obtain the revised deformation feature vector difference. The deformation feature vector of the geological unit node at the current time point is revised based on the difference in the revised deformation feature vector.

7. The landslide susceptibility analysis method based on graph neural networks according to claim 1, characterized in that, After obtaining the set of deformation feature vectors, the process further includes: If the deformation feature vector of a geological unit node is greater than the preset deformation feature vector threshold, mark the geological unit node and the corresponding time point. Based on the marked time points, extract the number of different geological unit nodes marked within a preset range at a set time. If the quantity value is greater than the preset quantity threshold, then the landslide susceptibility probability will be marked as 100% within the preset range.

Images