Method and device for predicting landslide deformation

Abstract

The application discloses a landslide deformation prediction method and device, and relates to the field of landslide deformation detection, wherein the method comprises the following steps: acquiring InSAR along-line-of-sight landslide deformation data of a to-be-predicted region and at least one adjacent region thereof; inputting the deformation data of the at least one adjacent region into a first layer LSTM of a double-layer LSTM landslide deformation prediction model to extract feature vectors of the at least one adjacent region; multiplying the deformation data of the to-be-predicted region by a to-be-predicted region influence coefficient matrix to obtain a first matrix, and multiplying the feature vectors of the at least one adjacent region by at least one adjacent region influence coefficient matrix to obtain a second matrix; splicing the first matrix and the second matrix to obtain a spliced matrix; inputting the spliced matrix into a second layer LSTM of the double-layer LSTM model to obtain a final landslide deformation prediction result. The application can accurately predict landslide deformation by combining the deformation correlation between the to-be-predicted region and the at least one adjacent region.

Description

Technical Field

[0001] This invention relates to the field of landslide deformation detection technology, and in particular to a method and apparatus for predicting landslide deformation. Background Technology

[0002] This section is intended to provide background or context for the invention described herein. The description herein is not intended to imply that it is prior art simply because it is included in this section.

[0003] Landslide deformation prediction is a method that utilizes advanced technologies such as remote sensing monitoring, ground observation equipment, machine learning, and deep learning to continuously analyze and predict the displacement changes and deformation trends of landslide bodies. Predicting landslide deformation provides crucial information for disaster early warning, emergency preparedness, and medium- to long-term prevention and control planning. Accurate landslide deformation prediction not only allows for early warning before landslides occur, buying valuable time for evacuation and protection in threatened areas, but also helps decision-makers formulate targeted disaster prevention and mitigation strategies, thereby minimizing casualties and economic losses. This has profound significance for improving the comprehensive prevention and control capabilities of regional geological disasters.

[0004] Currently, landslide deformation prediction using InSAR data combined with LSTM (Long Short-Term Memory) neural networks typically focuses only on time-series data of a single region, i.e., using only historical data of the target area to predict future landslide deformation. This method ignores information from the surrounding areas. In reality, landslides are geological phenomena involving a certain spatial range, and the deformation of the target area is often correlated with the deformation of its neighboring areas. Research on how to combine data from the target area and its neighbors for more accurate landslide deformation prediction is still limited. Therefore, existing landslide deformation prediction methods have low accuracy. Summary of the Invention

[0005] This invention provides a landslide deformation prediction method to accurately predict landslide deformation by leveraging the correlation between the deformation of the area to be predicted and at least one adjacent area. The method includes:

[0006] Acquire synthetic aperture radar interferometric (InSAR) landslide deformation data along the line of sight for the region to be predicted, and InSAR landslide deformation data along the line of sight for at least one adjacent region of the region to be predicted.

[0007] The InSAR landslide deformation data along the line of sight of at least one neighboring region of the area to be predicted is input into the first LSTM layer of the two-layer long short-term memory neural network LSTM landslide deformation prediction model to extract the feature vector of at least one neighboring region; the first LSTM layer is generated by prediction training based on the historical relationship sample data between the InSAR landslide deformation data along the line of sight of at least one neighboring region of the target area and the feature vector of at least one neighboring region.

[0008] The first matrix is ​​obtained by multiplying the InSAR landslide deformation data along the line of sight of the area to be predicted with the influence coefficient matrix of the area to be predicted. The second matrix is ​​obtained by multiplying the feature vector of at least one adjacent area with the influence coefficient matrix of at least one adjacent area. The first and second matrices are concatenated to obtain the concatenated matrix of the area to be predicted and at least one adjacent area. The influence coefficients in the matrix are obtained by self-learning during the model training process. The influence coefficient of each area represents the importance of that area to the landslide deformation prediction result.

[0009] The splicing matrix is ​​input into the second LSTM layer of the two-layer LSTM landslide deformation prediction model to obtain the final prediction result of the landslide deformation in the area to be predicted; the second LSTM layer is generated by prediction training based on the historical relationship sample data between the splicing matrix of the target area and at least one of its adjacent areas and the final landslide deformation prediction result of the target area.

[0010] The present invention also provides a landslide deformation prediction device for accurately predicting landslide deformation by means of the correlation between the deformation of a target node region and its surrounding neighboring node regions. The device includes:

[0011] The acquisition unit is used to acquire synthetic aperture radar interferometric InSAR landslide deformation data along the line of sight of the region to be predicted, and InSAR landslide deformation data along the line of sight of at least one adjacent region of the region to be predicted.

[0012] The first LSTM processing unit is used to input the InSAR landslide deformation data along the line of sight of at least one neighboring region of the area to be predicted into the first LSTM of the two-layer long short memory neural network LSTM landslide deformation prediction model, and extract the feature vector of at least one neighboring region; the first LSTM is generated by prediction training based on the historical relationship sample data between the InSAR landslide deformation data along the line of sight of at least one neighboring region of the target area and the feature vector of at least one neighboring region.

[0013] The processing unit is used to multiply the InSAR landslide deformation data along the line of sight of the area to be predicted with the influence coefficient matrix of the area to be predicted to obtain a first matrix, and to multiply the feature vector of at least one adjacent area with the influence coefficient matrix of at least one adjacent area to obtain a second matrix; the first matrix and the second matrix are concatenated to obtain a concatenated matrix of the area to be predicted and at least one adjacent area; the influence coefficients in the matrix are obtained by self-learning during the model training process, and the influence coefficient of each area represents the importance of that area to the landslide deformation prediction result;

[0014] The second-layer LSTM processing unit is used to input the splicing matrix into the second-layer LSTM of the double-layer LSTM landslide deformation prediction model to obtain the final prediction result of the landslide deformation in the area to be predicted; the second-layer LSTM is generated by prediction training based on the historical relationship sample data between the splicing matrix of the target area and at least one of its adjacent areas and the final landslide deformation prediction result of the target area.

[0015] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described method for predicting landslide deformation.

[0016] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for predicting landslide deformation.

[0017] The present invention also provides a computer program product, which includes a computer program that, when executed by a processor, implements the above-described method for predicting landslide deformation.

[0018] The landslide deformation prediction scheme in this invention, compared with the existing technology that only uses historical data of the target area to predict future landslide deformation while ignoring information about the surrounding areas, achieves the following: First, it acquires synthetic aperture radar interferometric (InSAR) landslide deformation data along the line of sight of the area to be predicted, and InSAR landslide deformation data along the line of sight of at least one adjacent area of ​​the area to be predicted; second, it inputs the InSAR landslide deformation data along the line of sight of the at least one adjacent area of ​​the area to be predicted into the first layer of a two-layer long short-term memory (LSTM) neural network landslide deformation prediction model to extract feature vectors of at least one adjacent area; third, it generates the first layer of the LSTM based on historical relationship sample data between the InSAR landslide deformation data along the line of sight of the target area and the feature vectors of at least one adjacent area; fourth, it generates the InSAR landslide deformation data along the line of sight of the area to be predicted. The deformation data is multiplied by the influence coefficient matrix of the region to be predicted to obtain the first matrix. The feature vector of at least one adjacent region is multiplied by the influence coefficient matrix of at least one adjacent region to obtain the second matrix. The first and second matrices are concatenated to obtain the concatenated matrix of the region to be predicted and at least one adjacent region. The influence coefficients in the matrix are obtained through self-learning during model training. The influence coefficient of each region represents the importance of that region to the landslide deformation prediction result. The concatenated matrix is ​​input into the second layer of the two-layer LSTM landslide deformation prediction model to obtain the final prediction result of the landslide deformation of the region to be predicted. The second layer LSTM is generated by prediction training based on the historical relationship sample data between the concatenated matrix of the target region and at least one adjacent region and the final landslide deformation prediction result of the target region. It can accurately predict landslide deformation by means of the correlation between the deformation of the region to be predicted and at least one adjacent region. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0020] Figure 1 This is a flowchart illustrating the landslide deformation prediction method in an embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram of the process of pre-training and generating a two-layer LSTM landslide deformation prediction model in an embodiment of the present invention;

[0022] Figure 3 This is a schematic diagram of the target area or the area to be predicted and its adjacent areas in an embodiment of the present invention;

[0023] Figure 4 This is a schematic diagram of the process for obtaining the influence coefficient of the region to be predicted in an embodiment of the present invention;

[0024] Figure 5 This is a schematic diagram illustrating the change of the influence coefficient W of the region to be predicted in the container in an embodiment of the present invention;

[0025] Figure 6 This is a schematic diagram of the container update for the region to be predicted in an embodiment of the present invention;

[0026] Figure 7 This is a graph showing the change in the influence coefficient of the region to be predicted during the model training process in this embodiment of the invention.

[0027] Figure 8 This is a schematic diagram of the process for obtaining the influence coefficient of the adjacent region of the region to be predicted in an embodiment of the present invention;

[0028] Figure 9 This is a schematic diagram illustrating the variation of the influence coefficient W of adjacent regions within the container in an embodiment of the present invention;

[0029] Figure 10 This is a graph showing the change in the influence coefficient of adjacent regions during model training in an embodiment of the present invention.

[0030] Figure 11 This is a schematic diagram of the landslide deformation prediction results in an embodiment of the present invention;

[0031] Figure 12 This is a schematic diagram of the landslide deformation prediction device in an embodiment of the present invention. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. Here, the illustrative embodiments of the present invention and their descriptions are used to explain the present invention, but are not intended to limit the present invention.

[0033] The acquisition, storage, use, and processing of data in this application all comply with relevant laws and regulations.

[0034] The inventors discovered a technical problem with existing landslide deformation prediction techniques: traditional LSTM models typically focus only on a single time series data point, using only historical data of the target node to predict future landslide deformation. This method ignores information about neighboring nodes surrounding the target node. However, landslides are geological phenomena involving a certain spatial range, and the deformation of the target point is often correlated with the deformation of its surrounding neighboring nodes. Therefore, considering this technical problem, the invention proposes a scheme that utilizes historical data of the target node and its neighboring nodes to predict landslide deformation, thereby improving prediction accuracy. This scheme is described in detail below.

[0035] Figure 1 This is a flowchart illustrating the landslide deformation prediction method in an embodiment of the present invention, as shown below. Figure 1 As shown, the method includes the following steps:

[0036] Step 101: Acquire synthetic aperture radar interferometric (InSAR) landslide deformation data along the line of sight for the region to be predicted, and InSAR landslide deformation data along the line of sight for at least one adjacent region of the region to be predicted.

[0037] Step 102: Input the InSAR landslide deformation data along the line of sight of at least one neighboring region of the region to be predicted into the first LSTM layer of the two-layer long short-term memory neural network LSTM landslide deformation prediction model to extract the feature vector of at least one neighboring region; the first LSTM layer is generated by prediction training based on the historical relationship sample data between the InSAR landslide deformation data along the line of sight of at least one neighboring region of the target region and the feature vector of at least one neighboring region.

[0038] Step 103: Multiply the InSAR landslide deformation data along the line of sight of the area to be predicted with the influence coefficient matrix of the area to be predicted to obtain the first matrix; multiply the feature vector of at least one adjacent area with the influence coefficient matrix of at least one adjacent area to obtain the second matrix; concatenate the first matrix and the second matrix to obtain the concatenated matrix of the area to be predicted and at least one adjacent area; the influence coefficients in the matrix are obtained through self-learning during model training, and the influence coefficient of each area represents the importance of that area to the landslide deformation prediction result;

[0039] Step 104: Input the splicing matrix into the second LSTM layer of the two-layer LSTM landslide deformation prediction model to obtain the final prediction result of the landslide deformation in the area to be predicted; the second LSTM layer is generated by prediction training based on the historical relationship sample data between the splicing matrix of the target area and at least one of its adjacent areas and the final landslide deformation prediction result of the target area.

[0040] The landslide deformation prediction method provided in this embodiment of the invention operates as follows: Synthetic Aperture Radar Interferometry (InSAR) landslide deformation data along the line of sight for the area to be predicted, and InSAR landslide deformation data along the line of sight for at least one adjacent area of ​​the area to be predicted; the InSAR landslide deformation data along the line of sight for at least one adjacent area of ​​the area to be predicted are input into the first layer LSTM of a two-layer Long Short-Term Memory (LSTM) landslide deformation prediction model to extract feature vectors of at least one adjacent area; the first layer LSTM is generated based on historical relationship sample data between the InSAR landslide deformation data along the line of sight for at least one adjacent area of ​​the target area and the feature vectors of at least one adjacent area; the InSAR landslide deformation data along the line of sight for the area to be predicted is then processed. The first matrix is ​​obtained by multiplying the data with the influence coefficient matrix of the region to be predicted. The second matrix is ​​obtained by multiplying the feature vector of at least one adjacent region with the influence coefficient matrix of at least one adjacent region. The first and second matrices are concatenated to obtain the concatenated matrix of the region to be predicted and at least one adjacent region. The influence coefficients in the matrix are obtained through self-learning during model training. The influence coefficient of each region represents the importance of that region to the landslide deformation prediction result. The concatenated matrix is ​​input into the second layer of the two-layer LSTM landslide deformation prediction model to obtain the final prediction result of the landslide deformation of the region to be predicted. The second layer LSTM is generated by prediction training based on the historical relationship sample data between the concatenated matrix of the target region and at least one adjacent region and the final landslide deformation prediction result of the target region.

[0041] Compared with existing technologies that rely solely on historical data of the target area to predict future landslide deformation, ignoring information about the surrounding areas, the landslide deformation prediction method provided in this invention can accurately predict landslide deformation by leveraging the correlation between the deformation of the area to be predicted and at least one adjacent area. The landslide deformation prediction method is described in detail below.

[0042] Therefore, based on the aforementioned existing problems, this invention proposes a self-learning landslide prediction method. This method uses LSTM (Long Short-Term Memory Neural Network) to predict landslide deformation through historical data of the target area and its surrounding areas. Furthermore, during training, the model automatically learns the influence coefficients of the central node and its neighboring nodes, thereby significantly improving prediction accuracy. A detailed description follows.

[0043] First, we will introduce the process of pre-training and generating a two-layer LSTM landslide deformation prediction model.

[0044] Figure 2 This is a schematic diagram illustrating the process of pre-training and generating a two-layer LSTM landslide deformation prediction model in one embodiment of the present invention. Figure 2As shown, the above-mentioned method for predicting landslide deformation can also include: pre-training a two-layer LSTM landslide deformation prediction model according to the following method:

[0045] Step 201: Acquire InSAR landslide deformation data along the line of sight for the target area and at least one adjacent area;

[0046] Step 202: Perform three-dimensional decomposition on the InSAR landslide deformation data of the target area and at least one adjacent area along the line of sight to obtain the InSAR deformation dataset along the slope direction of the target area and at least one adjacent area.

[0047] Step 203: Construct a two-layer LSTM landslide deformation prediction network, which includes a first layer LSTM and a second layer LSTM.

[0048] Step 204: Construct the first-layer LSTM InSAR slope deformation dataset by combining the InSAR slope deformation data of at least one adjacent region of the target area along the line of sight with the historical relationship sample data of the feature vectors of at least one adjacent region.

[0049] Step 205: Process the InSAR slope deformation dataset of the target area and at least one adjacent area into a spliced ​​matrix of the target area and at least one adjacent area, and construct the historical relationship sample data of the final landslide deformation prediction result of the target area to form the InSAR slope deformation dataset of the second layer LSTM.

[0050] Step 206: Train the first layer LSTM using the InSAR slope deformation dataset of the first layer LSTM, and train the second layer LSTM using the InSAR slope deformation dataset of the second layer LSTM to obtain the double-layer LSTM landslide deformation prediction model.

[0051] In practice, Figure 3 This is a schematic diagram of the target area or the area to be predicted and its adjacent areas in an embodiment of the present invention. In step 201 above, InSAR landslide deformation data along the line of sight of the target area and its neighborhood (at least one adjacent area) are collected, such as... Figure 3 As shown, the green area can be the target area (P5), and the yellow area can be the adjacent areas of the target area (P1-P4, P6-P9). The model ultimately uses historical data of the target area and its adjacent areas to predict the landslide deformation of the target area.

[0052] In specific implementation, in step 202 above, the InSAR landslide deformation data along the line of sight is decomposed into three dimensions and then into deformation data along the slope direction.

[0053] In specific implementation, in step 203 above, a two-layer LSTM landslide deformation prediction network is constructed, which includes a first layer LSTM and a second layer LSTM.

[0054] In specific implementation, in step 204 above, the InSAR slope deformation data along the line of sight of at least one adjacent region of the target area and the historical relationship sample data of the feature vectors of at least one adjacent region are used to form the InSAR slope deformation dataset of the first layer LSTM. In step 205 above, the InSAR slope deformation dataset of the target area and at least one adjacent region after the importance adjustment is used to form the InSAR slope deformation dataset of the second layer LSTM with the historical relationship sample data of the final landslide deformation prediction result of the target area.

[0055] In specific implementation, the method of pre-training and generating a two-layer LSTM landslide deformation prediction model may also include: dividing the InSAR slope deformation dataset of the first layer LSTM into a training set, a validation set, and a test set for the first layer LSTM, and dividing the InSAR slope deformation dataset of the second layer LSTM into a training set, a validation set, and a test set for the second layer LSTM.

[0056] In step 206 above, the first layer LSTM is trained using the InSAR slope deformation dataset of the first layer LSTM. Specifically, the first layer LSTM can be trained using the training set of the first layer LSTM to obtain the preliminary first layer LSTM. The preliminary first layer LSTM is then tested using the test set of the first layer LSTM to obtain the tested first layer LSTM. Finally, the tested first layer LSTM is validated using the validation set of the first layer LSTM, and the first layer LSTM of the two-layer LSTM landslide deformation prediction network model is obtained.

[0057] In step 206 above, the second-layer LSTM is trained using the InSAR slope deformation dataset of the second-layer LSTM. The detailed process of training the second-layer LSTM using the training set, validation set, and test set can be found in the previous paragraph and will not be repeated here.

[0058] In step 206 above, the dual-layer LSTM landslide deformation prediction model can be finally obtained.

[0059] In practice, during the pre-training of the two-layer LSTM landslide deformation prediction model, the target area is selected. After the two-layer LSTM landslide deformation prediction model is trained and generated, when using this model to predict the actual landslide deformation of the area to be predicted, the area to be predicted can be used. The process of pre-training the two-layer LSTM landslide deformation prediction model can be seen in the following process of using this model to predict the actual landslide deformation of the area to be predicted.

[0060] Second, we will introduce the process of using a trained two-layer LSTM landslide deformation prediction model to predict the actual landslide deformation of the area to be predicted.

[0061] In step 101 above, InSAR line-of-sight landslide deformation data of the area to be predicted and its neighborhood (at least one adjacent area) are collected, such as... Figure 3 As shown, the green area is the area to be predicted (P5), and the yellow area is the adjacent area of ​​the area to be predicted (P1-P4, P6-P9). The model finally uses historical data of the area to be predicted and its adjacent areas to predict the landslide deformation of the target area.

[0062] In step 101 above, the InSAR line-of-sight landslide deformation data can be further decomposed into three dimensions into slope-direction deformation data, which can then be used to predict landslide deformation.

[0063] Unlike existing technologies that use LSTM to predict landslide deformation in a single area, this invention uses a two-layer LSTM combined with historical data of the target area and its neighborhood to predict landslide deformation. The following describes the process in conjunction with... Figures 1 to 11 The process of using a two-layer LSTM to predict landslide deformation is explained.

[0064] In step 102 above, the neighborhood node data (InSAR line-of-sight landslide deformation data of at least one adjacent region of the region to be predicted, which can be slope-direction deformation data) is input into the first LSTM layer. The first LSTM layer has an input feature dimension of 1 and a hidden layer dimension of 64, and uses the sliding window method to generate the input sequence and target value. The window size is set to 5, that is, the first LSTM layer extracts neighborhood features from the neighborhood data (data of P1-P4, P6-P9) of the first 5 time steps to obtain the feature vector of at least one adjacent region (neighborhood).

[0065] In step 103 above, the output of the first LSTM layer (feature vector of each neighborhood point) is multiplied by the neighborhood influence coefficient matrix, the historical data of the center node is multiplied by the influence coefficient of the center node, and then concatenated. This concatenation realizes the fusion of the features of the neighborhood nodes (extracted by the first LSTM layer) and the features of the center node (which can be called the target region when training the model, and can be called the region to be predicted when using the trained model to predict actual landslide deformation), providing more comprehensive input information for the second LSTM layer.

[0066] To make it easier to understand, the following example illustrates the process of concatenating the following data:

[0067] Assume matrix A is the output of the first LSTM layer, with a shape of [18,2], and assume matrix B is the neighborhood influence coefficient matrix, with a shape of [8,1]:

[0068]

[0069] After transposing matrix A, we get:

[0070]

[0071] The first-layer LSTM output is fused with the domain influence coefficient matrix: A T B;

[0072] ;

[0073] Since the sliding window is 5, assume that the historical data of the central node is matrix C;

[0074] ;

[0075] Assume the influence coefficient of the central node is 0.5;

[0076] The historical data of the central node and the influence coefficient of the central node are combined as follows:

[0077] ,

[0078] D is the first matrix;

[0079] The weighted data of the neighboring nodes is concatenated with the weighted data of the center node:

[0080] Since the sliding window of the second LSTM layer is also 5, A T The shape of B is expanded from (1, 2) to (5, 2) to obtain matrix E (the second matrix):

[0081]

[0082] Connect matrices E and D TBy concatenating the matrices, we obtain matrix F (the concatenated matrix):

[0083]

[0084] The concatenated matrix F is used as the input to the second LSTM layer.

[0085] When training the model, a neighboring node can be referred to as at least one neighboring region of the target region. When using the trained model to predict actual landslide deformation, it can be referred to as at least one neighboring region of the region to be predicted. The influence coefficients of the central node and the neighboring nodes will be obtained through self-learning during the model training process. The specific implementation will be introduced in subsequent steps. The influence coefficient of each node represents the importance of that node to the prediction result.

[0086] In step 104 above, the data spliced ​​in step 103 above is input into the second LSTM layer. The input dimension of the second LSTM layer is 65 (64-dimensional features output by the first LSTM layer plus 1-dimensional features of the center node). The hidden layer dimension is 64, and finally the landslide deformation prediction result is obtained through the fully connected layer.

[0087] The key to predicting landslide deformation using a two-layer LSTM model based on target nodes and their neighboring nodes lies in how to obtain the influence coefficients of the central node (which can be called the influence coefficient of the target region during model training, and the influence coefficient of the region to be predicted during actual landslide deformation prediction using the trained model) and the influence coefficients of neighboring nodes. This invention innovatively proposes a self-learning device for the influence coefficients of the central node and a self-learning device for the influence coefficients of the neighboring nodes (which can be called the influence coefficient of at least one adjacent region of the target region during model training, and the influence coefficient of at least one adjacent region of the region to be predicted during actual landslide deformation prediction using the trained model). This enables the two-layer LSTM landslide deformation prediction model to self-learn and obtain the influence coefficients of each node during the training process. The following section combines... Figures 4 to 7 The self-learning process of the influence coefficient of the central node is described in detail.

[0088] Figure 4 This is a schematic diagram of the process for obtaining the influence coefficient of the region to be predicted in an embodiment of the present invention. In one embodiment, as shown... Figure 4 As shown, the above-mentioned method for predicting landslide deformation can also include obtaining the influence coefficient of the area to be predicted using the following method:

[0089] Step 301: Create a tuple (T1, W1) and a container to constrain the influence coefficients of the region to be predicted:

[0090] Step 302: When the container corresponding to the region to be predicted is not full, the influence coefficient W1 of the region to be predicted in the tuple is updated in a monotonically decreasing manner. After each round of training, the new tuple is inserted into the container to obtain the updated container. The tuples in the updated container are sorted according to the size of the T1 value in the tuple. The tuples with larger T1 values ​​are at the top of the container, and the tuples with smaller T1 values ​​are at the bottom of the container.

[0091] Step 303: When the container corresponding to the region to be predicted is full, remove a preset number of tuples at the top of the container that meet the preset removal conditions, reduce the upper limit of the container by a preset number, calculate the maximum value W1max and the minimum value W1min of W1 in the remaining tuples in the container, and update W1 in the tuples in a sinusoidal oscillation manner within the interval [W1max, W1min] to obtain the updated tuples; each training round generates an updated tuple and compares it with the tuple at the top of the container. If the T1 value of the new tuple is greater than the T1 value of the tuple at the top of the container, no operation is performed. If the T1 value of the updated tuple is less than the T1 value of the tuple at the top of the container, the updated tuple is inserted into the container, sorted, and the interval [W1max, W1min] is updated.

[0092] Step 304: For every preset number of rounds (which can be 50 rounds as mentioned below), remove a preset number of tuples from the top of the container of the region to be predicted, and reduce the upper limit of the container size of the region to be predicted by a preset number. When the upper limit of the container is less than or equal to twice the preset number, the container of the region to be predicted will no longer be reduced, and the optimal influence coefficient of the region to be predicted will be obtained.

[0093] In specific implementation, in step 301 above, a tuple (T1, W1) and a container are created to constrain the influence coefficients of the central node:

[0094] First, a tuple (T1, W1) is created. After each training round, the model calculates the loss values ​​on the training and validation sets. The loss on the training set is called Training Loss, and the loss on the validation set is called Val Loss. In the tuple (T1, W1), T1 is the weighted sum of Training Loss and Val Loss, used to quantify the quality of a training round. A larger T1 value indicates a poor training result, and a smaller T1 value indicates a better training result. W1 is the influence coefficient of the center node. W1 is updated after each training round, thus creating a new tuple for each training round.

[0095] Create a container to store the tuple (W1, T1). Set the container size and the initial value of W1. In practice, the container size can be 150 and W1 can be 3. The purpose of the tuple and the container is to continuously strengthen the constraint on the influence coefficient of the central node in order to find the optimal coefficient.

[0096] In specific implementation, in step 302 above, the update method for W1 when the container is not full is as follows:

[0097] When the container is not full, W1 is updated in a monotonically decreasing manner. After each round of training, a new tuple is inserted into the container. Then, the tuples in the container are sorted according to their T1 values, with the larger T1 value at the top of the container and the smaller T1 value at the bottom.

[0098] In specific implementation, in step 303 above, the W1 update method when the container is full is as follows:

[0099] Once the container is full, the top 10 tuples are first removed (removing poor results and eliminating the influence coefficient of poorly performing central nodes), and the upper limit of the container is reduced by 10. Then, W1max and W1min of the remaining tuples in the container are calculated. Afterward, W1 is updated in a sinusoidal oscillation pattern within the interval [W1max, W1min]. The central node (the region to be predicted) W1 is updated in a sinusoidal oscillation pattern within the interval [W1max, W1min] as follows. Figure 5 As shown, Figure 5 This is a schematic diagram illustrating the change of the influence coefficient W1 of the region to be predicted in the container in an embodiment of the present invention.

[0100] Each training round generates a new tuple, which is compared to the top tuple of the container. If the T value of the new tuple is greater than the T value of the top tuple (indicating that the training result in this round is worse than all the results in the container), no operation is performed. If it is less, the new tuple is inserted into the container, sorted, and [W1max, W1min] is updated. The operations on the new tuple and the updates to [W1max, W1min] in each training round are as follows: Figure 6 As shown, Figure 6 This is a schematic diagram of the container update of the central node (including the region to be predicted) in an embodiment of the present invention.

[0101] In practice, in step 304 above, the container changes during the training process:

[0102] Every 50 training rounds, the top 10 tuples (a preset number) of the container are removed, and the container size limit is reduced by 10. When the container size limit is less than or equal to 20, the container size is no longer reduced. This process gradually eliminates W1 values ​​that lead to poor training results and gradually narrows the range [W1max, W1min]. Furthermore, the updates to W1 are sinusoidally oscillated within the range [W1max, W1min], ultimately determining the optimal influence coefficient for the central node. In other words, each training round generates different W and T values ​​(the size of the T value reflects the quality of the training round) and sorts them within the container according to their T values. "Poor" refers to a comparison with other (W, T) tuples within the container; a larger T value (greater than the preset T value) indicates a poorer result.

[0103] In one embodiment, the above-mentioned method for predicting landslide deformation may further include: additional constraint processing for the influence coefficient of the area to be predicted: the influence coefficient of the area to be predicted is also subject to non-negative constraints and constraints that are not less than the influence coefficient of the adjacent area.

[0104] In practical implementation, other constraints on the influence coefficient of the central node are as follows:

[0105] The influence coefficient of the central node is further constrained by two constraints: "non-negative" and "not less than the influence coefficient of the neighboring nodes". Such constraints are consistent with the actual physical meaning and reality.

[0106] The changes in the influence coefficient of the central node during training are as follows: Figure 7 As shown, Figure 7 This is a graph showing the change in the influence coefficient of the central node (including the region to be predicted) during model training in this embodiment of the invention. As the container shrinks and W1 with poor training performance is eliminated, the range of [W1max, W1min] gradually shrinks, thus obtaining the optimal influence coefficient of the central node.

[0107] The following is combined Figures 8 to 10 This document details the self-learning device for the influence coefficients of neighboring nodes.

[0108] In one embodiment, such as Figure 8 As shown, Figure 8 The flowchart illustrating the process of obtaining the influence coefficient of adjacent regions (neighbor node influence coefficient) of the area to be predicted in an embodiment of the present invention is shown. The landslide deformation prediction method described above may further include obtaining the influence coefficient of at least one adjacent region according to the following method:

[0109] Step 401: Create a tuple (T2, W2) and a container to constrain the influence coefficients of adjacent regions;

[0110] Step 402: When the container corresponding to the adjacent region is not full, the adjacent region influence coefficient W2 in the tuple is updated during the training process with backpropagation. After each round of training, the new tuple is inserted into the container to obtain the updated container. The tuples in the updated container are sorted according to the size of the T2 value in the tuple. The tuples with larger T2 values ​​are at the top of the container, and the tuples with smaller T2 values ​​are at the bottom of the container.

[0111] Step 403: When the container corresponding to the adjacent region is full, remove the preset number of tuples at the top of the container that meet the preset removal conditions, reduce the upper limit of the container by the preset number, calculate the maximum value W2max and the minimum value W2min of W2 in the remaining tuples in the container, and update W2 in the tuples in a sinusoidal oscillation manner without strictly adhering to the interval [W2max, W2min].

[0112] Step 404: Every second round (which can be 100 rounds as mentioned below), remove a preset number of tuples from the top of the container corresponding to the adjacent region, and reduce the upper limit of the container size corresponding to the adjacent region by a preset number. When the upper limit of the container is less than or equal to five times the preset number, the container corresponding to the adjacent region will no longer be reduced, thus obtaining the optimal influence coefficient of the adjacent region.

[0113] In specific implementation, in step 401 above, a container and tuple (T2, W2) are created to constrain the influence coefficients of neighboring nodes:

[0114] First, a tuple and container are created for each node. Unlike the center node, the initial W2 of the neighboring nodes is 0 (because after multiple experiments, it was found that the optimal W2 of the neighboring nodes is basically around 0, so it is not necessary to monotonically decrease to search a larger range like the center node). Please refer to the definitions of T1 and W1 above. The definition of T2 and W2 is: to create a tuple (T2, W2). After each round of training, the model calculates the loss value of that round of training on the training set and the validation set. The loss value on the training set is Train Loss, and the loss value on the validation set is Val Loss. In the tuple (T2, W2), T2 refers to the weighted sum of Train Loss and Val Loss, which is used to quantify the quality of a round of training. The larger the T2 value, the worse the training result; the smaller the T2 value, the better the training result. W2 is the influence coefficient of the neighboring nodes. W2 is updated after each round of training, thus creating a new tuple for each round of training.

[0115] In specific implementation, the update method for the influence coefficient of neighboring nodes when the container is not full in step 402 above is as follows:

[0116] Unlike W1, which is updated monotonically when the container is not full, W2 can be initially 0. The neighborhood node influence coefficient W2 is updated during training with backpropagation. After each training round, the newly generated tuple (W2, T2) is inserted into the container and sorted according to the size of the T2 value. Tuples with larger T2 values ​​are placed at the top of the container and those with smaller T2 values ​​are placed at the bottom. Therefore, the container of the neighborhood node should also be as large as possible, such as 15, so that when the container is not full, the neighborhood node W2 can try more training effects with different influence coefficients.

[0117] In specific implementation, in step 403 above, the W2 update method for neighboring nodes when the container is full is as follows:

[0118] When the neighborhood node container is full, the top 10 tuples are first removed (removing poor results and eliminating the influence coefficients of poor neighborhood nodes), and the upper limit of the container is reduced by 10. Then, W2max and W2min of the remaining tuples in the container are calculated. Afterwards, W2 updates will be performed in a sinusoidal oscillation within the interval [W2max, W2min]. Unlike the center node, the neighborhood nodes do not strictly oscillate sinusoidally within the interval [W2max, W2min], but have a certain ability to "cross the boundary." This is because once a coefficient of the center node is eliminated by the container, that coefficient and all its neighbors are eliminated. However, the elimination of a coefficient of a neighborhood node does not completely negate the possibility that that coefficient and its neighbors are bad (perhaps the poor result was caused by the update of the center node's coefficient). Therefore, it is necessary to repeatedly verify the numbers near the neighborhood of that coefficient. The specific approach is as follows:

[0119] 1. When updating W2 using sinusoidal oscillation, a random offset should be added. In this embodiment of the invention, the range of the random offset can be [0, (W2max-W2min)]. 0.3], which means it can be [0, 0.3]. A random number between [W2max-W2min] .

[0120] 2. Instead of updating [W2max, W2min] in every training round, W2max and W2min are updated with a delay, only once every 20 or even 50 training rounds.

[0121] When the container is full, the changes in the update of the neighboring node W2 within the container are as follows: Figure 9 As shown, Figure 9 This is a schematic diagram of the change of the influence coefficient W2 of the adjacent region (neighboring node) in the container in an embodiment of the present invention. The update of the neighboring node is not strictly a sinusoidal oscillation in the interval [W2max, W2min], but a random offset and a delay in updating [W2max, W2min] are added when updating W2.

[0122] In specific implementation, the changes of the neighborhood node container during the training process in step 404 above are as follows:

[0123] The container constraints on neighboring nodes are also weaker than those on the center node. For example, the center node shrinks its container every 50 training iterations, while the neighboring nodes shrink it only once every 100 training iterations. The lower bound for the shrinkage of the center node's container is 20, while that of the neighboring nodes is 50 (five times the preset number of 10). Compared to stricter constraints, more lenient constraints on neighboring nodes have a better effect. This is because more lenient constraints require more training iterations for a certain coefficient before it is eliminated by the container. Random offsets and delayed updates [W2max, W2min] mean that a coefficient within a certain range may be eliminated by the container multiple times before it is completely eliminated. The container of the center node plays the role of continuously shrinking the range of values ​​to determine the optimal influence coefficient, while the container of the neighboring nodes is more inclined to guide the direction.

[0124] The changes in the influence coefficient of neighboring nodes during training are as follows: Figure 10 As shown, Figure 10 This is a graph showing the change in the influence coefficient of adjacent regions during model training in an embodiment of the present invention.

[0125] By employing a self-learning device for the influence coefficients of the central node and a self-learning device for the neighboring nodes, the model can self-learn to obtain the influence coefficients of each node during training, thereby obtaining the influence coefficient matrix. In step 103 above, the concatenated data is multiplied by the influence coefficient matrix, and the result is then input into the second LSTM layer to obtain the prediction result. The final prediction result is as follows: Figure 11 As shown, Figure 11 This is a schematic diagram of the landslide deformation prediction results in an embodiment of the present invention.

[0126] In specific implementation, other model optimizations and hyperparameter settings were as follows: Optimizer: Adam (learning rate 0.0001), model training epochs: 1200 epochs, gradient clipping: threshold set to 1.0 to prevent gradient explosion. L2 regularization: L2 regularization was applied to the center node coefficient W1 to prevent overfitting. Optimal model selection: records were made during training. T The model parameters with the minimum training loss and validation loss are used as the final model to avoid overfitting.

[0127] Unlike existing technologies that use historical data from a single region for prediction, this invention combines the target node and its neighboring node data to construct a two-layer LSTM for prediction. In addition, it introduces a self-learning algorithm for the influence coefficients of the target node and its neighboring nodes, which enables the model to automatically learn the importance of each node's data to the prediction result, thereby significantly improving the prediction accuracy of landslide deformation.

[0128] This invention also provides a landslide deformation prediction device, as described in the following embodiments. Since the principle behind this device is similar to that of the landslide deformation prediction method, its implementation can be referenced from the implementation of the landslide deformation prediction method; repeated details will not be elaborated further.

[0129] Figure 12 This is a schematic diagram of the landslide deformation prediction device in an embodiment of the present invention, as shown below. Figure 12 As shown, the device includes:

[0130] Acquisition unit 01 is used to acquire synthetic aperture radar interferometric InSAR landslide deformation data along the line of sight of the region to be predicted, and InSAR landslide deformation data along the line of sight of at least one adjacent region of the region to be predicted.

[0131] The first LSTM processing unit 02 is used to input the InSAR landslide deformation data along the line of sight of at least one neighboring region of the area to be predicted into the first LSTM of the two-layer long short memory neural network LSTM landslide deformation prediction model, and extract the feature vector of at least one neighboring region; the first LSTM is generated by prediction training based on the historical relationship sample data between the InSAR landslide deformation data along the line of sight of at least one neighboring region of the target area and the feature vector of at least one neighboring region.

[0132] Processing unit 03 is used to multiply the InSAR landslide deformation data along the line of sight of the area to be predicted with the influence coefficient matrix of the area to be predicted to obtain a first matrix, and to multiply the feature vector of at least one adjacent area with the influence coefficient matrix of at least one adjacent area to obtain a second matrix; the first matrix and the second matrix are concatenated to obtain a concatenated matrix of the area to be predicted and at least one adjacent area; the influence coefficients in the matrix are obtained by self-learning during the model training process, and the influence coefficients of each area represent the importance of that area to the landslide deformation prediction result;

[0133] The second-layer LSTM processing unit 04 is used to input the splicing matrix into the second-layer LSTM of the double-layer LSTM landslide deformation prediction model to obtain the final prediction result of the landslide deformation in the area to be predicted; the second-layer LSTM is generated by prediction training based on the historical relationship sample data between the splicing matrix of the target area and at least one of its adjacent areas and the final landslide deformation prediction result of the target area.

[0134] In one embodiment, the landslide deformation prediction device may further include a training unit for: pre-training a two-layer LSTM landslide deformation prediction model according to the following method:

[0135] Acquire InSAR landslide deformation data along the line of sight for the target area and at least one adjacent area;

[0136] The InSAR landslide deformation data of the target area and at least one adjacent area along the line of sight are decomposed into three dimensions to obtain the InSAR deformation dataset along the slope direction of the target area and at least one adjacent area.

[0137] A two-layer LSTM landslide deformation prediction network is constructed, which includes a first LSTM layer and a second LSTM layer.

[0138] The InSAR slope deformation dataset of the first layer LSTM is constructed by combining the InSAR slope deformation data of at least one adjacent region of the target area along the line of sight with the historical relationship sample data of the feature vectors of at least one adjacent region.

[0139] The InSAR slope deformation dataset of the target area and at least one adjacent area is processed into a spliced ​​matrix of the target area and at least one adjacent area, and the historical relationship sample data of the final landslide deformation prediction result of the target area constitutes the InSAR slope deformation dataset of the second layer LSTM.

[0140] The first-layer LSTM is trained using the InSAR slope deformation dataset of the first-layer LSTM, and the second-layer LSTM is trained using the InSAR slope deformation dataset of the second-layer LSTM, thus obtaining the two-layer LSTM landslide deformation prediction model.

[0141] In one embodiment, the landslide deformation prediction device may further include a prediction region influence coefficient determination unit, used to obtain the prediction region influence coefficient according to the following method:

[0142] Create a tuple (T1, W1) and a container to constrain the influence coefficients of the region to be predicted:

[0143] When the container corresponding to the region to be predicted is not full, the influence coefficient W1 of the region to be predicted in the tuple is updated in a monotonically decreasing manner. After each round of training, a new tuple is inserted into the container to obtain an updated container. The tuples in the updated container are sorted according to the size of the T1 value in the tuple. The tuples with larger T1 values ​​are at the top of the container, and the tuples with smaller T1 values ​​are at the bottom of the container.

[0144] When the container corresponding to the region to be predicted is full, a preset number of tuples that meet the preset removal conditions at the top of the container are removed, the upper limit of the container is reduced by a preset number, and the maximum value W1max and the minimum value W1min of W1 in the remaining tuples in the container are calculated. The update of W1 in the tuples will be updated in a sinusoidal oscillation manner within the interval [W1max, W1min] to obtain the updated tuples. The updated tuples generated in each training round are compared with the tuples at the top of the container. If the T1 value of the new tuple is greater than the T1 value of the tuple at the top of the container, no operation is performed. If the T1 value of the updated tuple is less than the T1 value of the tuple at the top of the container, the updated tuple is inserted into the container, sorted, and the interval [W1max, W1min] is updated.

[0145] In each preset first training round, a preset number of tuples at the top of the container of the region to be predicted are removed, and the upper limit of the container size of the region to be predicted is reduced by a preset number. When the upper limit of the container is less than or equal to twice the preset number, the container of the region to be predicted is no longer reduced, and the optimal influence coefficient of the region to be predicted is obtained.

[0146] In one embodiment, the above-mentioned unit for determining the influence coefficient of the region to be predicted is further used for: other constraint processing of the influence coefficient of the region to be predicted: the influence coefficient of the region to be predicted is also subject to non-negative constraints and constraints that are not less than the influence coefficient of the adjacent region.

[0147] In one embodiment, the landslide deformation prediction device may further include an adjacent area influence coefficient determination unit, used to obtain the influence coefficient of at least one adjacent area according to the following method:

[0148] Create a tuple (T2, W2) and a container to constrain the influence coefficients of adjacent regions;

[0149] When the container corresponding to the adjacent region is not full, the adjacent region influence coefficient W2 in the tuple is updated during the training process with backpropagation. After each round of training, the new tuple is inserted into the container to obtain the updated container. The tuples in the updated container are sorted according to the size of the T2 value in the tuple, with the larger T2 value at the top of the container and the smaller T2 value at the bottom of the container.

[0150] When the container corresponding to the adjacent region is full, a preset number of tuples that meet the preset removal conditions at the top of the container are removed, the upper limit of the container is reduced by a preset number, and the maximum value W2max and the minimum value W2min of W2 in the remaining tuples in the container are calculated. The update of W2 in the tuple will be updated in a sinusoidal oscillation manner within the interval [W2max, W2min] without being strictly limited to this interval.

[0151] In each second training round, a predetermined number of tuples at the top of the corresponding container in the adjacent region are removed, and the upper limit of the size of the corresponding container in the adjacent region is reduced by a predetermined number. When the upper limit of the container is less than or equal to five times the predetermined number, the corresponding container in the adjacent region is no longer reduced, thus obtaining the optimal influence coefficient of the adjacent region.

[0152] In one embodiment, the update of W2 in the tuple will not strictly follow a sinusoidal oscillation within the interval [W2max, W2min], and may include:

[0153] When W2 performs a sinusoidal oscillation update, a random offset should be added;

[0154] And instead of updating the interval [W2max, W2min] in each training round, W2max and W2min are updated with a delay.

[0155] This invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-described method for predicting landslide deformation.

[0156] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for predicting landslide deformation.

[0157] This invention also provides a computer program product, which includes a computer program that, when executed by a processor, implements the above-described method for predicting landslide deformation.

[0158] In this embodiment of the invention, the landslide deformation prediction scheme, compared with the prior art that only uses historical data of the target area to predict future landslide deformation and ignores information of the surrounding area, proceeds by: acquiring synthetic aperture radar interferometry (InSAR) landslide deformation data along the line of sight of the area to be predicted, and InSAR landslide deformation data along the line of sight of at least one adjacent area of ​​the area to be predicted; inputting the InSAR landslide deformation data along the line of sight of at least one adjacent area of ​​the area to be predicted into the first layer LSTM of the two-layer long short-term memory neural network (LSTM) landslide deformation prediction model, and extracting feature vectors of at least one adjacent area; the first layer LSTM is generated by prediction training based on historical relationship sample data between the InSAR landslide deformation data along the line of sight of at least one adjacent area of ​​the target area and the feature vectors of at least one adjacent area; and inputting the InSAR landslide deformation data along the line of sight of the area to be predicted into the first layer LSTM of the landslide deformation prediction model, and extracting feature vectors of at least one adjacent area; the first layer LSTM is generated by prediction training based on historical relationship sample data between the InSAR landslide deformation data along the line of sight of at least one adjacent area of ​​the target area and the feature vectors of at least one adjacent area; and the InSAR landslide deformation data along the line of sight of the area to be predicted is then processed. The landslide deformation data is multiplied by the influence coefficient matrix of the area to be predicted to obtain the first matrix. The feature vector of at least one adjacent area is multiplied by the influence coefficient matrix of at least one adjacent area to obtain the second matrix. The first and second matrices are concatenated to obtain the concatenated matrix of the area to be predicted and at least one adjacent area. The influence coefficients in the matrix are obtained through self-learning during model training. The influence coefficient of each area represents the importance of that area to the landslide deformation prediction result. The concatenated matrix is ​​input into the second layer of the two-layer LSTM landslide deformation prediction model to obtain the final prediction result of the landslide deformation of the area to be predicted. The second layer LSTM is generated by prediction training based on the historical relationship sample data between the concatenated matrix of the target area and at least one adjacent area and the final landslide deformation prediction result of the target area. It can accurately predict landslide deformation by means of the correlation between the deformation of the area to be predicted and at least one adjacent area.

[0159] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0160] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0161] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0162] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0163] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting landslide deformation, characterized in that, include: Acquire synthetic aperture radar interferometric (InSAR) landslide deformation data along the line of sight for the region to be predicted, and InSAR landslide deformation data along the line of sight for at least one adjacent region of the region to be predicted. The InSAR landslide deformation data along the line of sight of at least one neighboring region of the area to be predicted is input into the first LSTM layer of the two-layer long short-term memory neural network LSTM landslide deformation prediction model to extract the feature vector of at least one neighboring region; the first LSTM layer is generated by prediction training based on the historical relationship sample data between the InSAR landslide deformation data along the line of sight of at least one neighboring region of the target area and the feature vector of at least one neighboring region. The first matrix is ​​obtained by multiplying the InSAR landslide deformation data along the line of sight of the area to be predicted with the influence coefficient matrix of the area to be predicted. The second matrix is ​​obtained by multiplying the feature vector of at least one adjacent area with the influence coefficient matrix of at least one adjacent area. The first and second matrices are concatenated to obtain the concatenated matrix of the area to be predicted and at least one adjacent area. The influence coefficients in the matrix are obtained by self-learning during the model training process. The influence coefficient of each area represents the importance of that area to the landslide deformation prediction result. The splicing matrix is ​​input into the second LSTM layer of the two-layer LSTM landslide deformation prediction model to obtain the final prediction result of the landslide deformation in the area to be predicted. The second-layer LSTM is generated based on historical relationship sample data between the splicing matrix of the target region and at least one of its adjacent regions and the final landslide deformation prediction result of the target region.

2. The method as described in claim 1, characterized in that, It also includes pre-training a two-layer LSTM landslide deformation prediction model using the following method: Acquire InSAR landslide deformation data along the line of sight for the target area and at least one adjacent area; The InSAR landslide deformation data of the target area and at least one adjacent area along the line of sight are decomposed into three dimensions to obtain the InSAR deformation dataset along the slope direction of the target area and at least one adjacent area. A two-layer LSTM landslide deformation prediction network is constructed, which includes a first LSTM layer and a second LSTM layer. The InSAR slope deformation dataset of the first layer LSTM is constructed by combining the InSAR slope deformation data of at least one adjacent region of the target area along the line of sight with the historical relationship sample data of the feature vectors of at least one adjacent region. The InSAR slope deformation dataset of the target area and at least one adjacent area is processed into a spliced ​​matrix of the target area and at least one adjacent area, and the historical relationship sample data of the final landslide deformation prediction result of the target area constitutes the InSAR slope deformation dataset of the second layer LSTM. The first-layer LSTM is trained using the InSAR slope deformation dataset of the first-layer LSTM, and the second-layer LSTM is trained using the InSAR slope deformation dataset of the second-layer LSTM, thus obtaining the two-layer LSTM landslide deformation prediction model.

3. The method as described in claim 1, characterized in that, It also includes obtaining the influence coefficient of the region to be predicted using the following method: Create a tuple (T1, W1) and a container to constrain the influence coefficients of the region to be predicted: When the container corresponding to the region to be predicted is not full, the influence coefficient W1 of the region to be predicted in the tuple is updated in a monotonically decreasing manner. After each round of training, a new tuple is inserted into the container to obtain an updated container. The tuples in the updated container are sorted according to the size of the T1 value in the tuple. The tuples with larger T1 values ​​are at the top of the container, and the tuples with smaller T1 values ​​are at the bottom of the container. When the container corresponding to the region to be predicted is full, a preset number of tuples that meet the preset removal conditions at the top of the container are removed, the upper limit of the container is reduced by a preset number, and the maximum value W1max and the minimum value W1min of W1 in the remaining tuples in the container are calculated. The update of W1 in the tuples will be updated in a sinusoidal oscillation manner within the interval [W1max, W1min] to obtain the updated tuples. The updated tuples generated in each training round are compared with the tuples at the top of the container. If the T1 value of the new tuple is greater than the T1 value of the tuple at the top of the container, no operation is performed. If the T1 value of the updated tuple is less than the T1 value of the tuple at the top of the container, the updated tuple is inserted into the container, sorted, and the interval [W1max, W1min] is updated. In each preset first training round, a preset number of tuples at the top of the container of the region to be predicted are removed, and the upper limit of the container size of the region to be predicted is reduced by a preset number. When the upper limit of the container is less than or equal to twice the preset number, the container of the region to be predicted is no longer reduced, and the optimal influence coefficient of the region to be predicted is obtained.

4. The method as described in claim 3, characterized in that, Also includes: Other constraints on the influence coefficient of the region to be predicted: The influence coefficient of the region to be predicted is also subject to non-negativity constraints and constraints that it is not less than the influence coefficient of the adjacent region.

5. The method as described in claim 1, characterized in that, It also includes obtaining the influence coefficient of at least one adjacent region using the following method: Create a tuple (T2, W2) and a container to constrain the influence coefficients of adjacent regions; When the container corresponding to the adjacent region is not full, the adjacent region influence coefficient W2 in the tuple is updated during the training process with backpropagation. After each round of training, the new tuple is inserted into the container to obtain the updated container. The tuples in the updated container are sorted according to the size of the T2 value in the tuple, with the larger T2 value at the top of the container and the smaller T2 value at the bottom of the container. When the container corresponding to the adjacent region is full, a preset number of tuples that meet the preset removal conditions at the top of the container are removed, the upper limit of the container is reduced by a preset number, and the maximum value W2max and the minimum value W2min of W2 in the remaining tuples in the container are calculated. The update of W2 in the tuple will be updated in a sinusoidal oscillation manner within the interval [W2max, W2min] without being strictly limited to this interval. In each second training round, a predetermined number of tuples at the top of the corresponding container in the adjacent region are removed, and the upper limit of the size of the corresponding container in the adjacent region is reduced by a predetermined number. When the upper limit of the container is less than or equal to five times the predetermined number, the corresponding container in the adjacent region is no longer reduced, thus obtaining the optimal influence coefficient of the adjacent region.

6. The method as described in claim 5, characterized in that, The update of W2 in the tuple will not strictly follow a sinusoidal oscillation within the interval [W2max, W2min], including: When W2 performs a sinusoidal oscillation update, a random offset should be added; And instead of updating the interval [W2max, W2min] in each training round, W2max and W2min are updated with a delay.

7. A landslide deformation prediction device, characterized in that, include: The acquisition unit is used to acquire synthetic aperture radar interferometric InSAR landslide deformation data along the line of sight of the region to be predicted, and InSAR landslide deformation data along the line of sight of at least one adjacent region of the region to be predicted. The first LSTM processing unit is used to input the InSAR landslide deformation data along the line of sight of at least one neighboring region of the area to be predicted into the first LSTM of the two-layer long short memory neural network LSTM landslide deformation prediction model, and extract the feature vector of at least one neighboring region; the first LSTM is generated by prediction training based on the historical relationship sample data between the InSAR landslide deformation data along the line of sight of at least one neighboring region of the target area and the feature vector of at least one neighboring region. The processing unit is used to multiply the InSAR landslide deformation data along the line of sight of the area to be predicted with the influence coefficient matrix of the area to be predicted to obtain a first matrix, and to multiply the feature vector of at least one adjacent area with the influence coefficient matrix of at least one adjacent area to obtain a second matrix; the first matrix and the second matrix are concatenated to obtain a concatenated matrix of the area to be predicted and at least one adjacent area; the influence coefficients in the matrix are obtained by self-learning during the model training process, and the influence coefficient of each area represents the importance of that area to the landslide deformation prediction result; The second-layer LSTM processing unit is used to input the splicing matrix into the second-layer LSTM of the double-layer LSTM landslide deformation prediction model to obtain the final prediction result of the landslide deformation in the area to be predicted. The second-layer LSTM is generated based on historical relationship sample data between the splicing matrix of the target region and at least one of its adjacent regions and the final landslide deformation prediction result of the target region.

8. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method of any one of claims 1 to 6.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the method of any one of claims 1 to 6.

Images