Ground surface deformation space-time prediction method combining InSAR and graph neural network

Abstract

The invention discloses an InSAR (Interferometric Synthetic Aperture Radar) and graph neural network combined earth surface deformation space-time prediction method, which is suitable for space-time prediction of earth surface deformation. Based on DS-InSAR, obtaining radar sight line-to-time sequence earth surface deformation of the high-coherence measuring points; calculating the geographic distance between the measuring points and the mutual information between the deformation sequences corresponding to the points; the method comprises the following steps: judging connectivity between high-coherence measuring points according to a geographic distance between the measuring points and mutual information to generate an adjacent matrix, and organizing an original InSAR deformation monitoring result into graph structure data; and inputting the graph structure data into the LSTM-GCN model to predict and obtain the surface deformation of all the high-coherence measuring points in the whole area. The method is high in prediction precision and wide in application range, and can be effectively applied to the fields of space-time prediction of deformation of earth surfaces and buildings (structures) caused by mine closure, underground resource development, natural disasters and the like.

Description

Technical Field

[0001] The present invention designs a spatiotemporal prediction method for surface deformation by combining InSAR with graph neural network, which belongs to the field of geological disaster monitoring. Technical Background

[0002] Traditional InSAR monitoring methods rely too heavily on ground objects with strong scattering properties, resulting in an inability to identify sufficient monitoring points. Distributed Target Interferometry (DS-InSAR) technology, however, does not rely on stable targets on the ground, increasing the density of coherent points in complex areas.

[0003] Traditional surface deformation prediction models use only time series data from highly coherent measurement points, ignoring the spatial connections between these points. Surface subsidence often does not occur in isolation; subsidence in certain areas can affect neighboring areas. Therefore, treating each monitoring point as an independent entity can lose the spatial correlation of surface deformation. Furthermore, deformation predictions typically rely solely on the historical time series of a single measurement point, failing to fully consider the shared evolutionary trends of deformation sequences across different measurement points and overlooking the potential correlated patterns of change between highly connected points, impacting prediction accuracy in complex terrain.

[0004] Prior art CN116188969A discloses a surface deformation prediction method that integrates InSAR and spatiotemporal convolution. The method includes the following steps: acquiring surface deformation using PS-InSAR; constructing a surface deformation sample extraction model, using data from different years (Quaternary thickness, phreatic water level, first confined water level, second confined water level, and third confined water level) as inputs, and using surface deformation information as outputs to obtain a spatiotemporal sample set. This method requires a lot of computing power, is computationally complex, and has low efficiency and economical efficiency. Summary of the Invention

[0005] In response to the shortcomings of existing technologies, a spatiotemporal prediction method for surface deformation that combines InSAR with graph neural networks is provided. DS-InSAR technology is used to obtain deformation information of surface monitoring points in the survey area (no need to define surface monitoring points in advance). All surface monitoring points are traversed, and a comprehensive metric is generated based on the spatial distance between the two points and the amount of information in the deformation sequence to determine the connectivity of the two points, thereby constructing graph structure data. When constructing graph structure data, the time series data and spatial relationships between the measuring points are considered to more comprehensively reflect the dynamic changes and spatial distribution characteristics of surface deformation. By constructing a spatiotemporal graph structure, the graph neural network model can automatically learn the complex relationships between nodes in the graph structure, without relying on manually defined rules or assumptions, and can adaptively identify relevant measuring points, thereby improving the stability and generalization ability of the prediction results.

[0006] To achieve the above technical objectives, the present invention provides a spatiotemporal prediction method for surface deformation by combining InSAR with graph neural networks, which includes the following steps:

[0007] Step 1: Obtain InSAR data within the study time range of the survey area, including SAR images, precise orbit data, atmospheric correction data, and DEM data;

[0008] Step 2: Use the DS-InSAR solution method to obtain all high-coherence measurement points in the SAR image. The high-coherence measurement points include longitude and latitude, deformation rate, date, and corresponding time series deformation.

[0009] Step 3: Calculate the spatial distance D between any two highly coherent measurement points based on their longitude and latitude, and calculate the mutual information I between the deformation sequences of any two highly coherent measurement points. Define the comprehensive metric Q using the spatial distance D and the mutual information I.

[0010] Step 4: Set the comprehensive metric threshold to Q t , if Q>Q t , then it is determined that there is connectivity between the two high coherence measurement points, otherwise it is determined that there is no connectivity between the two points; traverse all high coherence measurement points and construct a graph structure based on the connectivity between the two high coherence measurement points;

[0011] In step 5, the weighted graph neural network model LSTM-GCN is used to directly obtain the deformation variables corresponding to high-coherence measurement points in the survey area within a certain continuous period in the future through graph structure data, thereby realizing the spatiotemporal prediction of surface deformation.

[0012] Furthermore, the steps for obtaining the time-series high-coherence surface deformation information of the measurement points in the line of sight of the SAR image radar using DS-InSAR are as follows:

[0013] Step 1.1, identifying homogeneous points: using the homogeneous point selection method based on confidence interval hypothesis test (HTCI) to compare the similarity of all SAR images, the central pixel with more than 20 homogeneous pixels is identified as a distributed scatterer target, i.e., a DS candidate point;

[0014] Step 1.2, coherence analysis: perform coherence analysis on the DS candidate points, and select the candidate points with temporal coherence greater than 0.6 as the final DS points;

[0015] Step 1.3, phase optimization: Optimize the phase of the DS point using the eigenvalue decomposition (EVD) method;

[0016] Step 1.4, select permanent scatterer targets: Based on the intensity of all SAR images, select the central pixel with amplitude deviation less than 0.4 as the permanent scatterer target, i.e., PS point;

[0017] Step 1.5, combine PS points and DS points to solve the surface deformation of the study area: combine all PS points and DS points to construct an irregular triangulation network, use the three-dimensional phase unwrapping algorithm to obtain the unwrapped phase of PS points and DS points, weaken the error factors through high-pass filtering in the time dimension and low-pass filtering in the space dimension, use digital elevation model data to remove the terrain phase, and use universal atmospheric correction data for correction to obtain the latitude and longitude, deformation rate and time series surface deformation of all PS points and DS points in the study area; all PS points and DS points are the high-coherence measurement points in the study area.

[0018] Furthermore, the spatial distance D between any two high coherence measurement points is calculated using the longitude and latitude of the high coherence measurement points, which can be expressed as follows:

[0019] D=R·b

[0020]

[0021] Where x and y are any two high-correlation measurement points, lat x ,lon x is the latitude and longitude of the high coherence measurement point x; lat y ,lon y are the latitude and longitude of the high-coherence measurement point y, R is the radius of the earth, which is 6371 km; atan2 represents the inverse tangent function.

[0022] Furthermore, the process of calculating the mutual information I between the temporal deformations of any two measurement points using the temporal deformations of high coherence measurement points is as follows:

[0023] Assume that the temporal deformation of high coherence measurement point x and high coherence measurement point y is expressed as X={x1,x2...,x n} and Y={y1,y2...,y m}, the information entropy of the nonlinear time series of high coherence measurement points x and y is expressed as:

[0024]

[0025] Among them, p(x i ), p(y j ) respectively indicate that X and Y are x i and y i The probability of X and Y is expressed as follows:

[0026]

[0027] The mutual information I between X and Y is calculated using the following formula:

[0028] I(X,Y)=H(X)+H(Y)-H(XY).

[0029] Furthermore, when using the comprehensive metric Q to determine whether there is spatiotemporal connectivity between any highly coherent measurement points x and y, the specific process of constructing graph structure data is as follows:

[0030] Combining the spatial distance D and the mutual information I of the deformation sequence to construct a comprehensive metric Q, let ω D and ω I are the spatial distance weight and mutual information weight between the high coherence measurement point x and the high coherence measurement point y, respectively, and satisfy ω D +ω I =1, then the comprehensive metric Q is expressed as:

[0031] Q=ω D ·f(D)+ω I I(X,Y)

[0032] Where f(D) is the Gaussian attenuation function of the spatial distance D. The connectivity between the high coherence measurement point x and the high coherence measurement point y decreases as the spatial distance D increases. The Gaussian attenuation function is expressed as:

[0033]

[0034] Among them, σ represents the parameter that controls the attenuation of spatial distance D, and its size determines the speed of change of the impact of distance on connectivity;

[0035] Calculate the comprehensive metric threshold Q using the mean and standard deviation t :

[0036] Q t =μ Q +kσ Q

[0037] Among them, μ Q is the mean of all Q values; σ Q is the standard deviation of all Q values; k is a hyperparameter, usually k = 1 or k = 2 to cover 70% to 95% of the data;

[0038] If Q>Q t , indicating that there is connectivity between high coherence measurement point x and high coherence measurement point y, otherwise there is no connectivity.

[0039] Furthermore, the graph structure data is a temporal deformation and adjacency matrix that records the connection between points. A value of 1 in the matrix indicates that there is a connection between the two points, and a value of 0 indicates that there is no connection between the two points.

[0040] The graph neural network model LSTM-GCN is used to learn the interactions and correlations between different surface monitoring points, capture the spatiotemporal connections between different high-coherence measurement points in the survey area, and improve the prediction of surface deformation. The graph neural network model LSTM-GCN includes an LSTM module responsible for learning the temporal information of high-coherence measurement points and a GCN module responsible for learning spatial information. The graph structure data input into the LSTM-GCN model first enters the LSTM module to encode and learn the temporal evolution information. Then, the GCN module aggregates the spatial features between the measurement points according to the complex relationships between the high-coherence measurement points defined in the adjacency matrix of the graph structure data and updates the temporal evolution information. It can adaptively identify the connectivity between high-coherence measurement points without relying on manually defined rules or assumptions, thereby improving the stability and generalization ability of the prediction results.

[0041] A computer device includes a processor and a memory, wherein the processor is electrically connected to the memory, the memory is used to store instructions and data, and the processor is used to execute a spatiotemporal prediction method for surface deformation combining InSAR and graph neural networks.

[0042] A computer-readable storage medium storing a computer program suitable for being loaded by a processor and executing the spatiotemporal prediction method of surface deformation combining InSAR and graph neural network as described in the claim.

[0043] Beneficial effects: The method of the present invention fully considers the geographical distance between surface monitoring points and the mutual information between their corresponding deformation sequences. Through the mutual constraints between the two, the robustness of the connectivity of the adjacency matrix is ​​improved, thereby obtaining more stable and information-rich graph structure data. Combined with the graph neural network model, the present invention can achieve accurate prediction of future deformations of all points in space. This method has high prediction accuracy and a wide range of applications. It can be effectively applied to fields such as spatiotemporal prediction of surface and building (structure) deformation caused by mine closures, underground resource development, natural disasters, etc. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 Schematic diagram of the process of spatiotemporal prediction of surface deformation by combining InSAR and graph neural network in an embodiment of the present invention.

[0045] Figure 2 Schematic diagram of the adjacency matrix in the measured area in an embodiment of the present invention.

[0046] Figure 3 This is a schematic diagram of the spatiotemporal prediction effect diagram of an embodiment of the present invention, with the prediction effect diagram on the left and the actual DS-InSAR monitoring diagram on the right for comparison.

[0047] Figure 4This is the error distribution diagram of the spatiotemporal prediction in an embodiment of the present invention; the left side is the error distribution diagram of the test area, and the right side is the result diagram of the error distribution on the left side. DETAILED DESCRIPTION

[0048] The present invention will be further described in detail below with reference to the figures and specific implementation process:

[0049] like Figure 1 As shown, a spatiotemporal prediction method for surface deformation combining InSAR and graph neural network of the present invention comprises the following steps:

[0050] Step 1: Obtain SAR images, precise orbit data, atmospheric correction data, and digital elevation models within a preset time period of the study area.

[0051] Step 2: The internal measurement area and research scope of the present invention can be customized by the user. It is only necessary to ensure that the files that meet the InSAR processing requirements are collected to obtain the longitude and latitude, deformation rate and time series deformation information corresponding to the high coherence measurement points in the corresponding area and time period.

[0052] Step 3: To obtain spatially complete and temporally continuous surface deformation information of the survey area and to increase the number and density of selected high-coherence measurement points, the surface deformation information of InSAR high-coherence measurement points is solved based on the HTCI homogeneous point identification method and the EVD phase optimization method. The main steps of data processing are as follows:

[0053] (1) SAR image preprocessing.

[0054] (2) Homogeneous point identification: HTCI test is used to select homogeneous pixels, and the central pixel with more than 20 homogeneous pixels is determined as the DS candidate point.

[0055] (3) Coherence analysis: Perform coherence analysis on the DS candidate points, and select the candidate points with temporal coherence greater than 0.6 as the final DS points.

[0056] (4) Phase optimization: The phase of the final DS point is optimized using the eigenvalue decomposition (EVD) phase optimization method.

[0057] (5) PS point selection: PS points are selected based on amplitude deviation (the threshold is generally 0.4), and points with amplitude deviation less than the threshold are taken as PS points.

[0058] (6) Deformation calculation by combining PS points and DS points: PS points and DS points are combined to construct a Delaunay triangulation network. The unwrapped phase is obtained by using a three-dimensional phase unwrapping algorithm. Error factors such as noise are weakened by high-pass filtering in the time dimension and low-pass filtering in the space dimension. GACOS data are introduced to remove atmospheric effects and obtain surface deformation information.

[0059] Step 4: Perform spatiotemporal prediction based on InSAR surface deformation monitoring points:

[0060] (1) Read the InSAR monitoring result file and calculate the geographical distance D between points based on the longitude and latitude between points using the Haversine formula:

[0061]

[0062] D=R·b

[0063] Where lat1 and lon1 are the latitude and longitude of the first point; lat2 and lon2 are the latitude and longitude of the second point; R is the radius of the Earth (taken as 6371 kilometers); and atan2 represents the inverse tangent function.

[0064] (2) For two nonlinear time series X = {x1, x2..., x n} and Y={y1,y2...,y m}, its information entropy is:

[0065]

[0066] Among them, p(x i ), p(y j ) indicates that the sequences X and Y take values ​​x i and y i The joint information entropy of X and Y can be expressed as:

[0067]

[0068] Then the mutual information between sequence X and sequence Y is:

[0069] I(X,Y)=H(X)+H(Y)-H(XY)

[0070] (3) Considering the geographical distance between points and the mutual information between the corresponding sequences of points, determine whether there is connectivity between two points.

[0071] The comprehensive metric Q combines the spatial distance D and the mutual information I of the deformation link, and defines it as a weighted function. Assume ω D and ω I The weights of time-space distance and mutual information respectively, satisfying ω D +ω I =1, then the comprehensive metric Q can be expressed as:

[0072] Q=ω D ·f(D)+ω I I(X,Y)

[0073] Where f(D) is the attenuation function of the spatial distance D, and I(X,Y) is the mutual information calculated in step 2. In addition, the connectivity weakens as the distance increases, and a Gaussian attenuation function is used, which is expressed as:

[0074]

[0075] Where σ is a parameter that controls distance attenuation, and its size determines the speed at which the effect of distance on connectivity changes.

[0076] If Q>Q t , indicating that there is connectivity between two monitoring points, otherwise there is no connectivity. Q is the mean of all Q values; σ Q is the standard deviation of all Q values; k is a hyperparameter, usually k = 1 or k = 2 to cover 70% to 95% of the data.

[0077] (4) Based on the connectivity between points, the graph structure data is determined and a graph structure dataset is generated. The graph structure data includes an adjacency matrix file that records the connectivity between points. Specifically, a value of 1 in the matrix indicates that there is a connection between two points, and a value of 0 indicates that there is no connection between the two points. It also includes a time series file that records the deformation sequence of the corresponding points.

[0078] (5) In the example of the present invention, the graph structure data generated above is input into the graph neural network (LSTM-GCN) spatiotemporal prediction model. The model will fully consider the contribution of the deformation of the points with spatiotemporal correlation to the deformation of the predicted point based on the connectivity between the points in the graph structure data, and combine the historical deformation information of the predicted point to make an accurate prediction of the future deformation of the predicted point.

[0079] (6) The graph structure data constructed above is used to construct a dataset, with 70% used as a training set and 30% as a validation set. The data is sliced ​​according to the user-defined time series length L. For example, if the time series length L is set to 5, the data organization format is [1,2,3,4,5,6], [2,3,4,5,6,7]..., which means that the deformation data on the 6th date is predicted using the deformation data on 5 historical dates, and the sixth data in each data set is used as the label. The validation set is processed in the same way.

[0080] (7) Build a graph neural network (LSTM-GCN) prediction model and debug the model hyperparameters.

[0081] The graph neural network mainly includes the temporal information extraction part and the spatial feature extraction part. The graph structure dataset constructed in (6) above is input into the tuned model for training to obtain the optimal LSTM-GCN model weights.

[0082] (8) In the LSTM-GCN prediction model, the encoding and extraction of time series information is implemented based on the long short-term memory network (LSTM) model. For the input data, the LSTM model performs the following operations:

[0083]

[0084] Among them, i represents the input gate, Indicates the current cell state, f t represents the forget gate, o t represents the output gate, C t Indicates long-term information, h t Represents short-term information. The memory cell in the LSTM model accepts two inputs, one is the output value h at the previous moment t-1 , one is the input value x at this moment t , h t-1 From the beginning of the sequence to time t-1, the characteristics of the target point deformation; x t Represents the target point feature at time t; the LSTM encoder is combined with h t-1 and x t , calculate the state at the predicted time t+1.

[0085] After these two parameters enter the forget gate, we get information f with smaller weight t And discard it, then enter the input gate to get the information i that needs to be updated t and the current cell state Finally, the output values ​​of the forget gate and the input gate are combined to obtain the long-term information C t and short-term information h t , the long and short time information is stored and transmitted downward. The input data is processed by the LSTM network and finally the temporal change law at the high coherence measurement point is obtained, which is recorded as G.

[0086] (9) Spatial information extraction between high-coherence measurement points is implemented based on the graph convolution module in the graph neural network. Specifically, the graph convolution module receives the temporal variation law G (obtained in (8)) output by the LSTM network, aggregates the spatial features between high-coherence measurement points according to the connection relationship recorded by the adjacency matrix in the graph structure data, and updates the spatiotemporal variation law G to obtain the spatiotemporal variation law G' considering the first-order neighborhood information of the high-coherence measurement points. If multi-order neighborhood information is to be extracted, it is only necessary to superimpose several convolution layers, so that the network can capture and fuse information from high-order neighborhoods (multi-hop neighbors) and realize the transmission of multi-order neighborhood information. The information transfer function of the convolution layer in the graph convolution module (GCN) is as follows:

[0087]

[0088] in, A represents the adjacency matrix, I n represents the identity matrix, represents the adjacency matrix including self-connections; represent The degree matrix, G (l-1) Represents the activation unit matrix of layer l-1, W (l-1) Represents the parameter matrix, which remains consistent in each layer. Where z represents the label value, Represents the predicted value of the model, which includes the time series deformation of the surface monitoring point in a certain period of time, including the date and the corresponding deformation amount.

[0089] (10) Inputting the graph structure data of the study area into the LSTM-GCN model after parameter adjustment and training can obtain the spatiotemporal prediction value of the study area, such as Figure 3 The predicted values ​​are compared with the actual DS-InSAR solution values.

[0090] (11) The spatiotemporal prediction results obtained in (10) above are evaluated based on the mean absolute value MAE, mean square error MSE and root mean square error RMSE, and the following is obtained: Figure 4 The error distribution results shown;

[0091]

[0092] in, is the difference between the true value and the predicted value on the test set.

Claims

1. A spatiotemporal prediction method for surface deformation combining InSAR and graph neural networks, characterized by: Here are the steps: Step 1: Obtain InSAR data within the study time range of the survey area, including SAR images, precise orbit data, atmospheric correction data, and DEM data; Step 2: Use the DS-InSAR solution method to obtain all high-coherence measurement points in the SAR image. The high-coherence measurement points include longitude and latitude, deformation rate, date, and corresponding time series deformation. Step 3: Calculate the spatial distance D between any two highly coherent measurement points based on their longitude and latitude, and calculate the mutual information I between the deformation sequences of any two highly coherent measurement points. Define the comprehensive metric Q using the spatial distance D and the mutual information I. Step 4: Set the comprehensive metric threshold to Q t , if Q>Q t , then it is determined that there is connectivity between the two high coherence measurement points, otherwise it is determined that there is no connectivity between the two points; traverse all high coherence measurement points and construct a graph structure based on the connectivity between the two high coherence measurement points; In step 5, the weighted graph neural network model LSTM-GCN is used to directly obtain the deformation variables corresponding to high-coherence measurement points in the survey area within a certain continuous period in the future through graph structure data, thereby realizing the spatiotemporal prediction of surface deformation.

2. The spatiotemporal prediction method for surface deformation combining InSAR and graph neural network according to claim 1, characterized in that: The steps for obtaining the time-series high-coherence surface deformation information of the measurement points in the line of sight of the SAR image radar using DS-InSAR are as follows: Step 1.1, identifying homogeneous points: using the homogeneous point selection method based on confidence interval hypothesis test (HTCI) to compare the similarity of all SAR images, the central pixel with more than 20 homogeneous pixels is identified as a distributed scatterer target, i.e., a DS candidate point; Step 1.2, coherence analysis: perform coherence analysis on the DS candidate points, and select the candidate points with temporal coherence greater than 0.6 as the final DS points; Step 1.3, phase optimization: Optimize the phase of the DS point using the eigenvalue decomposition (EVD) method; Step 1.4, select permanent scatterer targets: Based on the intensity of all SAR images, select the central pixel with amplitude deviation less than 0.4 as the permanent scatterer target, i.e., PS point; Step 1.5, combine PS points and DS points to solve the surface deformation of the study area: combine all PS points and DS points to construct an irregular triangulation network, use the three-dimensional phase unwrapping algorithm to obtain the unwrapped phase of PS points and DS points, weaken the error factors through high-pass filtering in the time dimension and low-pass filtering in the space dimension, use digital elevation model data to remove the terrain phase, and use universal atmospheric correction data for correction to obtain the latitude and longitude, deformation rate and time series surface deformation of all PS points and DS points in the study area; all PS points and DS points are the high-coherence measurement points in the study area.

3. The spatiotemporal prediction method for surface deformation combining InSAR and graph neural network according to claim 1, characterized in that: The spatial distance D between any two high coherence measurement points is calculated using the longitude and latitude of the high coherence measurement points, which can be expressed as follows: D=R·b Where x and y are any two high-correlation measurement points, lat x ,lon x is the latitude and longitude of the high coherence measurement point x; lat y ,lon y are the latitude and longitude of the high-coherence measurement point y, R is the radius of the earth, which is 6371 km; atan2 represents the inverse tangent function.

4. The spatiotemporal prediction method for surface deformation combining InSAR and graph neural network according to claim 3, characterized in that: The process of calculating the mutual information I between the temporal deformations of any two measurement points using the temporal deformations of high coherence measurement points is as follows: Assume that the temporal deformation of high coherence measurement point x and high coherence measurement point y are respectively expressed as X={x1,x2....,x n } and Y={y1,y2...,y m }, the information entropy of the temporal deformation of high coherence measurement point x and high coherence measurement point y is expressed as: Among them, p(x i ), p(y j ) respectively indicate that X and Y are x i and y i The probability of X and Y is expressed as follows: The mutual information I between X and Y is calculated using the following formula: I(X,Y)=H(X)+H(Y)-H(XY).

5. The spatiotemporal prediction method for surface deformation combining InSAR and graph neural network according to claim 4, characterized in that: When using the comprehensive metric Q to determine whether there is spatiotemporal connectivity between any highly coherent measurement points x and y, the specific process of constructing graph structure data is as follows: Combining the spatial distance D and the mutual information I of the deformation sequence to construct a comprehensive metric Q, let ω D and ω I are the spatial distance weight and mutual information weight between the high coherence measurement point x and the high coherence measurement point y, respectively, and satisfy ω D +ω I =1, then the comprehensive metric Q is expressed as: Q=ω D ·f(D)+ω I ·I(X,Y) Where f(D) is the Gaussian attenuation function of the spatial distance D. The connectivity between the high coherence measurement point x and the high coherence measurement point y decreases as the spatial distance D increases. The Gaussian attenuation function is expressed as: Among them, σ represents the parameter that controls the attenuation of spatial distance D, and its size determines the speed of change of the impact of distance on connectivity; Calculate the comprehensive metric threshold Q using the mean and standard deviation t : Q t =μ Q +kσ Q Among them, μ Q is the mean of all Q values; σ Q is the standard deviation of all Q values; k is a hyperparameter, usually k = 1 or k = 2 to cover 70% to 95% of the data; If Q>Q t , indicating that there is connectivity between high coherence measurement point x and high coherence measurement point y, otherwise there is no connectivity.

6. The spatiotemporal prediction method for surface deformation combining InSAR and graph neural network according to claim 1, characterized in that: The graph structure data is a temporal deformation and adjacency matrix that records the connection between points. A value of 1 in the matrix indicates that there is a connection between two points, and a value of 0 indicates that there is no connection between two points. The graph neural network model LSTM-GCN is used to learn the interactions and correlations between different surface monitoring points, capture the spatiotemporal connections between different high-coherence measurement points in the survey area, and improve the prediction of surface deformation. The graph neural network model LSTM-GCN includes an LSTM module responsible for learning the temporal information of high-coherence measurement points and a GCN module responsible for learning spatial information. The graph structure data input into the LSTM-GCN model first enters the LSTM module to encode and learn the temporal evolution information. Then, the GCN module aggregates the spatial features between the measurement points according to the complex relationships between the high-coherence measurement points defined in the adjacency matrix of the graph structure data and updates the temporal evolution information. It can adaptively identify the connectivity between high-coherence measurement points without relying on manually defined rules or assumptions, thereby improving the stability and generalization ability of the prediction results.

7. A computer device, characterized in that: It includes a processor and a memory, the processor is electrically connected to the memory, the memory is used to store instructions and data, and the processor is used to execute the spatiotemporal prediction method of surface deformation based on the combination of decomposition-based InSAR and graph neural network as described in any one of claims 1-6.

8. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, which is suitable for being loaded by a processor and executing the spatiotemporal prediction method for surface deformation combining InSAR and graph neural network as described in any one of claims 1 to 6.

Images