A method for predicting surface deformation by fusing InSAR and spatio-temporal convolution

Abstract

The present invention discloses a surface deformation prediction method that integrates InSAR and spatiotemporal convolution, comprising the following steps: acquiring surface deformation based on PS-InSAR; constructing a surface deformation sample extraction model, using five types of data (Quaternary thickness, phreatic water level, first confined water level, second confined water level, and third confined water level) from different years as input data for the model, and using surface deformation information as output data for the model, to obtain a spatiotemporal sample set. The present invention extracts a spatiotemporal sample set of surface deformation using the surface deformation sample extraction model. The feature search window combines attribute information and spatial location information unique to geographic data, providing a data foundation for the surface deformation prediction model. The spatiotemporal convolution model extracts multi-factor feature space and predicts surface deformation using multi-temporal data. The model has good prediction accuracy and relatively broad applicability.

Description

Technical Field

[0001] The present invention relates to the technical field of geological disaster simulation and prediction, and in particular to a surface deformation prediction method integrating InSAR and spatiotemporal convolution. Background Art

[0002] Land subsidence is a complex geological phenomenon caused by the interaction of multiple sources. Its continued development and evolution threatens the safety of urban infrastructure. Severe surface deformation can even cause ground fissures, hindering sustainable socioeconomic development. Over 60 countries worldwide are experiencing varying degrees of land subsidence. Exploring subsidence patterns and predicting its trends can provide a scientific basis for future disaster prevention and mitigation, as well as urban spatial planning.

[0003] Leveling is a traditional method for monitoring surface deformation, offering high accuracy and reliability. However, due to the limited number of leveling points, this method has become increasingly inapplicable in large-scale ground subsidence scenarios. The PS-InSAR method utilizes multiple SAR remote sensing images to calculate surface deformation information. Unaffected by inclement weather conditions such as clouds and fog, it is capable of monitoring surface deformation over long time series and over large areas. Its millimeter-level accuracy allows for the identification of minute deformations and has been widely used for monitoring surface deformation over large areas. PS-InSAR typically requires more than ten SAR images to achieve reliable deformation resolution. However, high-resolution SAR imagery is expensive, and the deformation resolution process is complex. Leveraging existing time-series deformation results to predict future trends can further maximize the value of SAR imagery. Furthermore, PS-InSAR typically generates hundreds of thousands or even tens of millions of PS points with time-series information, a sample size that meets the basic requirements for deformation prediction.

[0004] Existing deformation prediction models can be divided into three categories: mathematical statistics models, soil and water models, and machine learning models. Mathematical statistics models suffer from multicollinearity issues when calculating high-dimensional data. Some studies construct models based solely on settlement data, ignoring the constitutive relationships of the geotechnical medium. The construction of soil and water models requires a large amount of hydrogeological data and involves setting numerous parameters based on empirical experience, which is highly subjective. Limited by the variability of hydrogeological conditions, soil and water models are only applicable to deformation simulations in relatively small areas. Their high construction costs and poor applicability make them difficult to apply on a large scale. Machine learning models have strong nonlinear fitting capabilities and are not limited by hydrogeological data. Machine learning-based algorithms such as support vector machines, artificial neural networks, and random forests have achieved some success in deformation prediction. Existing machine learning-based surface deformation prediction models still have two shortcomings: First, surface deformation information is spatial data. According to the first law of geography, similar elements are more closely related. Prediction models that only consider the temporal information of deformation and its influencing factors while ignoring spatial distribution patterns have certain limitations. Secondly, when constructing the model sample set, the data organization formats of the input layer and the output layer are different. How to use data in different formats to construct a sample set that contains both temporal and spatial features is of great significance for the subsequent training of the deformation prediction model. Summary of the Invention

[0005] In response to the above technical problems in the related art, the present invention proposes a surface deformation prediction method that integrates InSAR and spatiotemporal convolution, which can overcome the above shortcomings of the existing technology.

[0006] To achieve the above technical objectives, the technical solution of the present invention is implemented as follows:

[0007] A surface deformation prediction method integrating InSAR and spatiotemporal convolution includes the following steps:

[0008] S1 obtains surface deformation based on PS-InSAR;

[0009] S2 builds a surface deformation sample extraction model, takes the Quaternary thickness, phreatic water level, first confined water level, second confined water level, and third confined water level of different years as the input data of the model, and takes the surface deformation information as the output data of the model to obtain a spatiotemporal sample set;

[0010] S21 Missing Value Processing: When there is missing data in the input layer or output layer, the spline interpolation method is used to supplement the missing values according to the characteristics of the existing time series or regional data;

[0011] S22 raster data resampling: Calculate the weight based on the distance between each sampling point and the surrounding four pixels, and interpolate in the X and Y directions to obtain the pixel value; Assume that point P is the sampling point and the four surrounding points Q12 , Q 22 , Q 11 , Q 21 The values of are (X1, Y2), (X2, Y2), (X1, Y1), (X2, Y1), then the resampled value of point P(X, Y) is:

[0012]

[0013] S23 acquisition of spatiotemporal sample set of surface deformation: Each PS point (subsidence point) corresponds to the Quaternary thickness Q t , diving water level L s , the first pressure water level L1, the second pressure water level L2, and the third pressure water level L3 grid data, first determine which pixel of the corresponding grid the PS point falls in, and record the pixel position, and then extract the input layer sample;

[0014] S3 builds a spatiotemporal convolution model to predict surface deformation. The spatiotemporal convolution model includes a multi-factor feature extraction module based on spatial information and a multi-temporal surface deformation prediction module based on historical data.

[0015] S31 Obtaining surface deformation feature map RM SS :Fusing the Quaternary thickness Q of different years by concat method t , diving water level L s , the first pressure water level L1, the second pressure water level L2, the third pressure water level L3 five characteristic band data, the surface deformation characteristic map RM is obtained SS ;

[0016] S32 uses a multi-factor feature extraction module to obtain a surface deformation correlation feature map;

[0017] S33 inputs the multi-temporal surface deformation feature vector into the time-based temporal feature extraction module to obtain a temporal feature vector, passes the temporal feature vector through the fully connected layer, and finally maps it into the predicted surface deformation information.

[0018] Furthermore, the specific steps of S1 are as follows:

[0019] S11 selects one of the M equally spaced SAR images as the main image and the remaining M-1 images as auxiliary images. At this time, the phase information of the echo signal of each image is It can be expressed as:

[0020]

[0021] Where λ is the radar wavelength, R is the slant distance from the observation point to the target object, arg represents the argument calculation, and U is the phase contribution value of the object;

[0022] S12 performs interference processing on the main image and the auxiliary image to generate INT, that is, the two images are conjugate multiplied:

[0023]

[0024] Where A1 and A2 are the amplitude information of the main and auxiliary images, and i is the imaginary unit

[0025] S13 uses the amplitude angle function to extract the interference phase This interference phase includes the reference ellipsoid phase Terrain relief phase Surface deformation phase Atmospheric delay phase and noise phase

[0026]

[0027]

[0028] S14 expresses the reference ellipsoid phase according to the baseline distance B, radar viewing angle θ, and baseline inclination α as:

[0029]

[0030] S15 When the target object is not on the reference ellipsoid surface, but on the earth surface with a certain elevation value, according to the baseline vertical component B v and elevation h, the terrain relief phase is expressed as:

[0031]

[0032] S16 will interfere with the phase minus and The differential interference phase is obtained and combined with the PS points extracted based on the amplitude deviation index to obtain the PS time series differential interference model;

[0033] S17 obtains elevation error and linear surface deformation value based on PS network model, and uses filtering method to remove atmospheric delay and noise value to obtain nonlinear deformation information;

[0034] S18 superimposed linear and nonlinear deformations are the final surface deformation values obtained.

[0035] Furthermore, in step S21, the interpolation polynomial and the boundary conditions satisfied are:

[0036] S i (x) = a i +b i x+c i x2 +d i x 3

[0037] S″(x0)=S″(x n )=0

[0038] S′(x0)=c1,S′(x n )=c2

[0039] S′(x0)=S′(x n ), S″(x0)=S″(x n )

[0040] S″′(x0)=S″′(x1), S″′(x n-1 )=S″′(x n )

[0041] Where a i 、b i 、c i d i is the coefficient to be determined, and x is the corresponding data value.

[0042] Furthermore, the specific sample extraction method in step S23 is: with the pixel where the PS point is located as the center, search for the input data of 21×21 pixel blocks around it one by one. The data shape of the input layer corresponding to each PS point in the final spatiotemporal sample set is M×5×21×21 (number of SAR images×number of input feature bands×pixel block length×pixel block width).

[0043] Furthermore, the surface deformation feature map RM in step S31 SS The formula is:

[0044] RM SS =concat(Q t , L s , L1, L2, L s ).

[0045] Furthermore, the specific steps of step S32 are:

[0046] S321 multi-factor feature extraction module construction: the surface deformation feature map RM SS The input is sent to the multi-factor feature extraction module based on spatial information for feature extraction, wherein the multi-factor feature extraction module is represented as follows:

[0047]

[0048] h′ i，j (l) = relu[h i，j (ml), ..., h i，j(ml+n-1)]

[0049] Its i，j represents the jth feature map obtained from the i convolutional layer, b i，j is the deviation value, w i，kj represents the convolution kernel, K represents the number of feature maps, m represents the stride of the convolution, and n is the number of output results; "relu" is the activation function of the model, and the formula is as follows:

[0050]

[0051] x is the input value, f(x) is the output value;

[0052] S322 will surface deformation feature map RM SS Input into the surface deformation correlation layer and output the surface deformation correlation feature map;

[0053] S323 inputs the surface deformation correlation feature map into the spatial correlation layer and outputs the spatial surface deformation feature map.

[0054] Furthermore, the specific steps of step S33 are:

[0055] S331 Multi-temporal Surface Deformation Prediction Module Construction: The multi-temporal surface deformation feature vector is input into the time-based temporal feature extraction module to obtain the temporal feature vector. This module uses a gate mechanism to control the introduction and loss of information and connects features at different times. The connection method can be expressed as follows:

[0056] f k =σ(W fx x k +W fh q k-1 +b f )

[0057]

[0058] c k Represented as feature information at time k, c k-1 is the k-1 moment information, and its σ activation function is W fx 、W fh q k-1 represents the weight matrix, f k is the surface deformation weight matrix, is the surface deformation deviation matrix;

[0059] S332 inputs the temporal feature vector into the fully connected layer and outputs the spatiotemporal feature vector;

[0060] S333 inputs the spatiotemporal feature vector into the output node, which is ultimately mapped into predicted surface deformation information.

[0061] The beneficial effects of the present invention are as follows: the present invention realizes the extraction of spatiotemporal sample sets of surface deformation through the surface deformation sample extraction model; the feature search window combines the attribute information and spatial location information unique to geographic data, and can provide a data basis for the surface deformation prediction model; the multi-factor feature space extraction and multi-phase data surface deformation prediction are realized through the spatiotemporal convolution model, and the model has good prediction accuracy and relatively wide applicability. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0063] Figure 1 1 is a flowchart of a surface deformation prediction method integrating InSAR and spatiotemporal convolution according to an embodiment of the present invention;

[0064] Figure 2 2 is a schematic diagram of a surface deformation sample extraction model of a surface deformation prediction method integrating InSAR and spatiotemporal convolution according to an embodiment of the present invention;

[0065] Figure 3 2 is a schematic diagram of a spatiotemporal convolution model framework of a surface deformation prediction method integrating InSAR and spatiotemporal convolution according to an embodiment of the present invention;

[0066] Figure 4 3 is a structural diagram of a spatiotemporal convolution model of a surface deformation prediction method integrating InSAR and spatiotemporal convolution according to an embodiment of the present invention. DETAILED DESCRIPTION

[0067] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention are within the scope of protection of the present invention.

[0068] like Figure 1-4As shown, according to a surface deformation prediction method that integrates InSAR and spatiotemporal convolution described in an embodiment of the present invention, the present invention aims to solve two problems: 1) Surface deformation information is spatial data. According to the first law of geography, similar elements are more closely related. Existing prediction models only consider the temporal information of deformation influencing factors and ignore their spatial distribution patterns, which has certain limitations; 2) When constructing a model sample set, the data organization formats of the input layer and the output layer are different. How to use data in different formats to construct a sample set that contains both temporal and spatial features is of great significance to the training of subsequent deformation prediction models.

[0069] This method is divided into three parts: obtaining surface deformation based on PS-InSAR, obtaining spatiotemporal sample sets using the surface deformation sample extraction model, and predicting surface deformation based on the spatiotemporal convolution model.

[0070] Using multiple equally spaced SAR images, we acquire surface deformation information using the PS-InSAR method. This method overcomes the effects of temporal and spatial decorrelation and allows us to obtain deformation information with millimeter-level accuracy.

[0071] The thickness of the Quaternary system provides the original geological background information for predicting surface deformation. Fluctuations in the phreatic water level, the first artesian water level, the second artesian water level, and the third artesian water level significantly affect the effective stress in the soil. The Quaternary system thickness, phreatic water level, the first artesian water level, the second artesian water level, and the third artesian water level from different years are used as input to the model, and surface deformation information is used as output. A surface deformation sample extraction model is constructed based on the characteristics of the input and output data. The surface deformation sample extraction model has three functions: missing data processing, raster data resampling, and the creation of a spatiotemporal surface deformation sample set.

[0072] When model input or output data are missing, spline interpolation methods can be used to supplement them, and the missing values can be supplemented according to the characteristics of the existing time series or regional data.

[0073] Considering the inconsistent resolution of the input layer raster data, which affects the subsequent extraction of pixel blocks, the raster data needs to be resampled. The weight is calculated based on the distance between each sampling point and the four surrounding pixels, and the pixel value is interpolated in the X and Y directions respectively.

[0074] Spatial data contains both attribute and geographic information. When constructing the sample set, it is necessary to consider not only the attribute values of the input layer data at each PS point (subsidence point) but also the attribute information of the surrounding input data. Based on the spatial correlation between the subsidence point and the surrounding hydrogeological conditions, a 21×21 feature search window is used to extract input layer samples. This ensures low data redundancy while meeting the required accuracy for surface deformation prediction. Because the PS point vectors and the input layer raster data have different organizational formats, it is necessary to first determine which pixel of the corresponding raster the PS point falls into and record the pixel location before extracting input layer samples. The final input layer data shape for each PS point in the spatiotemporal sample set is M×5×21×21 (number of SAR images × number of input feature bands × pixel block length × pixel block width).

[0075] The spatiotemporal convolutional prediction model predicts future deformation trends by training on data from a spatiotemporal sample set of surface deformation. The spatiotemporal convolutional model primarily consists of a multi-factor feature extraction module based on spatial information and a multi-temporal surface deformation prediction module based on historical data. Five characteristic bands, representing Quaternary thickness, groundwater, primary confined water, secondary confined water, and tertiary confined water, are fused via a concatenation method to generate a surface deformation feature map. The multi-factor feature extraction module generates a surface deformation correlation feature map, further strengthening the connection between different PS points and their surrounding spatial regions. The multi-temporal surface deformation feature vector is input into a time-based temporal feature extraction module to generate a temporal feature vector. The temporal feature vector is then passed through a fully connected layer and ultimately mapped into predicted surface deformation information.

[0076] In order to facilitate understanding of the above technical solutions of the present invention, the above technical solutions of the present invention are described in detail below through specific usage methods.

[0077] In specific use, according to the surface deformation prediction method integrating InSAR and spatiotemporal convolution described in the present invention, the method is divided into three parts: obtaining surface deformation based on PS-InSAR, obtaining spatiotemporal sample sets using the surface deformation sample extraction model, and predicting surface deformation based on the spatiotemporal convolution model.

[0078] Step 1: Acquisition of surface deformation data.

[0079] S1: Among M equally spaced SAR images, select one main image and the remaining M-1 images as auxiliary images. At this time, the phase information of the echo signal of each image is It can be expressed as:

[0080]

[0081] Where λ is the radar wavelength, R is the slant distance from the observation point to the target object, arg represents the argument calculation, and U is the phase contribution value of the object.

[0082] S2: Interfere the main image and the auxiliary image to generate INT, that is, conjugate multiplication of the two images:

[0083]

[0084] Where A1 and A2 are the amplitude information of the main and auxiliary images, and i is the imaginary unit

[0085] S3: Extract the interference phase using the amplitude function, which includes the reference ellipsoid phase Terrain relief phase Surface deformation phase Atmospheric delay phase and noise phase

[0086]

[0087]

[0088] S4: Based on the baseline distance B, radar viewing angle θ, and baseline inclination α, the reference ellipsoid phase can be expressed as:

[0089]

[0090] S5: When the target object is not on the reference ellipsoid surface, but on the earth surface with a certain elevation value, according to the baseline vertical component B v and elevation h, the terrain relief phase can be expressed as:

[0091]

[0092] S6: Interference phase minus and The differential interference phase is obtained and combined with the PS points extracted based on the amplitude deviation index to obtain the PS timing differential interference model.

[0093] S7: Calculate the elevation error and linear surface deformation value based on the PS network model; use the filtering method to remove the atmospheric delay and noise value to obtain nonlinear deformation information.

[0094] S8: The superposition of linear and nonlinear deformations is the final surface deformation value.

[0095] Step 2: Construct a surface deformation sample extraction model.

[0096] S1: Missing value processing. When there is missing data in the input layer or output layer, the spline interpolation method is used to supplement the missing values based on the characteristics of the existing time series or regional data. The interpolation polynomial and the boundary conditions satisfied are:

[0097] S i (x) = a i +b i x+c i x 2 +d i x 3

[0098] S″(x0)=S″(x n )=0

[0099] S′(x0)=c1,S′(x n )=c2

[0100] S′(x0)=S′(x n )S″(x0)=S″(x n )

[0101] S″′(x0)=S″′(x1), S″′(x n-1 )=S″′(x n )

[0102] Where a i 、b i 、c i d i is the coefficient to be determined, and x is the corresponding data value.

[0103] S2: Resampling of raster data. Considering the inconsistent resolution of input layer data, which affects the subsequent extraction of pixel blocks, the input layer raster data needs to be resampled. The weight is calculated based on the distance between each sampling point and the surrounding four pixels, and the pixel value is obtained by interpolation in the X and Y directions. Assume that point P is the sampling point and the four surrounding points Q are 12 , Q 22 , Q 11 , Q 21 The values of are (X1, Y2), (X2, Y2), (X1, Y1), (X2, Y1), then the resampled value of point P(X, Y) is:

[0104]

[0105] S3_1: Acquisition of spatiotemporal sample sets of surface deformation. Considering that the Quaternary thickness can provide original geological background information for the prediction of surface deformation. The rise and fall of the groundwater level, the first confined water level, the second confined water level, and the third confined water level will significantly affect the effective stress of the soil. Therefore, the Quaternary thickness Q of different years is calculated. t, diving water level L s , the first confined water level L1, the second confined water level L2, and the third confined water level L3 are used as the input data of the model. Spatial data contains both attribute information and geographic location information. Therefore, when constructing the sample set, it is necessary to consider not only the attribute value of the input layer data at each PS point (settlement point) but also the attribute information of the surrounding input data. According to the spatial correlation between the settlement point and the surrounding hydrogeological conditions, a 21×21 feature search window is used to construct the input layer sample. While meeting the accuracy requirements of surface deformation prediction, low data redundancy is guaranteed. The coverage range of each input data is larger than the PS point range, which facilitates the extraction of edge PS point pixel blocks.

[0106] S3_2: Each PS point corresponds to the Quaternary thickness Q t , diving water level L s , the first pressure water level L1, the second pressure water level L2, and the third pressure water level L3 raster data. Due to the difference in the organization format of point vector and raster data, it is necessary to first determine which pixel of the corresponding raster the PS point falls in and record the pixel position before extracting the input layer sample. With the pixel where the PS point is located as the center, search the input data of the 21×21 pixel block around it one by one. Figure 2 As shown in the figure, the pixel position corresponding to a PS point at time t1 in the Quaternary thickness grid data is obtained, and a 21×21 pixel block around the pixel is cut and saved. Then, the pixel positions at time t2, t3…t m After the Quaternary thickness data were extracted, the groundwater, primary confined water, secondary confined water, and tertiary confined water datasets were extracted in sequence. The input layer data shape for each PS point in the final spatiotemporal sample set was M × 5 × 21 × 21 (number of SAR images × number of input feature bands × pixel block length × pixel block width).

[0107] Step 3: Construct a spatiotemporal convolution model to predict surface deformation.

[0108] The spatiotemporal convolution model mainly includes a multi-factor feature extraction module based on spatial information and a multi-temporal surface deformation prediction module based on historical data. The network model framework diagram is as follows: Figure 3 As shown, the spatiotemporal convolution model structure is as follows Figure 4 shown.

[0109] S1: Obtain the surface deformation characteristic map. The Quaternary thickness Q of different years is t , diving water level L s The five data of the first pressure water level L1, the second pressure water level L2, and the third pressure water level L3 are combined as characteristic bands through concat to obtain a surface deformation feature map RM with a size of 21*21*5. SSThe concat operation can be used to fuse data while preventing the loss of feature information in the data.

[0110] RM SS =concat(Q t , L s , L1, L2, L s )

[0111] S2_1: Multi-factor feature extraction module construction. SS The input is sent to the multi-factor feature extraction module based on spatial information for feature extraction. The multi-factor feature extraction module can be represented as follows:

[0112]

[0113] h′ i，j (l) = relu[h i，j (ml), ..., h i，j (ml+n-1)]

[0114] Its i，j represents the jth feature map obtained from the i convolutional layer, b i，j is the deviation value, w i，kj Represents the convolution kernel, K represents the number of feature maps, m represents the convolution step, and n is the number of output results. "relu" is the activation function of the model, and the formula is as follows:

[0115]

[0116] x is the input value and f(x) is the output value.

[0117] S2_2: Surface deformation feature map RM SS This is input into the surface deformation association layer, which outputs a 21*21*16 surface deformation association feature map. The surface deformation association layer mainly includes a 1*1*9 and a 1*1*16 convolution kernel. The purpose is to enhance the feature correlation between feature bands through information interaction between different feature bands.

[0118] S2_3: Input the surface deformation correlation feature map into the spatial correlation layer, which outputs a 9*9*16 spatial surface deformation feature map. The spatial constraint layer uses two 5*5 convolution kernels connected in series with a global pooling layer. This layer uses nonlinear transformations to strengthen the connection between different PS points in the surface deformation correlation feature map and their surrounding spatial regions. Global pooling is also used to compress the feature dimension and adjust the input format to the multi-temporal surface deformation prediction module.

[0119] S3_1: Construction of a multi-temporal surface deformation prediction module. The multi-temporal surface deformation feature vector is input into the time-based temporal feature extraction module to obtain a temporal feature vector. This module uses a gate mechanism to control the introduction and loss of information and connects features at different times. The connection method can be expressed as:

[0120] f k =σ(W fx x k +W fh q k-1 +b f )

[0121]

[0122] c k Represented as feature information at time k, c k-1 is the k-1 moment information, and its σ activation function is W fx 、W fh q k-1 represents the weight matrix, f k is the surface deformation weight matrix, For the surface deformation deviation matrix, in order to more effectively extract the time correlation information in the input data, the nodes of the time series feature extraction module are set to 32. At the same time, the long-term dependence problem of the output features can be reduced and the prediction accuracy of the model can be improved.

[0123] S3_2: Input the time series feature vector into a fully connected layer with 16 nodes, outputting a 16-dimensional spatiotemporal feature vector. This connection layer initially maps the distributed features learned by the model to the sample space, compressing the feature vector dimension and transferring the time series feature vector.

[0124] S3_3: Input the spatiotemporal feature vector to the output node, ultimately mapping it to predicted surface deformation information. This output layer uses a fully connected architecture to fully capture the information extracted by the previous network, and the output value is the predicted surface deformation information.

[0125] In summary, with the help of the above-mentioned technical scheme of the present invention, the extraction of spatiotemporal sample sets of surface deformation is realized through the surface deformation sample extraction model. The feature search window combines the attribute information and spatial location information unique to geographic data, which can provide a data basis for the surface deformation prediction model; the multi-factor feature space extraction and multi-temporal data surface deformation prediction are realized through the spatiotemporal convolution model. The model has good prediction accuracy and relatively wide applicability.

[0126] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A surface deformation prediction method integrating InSAR and spatiotemporal convolution, characterized in that: The steps include: S1 obtains surface deformation based on PS-InSAR; S2 builds a surface deformation sample extraction model, takes the Quaternary thickness, phreatic water level, first confined water level, second confined water level, and third confined water level of different years as the input data of the model, and takes the surface deformation information as the output data of the model to obtain a spatiotemporal sample set; S21 Missing Value Processing: When there is missing data in the input layer or output layer, the spline interpolation method is used to supplement the missing values according to the characteristics of the existing time series or regional data; S22 raster data resampling: Calculate the weight based on the distance between each sampling point and the surrounding four pixels, and interpolate in the X and Y directions to obtain the pixel value; Assume that point P is the sampling point and the four surrounding points Q 12 , Q 22 , Q 11 , Q 21 The values of are (X1, Y2), (X2, Y2), (X1, Y1), (X2, Y1), then the resampled value of point P(X, Y) is: S23 acquisition of spatiotemporal sample set of surface deformation: Each PS point corresponds to the Quaternary thickness Q t , diving water level L s , the first pressure water level L1, the second pressure water level L2, and the third pressure water level L3 grid data, first determine which pixel of the corresponding grid the PS point falls in, and record the pixel position, and then extract the input layer sample; S3 builds a spatiotemporal convolution model to predict surface deformation. The spatiotemporal convolution model includes a multi-factor feature extraction module based on spatial information and a multi-temporal surface deformation prediction module based on historical data. S31 Obtaining surface deformation feature map RM SS :Fusing the Quaternary thickness Q of different years by concat method t , diving water level L s , the first pressure water level L1, the second pressure water level L2, the third pressure water level L3 five characteristic band data, the surface deformation characteristic map RM is obtained SS ; S32 uses a multi-factor feature extraction module to obtain a surface deformation correlation feature map; S33 inputs the multi-temporal surface deformation feature vector into the time-based temporal feature extraction module to obtain a temporal feature vector, passes the temporal feature vector through the fully connected layer, and finally maps it into the predicted surface deformation information.

2. The surface deformation prediction method integrating InSAR and spatiotemporal convolution according to claim 1 is characterized in that: The specific steps of step S1 are as follows: S11 selects one of the M equally spaced SAR images as the main image and the remaining M-1 images as auxiliary images. At this time, the phase information of the echo signal of each image is Expressed as: Where λ is the radar wavelength, R is the slant distance from the observation point to the target object, arg represents the argument calculation, and U is the phase contribution value of the object; S12 performs interference processing on the main image and the auxiliary image to generate INT, that is, the two images are conjugate multiplied: Where A1 and A2 are the amplitude information of the main and auxiliary images, and i is the imaginary unit ; S13 uses the amplitude angle function to extract the interference phase This interference phase includes the reference ellipsoid phase Terrain relief phase Surface deformation phase Atmospheric delay phase and noise phase S14 expresses the reference ellipsoid phase according to the baseline distance B, radar viewing angle θ, and baseline inclination α as: S15 When the target object is not on the reference ellipsoid surface, but on the earth surface with a certain elevation value, according to the baseline vertical component B v and elevation h, the terrain relief phase is expressed as: S16 will interfere with the phase minus and The differential interference phase is obtained and combined with the PS points extracted based on the amplitude deviation index to obtain the PS time series differential interference model; S17 obtains elevation error and linear surface deformation value based on PS network model, and uses filtering method to remove atmospheric delay and noise value to obtain nonlinear deformation information; S18 superimposed linear and nonlinear deformations are the final surface deformation values obtained.

3. The surface deformation prediction method integrating InSAR and spatiotemporal convolution according to claim 1 is characterized in that: In step S21, the interpolation polynomial and the boundary conditions satisfied are: S i (x)=a i +b i x+c i x 2 +d i x 3 S″(x0)=S″(x n )=0S′(x0)=c1,S′(x n )=c2S′(x0)=S′(x n ),S″(x0)=S″(x n ) S″′(x0)=S″′(x1),S″′(x n-1 )=S″′(x n ) Where a i 、b i 、c i d i is the coefficient to be determined, and x is the corresponding data value.

4. The surface deformation prediction method integrating InSAR and spatiotemporal convolution according to claim 1, characterized in that: The specific sample extraction method in step S23 is: with the pixel where the PS point is located as the center, search for the input data of 21×21 pixel blocks around it one by one. The data shape of the input layer corresponding to each PS point in the final spatiotemporal sample set is M×5×21×21, that is, the number of SAR images×the number of input feature bands×the length of the pixel block×the width of the pixel block.

5. The surface deformation prediction method integrating InSAR and spatiotemporal convolution according to claim 1, characterized in that: The surface deformation feature map RM in step S31 SS The formula is: RM SS =concat(Q t ,L s ,L1,L2,L s )。 6. The surface deformation prediction method integrating InSAR and spatiotemporal convolution according to claim 1, characterized in that: The specific steps of step S32 are: S321 multi-factor feature extraction module construction: the surface deformation feature map RM SS The input is sent to the multi-factor feature extraction module based on spatial information for feature extraction, wherein the multi-factor feature extraction module is represented as follows: h′ i，j (l)=relu[h i，j (ml),...,h i，j (ml+n-1)] Its i，j represents the jth feature map obtained from the i convolutional layer, b i，j is the deviation value, w i，kj represents the convolution kernel, K represents the number of feature maps, m represents the convolution step, and n is the number of output results; "relu" is the activation function of the model, and the formula is as follows: x is the input value, f(x) is the output value; S322 will surface deformation feature map RM SS Input into the surface deformation correlation layer and output the surface deformation correlation feature map; S323 inputs the surface deformation correlation feature map into the spatial correlation layer and outputs the spatial surface deformation feature map.

7. The surface deformation prediction method integrating InSAR and spatiotemporal convolution according to claim 1, characterized in that: The specific steps of step S33 are: S331 Multi-temporal Surface Deformation Prediction Module Construction: The multi-temporal surface deformation feature vector is input into the time-based temporal feature extraction module to obtain the temporal feature vector. This module uses a gate mechanism to control the introduction and loss of information and connects features at different times. The connection method is expressed as follows: f k =σ(W fx x k +W fh q k-1 +b f ) c k Represented as feature information at time k, c k-1 is the k-1 moment information, and its σ activation function is W fx 、W fh q k-1 represents the weight matrix, f k is the surface deformation weight matrix, is the surface deformation deviation matrix; S332 inputs the temporal feature vector into the fully connected layer and outputs the spatiotemporal feature vector; S333 inputs the spatiotemporal feature vector into the output node, which is ultimately mapped into predicted surface deformation information.

Images