InSAR inconsistent phase correction method and system based on coherence matrix guidance, terminal and storage medium

Abstract

The invention belongs to the technical field of data processing, and discloses an InSAR non-consistency phase correction method and system based on coherence matrix guidance, a terminal and a storage medium, and the method comprises the steps: obtaining ground feature category data of a target research region, and calculating a symmetric phase matrix and a symmetric coherence time matrix according to the ground feature category data; performing phase reconstruction processing on the symmetric phase matrix to obtain a target closed phase, and constructing a target data set according to the target closed phase and the symmetric coherence time matrix; determining a preset model, and performing model training and model optimization on the preset model according to the target data set to obtain an earth surface deformation displacement test model; and obtaining target ground object category data of the target test area, and inputting the target ground object category data into the ground surface deformation displacement test model to obtain a target ground surface deformation displacement test result. By constructing the earth surface deformation displacement test model, the test efficiency and the test precision of the earth surface deformation displacement test result can be effectively improved.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to an InSAR non-uniform phase correction method, system, terminal, and computer-readable storage medium based on coherence matrix guidance. Background Technology

[0002] InSAR (Interferometric Synthetic Aperture Radar) time series analysis is a technique for monitoring changes in Earth's surface displacement over time. Due to its advantages such as all-weather, all-time accuracy, high precision, and high resolution, InSAR technology has rapidly developed into a mature space-based Earth observation technique, capable of mapping global topography with meter-level precision. In other words, the accuracy of InSAR time series analysis is crucial for monitoring Earth's surface deformation and displacement.

[0003] While existing multi-view techniques can reduce phase noise in interferograms, they introduce phase bias (i.e., closed phase). Since the behavior of closed phase is related to various factors, such as soil moisture, vegetation water content, and land cover classification, its behavior cannot be fully encompassed when modeling closed phase, resulting in low accuracy of synthetic aperture radar interferometry and failing to meet measurement requirements.

[0004] Therefore, existing technologies still need to be improved and developed. Summary of the Invention

[0005] The main objective of this invention is to provide a method, system, terminal, and computer-readable storage medium for InSAR non-uniform phase correction based on coherence matrix guidance. This aims to address the problem that while multi-look techniques in existing technologies can reduce phase noise in interferograms, they introduce phase deviation (i.e., closed phase). Since the behavior of closed phase is related to various factors, such as soil moisture, vegetation water content, and land cover classification, modeling closed phase cannot fully encompass its behavior, resulting in low accuracy in synthetic aperture radar interferometry and failing to meet measurement requirements.

[0006] To achieve the above objectives, this invention provides an InSAR non-uniform phase correction method guided by a coherence matrix, which includes the following steps: Obtain land cover category data for the target study area, and calculate the symmetric phase matrix and symmetric coherence time matrix based on the land cover category data; The symmetric phase matrix is ​​subjected to phase reconstruction processing to obtain the target closed phase, and the target dataset is constructed based on the target closed phase and the symmetric coherence time matrix. A preset model is determined, and the preset model is trained and optimized based on the target dataset to obtain a surface deformation and displacement test model; The target test area is determined, and the target land cover category data of the target test area is obtained. The target land cover category data is input into the surface deformation and displacement test model to obtain the target surface deformation and displacement test results of the target test area.

[0007] Optionally, the InSAR non-uniform phase correction method based on coherence matrix guidance, wherein acquiring land cover category data of the target study area and calculating a symmetric phase matrix and a symmetric coherence time matrix based on the land cover category data specifically includes: The target study area is determined, global land cover category data is obtained, and the global land cover category data is cropped according to the target study area to obtain a cropped image; Acquire a preset single-view complex image and digital elevation model data in the target study area, and register the cropped image with the preset single-view complex image and digital elevation model data to obtain a differential interferogram; The differential interferogram is subjected to adaptive spectral filtering to obtain a smooth interferogram; The smoothed interferogram is subjected to phase unrolling to obtain the interference phase; Calculate the coherence value corresponding to the interference phase, obtain the symmetric phase matrix based on the interference phase, and obtain the symmetric coherence time matrix based on the coherence value.

[0008] Optionally, the InSAR non-uniform phase correction method based on coherence matrix guidance, wherein the adaptive spectral filtering of the differential interferogram to obtain a smooth interferogram specifically includes: The differential interferogram is subjected to local window Fourier transform processing to obtain the spectral components of the signal and noise separation. The signal spectrum is obtained by adaptively calculating the spectral components. The signal spectrum is spatially transformed using the inverse Fourier transform method to obtain a smooth interferogram.

[0009] Optionally, the InSAR non-uniform phase correction method based on coherence matrix guidance, wherein the step of performing phase reconstruction processing on the symmetric phase matrix to obtain the target closed phase, and constructing the target dataset based on the target closed phase and the symmetric coherence time matrix, specifically includes: Determine the first preset waveband and the second preset waveband, and use the least squares method to solve the phase consistency of the symmetric phase matrix based on the first preset waveband and the second preset waveband to obtain the first phase sequence and the second phase sequence; Subtract the first phase sequence from the second phase sequence to obtain the target closed phase; The symmetric coherent time matrix is ​​used as the model input, the target closed phase is used as the input label, and the target dataset is constructed based on the symmetric coherent time matrix and the target closed phase.

[0010] Optionally, the InSAR non-uniform phase correction method based on coherence matrix guidance, wherein determining a preset model and training and optimizing the preset model according to the target dataset to obtain a surface deformation displacement test model specifically includes: A preset model is determined, wherein the preset model is the ViT model; The symmetric coherence time matrix in the target dataset is input into the ViT model. The ViT model performs block processing on the symmetric coherence time matrix to obtain image blocks, and the image blocks are converted into block matrices. Obtain a preset learnable tag, and perform addition and matrix addition processing on the preset learnable tag and the block matrix to obtain the embedded block; A preset transformer encoder and MLP model are determined. The embedding block is encoded and closed phase is predicted using the preset transformer encoder and the MLP model to obtain the predicted closed phase. The model training is completed, and the initial training model is obtained. The error between the predicted closed phase and the target closed phase is calculated using the MSE loss function, and the initial training model is optimized based on the error to obtain the surface deformation and displacement test model.

[0011] Optionally, the InSAR non-uniform phase correction method based on coherence matrix guidance, wherein determining a preset transformer encoder and MLP model, and performing encoding and closure phase prediction processing on the embedded block using the preset transformer encoder and the MLP model to obtain the predicted closure phase, specifically includes: A preset transformer encoder is determined, and the embedding block is input to the preset transformer encoder. The preset transformer encoder performs encoding and matrix transformation processing on the embedding block to obtain a transformed embedding block. Self-attention calculation is performed on the transformed embedding block to obtain the self-attention calculation result; The self-attention calculation result is input into the linear layer to obtain the final output result of the preset transformer encoder; The MLP model is determined, and the final output result is input into the MLP model. The MLP model is then used to perform closed-phase prediction processing on the final output result to obtain the predicted closed phase.

[0012] Optionally, the InSAR non-uniform phase correction method based on coherence matrix guidance, wherein the step of determining a preset model and training and optimizing the preset model according to the target dataset to obtain a surface deformation displacement test model, further includes: Determine a preset study region and obtain a preset temporal coherence matrix for the preset study region; The preset time coherence matrix is ​​input into the surface deformation displacement test model to obtain the test closed phase sequence; Based on the test closed phase sequence, the first phase sequence and the second phase sequence are subjected to phase deviation correction processing to obtain the target phase deviation; The annual deformation displacement rate corresponding to each pixel in the preset study area is calculated using the polyfit function based on the target phase deviation. The results output by the surface deformation displacement test model are then compared and verified based on the annual deformation displacement rate to obtain the verification results.

[0013] Furthermore, to achieve the above objectives, the present invention also provides an InSAR non-uniform phase correction system guided by a coherence matrix, wherein the InSAR non-uniform phase correction system guided by a coherence matrix includes: The land cover category data processing module is used to acquire land cover category data of the target study area and calculate the symmetric phase matrix and symmetric coherence time matrix based on the land cover category data; The dataset construction module is used to perform phase reconstruction processing on the symmetric phase matrix to obtain the target closed phase, and construct the target dataset based on the target closed phase and the symmetric coherence time matrix; The test model construction module is used to determine the preset model and perform model training and optimization on the preset model based on the target dataset to obtain the surface deformation and displacement test model. The test result output module is used to determine the target test area, obtain the target land cover category data of the target test area, input the target land cover category data into the surface deformation and displacement test model, and obtain the target surface deformation and displacement test results of the target test area.

[0014] Furthermore, to achieve the above objectives, the present invention also provides a terminal, wherein the terminal includes: a memory, a processor, and an InSAR non-uniform phase correction program based on coherence matrix guided by the memory and executable on the processor, wherein when the InSAR non-uniform phase correction program based on coherence matrix guided by the processor is executed, it implements the steps of the InSAR non-uniform phase correction method based on coherence matrix guided by the processor as described above.

[0015] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores an InSAR non-uniform phase correction program guided by a coherence matrix, wherein when the InSAR non-uniform phase correction program guided by a coherence matrix is ​​executed by a processor, it implements the steps of the InSAR non-uniform phase correction method guided by a coherence matrix as described above.

[0016] In this invention, land cover category data of the target study area is acquired, and a symmetric phase matrix and a symmetric coherence time matrix are calculated based on the land cover category data. Phase reconstruction processing is performed on the symmetric phase matrix to obtain a target closed phase, and a target dataset is constructed based on the target closed phase and the symmetric coherence time matrix. A preset model is determined, and model training and optimization are performed on the preset model based on the target dataset to obtain a surface deformation displacement test model. A target test area is determined, and target land cover category data of the target test area is acquired. The target land cover category data is input into the surface deformation displacement test model to obtain the target surface deformation displacement test results for the target test area. This invention constructs a symmetric phase matrix and a symmetric coherence time matrix by acquiring land cover category data of the study area, and then performs phase reconstruction processing based on the symmetric phase matrix and the symmetric coherence time matrix to obtain a target closed phase, thus realizing the construction of a target dataset. Furthermore, model training is performed on the preset model based on the target dataset to obtain a surface deformation displacement test model, which can achieve accurate testing of surface deformation displacement and effectively improve the testing efficiency and accuracy of surface deformation displacement test results. Attached Figure Description

[0017] Figure 1 This is a flowchart of a preferred embodiment of the InSAR non-uniform phase correction method guided by the coherence matrix of the present invention; Figure 2 This is a schematic diagram of the overall process of a preferred embodiment of the InSAR non-uniform phase correction method guided by coherence matrix of the present invention. Figure 2 In this invention, ViT stands for Visual Transformer, SVD stands for Singular Value Decomposition, bw2 is the first preset bandwidth, and bw156 is the second preset bandwidth. Figure 3 This is a structural diagram of a preferred embodiment of the InSAR non-uniform phase correction system guided by the coherence matrix of the present invention; Figure 4 This is a structural diagram of a preferred embodiment of the terminal of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0019] InSAR (Interferometric Synthetic Aperture Radar) time series analysis is a technique for monitoring changes in Earth's surface displacement over time. Due to its advantages such as all-weather, all-time capability, high precision, and high resolution, InSAR technology has rapidly developed into a mature space-based Earth observation technique, capable of mapping global topography with meter-level accuracy. The accuracy of InSAR time series analysis is crucial for monitoring surface deformation and displacement. PSI (Persistent Scatter Interferometry) and DSI (Distributed Scatter Interferometry) techniques are the main methods for improving the accuracy of InSAR deformation monitoring. Although PSI does not have the limitation of InSAR signal decorrelation, PS (Persistent Scatter) only exists in urban areas. In vegetated areas, PS density is low, and the natural scatterers in these areas exhibit signal decorrelation. Therefore, DSI technology was subsequently introduced. Since natural scatterers in vegetated areas introduce phase noise due to signal decorrelation, all DS (Distributed Scatter) techniques employ spatially homogeneous patch (SHP) or multi-look techniques to reduce phase noise in the interferogram. However, the study found that after multi-view processing, a phase component appeared that was not caused by statistical error and disrupted the consistency of the phase triplet. A phase triplet refers to three images that can interfere with each other, forming three interferograms with their phases. If these phases form a closed loop, and the interferogram is spatially homogeneous, the value of this closed loop is 0, a situation called phase consistency. When the closed loop value is not 0, phase inconsistency occurs. The value of the closed loop is called the closed phase, and a non-zero closed phase value is also called a non-zero closed phase. Therefore, reconstructing phase consistency means adding a component to the phase so that the closed loop formed by the phase triplet is 0. In interferometric models, this phase consistency has always been assumed due to pure propagation effects. To reduce the number of interferograms by using the most coherent and highest signal-to-noise ratio observations in the time series, short-time baseline interferograms are typically used when processing large amounts of data. However, the study found that phase inconsistency caused by multi-view processing affects short-time baseline interferograms, introducing significant errors into surface deformation and displacement. The study found that closed phase is a transient signal that decays over time. It also found that even small closed phases can cause the accuracy of surface displacement rate maps to increase from sub-millimeter to centimeter per year during long-term series propagation. This error will have a serious impact on the accuracy of deformation measurements. Therefore, the study of closed phase and the reconstruction of phase consistency are very necessary.

[0020] Existing technologies have identified this phase inconsistency phenomenon and attempted to explain closed phases. Based on a proposed soil moisture variation model, studies have found that changes in soil moisture may be a contributing factor to closed phases. Further observations suggest that phase inconsistency may also be related to variations in the dielectric constant at different heights due to changes in the water content of the trees themselves. However, after simultaneously studying two experimental sites, it was found that the closed phase sign in the vegetated areas of these two sites did not match the moisture model. The vegetation at both test sites was predominantly sparse grassland and crop fields, which differed from the closed phase behavior observed in densely vegetated areas using the moisture model. Therefore, a complex-valued decorrelation model was proposed to model the closed phase and explain its behavior. This model incorporates three scattering behaviors: a continuous coherent scattering and two decorrelation scatterings with different time constants. While this model helps explain the physical phenomena behind closed phases, research on closed phases related to biomass remains lacking, and further comprehensive studies on closed phases are needed. Currently, the optimal method for reconstructing phase consistency remains the phase linking method.

[0021] Phase linking methods have traditionally been used for phase estimation optimization, but they are also an efficient method for phase consistency reconstruction. Phase linking estimates a common interferometric master phase sequence from all interferograms in a time series of SAR data. That is, given a SAR image, phase linking estimates a long sequence of independent phases from this interferogram. Due to the complexity of closed-phase modeling, phase linking does not need to consider a specific model for closed phases; instead, it achieves robustness in calculating closed phases by utilizing a sufficient number of interferograms over a long time series. Phase linking methods used for reconstructing phase consistency include PTA (Phase triangulation algorithm), EVD (Eigenvalue decomposition), EMI (Eigenvalue decomposition-based maximum likelihood), and SBW (Short Bandwidth). PTA decomposes the coherence matrix into an eigenvalue-like form, consisting of multiplying a diagonal matrix by a real symmetric matrix, and then inverts the optimal phase satisfying phase consistency from the coherence matrix by maximizing the likelihood function. The PTA algorithm assumes that the full-rank real symmetric matrix estimated from coherence values ​​has high accuracy; therefore, the accuracy of phase consistency reconstruction depends on the estimated phase sequence. The EVD-based phase linking method performs eigenvalue decomposition on the covariance matrix of multi-baseline SAR data, selecting eigenvectors corresponding to larger eigenvalues. The phase components corresponding to these vectors have strong correlation and stability, effectively filtering out phase noise caused by decoherence. Finally, the orthogonality of EVD is utilized, and the interferogram is reconstructed based on the selected eigenvectors, ensuring phase consistency. Although EVD is computationally efficient, it is not the optimal estimate, and its phase accuracy is lower than that using maximum likelihood estimation. EMI combines the advantages of both algorithms, performing maximum likelihood estimation of the interferometric phase based on eigenvalue decomposition. To correct coherence errors, this algorithm introduces a calibration term. The methods described above use full interferometric pairs, while the SBW method uses only partial interferometric pairs, such as short-baseline interferometric pairs. It performs phase unwrapping on the interferometric pairs in the subset, obtaining the relative phase difference of each pair. Finally, using a fixed SLC (simple look complex) as a common reference, it fuses the local deformation information of all subsets using the least squares method to establish the absolute deformation of the entire time series, ensuring global phase consistency. This method, however, only uses short-baseline interferometric pairs, resulting in the loss of high-coherence signals from long-baseline interferometric pairs, thus reducing accuracy.

[0022] In summary, the closed phase exhibits an exponential decay trend with increasing average time baseline, and this decay also follows the trend of average coherence. Before proceeding with this invention, an attempt was made to correct the phase deviation calculated based on the SBW method using a coherent time matrix. This method attempts to assume that the ratio between the phase deviation of the first diagonal element and the closed phase of the triplet is... So how are the coefficients obtained in this method? This is the most crucial point. Since changes in coherence are related to the decay of the closed phase, this invention estimates coherence changes based on these changes. First, the coherence variation is fitted using the average coherence value, as shown below: ; in, , N For SAR image scene number, For the first Jinghe Coherence between SAR images and This is the general formula for a power function. The function is then used. To estimate the ratio As shown below: .

[0023] In the simulation experiment, a baseline constraint was used as a reference. The proposed method, along with SVD (Singular Value Decomposition) and SVD-NL, effectively corrected the phase bias. Phase sequences obtained using different bw (bandwidth) values ​​showed a similar trend, but the correction effect was poor for nonlinearly growing, inconsistent phases. When using SVD-NL, the phase biases estimated by different bw-n values ​​largely overlapped. However, SVD, due to its use of least squares, exhibited significant biases in different bw-n phase sequences. In contrast, the proposed method effectively corrected phase bias in short-baseline analysis, maintained high consistency across different bw-n values, and performed well in nonlinear phase bias analysis. However, in the experiments of this invention, it was found that while coherence changes and phase bias changes showed the same trend with increasing average time baseline, there was no linear relationship. Furthermore, in some scenarios, the phase bias did not decrease with increasing time baseline, making it impossible to use coherence changes to estimate coefficients. In the formula derivation, this invention discovered that the cube root of the first derivative of the triplet closed phase is proportional to the phase deviation sequence. However, it is difficult to invert the phase deviation solely based on the triplet closed phase because the coefficient related to the inconsistency process intensity is an unknown parameter. Therefore, this invention uses a deep learning-based method to skip the process of obtaining this unknown parameter and establish a nonlinear relationship between the coherence time matrix and the phase deviation.

[0024] The study area of ​​this invention is located in a dry lake region of a desert. Because this region belongs to an arid climate zone, soil moisture variation is likely the main factor causing phase inconsistencies, while other factors that might complicate the analysis are absent. The study area is characterized by a desert landscape and contains several dry lakes.

[0025] The preferred embodiment of the InSAR non-uniform phase correction method based on coherence matrix guidance described in this invention, such as... Figure 1 As shown, the InSAR non-uniform phase correction method guided by the coherence matrix includes the following steps: Step S10: Obtain land cover category data for the target study area, and calculate the symmetric phase matrix and symmetric coherence time matrix based on the land cover category data.

[0026] Accurate InSAR measurements of surface deformation and displacement are a prerequisite for studying natural disasters such as volcanic eruptions, landslides, and earthquakes. While multi-view techniques can reduce phase noise in interferograms, they introduce phase bias, i.e., closed phases. Since the behavior of closed phases is related to various factors, such as soil moisture, vegetation water content, and land cover classification, modeling closed phases cannot fully encompass their behavior. Currently, phase linking methods such as PTA, EMI, and EVD reconstruct phase consistency, but these methods are computationally expensive due to the use of the full covariance matrix. While SBW uses only short baseline interferogram pairs to reconstruct consistent phases, its accuracy is low. To address these limitations, this invention proposes using deep learning to directly derive closed phases from the coherence time matrix to correct SBW. This invention uses a ViT (Vision Transformer) model to extract features from the coherence time matrix, establishes a nonlinear relationship between the coherence matrix and the closed phase, and directly predicts the closed phase using a trained model to obtain the reconstructed phase sequence. Since dense vegetation and water bodies can cause interferogram decoherence, this invention conducts experiments in an arid region with a large amount of bare land, specifically a dry lake area. This invention initially uses bare land areas for training and prediction. The predicted closed-phase sequences are evaluated using indicators such as R, RMSE, and MSE. The results show that the method of this invention is effective in predicting in a bare land area of ​​a lake and can provide some improvement to the SBW method.

[0027] like Figure 2As shown, the present invention consists of four parts: data collection and preprocessing, data modeling and training, evaluation of model effectiveness, and measurement using the trained model.

[0028] The data collection and preprocessing process is as follows: Figure 2 As shown, the present invention requires a total of three types of data: land cover category data of the study area, coherence time matrix of each pixel in the study area (i.e., symmetric coherence time matrix in the present invention), and phase matrix of each pixel in the study area (i.e., symmetric phase matrix in the present invention).

[0029] Specifically, a target study area is determined, global land cover category data is acquired, and the global land cover category data is cropped according to the target study area to obtain a cropped image; a preset single-view complex image and digital elevation model data are acquired in the target study area, and the cropped image is registered with the preset single-view complex image and the digital elevation model data to obtain a differential interferogram.

[0030] This invention is based on global land cover category data for a specific year provided by Sentinel-1 and Sentinel-2 data, with a spatial resolution of 10m. This invention uses the range-based raster clipping function in QGIS software to clip the global land cover category data according to the scope of the study area (i.e., the target study area in this invention). Specifically, 156 Sentinel-1 down-orbit SLC images were collected for the study area (i.e., the target study area in this invention), covering the period from May 10, 2017 to January 31, 2021. Because the original SLC image was too large, this invention cropped the original SLC image to an image with 4600 rows in the azimuth direction and 44000 columns in the range direction. This image was then registered with an SLC image dated October 15, 2019, and a 1-arc second SRTM DEM (Digital Elevation Model (DEM) data with a resolution of 1 arcsecond acquired by the Shuttle Radar Topography Mission (SRTM)). This registration step was to enable the conversion between geographic coordinates and radar coordinates. (In practical applications, the calibration process has been integrated into the S1_proc_DEM command line of the GAMMA software; this invention uses GAMMA software to process the SLC image data.) In generating the interferogram, 40 and 10 multi-look operations were performed in the range and azimuth directions, respectively (i.e., spatial averaging of the image was performed using a sliding window of size 40 in the range direction and 10 in the azimuth direction), resulting in an interferogram with 460 rows in the azimuth direction and 1100 columns in the range direction in the radar coordinate system. The SRTM digital elevation model was also used to simulate terrain phase, where the differential interferometric phase was obtained by subtracting the terrain phase from the interferometric phase.

[0031] The differential interferogram is subjected to local window Fourier transform processing to obtain the spectral components of the signal and noise separation; the spectral components are adaptively calculated to obtain the signal spectrum; the signal spectrum is spatially transformed using the inverse Fourier method to obtain a smooth interferogram.

[0032] Furthermore, the differential interferogram is processed using Adaptive Spectral Filtering (ADF). The specific processing steps are as follows: This step is inherited from the ADF command in the GAMMA software. First, a local window Fourier transform is performed on the differential interferogram to separate the spectral components of the signal and noise. Then, noise suppression weights are adaptively calculated to preserve the signal spectrum and attenuate the noise spectrum. Finally, an inverse Fourier transform is used to return the signal to the spatial domain, and the smoothed interferogram is output.

[0033] The smoothed interferogram is subjected to phase expansion processing to obtain the interference phase.

[0034] After ADF processing, phase unwrapping is performed using MCF (Minimum Cost Flow). The specific implementation steps are as follows: This step is inherited from the mcf command of the GAMMA software. Interferogram pixels are regarded as network nodes, and adjacent pixels are regarded as edges. Edges are only allowed to flow once, and the capacity is set to 1. The cost of an edge is determined by the entangled phase difference of adjacent pixels. The cost needs to minimize the true phase difference of adjacent pixels. Pixels in the interferogram whose entangled phase difference does not meet the continuity condition are identified. Pixels with positive residuals are set as source points, and pixels with negative residuals are set as sink points. Finally, the minimum cost flow path from the source point to the sink point is found through an optimization algorithm. The flow on the path corresponds to the correction amount that is an integer multiple of 2 required for phase unwrapping. Finally, the true phase is recovered, and the interferometric phase is obtained.

[0035] Calculate the coherence value corresponding to the interference phase, obtain the symmetric phase matrix based on the interference phase, and obtain the symmetric coherence time matrix based on the coherence value.

[0036] This invention, upon obtaining the interferometric phase, can calculate the coherence value between different dates based on the interferometric phase. This invention constructs a symmetric phase matrix and a symmetric coherence time matrix from the interferometric phases and coherence values ​​of all interferometric pairs at each point. The first row of both matrices represents the unwrapped phase and coherence obtained by differential interferometry between the first SAR image and other SAR images, and so on for the remaining rows.

[0037] Step S20: Perform phase reconstruction processing on the symmetric phase matrix to obtain the target closed phase, and construct the target dataset based on the target closed phase and the symmetric coherent time matrix.

[0038] Because vegetation has poor coherence, this invention only studies bare land. Based on land cover category data, this invention extracts the coherence time matrix and phase matrix of points in bare land or sparse vegetation for subsequent data preprocessing.

[0039] Specifically, a first preset waveband and a second preset waveband are determined. The least squares method is used to solve for the phase consistency of the symmetric phase matrix based on the first preset waveband and the second preset waveband to obtain a first phase sequence and a second phase sequence. The first phase sequence and the second phase sequence are subtracted to obtain the target closed phase. The symmetric coherent time matrix is ​​used as the model input, the target closed phase is used as the input label, and a target dataset is constructed based on the symmetric coherent time matrix and the target closed phase.

[0040] After obtaining the phase matrix (i.e., the symmetric coherent time matrix in this invention), this invention uses the Singular Value Decomposition (SVD) method to reconstruct the consistent phase using bw=2 (i.e., the first preset bandwidth in this invention) and bw=156 (i.e., the second preset bandwidth in this invention) respectively (solving the phase consistency constraint equation by least squares). In this invention, the phase reconstructed by bw=156 is considered to be the accuracy achieved after deep learning correction. The phase reconstructed by bw=2 still has a large error compared to bw=156, but the closed phase obtained by subtracting the two phase sequences (i.e., the target closed phase in this invention) can be used as the label of the deep learning model. To enable faster and better model convergence, this invention normalizes and scales the labels to the range of 0-1, and the coherent time matrix (i.e., the symmetric coherent time matrix in this invention) is used as the input of the deep learning model, together with the target closed phase, to form the input dataset (i.e., the target dataset in this invention).

[0041] Step S30: Determine the preset model, and perform model training and optimization on the preset model based on the target dataset to obtain the surface deformation and displacement test model.

[0042] The modeling and training process is as follows: Since this invention aims to predict the closed phase based on the features in the coherent time matrix, it uses the ViT model (i.e., the preset model in this invention) to establish the relationship between the two.

[0043] Specifically, a preset model is determined, wherein the preset model is a ViT model; the symmetric coherence time matrix in the target dataset is input into the ViT model, the symmetric coherence time matrix is ​​divided into blocks by the ViT model to obtain image blocks, and the image blocks are converted into block matrices.

[0044] The bare land dataset obtained in this invention comprises a total of 435,060 pairs. Considering training time, this invention randomly selects a portion of the data to input into the ViT model for training. The symmetric coherence time matrix is ​​input into the ViT model as a single-channel image in the form of a numerical matrix. The ViT model ultimately outputs the predicted closed phase. Based on the predicted closed phase, this invention can reduce the error between the phase reconstructed based on bw=2 and the phase reconstructed based on bw=156. Figure 2 As shown, the learning and prediction of the closed phase using the ViT model comprises five stages: block partitioning of the coherent time matrix, block embedding, position embedding, transformer encoder, and prediction of the closed phase. In the block partitioning of the time matrix, the input is first... The single-channel time coherence matrix of size is divided into sizes. Non-overlapping image patches, each image patch is also a single-channel image, and the value of each point in the image patch represents the coherence value. During the block embedding stage, Each block is flattened into one dimension, and the input block is transformed into a linear layer. A block matrix P of dimension is then passed through a linear layer to map the dimensions of the block matrix P to... .

[0045] Obtain a preset learnable tag, and perform addition and matrix addition processing on the preset learnable tag and the block matrix to obtain the embedded block.

[0046] In the location embedding stage, each block sequence is first associated with a learnable label in the block matrix P. The summation results in a final output that is a global feature representation of the input phase matrix. This invention uses this global representation as the result of feature extraction to predict the closed phase. Subsequently, a dimension is added after the block matrix P. matrix This matrix is ​​used to preserve positional information, ultimately resulting in the embedded block. .

[0047] A preset transformer encoder is determined, and the embedding block is input to the preset transformer encoder. The preset transformer encoder performs encoding and matrix transformation processing on the embedding block to obtain a transformed embedding block.

[0048] The transformer encoder consists of N sets of encoders with the same structure. Each set of encoders processes the encoder generated by the previous set, which includes position markers. Each encoder consists of two main sub-blocks: a feedforward module and a multi-head attention module. The feedforward module comprises two dense layers and one GeLU activation layer. In the multi-head attention module, matrix multiplication is used to perform self-attention calculations before the self-attention calculations are performed. This is transformed into three distinct matrices Q, K, and V, which represent the query, key, and value, respectively, where dQ, dK, and dV represent their dimensions.

[0049] Self-attention calculation is performed on the transformation embedding block to obtain the self-attention calculation result.

[0050] Understandably, self-attention computation is computation. The similarity between different blocks is calculated using the following formula: ; in, For similarity results, Typically set to k is the number of heads in the multi-head attention module. This is a transpose.

[0051] The self-attention calculation result is input into the linear layer to obtain the final output result of the preset transformer encoder; the MLP model is determined, the final output result is input into the MLP model, and the MLP model is used to perform closed phase prediction processing on the final output result to obtain the predicted closed phase. The model training is completed, and the initial training model is obtained.

[0052] Multi-head attention computation involves multiple self-attention modules computing independently. The results of all self-attention modules are then concatenated and input into a linear layer to maintain the same input and output dimensions for each multi-head attention module. After multi-head attention module computation, a residual connection is used to add the input Z to the output of the multi-head attention module as the input to subsequent sub-blocks. The encoder contains two layer normalization layers. The input of each sub-block must pass through a normalization layer, which plays a role in stabilizing the loss gradient during backpropagation. The final result of the encoder is the output of the Nth layer, which is then fed into an MLP (Multi-layer Perceptron) with two hidden layers to predict the corresponding closure phase, as shown in the following formula: ; in, The predicted closed phase; This is the final output of the encoder; , , They consist of three fully connected layers; ReLU activation function, used to increase model nonlinearity; Dropout regularization is used to prevent overfitting in MLPs; after each fully connected layer, the ReLU activation function is applied first, followed by Dropout regularization.

[0053] The error between the predicted closed phase and the target closed phase is calculated using the MSE loss function, and the initial training model is optimized based on the error to obtain the surface deformation and displacement test model.

[0054] Understandably, this invention uses the MSE (Mean Squared Error) loss function in the model to calculate the error between the predicted closed phase and the true label, as shown in the following formula: ; Where i is the i-th sample, and n is the total number of samples. For the i-th predicted sample, Let i be the i-th real label.

[0055] Furthermore, this invention selects three indicators—MSE, RMSE (Root Mean Squared Error), and correlation coefficient R (Pearson Correlation Coefficient)—to evaluate the predictive performance of the model.

[0056] Further, a preset study area is determined, and a preset temporal coherence matrix of the preset study area is obtained; the preset temporal coherence matrix is ​​input into the surface deformation displacement test model to obtain a test closed phase sequence; the first phase sequence and the second phase sequence are subjected to phase deviation correction processing according to the test closed phase sequence to obtain a target phase deviation; the polyfit function is used to calculate the annual deformation displacement rate corresponding to each pixel in the preset study area according to the target phase deviation, and the results output by the surface deformation displacement test model are compared and verified according to the annual deformation displacement rate to obtain a verification result.

[0057] The specific process for evaluating model effectiveness is as follows: To test the model's generalization ability, this invention uses the trained model to predict on all bare land or sparsely vegetated points in the study area. The temporal coherence matrix of all bare land or sparsely vegetated points is used as the model input, and the closed phase sequence of each bare land or sparsely vegetated point is output. In addition, this invention selects an area with a similar land cover category to the study area (i.e., the preset study area in this invention), and uses the trained model to predict on bare land or sparsely vegetated points in this area. The temporal coherence matrix of each point in this area is used as the model input, and the closed phase sequence of each point is output. This invention adds the predicted closed phase sequence of each point back to the biased phase sequence (i.e., the phase sequence obtained by calculating bw=2 using SVD) for phase bias correction. Then, the annual deformation displacement rate of each point is calculated using the polyfit function on the corrected phase sequence. Finally, this invention analyzes three evaluation indicators and compares the annual displacement rate map after phase bias correction by the deep learning model with the annual displacement rate map after ideal phase bias correction (i.e., the phase sequence obtained by calculating bw=156 using SVD). The comparison yields the verification results.

[0058] Step S40: Determine the target test area and obtain the target land cover category data of the target test area. Input the target land cover category data into the surface deformation and displacement test model to obtain the target surface deformation and displacement test results of the target test area.

[0059] Specifically, once the surface deformation displacement test model is constructed, it can be used to test the surface deformation displacement of the target test area, which can effectively improve the testing efficiency and accuracy of surface deformation displacement.

[0060] In summary, current research on closed phases in existing technologies follows two main directions: one is mathematical modeling of the closed phase, but since the behavior of closed phases varies across different surfaces and environments, direct modeling is not comprehensive; the other is using phase linking methods to reconstruct phase consistency, such as PTA, EMI, and EVD. However, these methods use full interferometric pairs, which, while highly accurate, are computationally expensive, especially when studying interferometric pairs over longer periods. While SBW uses only short-baseline interferometric pairs, its estimated phase sequence deviates significantly from methods like PTA. Based on the relationship between coherence variation and phase deviation discussed above, current technologies do not yet utilize deep learning to establish the relationship between the coherence time matrix and the closed phase. This invention uses deep learning, requiring only a trained ViT model (i.e., a surface deformation displacement test model) to predict the area to be predicted, directly obtaining the closed phase. This overcomes the incompleteness of the aforementioned mathematical modeling and achieves faster speed and higher accuracy in prediction.

[0061] In addition, the present invention specifies the following possible design changes: 1. Currently, the technical solution of this invention only uses the ViT model. In the future, cross-attention mechanism or other mechanisms can be added to the ViT model as a decoder to further learn the relationship between coherence and closed phase.

[0062] 2. In addition to making predictions in bare land or sparsely vegetated areas in arid climate zones, predictions can also be made for other land cover types such as shrubs and grasslands.

[0063] 3. We can try to train predictions for non-arid climate zones, such as semi-arid zones.

[0064] 4. The encoder can use other image feature extraction models instead of the ViT model, such as VGG16, ResNet18, etc.; the decoder MLP can also be replaced by other time-series models, such as LSTM, RNN, GRU, etc.

[0065] Furthermore, such as Figure 3 As shown, based on the above-described InSAR non-uniform phase correction method guided by coherence matrix, this invention also provides an InSAR non-uniform phase correction system guided by coherence matrix, wherein the InSAR non-uniform phase correction system guided by coherence matrix includes: The land cover category data processing module 51 is used to acquire land cover category data of the target study area and calculate the symmetric phase matrix and symmetric coherence time matrix based on the land cover category data; The dataset construction module 52 is used to perform phase reconstruction processing on the symmetric phase matrix to obtain the target closed phase, and construct the target dataset based on the target closed phase and the symmetric coherence time matrix; The test model construction module 53 is used to determine a preset model and to train and optimize the preset model based on the target dataset to obtain a surface deformation and displacement test model. The test result output module 54 is used to determine the target test area, obtain the target land cover category data of the target test area, input the target land cover category data into the surface deformation displacement test model, and obtain the target surface deformation displacement test result of the target test area.

[0066] Furthermore, such as Figure 4 As shown, based on the above-mentioned InSAR non-uniform phase correction method and system guided by coherence matrix, the present invention also provides a terminal, which includes a processor 10, a memory 20 and a display 30. Figure 4 Only some of the terminal components are shown; however, it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0067] In some embodiments, the memory 20 may be an internal storage unit of the terminal, such as a hard disk or memory. In other embodiments, the memory 20 may be an external storage device of the terminal, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc. Further, the memory 20 may include both internal and external storage devices. The memory 20 is used to store application software and various types of data installed on the terminal, such as the program code installed on the terminal. The memory 20 can also be used to temporarily store data that has been output or will be output. In one embodiment, the memory 20 stores an InSAR non-uniform phase correction program 40 guided by a coherence matrix, which can be executed by the processor 10 to implement the InSAR non-uniform phase correction method guided by a coherence matrix in this application.

[0068] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor, or other data processing chip, used to run program code stored in the memory 20 or process data, such as executing the InSAR non-uniform phase correction method guided by the coherence matrix.

[0069] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. The display 30 is used to display information on the terminal and to display a visual user interface.

[0070] In one embodiment, when the processor 10 executes the InSAR non-uniform phase correction program 40 guided by the coherence matrix in the memory 20, it implements the steps of the InSAR non-uniform phase correction method guided by the coherence matrix as described above.

[0071] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores an InSAR non-uniform phase correction program guided by a coherence matrix, wherein when the InSAR non-uniform phase correction program guided by a coherence matrix is ​​executed by a processor, it implements the steps of the InSAR non-uniform phase correction method guided by a coherence matrix as described above.

[0072] In summary, this invention provides a method, system, terminal, and storage medium for InSAR non-consistent phase correction based on coherence matrix guidance. The method includes: acquiring land cover category data of a target study area, and calculating a symmetric phase matrix and a symmetric coherence time matrix based on the land cover category data; performing phase reconstruction processing on the symmetric phase matrix to obtain a target closed phase, and constructing a target dataset based on the target closed phase and the symmetric coherence time matrix; determining a preset model, and training and optimizing the preset model based on the target dataset to obtain a surface deformation displacement test model; determining a target test area, acquiring target land cover category data of the target test area, inputting the target land cover category data into the surface deformation displacement test model, and obtaining the target surface deformation displacement test results of the target test area. This invention constructs a symmetric phase matrix and a symmetric coherence time matrix by acquiring land cover category data of the study area, and then performs phase reconstruction processing based on the symmetric phase matrix and the symmetric coherence time matrix to obtain the target closed phase, thus realizing the construction of a target dataset. Furthermore, by training the preset model based on the target dataset, a surface deformation displacement test model is obtained, which can accurately test the surface deformation displacement and effectively improve the testing efficiency and accuracy of the surface deformation displacement test results.

[0073] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal that includes that element.

[0074] Of course, those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware (such as a processor, controller, etc.). The program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The computer-readable storage medium can be a memory, magnetic disk, optical disk, etc.

[0075] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A method for InSAR non-uniform phase correction based on coherence matrix guidance, characterized in that, The InSAR non-uniform phase correction method guided by the coherence matrix includes: Obtain land cover category data for the target study area, and calculate the symmetric phase matrix and symmetric coherence time matrix based on the land cover category data; The symmetric phase matrix is ​​subjected to phase reconstruction processing to obtain the target closed phase, and the target dataset is constructed based on the target closed phase and the symmetric coherence time matrix. A preset model is determined, and the preset model is trained and optimized based on the target dataset to obtain a surface deformation and displacement test model; The target test area is determined, and the target land cover category data of the target test area is obtained. The target land cover category data is input into the surface deformation and displacement test model to obtain the target surface deformation and displacement test results of the target test area.

2. The InSAR non-uniform phase correction method based on coherence matrix guidance according to claim 1, characterized in that, The acquisition of land cover category data for the target study area, and the calculation of the symmetric phase matrix and symmetric coherence time matrix based on the land cover category data, specifically includes: The target study area is determined, global land cover category data is obtained, and the global land cover category data is cropped according to the target study area to obtain a cropped image; Acquire a preset single-view complex image and digital elevation model data in the target study area, and register the cropped image with the preset single-view complex image and digital elevation model data to obtain a differential interferogram; The differential interferogram is subjected to adaptive spectral filtering to obtain a smoothed interferogram; The smoothed interferogram is subjected to phase unrolling to obtain the interference phase; Calculate the coherence value corresponding to the interference phase, obtain the symmetric phase matrix based on the interference phase, and obtain the symmetric coherence time matrix based on the coherence value.

3. The InSAR non-uniform phase correction method based on coherence matrix guidance according to claim 2, characterized in that, The adaptive spectral filtering process performed on the differential interferogram to obtain a smooth interferogram specifically includes: The differential interferogram is subjected to local window Fourier transform processing to obtain the spectral components of the signal and noise separation. The signal spectrum is obtained by adaptively calculating the spectral components. The signal spectrum is spatially transformed using the inverse Fourier transform method to obtain a smooth interferogram.

4. The InSAR non-uniform phase correction method based on coherence matrix guidance according to claim 1, characterized in that, The step of performing phase reconstruction processing on the symmetric phase matrix to obtain the target closed phase, and constructing the target dataset based on the target closed phase and the symmetric coherence time matrix, specifically includes: Determine the first preset waveband and the second preset waveband, and use the least squares method to solve the phase consistency of the symmetric phase matrix based on the first preset waveband and the second preset waveband to obtain the first phase sequence and the second phase sequence; Subtract the first phase sequence from the second phase sequence to obtain the target closed phase; The symmetric coherent time matrix is ​​used as the model input, the target closed phase is used as the input label, and the target dataset is constructed based on the symmetric coherent time matrix and the target closed phase.

5. The InSAR non-uniform phase correction method based on coherence matrix guidance according to claim 4, characterized in that, The process of determining a preset model and training and optimizing the preset model based on the target dataset to obtain a surface deformation and displacement test model specifically includes: A preset model is determined, wherein the preset model is the ViT model; The symmetric coherence time matrix in the target dataset is input into the ViT model. The ViT model performs block processing on the symmetric coherence time matrix to obtain image blocks, and the image blocks are converted into block matrices. Obtain a preset learnable tag, and perform addition and matrix addition processing on the preset learnable tag and the block matrix to obtain the embedded block; A preset transformer encoder and MLP model are determined. The embedding block is encoded and closed phase is predicted using the preset transformer encoder and the MLP model to obtain the predicted closed phase. The model training is completed, and the initial training model is obtained. The error between the predicted closed phase and the target closed phase is calculated using the MSE loss function, and the initial training model is optimized based on the error to obtain the surface deformation and displacement test model.

6. The InSAR non-uniform phase correction method based on coherence matrix guidance according to claim 5, characterized in that, The process of determining a preset transformer encoder and an MLP model, and then encoding and predicting the closure phase of the embedding block using the preset transformer encoder and the MLP model to obtain the predicted closure phase, specifically includes: A preset transformer encoder is determined, and the embedding block is input to the preset transformer encoder. The preset transformer encoder performs encoding and matrix transformation processing on the embedding block to obtain a transformed embedding block. Self-attention calculation is performed on the transformed embedding block to obtain the self-attention calculation result; The self-attention calculation result is input into the linear layer to obtain the final output result of the preset transformer encoder; The MLP model is determined, and the final output result is input into the MLP model. The MLP model is then used to perform closed-phase prediction processing on the final output result to obtain the predicted closed phase.

7. The InSAR non-uniform phase correction method based on coherence matrix guidance according to claim 4, characterized in that, The process of determining a preset model, training and optimizing the preset model based on the target dataset to obtain a surface deformation and displacement test model, further includes: Determine a preset study region and obtain a preset temporal coherence matrix for the preset study region; The preset time coherence matrix is ​​input into the surface deformation displacement test model to obtain the test closed phase sequence; Based on the test closed phase sequence, the first phase sequence and the second phase sequence are subjected to phase deviation correction processing to obtain the target phase deviation; The annual deformation displacement rate corresponding to each pixel in the preset study area is calculated using the polyfit function based on the target phase deviation. The results output by the surface deformation displacement test model are then compared and verified based on the annual deformation displacement rate to obtain the verification results.

8. A non-uniform phase correction system for InSAR based on coherence matrix guidance, characterized in that, The InSAR non-uniform phase correction system guided by the coherence matrix includes: The land cover category data processing module is used to acquire land cover category data of the target study area and calculate the symmetric phase matrix and symmetric coherence time matrix based on the land cover category data; The dataset construction module is used to perform phase reconstruction processing on the symmetric phase matrix to obtain the target closed phase, and construct the target dataset based on the target closed phase and the symmetric coherence time matrix; The test model construction module is used to determine the preset model and perform model training and optimization on the preset model based on the target dataset to obtain the surface deformation and displacement test model. The test result output module is used to determine the target test area, obtain the target land cover category data of the target test area, input the target land cover category data into the surface deformation and displacement test model, and obtain the target surface deformation and displacement test results of the target test area.

9. A terminal, characterized in that, The terminal includes: a memory, a processor, and an InSAR non-uniform phase correction program based on coherence matrix guided by the memory that is stored in the memory and can run on the processor. When the InSAR non-uniform phase correction program based on coherence matrix guided by the processor is executed, it implements the steps of the InSAR non-uniform phase correction method based on coherence matrix guided by any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores an InSAR non-uniform phase correction program guided by a coherence matrix, which, when executed by a processor, implements the steps of the InSAR non-uniform phase correction method guided by a coherence matrix as described in any one of claims 1-7.

Images