An InSAR regional network adjustment method and device based on physical driving and data learning

Abstract

The application discloses an InSAR regional network adjustment method based on physical driving and data learning, belongs to the field of satellite image processing, and effectively extracts common nodes of multi-strip DEM data based on an optimal registration method of minimum height difference, so that the robustness of the common node constraint condition is avoided to be reduced due to the position deviation of the DEM; in combination with the constraint conditions of the common nodes of the multi-strip DEM data and height control points, a spatial error optimization model based on a polynomial is established to effectively fit the error joint distribution of the multi-strip DEM data; the common nodes of the multi-strip DEM data and the height control points are used as constraint conditions of adjustment equations to perform InSAR regional network adjustment, which is beneficial to unifying the system difference between the DEM data obtained from different strips and ensuring the consistency of the error distribution of the DEM product; and the common nodes of the multi-strip DEM data are combined to perform regional network adjustment, which is beneficial to reducing the requirement for the density of the ground control points and improving the numerical stability of the adjustment equations.

Description

Technical Field

[0001] This invention belongs to the field of satellite image processing, and in particular relates to an InSAR regional network adjustment method and apparatus based on physical drive and data learning. Background Technology

[0002] Synthetic Aperture Radar Interferometry (InSAR) technology, with its all-weather, all-time, and high-precision Earth observation capabilities, has become an important tool in fields such as surface deformation monitoring and topographic height measurement, and is widely used in the generation and construction of Digital Elevation Models (DEMs) and Digital Surface Models (DSMs). As the application scope of DEM products continues to expand, higher requirements are being placed on the accuracy of DEM products generated by InSAR systems. Errors between multi-band DEMs are mainly divided into random errors and systematic errors. Random errors can be reduced in terms of accuracy through image preprocessing and filtering; however, the impact of systematic errors cannot be reduced by preprocessing or filtering. Therefore, it is necessary to utilize regional network adjustment to correct the systematic errors of multi-band DEMs.

[0003] A comprehensive comparison of existing InSAR regional network adjustment algorithms, both domestically and internationally, reveals two main categories of methods: those based on imaging models and those based on DEM products. The DEM-based method considers the elevation of common nodes and control points in overlapping adjacent strips as constraints. It models the error distribution of DEM data and solves optimization equations, performing joint adjustment on regions composed of DEM data from different strips to ensure consistent output DEM product performance within a certain range. However, due to the differences in elevation control points and common nodes in overlapping areas, the numerical stability of the optimization problem may decrease in complex terrain, leading to ill-conditioned problems in the adjustment equations. Positional offsets in overlapping DEM data can introduce errors in the common node constraints, affecting the accuracy of the image adjustment model parameter estimation. Furthermore, polynomial-based spatial error models cannot accurately fit complex error distributions in areas with large terrain variations, resulting in decreased stability and accuracy of the regional network adjustment. Data-based spatial error model construction methods rely heavily on the number and accuracy of external control points, and are prone to insufficient generalization ability due to overfitting.

[0004] Various factors lead to the following technical challenges in InSAR regional network adjustment based on DEM products: 1) How to select appropriate common nodes of adjacent DEM strips as constraints to improve the numerical stability of the adjustment equations; 2) How to establish a physics-driven spatial error model to effectively fit the joint error distribution of multiple DEM data; 3) How to establish a data-learning-based error regression prediction model to compensate for the shortcomings of multinomial models in complex terrain areas and achieve accurate prediction and correction of elevation errors. Summary of the Invention

[0005] This invention aims to solve the above-mentioned technical problems and provides an InSAR regional network adjustment method and apparatus based on physical drive and data learning.

[0006] In a first aspect, the present invention provides an InSAR regional network adjustment method based on physical driving and data learning, comprising the following steps:

[0007] Step 1: Based on the ground sampling interval of the DEM data, divide the overlapping areas of multiple DEM data into blocks of a preset size; calculate the elevation difference between each pair of image blocks, optimize the longitude and latitude offset of each pair of image blocks according to the minimum elevation difference criterion, and complete the DEM registration.

[0008] Step 2: Extract the center coordinates of each pair of image patches as the common node of the multi-band data. TP Elevation control points were obtained using bicubic interpolation. GCP The DEM elevation value at the coordinate location; the constraint condition is that the elevation values ​​of the DEM data of each strip are equal at the common node.

[0009] Step 3: Combine the common nodes of multiple DEM strips TP and elevation control points GCP Construct a spatial error model; construct error observation equations based on constraints; and then solve for the optimal parameter matrix of the spatial error model.

[0010] Step 4: Based on elevation control points GCP Create an elevation control point feature dataset using longitude, latitude, slope, aspect, and surface roughness at the coordinate location;

[0011] Step 5: Construct a DEM error regression prediction model based on the Transformer architecture, input the elevation control point feature dataset for feature encoding and decoding, and use the optimizer to complete model training;

[0012] Step 6: Combine the polynomial-based spatial error model and the Transformer-based DEM error regression prediction model, use learnable parameter vectors to achieve weighted fusion of different models, output the predicted elevation difference, and use the superposition method to complete the correction and compensation of DEM error.

[0013] Furthermore, in the InSAR regional network adjustment method based on physics-driven and data learning described in this invention, when optimizing the longitude and latitude offsets of each pair of image patches according to the minimum elevation difference criterion in step 1, the optimal translation parameters are solved by minimizing the elevation difference between the two DEMs in the overlapping area.

[0014] ;

[0015] In the formula, N This indicates the total number of pixels in the DEM data block. Z 1 and Z 2 represents the two DEM data blocks to be registered, Δ x Δ y This indicates the offset along the longitude and latitude directions. x and y These represent longitude and latitude coordinates, respectively.

[0016] Furthermore, in the InSAR regional network adjustment method based on physics-driven and data learning described in this invention, the error observation equation in step 3 is as follows:

[0017] H ERROR =HC+E;

[0018] In the formula, H represents the observation matrix, C represents the coefficient column vector, and E represents the error matrix.

[0019] Furthermore, in the InSAR regional network adjustment method based on physics-driven and data learning described in this invention, the optimal parameter matrix of the spatial error model in step 3 is:

[0020] C WLS =(H T WH) -1 H T WH ERROR ;

[0021] In the formula, C WLS H represents the spatial error model parameter matrix, H represents the observation matrix, C represents the coefficient column vector, and H... ERROR Let W represent the error matrix, and W be the weighting matrix.

[0022] Furthermore, in the InSAR regional network adjustment method based on physics-driven and data learning described in this invention, when creating the elevation control point feature dataset in step 4, a linear layer is first used to encode the features of each sample, so as to map from the feature space to a high-dimensional latent space:

[0023] Y = XW + b;

[0024] In the formula, X∈ R m×n Indicates input features, n Indicates the dimension of the input features. m Indicates the number of samples; W∈ R n×p Represents the weight matrix. p Y ∈ [dimensionality of the high-dimensional implicit space] R m×p Represents the latent features of the output, b∈ R m×p This represents the bias vector.

[0025] Furthermore, in the InSAR regional network adjustment method based on physics-driven and data learning described in this invention, the DEM error regression prediction model based on the Transformer architecture in step 5 is as follows:

[0026] Y out =MLP(Y attn );

[0027] In the formula, Y attn ∈ R m×p Y represents the feature matrix encoded by the self-attention mechanism. out ∈ R m×1 represents the output error vector, and MLP(·) represents the multilayer perceptron mapping function.

[0028] Furthermore, in the InSAR regional network adjustment method based on physics-driven and data learning described in this invention, the weighted fusion form of different models in step 6 is as follows:

[0029] h errror =wY out +(1-w)HC;

[0030] In the formula, H and C represent the observation matrix and coefficient column vector of the polynomial model, respectively. h errror Let w represent the final predicted elevation difference, and w be a learnable parameter vector.

[0031] Secondly, the present invention provides an InSAR regional network adjustment system based on physical drive and data learning, including a DEM registration module, a spatial error model construction module, an error regression prediction model construction module, and a model fusion module;

[0032] The DEM registration module is used to divide the overlapping areas of multiple DEM data strips into blocks of a preset size according to the ground sampling interval of the DEM data; calculate the elevation difference between each pair of image blocks, optimize the longitude and latitude offsets of each pair of image blocks according to the minimum elevation difference criterion, and complete the DEM registration; and extract the center coordinates of each pair of image blocks as the common node of the multiple data strips. TP Elevation control points were obtained using bicubic interpolation. GCP The DEM elevation value at the coordinate location; the constraint condition is that the elevation values ​​of the DEM data of each strip are equal at the common node.

[0033] The spatial error model construction module is used to combine the common nodes of multiple DEM strips. TP and elevation control points GCP Construct a spatial error model; construct error observation equations based on constraints; solve for the optimal parameter matrix of the spatial error model; and then, based on elevation control points... GCP Create an elevation control point feature dataset using longitude, latitude, slope, aspect, and surface roughness at the coordinate location;

[0034] The error regression prediction model construction module is used for DEM error regression prediction model based on Transformer architecture. It inputs the elevation control point feature dataset for feature encoding and decoding, and uses an optimizer to complete model training.

[0035] The model fusion module is used to combine a polynomial-based spatial error model and a DEM error regression prediction model based on the Transformer architecture. It uses learnable parameter vectors to achieve weighted fusion of different models, outputs the predicted elevation difference, and uses a superposition method to complete the correction and compensation of DEM error.

[0036] Thirdly, the present invention provides an InSAR regional network adjustment device based on physical drive and data learning, comprising a memory and a processor; the memory is used to store a computer program; the processor is used to implement the InSAR regional network adjustment method based on physical drive and data learning as described in any one of the first aspects when the computer program is executed.

[0037] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the InSAR regional network adjustment method based on physical drive and data learning as described in any one of the first aspects.

[0038] The InSAR regional network adjustment method and apparatus based on physics-driven and data learning described in this invention have the following technical effects:

[0039] 1) The optimized registration method based on minimum elevation difference effectively extracts the common nodes of multiple DEM data, avoiding the decrease in the robustness of the common node constraint conditions caused by DEM position offset;

[0040] 2) Combining the constraints of common nodes and elevation control points of multi-band DEM data, a polynomial-based spatial error optimization model is established to effectively fit the joint error distribution of multi-band DEM data; the weighted least squares method is used to solve the adjustment equation, and different weights are assigned to the common nodes and elevation control points of multi-band DEM data to avoid the decrease in accuracy caused by the differences in different constraints.

[0041] 3) By combining the Transformer encoder in the neural network, the features initially manually selected are encoded, the self-attention mechanism is used to complete the effective feature weighting, and the multilayer perceptron is used to realize the final feature aggregation and decoding reconstruction, so as to achieve effective regression prediction of elevation error in complex terrain areas.

[0042] 4) At the same time, using the common nodes and elevation control points of multi-band DEM data as constraints for adjustment equations in InSAR regional network adjustment is beneficial to unify the systematic differences between DEM data acquired from different bands and ensure the consistency of DEM product error distribution; combining the common nodes of multi-band DEM data for regional network adjustment helps to reduce the requirements for ground control point density and improve the numerical stability of adjustment equations. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the InSAR regional network adjustment method based on physics-driven and data learning as described in an embodiment of the present invention;

[0044] Figure 2 This is a schematic diagram showing the distribution of control points and common nodes according to an embodiment of the present invention;

[0045] Figure 3 This is a schematic diagram of the optical images and data partitioning of the experimental area for the joint adjustment of multiple strip datasets according to an embodiment of the present invention;

[0046] Figure 4 This is a schematic diagram of the DEM mosaicking result before adjustment and the result of winding to the phase domain as described in an embodiment of the present invention;

[0047] Figure 5 This is a schematic diagram of the DEM mosaicking result after adjustment and the result of winding into the phase domain as described in the embodiment of the present invention. Detailed Implementation

[0048] The InSAR regional network adjustment method and apparatus based on physics-driven and data learning described in this invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0049] This embodiment discloses an InSAR regional network adjustment method based on physics-driven and data learning, such as... Figure 1 As shown, it includes the following steps:

[0050] Step 1: Based on the ground sampling interval of the DEM data, divide the overlapping areas of multiple DEM data into blocks of a preset size; calculate the elevation difference between each pair of image blocks, optimize the longitude and latitude offset of each pair of image blocks according to the minimum elevation difference criterion, and complete the DEM registration.

[0051] Step 2: Extract the center coordinates of each pair of image patches as the common node of the multi-band data. TP Elevation control points were obtained using bicubic interpolation. GCP The DEM elevation value at the coordinate location; the constraint condition is that the elevation values ​​of the DEM data of each strip are equal at the common node.

[0052] In this embodiment, the overlapping area of ​​multiple DEM data blocks is divided into data blocks of 1 km × 1 km size, and then the DEM data blocks on adjacent strips are registered. After registration, the center pixel elevation of the two DEM blocks is used as a common node constraint condition.

[0053] In the joint adjustment of multi-strip data, the adjustment model parameters for each strip data, in addition to satisfying the elevation constraints at the control points, include, for example: Figure 2 As shown, the adjustment models for adjacent strips (strip 1, strip 2) also need to satisfy the common node constraint, that is, the elevations of different strips are equal at the nodes. The common node is a virtual control point constructed to correct the relative errors between different strips.

[0054] After obtaining multiple pairs of DEM data blocks, registration is performed on the DEM data blocks based on the minimum elevation difference criterion. Since DEM data represents surface height information, the DEM registration problem only needs to handle the offsets in the longitude and latitude directions, without needing to handle affine transformations such as rotation and scaling. DEM registration based on the minimum elevation difference criterion solves for the optimal translation parameters by minimizing the elevation difference between the two DEMs in the overlapping area, and is modeled as follows:

[0055] ;

[0056] In the formula, N This indicates the total number of pixels in the DEM data block. Z 1 and Z2 represents the two DEM data blocks to be registered, Δ x Δ y This indicates the offset along the longitude and latitude directions. x and y These represent longitude and latitude coordinates, respectively.

[0057] To address the optimization problem, this embodiment employs gradient descent to obtain sub-pixel-level offsets. After registration, the center position of the DEM data block is selected as a common node, with the constraint that the elevation values ​​of each strip of DEM data at the common node are equal.

[0058] Step 3: Combine the common nodes of multiple DEM strips TP and elevation control points GCP Construct a spatial error model; construct error observation equations based on constraints; and then solve for the optimal parameter matrix of the spatial error model.

[0059] By combining the common nodes and elevation control points of multi-band DEM data as constraints, a physical-driven spatial error polynomial model is established, and the weighted least squares method is used to measure the differences between different constraints, so as to effectively fit the joint error distribution of multi-band DEM data and solve the adjustment equation.

[0060] In this embodiment, specifically during the DEM production process, a 1°×1° DEM (approximately 100 km × 100 km) product requires at least 2×4=8 scenes (4 strips) of interferometric data mosaicking. Within the same strip, the systematic error over a 100 km length is almost constant, and the corresponding DEM error variation is also very small. Due to differences in systematic errors, and even opposite signs, between data acquired from different strips, relative errors exist between DEMs. Therefore, the constructed spatial error model needs to consider both the absolute and relative errors between data acquired from different strips, scientifically fit the joint error distribution of DEM data from different strips, and correct the relative errors between datasets from different strips, thereby ensuring the consistency of errors within a single DEM product.

[0061] This embodiment uses the joint adjustment of two data strips as an example. J and strips K The DEM data need to meet their respective control point constraints:

[0062] ;

[0063] At the same time, the two adjustment models must also satisfy the common node constraint:

[0064] h J_DEM ( x J_TP , y J_TP )- g J ( x J_TP , y J_TP )= h K_DEM ( x K_TP , y K_TP )- g K ( x K_TP , y K_TP );

[0065] In the formula, h J_DEM ( x Ji , y Ji ), h J_GCP ( x Ji , y Ji ) respectively represent the position ( x Ji , y Ji ) stripe J DEM elevation and control point elevation, h K_DEM ( x Ki , y Ki ), h K_GCP ( x Ki , y Ki ) respectively represent the position ( x Ki , y Ki ) stripe K DEM elevation and control point elevation, ( x J_TP , y J_TP ), ( xK_TP , y K_TP ) represent the common nodes in the stripe. J , K The position in the middle, g J ( x Ji , y Ji ), g k ( x Ki , y Ki ) represent stripes respectively J , K The adjustment model after substituting the control points into the model, g J ( x J_TP , y J_TP ), g k ( x K_TP , y K_TP ) represent stripes respectively J , K The adjustment model after substituting common nodes;

[0066] g J ( x Ji , y Ji )= a J0 + a J1 · x Ji + a J2 · x Ji 2 + a J3 · x Ji 3 + b J1 · y Ji + c J · x Ji ·y Ji ;

[0067] g K ( x Ki , y Ki )= a K0 + a K1 · x Ki + a K2 · x Ki 2 + a K3 · x Ki 3 + b K1 · y Ki + c K · x Ki · y Ki ;

[0068] In the formula, the parameter a , b , c The parameters of the spatial error model are represented by: By introducing high-precision reference data and combining it with constraints, the observation equation can be established as follows:

[0069] H ERROR =HC+E;

[0070] In the formula, H, C, and E represent the observation matrix, coefficient column vector, and error matrix, respectively. Specifically:

[0071] ;

[0072] C=[ a J0 a J1 a J2 a J3 b J1 k J a K0 a K1 aK2 a K3 b K1 k K ] T ;

[0073] E=[ e (1) e (2) … e (i) … e ( m + n +l)] T ;

[0074] e This represents the error components determined by the control point constraints and the common node constraints;

[0075] For matrix H, it contains the following components:

[0076] ;

[0077] ;

[0078] ;

[0079] ;

[0080] In the formula, m , n Strips J, K The number of control points in the system l This represents the number of common nodes.

[0081] After constructing the spatial error model, the optimal coefficient matrix C of the model needs to be estimated using the weighted least squares method, i.e., to find the objective function. J Minimum optimization parameters:

[0082] ;

[0083] In the formula, W is the weighting matrix:

[0084] ;

[0085] Considering the differences between elevation control points and common nodes in overlapping areas, w i There are two cases: the location of control points and the location of common nodes in overlapping areas. The weighting coefficient for control point locations is determined by both the DEM variance and the control point variance, while the weighting coefficient for common nodes is the sum of the variances of the different strip DEMs.

[0086] ;

[0087] in, Indicates the first i The variance of the DEM corresponding to each control point Indicates the first i Variance at each control point and They represent stripes respectively. J , K The variance of the common nodes.

[0088] This shows that sample points with large variance receive smaller weights to reduce their overall weight, while sample points with small variance receive larger weights to increase their role in the objective function. The variance of the DEM is primarily determined by the random interferometric phase. Regarding the objective function... J coefficient column vector C By taking the derivative and setting it equal to zero, we can obtain the analytical solution for the coefficient column vector:

[0089] C WLS =(H T WH) -1 H T WH ERROR ;

[0090] In the formula, C WLS H represents the spatial error model parameter matrix, H represents the observation matrix, C represents the coefficient column vector, and H... ERROR Let W represent the error matrix, and W be the weighting matrix.

[0091] When acquiring global DEMs using spaceborne InSAR systems, the primary approach is joint adjustment of multi-strip datasets. Joint adjustment of multi-strip datasets not only ensures greater consistency in the error characteristics of global DEM products, but more importantly, it reduces the requirements for control point distribution density.

[0092] Step 4: Based on elevation control points GCP Create an elevation control point feature dataset using longitude, latitude, slope, aspect, and surface roughness at the coordinate location.

[0093] Step 5: Construct a DEM error regression prediction model based on the Transformer architecture. Input the elevation control point feature dataset for feature encoding and decoding, and use the optimizer to complete model training.

[0094] In this embodiment, a control point feature dataset in tabular form is first created, containing features such as longitude, latitude, slope, aspect, and surface roughness. A linear layer is then used to encode the features of each sample, mapping them from the feature space to a high-dimensional latent space.

[0095] Y = XW + b;

[0096] In the formula, X∈ R m×n Indicates input features, n Indicates the dimension of the input features. m Indicates the number of samples; W∈ R n×p Represents the weight matrix. p Y ∈ [dimensionality of the high-dimensional implicit space] R m×p Represents the latent features of the output, b∈ R m×p This represents the bias vector.

[0097] Then, a Transformer neural network based on the self-attention mechanism is constructed to perform regression prediction of the error.

[0098] The latent feature Y is linearly transformed to generate the query, key, and value matrices Q, K, and V required for calculating the attention mechanism. Attention scores are calculated by determining the correlation between the query matrix Q and the key matrix K to assign different weighting coefficients, thereby weighting V as a feature.

[0099] ;

[0100] In the formula, represents the scaling factor, and softmax(·) represents the activation function. Since the attention mechanism is essentially a linear weighting of features, a multilayer perceptron is needed to learn the nonlinear mapping relationship. Finally, feature aggregation and dimensionality reduction are used to regress and predict spatial errors.

[0101] Y out =MLP(Y attn );

[0102] In the formula, Y attn ∈ R m×p Y represents the feature matrix encoded by the self-attention mechanism. out ∈ R m×1 The output error vector is represented by MLP(·), which represents the multilayer perceptron mapping function. It includes multilayer feature dimensionality upscaling and feature dimensionality downscaling processes to achieve effective feature aggregation and decoding, and complete effective elevation difference prediction.

[0103] Step 6: Combine the polynomial-based spatial error model and the Transformer-based DEM error regression prediction model, use learnable parameter vectors to achieve weighted fusion of different models, output the predicted elevation difference, and use the superposition method to complete the correction and compensation of DEM error.

[0104] In this embodiment, a weighted fusion of different models is achieved by combining a polynomial-based spatial error model and utilizing a learnable parameter vector w:

[0105] h errror =wY out +(1-w)HC;

[0106] In the formula, H and C represent the observation matrix and coefficient column vector of the polynomial model, respectively. h errror This represents the final predicted elevation difference.

[0107] In this embodiment, the following specific verification process is described:

[0108] like Figure 3 As shown, the data comprises three strips (①, ②, and ③), all derived from DEM data generated by the TH-2 interferometry system. The three strips overlap by 20%, and each strip is approximately 50 × 30 km in size. Each strip contains 15 control points, and the initial number of common nodes in the overlapping area is 10. The mosaic result of the three strip DEM data before regional network adjustment is shown in the figure. Furthermore, to better illustrate the differences between the DEM data before adjustment, the data is arranged according to ambiguity height... h amb =100 meters to wrap the DEM data into the phase domain.

[0109] like Figure 4 As shown, before the regional network adjustment, the DEM data of multiple strips exhibit significant abrupt changes in the phase domain. Considering the impact of positioning errors, common nodes in overlapping areas are divided into blocks for registration processing. Constraints on common nodes are determined according to the registration offset, and combined with the elevation constraints of existing control points. The resulting DEM data mosaic after joint adjustment of the three strips is shown below. Figure 5 As shown.

[0110] After the above regional network adjustment process, as follows: Figure 5As shown, no obvious discontinuities are visible in the DEM, indicating that the adjustment process improved the DEM accuracy. Considering the impact of positioning errors on the constraints of common nodes, the accuracy of model parameter estimation is improved after obtaining the relative offsets of common nodes through block registration in overlapping areas, resulting in higher DEM accuracy. To better compare the adjustment results, the elevation error of the adjusted DEM was evaluated. The absolute elevation accuracy of the DEM before adjustment was 16.972 meters, and the absolute elevation accuracy after regional network adjustment was 4.193 meters. The experimental results show that the InSAR regional network adjustment technique based on the spatial error model effectively improves the accuracy of multi-band DEM data.

[0111] This second embodiment discloses an InSAR regional network adjustment system based on physical drive and data learning, including a DEM registration module, a spatial error model construction module, an error regression prediction model construction module, and a model fusion module;

[0112] The DEM registration module is used to divide the overlapping areas of multiple DEM data strips into blocks of a preset size according to the ground sampling interval of the DEM data; calculate the elevation difference between each pair of image blocks, optimize the longitude and latitude offsets of each pair of image blocks according to the minimum elevation difference criterion, and complete the DEM registration; and extract the center coordinates of each pair of image blocks as the common node of the multiple data strips. TP Elevation control points were obtained using bicubic interpolation. GCP The DEM elevation value at the coordinate location; the constraint condition is that the elevation values ​​of the DEM data of each strip are equal at the common node.

[0113] The spatial error model construction module is used to combine the common nodes of multiple DEM strips. TP and elevation control points GCP Construct a spatial error model; construct error observation equations based on constraints; solve for the optimal parameter matrix of the spatial error model; and then, based on elevation control points... GCP Create an elevation control point feature dataset using longitude, latitude, slope, aspect, and surface roughness at the coordinate location;

[0114] The error regression prediction model construction module is used for DEM error regression prediction model based on Transformer architecture. It inputs the elevation control point feature dataset for feature encoding and decoding, and uses an optimizer to complete model training.

[0115] The model fusion module is used to combine a polynomial-based spatial error model and a DEM error regression prediction model based on the Transformer architecture. It uses learnable parameter vectors to achieve weighted fusion of different models, outputs the predicted elevation difference, and uses a superposition method to complete the correction and compensation of DEM error.

[0116] The specific operation steps of the InSAR regional network adjustment system based on physical driving and data learning described in this embodiment are the same as those of the InSAR regional network adjustment method based on physical driving and data learning described in Embodiment 1 above, and will not be repeated here.

[0117] This embodiment three discloses an InSAR regional network adjustment device based on physical drive and data learning, including a memory and a processor; the memory is used to store a computer program; the processor is used to implement the InSAR regional network adjustment method based on physical drive and data learning described in embodiment one when the computer program is executed. The specific adjustment method steps are the same as those in the aforementioned embodiment one, and will not be repeated here.

[0118] This embodiment four discloses a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the InSAR regional network adjustment method based on physical drive and data learning as described in embodiment one. The specific adjustment method steps are the same as those in the aforementioned embodiment one, and will not be repeated here.

[0119] The computer described in this application embodiment can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. The computer-readable storage medium can be any usable medium that a computer can read, or a data storage device such as a server or data center that integrates one or more usable media. The usable medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., digital versatile optical disc (DVD)), or a semiconductor medium (e.g., solid-state drive (SSD)). The software formed by the computer's stored code can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other storage media that are mature in the art.

[0120] In the various embodiments of this application, the functional modules can be integrated into one processing unit or module, or each module can exist physically separately, or two or more modules can be integrated into one unit or module. In the above embodiments, they can be implemented entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, they can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated.

[0121] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for adjusting InSAR regional networks based on physics-driven and data learning, characterized in that, Includes the following steps: Step 1: Based on the ground sampling interval of the DEM data, divide the overlapping areas of multiple DEM data into blocks of a preset size; Calculate the elevation difference between each pair of image blocks, optimize the longitude and latitude offsets of each pair of image blocks according to the minimum elevation difference criterion, and complete the DEM registration; Step 2: Extract the center coordinates of each pair of image patches as the common node of the multi-band data. TP Elevation control points were obtained using bicubic interpolation. GCP DEM elevation values ​​at the coordinate locations; The constraint condition is that the elevation values ​​of the DEM data of each strip are equal at the common node. Step 3: Combine the common nodes of multiple DEM strips TP and elevation control points GCP Construct a spatial error model; construct error observation equations based on constraints; and then solve for the optimal parameter matrix of the spatial error model. Step 4: Based on elevation control points GCP Create an elevation control point feature dataset using longitude, latitude, slope, aspect, and surface roughness at the coordinate location; Step 5: Construct a DEM error regression prediction model based on the Transformer architecture, input the elevation control point feature dataset for feature encoding and decoding, and use the optimizer to complete model training; Step 6: Combine the polynomial-based spatial error model and the Transformer-based DEM error regression prediction model, use learnable parameter vectors to achieve weighted fusion of different models, output the predicted elevation difference, and use the superposition method to complete the correction and compensation of DEM error.

2. The InSAR regional network adjustment method based on physics-driven and data learning according to claim 1, characterized in that: In step 1, when optimizing the longitude and latitude offsets of each pair of image patches based on the minimum elevation difference criterion, the optimal translation parameters are obtained by minimizing the elevation difference between the two DEMs in the overlapping area. ； In the formula, N This indicates the total number of pixels in the DEM data block. Z 1 and Z 2 represents the two DEM data blocks to be registered, Δ x Δ y This indicates the offset along the longitude and latitude directions. x and y These represent longitude and latitude coordinates, respectively.

3. The InSAR regional network adjustment method based on physics-driven and data learning according to claim 1, characterized in that, The error observation equation described in step 3 is: H ERROR =HC+E; In the formula, H represents the observation matrix, C represents the coefficient column vector, and E represents the error matrix.

4. The InSAR regional network adjustment method based on physics-driven and data learning according to claim 3, characterized in that, The optimal parameter matrix of the spatial error model described in step 3 is: C WLS =(H T WH) -1 H T WH ERROR ; In the formula, C WLS H represents the spatial error model parameter matrix, H represents the observation matrix, C represents the coefficient column vector, and H... ERROR Let W represent the error matrix, and W be the weighting matrix.

5. The InSAR regional network adjustment method based on physics-driven and data learning according to claim 1, characterized in that, When creating the elevation control point feature dataset in step 4, a linear layer is first used to encode the features of each sample, mapping the feature space to a high-dimensional latent space: Y = XW + b; In the formula, X∈ R m×n Indicates input features, n Indicates the dimension of the input features. m Indicates the number of samples; W∈ R n×p Represents the weight matrix. p Y ∈ [dimensionality of the high-dimensional implicit space] R m×p Represents the latent features of the output, b∈ R m×p This represents the bias vector.

6. The InSAR regional network adjustment method based on physics-driven and data learning according to claim 5, characterized in that, The DEM error regression prediction model based on the Transformer architecture described in step 5 is as follows: AND out =MLP(Y attn ); In the formula, Y attn ∈ R m×p Y represents the feature matrix encoded by the self-attention mechanism. out ∈ R m×1 represents the output error vector, and MLP(·) represents the multilayer perceptron mapping function.

7. The InSAR regional network adjustment method based on physics-driven and data learning according to claim 6, characterized in that, The weighted fusion form of the different models described in step 6 is as follows: h errror =wY out +(1-w)HC; In the formula, H and C represent the observation matrix and coefficient column vector of the polynomial model, respectively. h errror Let w represent the final predicted elevation difference, and w be a learnable parameter vector.

8. An InSAR regional network adjustment system based on physics-driven and data learning, characterized in that: It includes a DEM registration module, a spatial error model building module, an error regression prediction model building module, and a model fusion module; The DEM registration module is used to divide the overlapping areas of multiple DEM data strips into blocks of a preset size according to the ground sampling interval of the DEM data; calculate the elevation difference between each pair of image blocks, optimize the longitude and latitude offsets of each pair of image blocks according to the minimum elevation difference criterion, and complete the DEM registration; and extract the center coordinates of each pair of image blocks as the common node of the multiple data strips. TP Elevation control points were obtained using bicubic interpolation. GCP DEM elevation values ​​at the coordinate locations; The constraint condition is that the elevation values ​​of the DEM data of each strip are equal at the common node. The spatial error model construction module is used to combine the common nodes of multiple DEM strips. TP and elevation control points GCP Construct a spatial error model; construct error observation equations based on constraints; solve for the optimal parameter matrix of the spatial error model; and then, based on elevation control points... GCP Create an elevation control point feature dataset using longitude, latitude, slope, aspect, and surface roughness at the coordinate location; The error regression prediction model construction module is used for DEM error regression prediction model based on Transformer architecture. It inputs the elevation control point feature dataset for feature encoding and decoding, and uses an optimizer to complete model training. The model fusion module is used to combine a polynomial-based spatial error model and a DEM error regression prediction model based on the Transformer architecture. It uses learnable parameter vectors to achieve weighted fusion of different models, outputs the predicted elevation difference, and uses a superposition method to complete the correction and compensation of DEM error.

9. An InSAR regional network adjustment device based on physical drive and data learning, characterized in that: It includes a memory and a processor; the memory is used to store a computer program; the processor is used to implement the InSAR regional network adjustment method based on physical drive and data learning as described in any one of claims 1-7 when the computer program is executed.

10. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, which, when executed by a processor, implements the InSAR regional network adjustment method based on physical drive and data learning as described in any one of claims 1-7.

Images