Method for predicting 3d deformation of mine surface by combining insar and neural network

Abstract

This invention discloses a method for predicting three-dimensional surface deformation in mining areas using a combination of InSAR and neural networks. The method includes: obtaining the three-dimensional deformation at any point on the surface during mining based on a mining subsidence model; acquiring the LOS-direction deformation phase; obtaining a simulated interferogram based on the LOS-direction deformation; dividing the simulated interferogram into training, validation, and test sets; calculating the variance of the InSAR phase derivative based on the entanglement phase; constructing a high-resolution three-dimensional network; inputting the training set into the high-resolution three-dimensional network; iteratively training the high-resolution three-dimensional network using a combined loss function; evaluating the trained high-resolution three-dimensional network using a validation set; and testing the network with the highest evaluation score using a test set; finally, acquiring a real mining area interferogram and inputting it into the tested high-resolution three-dimensional network to obtain the predicted three-dimensional deformation. This invention achieves a deformation solution with good stability and high accuracy through InSAR data and deep learning methods.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional surface deformation prediction in mining areas, and in particular to a method for predicting three-dimensional surface deformation in mining areas by combining InSAR and neural networks. Background Technology

[0002] Due to the complexity of underground coal mining conditions and the diverse stress variations of the overlying rock mass, large-gradient local deformations easily occur on the surface during mining. Phase unwrapping results in line-of-sight deformation, making it difficult to recover the correct surface deformation by directly unwrapping the interferogram. Existing methods, such as multi-source SAR data fusion for temporal deformation solutions, methods combining single-track InSAR data with prior models, and offset tracking (OT), suffer from difficulties in multi-source data fusion, acquisition of multi-source SAR data, poor stability in deformation solutions due to a lack of redundant observations, and low accuracy in deformation solutions. Summary of the Invention

[0003] The purpose of this invention is to provide a method for predicting three-dimensional surface deformation in mining areas by combining InSAR and neural networks, aiming to solve the problem of predicting three-dimensional surface deformation in mining areas.

[0004] This invention provides a method for predicting three-dimensional surface deformation in mining areas, comprising:

[0005] S1. Based on the mining subsidence model, obtain the three-dimensional deformation of any point on the surface during mining. Convert the three-dimensional deformation of any point on the surface into the LOS deformation phase. Generate a winding phase based on the LOS deformation phase. Add noise to the winding phase to obtain the simulated real interference phase. Obtain the simulated interferogram based on the simulated real interference phase. Divide the simulated interferogram into training set, validation set and test set.

[0006] S2. Calculate the InSAR phase derivative variance based on the entanglement phase, construct a high-resolution three-dimensional network, input the training set and the InSAR phase derivative variance into the high-resolution three-dimensional network, and use the combined loss function to iteratively train the high-resolution three-dimensional network. The combined loss function is the sum of the absolute phase loss function and the loss function between the network output three-dimensional deformation result and the mining subsidence model simulated three-dimensional deformation result.

[0007] S3. The high-resolution 3D network with the training completed is evaluated using the validation set, and the network with the highest evaluation score is tested using the test set.

[0008] S4. Obtain the actual interferogram of the mining area and input it into the high-resolution 3D network after the test to obtain the predicted 3D deformation.

[0009] Using the embodiments of the present invention, deformation solutions with good stability and high accuracy are obtained by combining InSAR data with deep learning methods.

[0010] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0011] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0012] Figure 1 This is a flowchart of the method for predicting three-dimensional surface deformation in mining areas by combining InSAR and neural networks according to an embodiment of the present invention.

[0013] Figure 2 This is a schematic diagram of the coal mine unit mining plane coordinate system and spatial coordinate system for the three-dimensional deformation prediction method of the surface in a mining area, which combines InSAR and neural networks according to an embodiment of the present invention. Figure 2 (a) is a schematic diagram of surface subsidence in the direction of unit mining. Figure 2 (b) is a schematic diagram of surface subsidence in the downdip direction of the unit mining area. Figure 2 (c) is a schematic diagram of the subsidence in the three-dimensional space of the surface under unit mining;

[0014] Figure 3 This is a schematic diagram of the high-resolution three-dimensional network framework of the method for predicting three-dimensional deformation of the surface in mining areas by combining InSAR and neural networks according to an embodiment of the present invention.

[0015] Figure 4 This is a detailed schematic diagram of the feature extraction module, feature fusion module, and upsampling module of the high-resolution 3D network in the combined InSAR and neural network method for predicting 3D surface deformation in mining areas, according to an embodiment of the present invention. Figure 4 (a) For the feature extraction module, Figure 4 (b) For the upsampling module, Figure 4 (c) is the feature fusion module. Detailed Implementation

[0016] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Example

[0018] According to embodiments of the present invention, a method for predicting three-dimensional surface deformation in mining areas by combining InSAR and neural networks is provided, such as... Figure 1 As shown, it specifically includes:

[0019] S1. Based on the mining subsidence model, obtain the three-dimensional deformation of any point on the surface during mining. Convert the three-dimensional deformation of any point on the surface into the LOS deformation phase. Generate a winding phase based on the LOS deformation phase. Add noise to the winding phase to obtain the simulated real interference phase. Obtain the simulated interferogram based on the simulated real interference phase. Divide the simulated interferogram into training set, validation set and test set.

[0020] S2. Calculate the InSAR phase derivative variance based on the entanglement phase, construct a high-resolution three-dimensional network, input the training set and the InSAR phase derivative variance into the high-resolution three-dimensional network, and use the combined loss function to iteratively train the high-resolution three-dimensional network. The combined loss function is the sum of the absolute phase loss function and the loss function between the network output three-dimensional deformation result and the mining subsidence model simulated three-dimensional deformation result.

[0021] S3. The high-resolution 3D network with the training completed is evaluated using the validation set, and the network with the highest evaluation score is tested using the test set.

[0022] S4. Obtain the actual interferogram of the mining area and input it into the high-resolution 3D network after the test to obtain the predicted 3D deformation.

[0023] The three-dimensional deformation of any point on the surface during mining is obtained based on the mining subsidence model.

[0024] In this embodiment of the invention, more than 250 sets of real mining parameters from mining areas in my country are used to simulate three-dimensional surface deformation, and the Stochastic Medium Theory Model (SMTM) is used as the mining subsidence model.

[0025] Figure 2 (a) is a schematic diagram of surface subsidence in the direction of unit mining. Figure 2 (b) is a schematic diagram of surface subsidence in the downdip direction of the unit mining area. Figure 2 (c) is a schematic diagram of the subsidence in the three-dimensional space of the surface under unit mining; such as Figure 2As shown in (a), underground coal seam units are located in two-dimensional space. When mining is carried out, the corresponding surface point The lateral vertical deformation is:

[0026] Formula 1;

[0027] In the formula, For surface points The lateral and vertical deformation, r is the coal seam unit being mined. The main radius of influence, H represents the mining depth. As the main influencing angle The tangent value, Let A be the x-coordinate of point A on the Earth's surface. For mining units The x-coordinate.

[0028] according to Figure 2 The geometric relationship shown in (c) is as follows, where The coordinates of point A on the ground can be obtained using the ground point. express, For surface points The coordinates of the corresponding mining unit B can be obtained using This indicates that, by extending Equation 1 to three-dimensional space, the coal seam unit can be obtained. Corresponding surface point Vertical deformation for:

[0029] Formula 2;

[0030] In Formula 2, t represents the mining unit. ordinate, Coal seam unit in three-dimensional space Corresponding surface point Vertical deformation.

[0031] To obtain the vertical deformation of the entire subsidence basin, simply integrate the subsidence caused by the mining of the coal seam unit as shown in Formula 2.

[0032] Formula 3;

[0033] In Formula 3, W(x,y) represents the vertical deformation of the entire subsidence basin. , Represents the thickness of the coal seam. The subsidence coefficient represents the coal seam mining process. This is the dip angle of the coal seam.

[0034] According to the SMTM theory, the horizontal movement caused by coal seam unit mining can be deduced as follows:

[0035] Formula 4;

[0036] In the formula, B is the horizontal movement coefficient.

[0037] Due to tilt It is vertical deformation The first derivative, therefore the surface point The inclination at that point can be expressed as:

[0038] Formula 5;

[0039] In Formula 5 Representing surface points The slope in the north-south direction, Representing surface points Inclination in the east-west direction, It is a surface point Vertical deformation, It is a surface point Vertical deformation, It is a surface point and adjacent points The distance in the east-west direction, It is a surface point and The distance in the north-south direction, therefore, combining formulas 3 and 5, and Figure 2 Surface points can be derived Horizontal movement in the north-south and east-west directions:

[0040] Formula 6;

[0041] In the above formula, Representing surface points The horizontal distance traveled in the north-south direction, Representing surface points B represents the horizontal distance traveled in the east-west direction, where B is the horizontal movement coefficient. The angle of influence of mining.

[0042] By deriving the above formula, we can obtain the location of any surface point during mining operations. The formula for calculating three-dimensional deformation. = For vertical deformation, = For east-west deformation, = It is a north-south oriented deformation.

[0043] Converting the 3D deformation of any point on the Earth's surface into the LOS-oriented deformation phase specifically includes:

[0044] The following formula is used to convert the three-dimensional deformation into the LOS-oriented deformation phase;

[0045] ;

[0046] Where λ is the radar wavelength. For vertical deformation, For east-west deformation, It is a north-south oriented deformation. The radar incident angle, The azimuth angle of the satellite during flight. For LOS-directed deformation phase.

[0047] In this embodiment of the invention, the radar incident angle is... Set to 34°, azimuth angle Set to 349°.

[0048] In this embodiment of the invention, the generation of the winding phase based on the LOS-directed deformation phase in S1 specifically includes: generating the winding phase by rewinding the LOS-directed deformation phase.

[0049] In this embodiment of the invention, S1, which describes adding noise to the entangled phase to obtain a simulated real interference phase, specifically includes: adding random complex Gaussian noise to simulate real noise, adding Berlin noise to simulate the real atmospheric phase, and adding water body noise to simulate the decoherent region to obtain the interference phase. The following formula is used for calculation:

[0050] ;

[0051] Where mod is the modulo operation, and n is the sum of complex Gaussian noise, Berlin noise and water body noise.

[0052] In this embodiment of the invention, 10,000 simulated interferograms are generated based on the simulated real interference phase, with a sample size of 128*128. The simulated interferograms are normalized, with 70% used as the training set, 20% as the validation set, and 10% as the test set.

[0053] In this embodiment of the invention, the calculation of the InSAR phase derivative variance based on the winding phase in S2 specifically includes:

[0054] The variance of the InSAR phase derivative is calculated using the following formula:

[0055] ;

[0056] in, Let k represent the variance of the InSAR phase derivative of the center pixel (m,n) of the window, where k is the window size. The center pixel of the window, For pixels within the window, and Let be the horizontal and vertical partial derivatives of the winding phase of the center pixel (m,n) of the window. and It is the average of the horizontal and vertical partial derivatives of the winding phase of all pixels within the window.

[0057] In this embodiment of the invention, the InSAR phase derivative variance (PDV) is... The quality of interferograms is evaluated by assessing the local statistical characteristics of the interferometric phase, thus determining the quality of the phase data. Compared to quality maps such as pseudo-correlation coefficient maps and maximum phase gradient maps (REF_Ref194328615), the InSAR phase derivative variance map is more reliable and accurate when accurate correlation coefficient maps are difficult to obtain. By introducing the InSAR phase derivative variance map, the network can focus more on regions of phase gradient abrupt changes during the learning process.

[0058] In this embodiment of the invention, the high-resolution three-dimensional network in S2 includes: a feature extraction module, a feature fusion module, and an upsampling module;

[0059] The feature extraction module is used to extract the phase gradient jump features in the input image and input them into the feature fusion module. The feature fusion module is used to obtain the weight information based on the phase gradient jump features and obtain the fused features based on the weight information. The fused features are then input into the upsampling module. After upsampling, the fused features are obtained into a fused feature map. The fused feature map and the phase gradient jump features are then stitched together to obtain the three-dimensional deformation.

[0060] like Figure 3 As shown in the diagram, the neural network framework used is based on the constructed HRNet-TD framework. In the network design diagram, the purple, green, and blue squares represent the Feature Extraction Module (FEM), Feature Fusion Module (FFM), and Upsample Module (UM), respectively, while the pink squares represent convolutional blocks. The internal framework of each module is shown below. Figure 4 (a) Figure 4 (b) Figure 4 As shown in (c).

[0061] like Figure 4As shown in (a), FEM adjusts the number of channels by using 1×1 convolutions on the input data, and then performs convolutions using four sets of convolution kernels of different sizes to extract phase gradient transition features at different scales under different receptive fields and concatenate them. For each set of convolution results, 1 / 4 of the feature map is randomly selected, and a squeeze and excitation (SE) attention mechanism is used to concatenate the four SE results to obtain attention weights for feature maps at different scales, which are then normalized, fusing global information at different scales and achieving better pixel-level attention. Residual connections are used to combine the original input with the ReLU activation result, which is then divided into three branches. Each branch undergoes different operations, and the results are concatenated and output using the Silu activation function.

[0062] like Figure 4 As shown in (b), after upsampling the input result, UM first performs bilinear interpolation and then convolutions of size 1, 3, and 5 respectively, and concatenates the three convolution results. After average pooling of the upsampled result, it is convolved with a 1×1 convolution. The convolution result is then used with the Sigmoid function to generate feature weights. The feature weights are multiplied with the upsampled result and concatenated with the convolution result in the channel dimension to obtain the output result of UM.

[0063] like Figure 4 As shown in (c), after the output of the upper encoder is convolved, regularized and activated by 3×3, the FFM compresses it to a size of 1×1 using adaptive average pooling and then passes it through two fully connected layers (FC) to integrate the local features extracted by the pooling layer. Finally, the Sigmoid function maps the features of each channel of the feature map to the range [0,1] and multiplies it with the feature map output by the lower encoder to enhance detail recovery.

[0064] The high-resolution 3D network incorporates phase derivative variance and interferograms as inputs. The interferogram guides the network to learn phase gradient abrupt change features in the interference fringes, while the phase derivative variance guides the network to focus on abrupt phase gradient change regions in the interferogram. A loss function is designed to constrain the network for the task of predicting 3D deformation. After learning by the FEM, UM, and FFM modules, the network feature outputs are integrated to obtain the 3D deformation prediction results of the mining area by HRNet-TD. The feature extraction module is applied to the encoder part to extract the latent phase gradient abrupt change features in the feature map. The skip connections of U-Net, which directly copy the feature map of each encoder layer to the upsampling part, may introduce some errors. Therefore, a feature fusion module is designed. For the feature map output of the upper encoder layer, weight information is generated by FFM and then applied to the upsampling module. Finally, the outputs of each encoder layer are upsampled to obtain feature maps, and the feature maps are integrated to obtain the final network output result.

[0065] The simulated interferogram and InSAR phase derivative variance are used as inputs to the high-resolution 3D network, and the 3D deformation is used as the output. The training set is input into the high-resolution 3D network, and the combined loss function is used to iteratively train the high-resolution 3D network. The combined loss function is the sum of the absolute phase loss function and the loss function between the network output 3D deformation result and the simulated 3D deformation result of the mining subsidence model.

[0066] The combined loss function is the sum of the absolute phase loss function and the loss function between the network output three-dimensional deformation result and the mining subsidence model simulated three-dimensional deformation result. Specifically, it includes:

[0067] Combined loss function ;

[0068] Loss function of absolute phase for:

[0069] ,in, Represents the combined loss function. The loss function representing the absolute phase. This represents the loss function between the network output of the three-dimensional deformation results and the simulation of the three-dimensional deformation results by the mining subsidence model. This represents the predicted gradient for the pixel with coordinates (i,j) in the image. A pixel is the smallest unit of analysis for 3D deformation data. Let be the ground truth gradient of the pixel with coordinates (i,j) in the image; where The coordinates in the image are ( The predicted absolute phase value, The coordinates in the image are ( The true absolute phase value of ) The coordinates in the image are The true absolute phase value;

[0070] ;

[0071] Let be the predicted absolute phase value of the leftmost pixel at coordinates (i,j) in the image. Let be the predicted absolute phase value of the rightmost pixel at coordinates (i,j) in the image. For the remaining cases.

[0072] ;

[0073] Let be the true absolute phase value of the leftmost pixel at coordinates (i,j) in the image.

[0074] The true absolute phase value of the rightmost pixel at coordinates (i,j) in the image.

[0075] For the remaining cases.

[0076] In their paper "A novel knowledge-learning coupling method for InSAR phase unwrapping of large surface displacements in coal mining areas," Chen et al. proposed a novel vertical deformation loss function based on the relationship between absolute phase and mining subsidence models. This loss function is used to compare the 3D deformation results output by the high-resolution 3D network with the 3D deformation results simulated by the mining subsidence model. ;

[0077] In the formula, N is the number of training sets. , , These are the vertical, east-west, and north-south deformation results output by the high-resolution 3D network, respectively. , , These represent the vertical, east-west, and north-south deformation phases simulated by the mining subsidence model. For interferograms, For the vertical deformation of a region, Here, F represents the weight parameters of a high-resolution 3D network, and F is the Frobenius norm, which represents the square root of the sum of the squares of all elements in the matrix.

[0078] In this embodiment of the invention, InSAR phase derivative variance information is used to guide HRNet-TD to focus on regions with abrupt phase gradient changes. The combined loss function is coupled with knowledge of mining mechanisms and InSAR phase gradients. Knowledge of mining mechanisms helps the network output results to better conform to the mining subsidence pattern, while InSAR phase gradient knowledge can reduce the network's sensitivity to absolute phase values ​​and reduce the impact of phase gradient abrupt changes, thereby improving the network's generalization performance.

[0079] HRNet-TD was implemented using Python 3.9 and PyTorch 1.13.0, and trained on an Intel(R) i7-12700KF (CPU), 32GB RAM, and an NVIDIA GeForce RTX4090 (GPU). The initial learning rate of the network was set to 1e. -4 The learning rate was dynamically adjusted using the cosine annealing algorithm. The total number of training rounds was 100, the batch size was set to 16, and the overall training time of the network was about 4 hours.

[0080] The high-resolution 3D network trained is evaluated using a validation set. In this embodiment of the invention, RMSE is used as the evaluation metric, and the network weights with the highest accuracy on the validation dataset are selected as the network with the highest score and saved.

[0081] The high-resolution 3D network with the highest evaluation score was tested using a test set;

[0082] Obtain an interferometric map of the actual mining area and input it into a high-resolution 3D network that has completed testing to obtain the predicted 3D deformation.

[0083] Experiment 1

[0084] Experiment 1 designed experiments to predict 3D deformation using MCF (Multi-Scale Context Fusion Module), QGPU (Quantization Graph Processing Module), PUNet (Pyramid U-Net), FCNet-CA (Fully Convolutional Network with Channel Attention), HRPNet (High-Resolution Pyramid Network), HRNet-TD (Temporal Modeling Module), and KLCNet (Knowledge Guided Constraint Network). MCF, QGPU, PUNet, FCNet-CA, HRPNet, and KLCNet ignored horizontal displacement and converted it into vertical deformation. The accuracy of the above methods is represented by RMSE, as shown in Table 1.

[0085] Table 1

[0086]

[0087] As can be seen, HRNet-TD achieves the highest accuracy, with an RMSE of only 9.22 mm. Among the other methods, HRPNet and KLCNet perform well, with RMSE results of 18.57 mm and 21.77 mm, respectively. HRNet-TD improves upon HRPNet and KLCNet by 50.33% and 57.63%, respectively. Since HRPNet and KLCNet both ignore horizontal movement and directly convert LOS to vertical deformation results, it is evident that when the deformation is large, ignoring the influence of horizontal movement will lead to a certain difference between the obtained deformation results and the measured results. HRNet-TD, however, uses a mining subsidence model to simulate 3D deformation as training samples, thus considering the influence of horizontal movement, resulting in higher deformation recovery accuracy.

[0088] Experiment 2

[0089] Experiment 2 selected interferograms with smaller deformation levels as the conditions for predicting 3D deformation using MCF, QGPU, PUNet, FCNet-CA, HRPNet, HRNet-TD, and KLCNet. The accuracy of these methods was represented by RMSE, as shown in Table 2.

[0090]

[0091] HRNet-TD achieved an RMSE of only 4.17, which is 53.2%, 49.6%, 48.9%, and 46.1% higher than PUNet, FCNet-CA, HRPNet, and KLCNet, respectively.

[0092] Experiment 3

[0093] Experiment 3 was designed to simulate an increasing decoherence region in the lower left corner of the mining area as mining progressed, leading to a corresponding increase in the variance of the InSAR phase derivative. Three-dimensional deformation was predicted using MCF, QGPU, PUNet, FCNet-CA, HRPNet, HRNet-TD, and KLCNet. The accuracy of these methods was represented by RMSE, as shown in Table 3.

[0094]

[0095] In this embodiment of the invention, as shown in Table 3, the RMSE of HRNet-TD is significantly lower than that of other methods, at only 5.77. Among the other methods, KLCNet and HRPNet show the best RMSE results, with KLCNet having an RMSE of 10.57 and HRPNet having an RMSE of 15.83. The RMSE result of HRNet-TD is 45.4% and 63.6% higher than that of KLCNet and HRPNet, respectively. This result also verifies that considering horizontal deformation in the deformation of the mining area significantly improves the accuracy of the final obtained vertical deformation.

[0096] Experiment 4

[0097] To test the sensitivity of HRNet-TD to different noise levels and deformation gradients, a mining subsidence model was used to generate a three-dimensional deformation sample set. The mining subsidence model parameters were set as follows: the subsidence coefficient increased in increments of 0.01 from 0.01 to 0.9; the tangent of the main influence angle was set to 2; the inflection point offset was set to 20 m; the coal seam dip angle was 5°; the mining depth was set to 420 m; and the mining thickness was 7 m. Satellite parameters were set as follows: wavelength 5.6 cm, incident angle 37°, and resolutions of 10 m and 20 m. Noise levels from 0 to 2π were added to the interferograms of different deformation gradients at intervals of 0.5π, resulting in a total of 900 simulated interferograms.

[0098] For the same subsidence coefficient, the maximum phase gradient is 15 rad at a resolution of 10 m, and reaches 30 rad at a resolution of 20 m.

[0099] At a resolution of 10 m, when the noise level reaches 1.5π, the RMSE results of HRNet-TD in the W, EW, and NS directions all begin to increase uncontrollably. However, when the noise level does not exceed π, the RMSE of the 3D deformation predicted by HRNet-TD is better, with the RMSE in the W direction being less than 0.5 rad and the RMSE in the EW and NS directions being less than 0.2 rad, indicating high accuracy. Even when the maximum phase gradient reaches 15 rad / pixel, HRNet-TD can still obtain high-accuracy 3D deformation prediction results.

[0100] When the resolution is 20 m, HRNet-TD can only maintain good RMSE results within the noise level of [0, π]. Once the noise level reaches 1.5π, the RMSE results of HRNet-TD begin to spread in the W, EW and NS directions and no longer have a regular pattern. When the noise level does not exceed π, even if the maximum phase gradient reaches 30 rad / pixel, the RMSE distribution of HRNet-TD is still very regular and only increases with the increase of the maximum phase gradient.

[0101] In summary, the analysis shows that, regardless of whether the resolution is 10 m or 20 m, when the noise level is between 0 and π, the 3D deformation prediction results of HRNet-TD are basically unaffected by noise. However, as the maximum phase gradient increases, the RMSE of the 3D deformation results of HRNet-TD gradually increases; the RMSE values ​​in the W, EW, and NS directions all exhibit discrete and jump-like phenomena. Therefore, HRNet-TD has good robustness when the noise level is below π and still maintains good performance even when the maximum phase gradient reaches 30 rad / pixel (at a resolution of 20 m).

[0102] Beneficial effects:

[0103] This invention uses a mining subsidence model to simulate three-dimensional deformation as training samples and takes into account the influence of horizontal movement, thus the deformation recovery accuracy of the network is higher.

[0104] By introducing the InSAR phase derivative variance map, the network can focus more on regions of phase gradient abrupt changes during the learning process.

[0105] An attention module is introduced to construct the network. The combined loss function is constructed using prior knowledge of phase gradient and knowledge of mining subsidence mechanism to constrain the network. The combined loss function is coupled with knowledge of mining mechanism and InSAR phase gradient. Knowledge of mining mechanism helps the network output results to better conform to the mining subsidence law, while InSAR phase gradient knowledge can reduce the network's sensitivity to absolute phase value and reduce the impact of phase gradient jump, thereby improving the network's generalization performance.

[0106] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions to the technical solutions of the embodiments of the present invention do not cause the essence of the corresponding technical solutions to deviate from the scope of the present solution.

Claims

1. A method for predicting three-dimensional surface deformation in mining areas by combining InSAR and neural networks, characterized in that, include, S1. Based on the mining subsidence model, obtain the three-dimensional deformation of any point on the surface during mining. Convert the three-dimensional deformation of any point on the surface into the LOS deformation phase. Generate a winding phase based on the LOS deformation phase. Add noise to the winding phase to obtain the simulated real interference phase. Obtain the simulated interferogram based on the simulated real interference phase. Divide the simulated interferogram into training set, validation set and test set. S2. Calculate the InSAR phase derivative variance based on the entanglement phase, construct a high-resolution three-dimensional network, input the training set and the InSAR phase derivative variance into the high-resolution three-dimensional network, and use the combined loss function to iteratively train the high-resolution three-dimensional network. The phase derivative variance is used to focus the network on the region of phase gradient jump during the learning process. The combined loss function is the sum of the absolute phase loss function and the loss function between the network output three-dimensional deformation result and the three-dimensional deformation result simulated by the mining subsidence model. S3. The high-resolution 3D network with the training completed is evaluated using the validation set, and the network with the highest evaluation score is tested using the test set. S4. Obtain the actual interferogram of the mining area and input it into the high-resolution 3D network after the test to obtain the predicted 3D deformation.

2. The method according to claim 1, characterized in that, The specific steps of converting three-dimensional deformation into LOS-oriented deformation phase include: The following formula is used to convert the three-dimensional deformation into the LOS-oriented deformation phase; ； in, Let LOS be the deformation phase, and λ be the radar wavelength. For vertical deformation, For east-west deformation, It is a north-south oriented deformation. The radar incident angle, This is the azimuth angle of the satellite during its flight.

3. The method according to claim 1, characterized in that, The generation of the winding phase based on the LOS-directed deformation phase specifically includes: generating the winding phase by rewinding the LOS-directed deformation phase.

4. The method according to claim 1, characterized in that, The process of adding noise to the entangled phase to obtain a simulated real interference phase specifically includes: adding random complex Gaussian noise to simulate real noise, adding Berlin noise to simulate the real atmospheric phase, and adding water body noise to simulate the decoherent region to obtain the interference phase.

5. The method according to claim 1, characterized in that, The calculation of the InSAR phase derivative variance based on the entangled phase specifically includes: The variance of the InSAR phase derivative is calculated using the following formula: ； in, Let k represent the variance of the InSAR phase derivative of the center pixel (m,n) of the window, where k is the window size. The center pixel of the window, For pixels within the window, and Let be the horizontal and vertical partial derivatives of the winding phase of the center pixel (m,n) of the window. and It is the average of the horizontal and vertical partial derivatives of the winding phase of all pixels within the window.

6. The method according to claim 1, characterized in that, The high-resolution 3D network includes: a feature extraction module, a feature fusion module, and an upsampling module; The feature extraction module is used to extract the phase gradient jump features in the variance of the InSAR phase derivative. The feature fusion module is used to obtain weight information based on the phase gradient jump features, obtain fused features based on the weight information, input the fused features into the upsampling module, obtain the fused feature map after upsampling the fused features, and stitch the fused feature map and the phase gradient jump features together to obtain the three-dimensional deformation.

7. The method according to claim 1, characterized in that, The combined loss function is the sum of the absolute phase loss function and the loss function between the network output three-dimensional deformation result and the mining subsidence model simulated three-dimensional deformation result. Specifically, it includes: ； ； in, Represents the combined loss function. The loss function representing the absolute phase. This represents the loss function between the network output of the three-dimensional deformation results and the simulation of the three-dimensional deformation results by the mining subsidence model. Let be the predicted gradient of the pixel with coordinates (i,j) in the image. Let be the ground truth gradient of the pixel with coordinates (i,j) in the image; where The coordinates in the image are ( The predicted absolute phase value, The coordinates in the image are ( The true absolute phase value of ) The coordinates in the image are The true absolute phase value; ； Let be the predicted absolute phase value of the leftmost pixel at coordinates (i,j) in the image. Let be the predicted absolute phase value of the rightmost pixel at coordinates (i,j) in the image. For the remaining cases; ； Let be the true absolute phase value of the leftmost pixel at coordinates (i,j) in the image. The true absolute phase value of the rightmost pixel at coordinates (i,j) in the image. For the remaining cases; The loss function between the high-resolution 3D network output 3D deformation results and the 3D deformation results simulated by the mining subsidence model ; In the formula, N is the number of training sets. , , These are the vertical, east-west, and north-south deformation results output by the high-resolution 3D network, respectively. , , These represent the vertical, east-west, and north-south deformation phases simulated by the mining subsidence model. For interferograms, For vertical deformation, Here, F represents the weight parameters of a high-resolution 3D network, and F is the Frobenius norm, which represents the square root of the sum of the squares of all elements in the matrix.

Images