Coal mining subsidence InSAR phase filtering method and model based on deep learning

Abstract

The invention discloses a coal mining subsidence InSAR phase filtering method and model based on deep learning, and belongs to the field of phase filtering. According to the method, an MCR-PFNet model is constructed, phase filtering is carried out for the noise problem in coal mining subsidence InSAR data, a multi-head self-attention module and a CBAM attention module are introduced into coding and decoding paths through the MCR-PFNet model, the capturing capability of space and deformation characteristics is enhanced, the convolution structure of the network is optimized through a residual block, and the method is suitable for the real-time detection of coal mining subsidence InSAR data. And high precision and robustness of a filtering result are ensured. Through verification of multiple groups of simulation and actual data, the method is superior to a traditional method in phase filtering precision and noise elimination effect, and surface deformation information can be effectively recovered. The method has the advantages of simple steps and high efficiency, is suitable for processing the InSAR data of the coal mining subsidence area, and meets the requirements of the fields of surface deformation monitoring, disaster prevention, resource management and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of interferometric phase filtering, and specifically to a coal mining subsidence InSAR phase filtering method and model based on deep learning. Technical Background

[0002] Interferometric Synthetic Aperture Radar (InSAR) technology has been widely used in the monitoring of surface deformation, such as earthquakes, volcanoes, glacier movements, and subsidence caused by mining activities, due to its high-precision, all-day, and all-weather monitoring capabilities. Currently, when using InSAR technology to monitor coal mining subsidence, in practical applications, due to the complex geological structure of underground mining areas and the dynamics of coal mining operations, the surface deformation in coal mining subsidence areas has characteristics such as non-linearity and local mutations. Moreover, the phase data collected by InSAR is usually affected by factors such as noise, interference fringes, and terrain effects, resulting in a decrease in its accuracy. Especially in the phase unwrapping link, noise will make the unwrapping accuracy unstable, thereby affecting the accuracy of subsidence monitoring results.

[0003] Summarizing the prior art, the research on InSAR data filtering mainly focuses on the suppression of phase noise and the improvement of unwrapping accuracy. Traditional phase filtering methods, such as Goldstein filtering, Lee filtering, etc., rely on local statistical characteristics. Although they can smooth phase noise to a certain extent, they are prone to losing phase information when dealing with non-uniform regions. In addition, most existing deep learning models lack sufficient consideration of the periodicity and local mutations of phase data and cannot effectively handle the phase noise problem in the complex scenario of coal mining subsidence.

[0004] The prior art CN115015929A provides an efficient and high-precision InSAR phase filtering network (SMD-Net) driven by a sparse model. It first establishes a sparse reconstruction model for interferometric phase filtering; then, by expanding the iterative shrinkage threshold algorithm (ISTA) for solving the interferometric phase filtering model into a block, SMD-Net is designed as an iterative-based convolutional neural network (CNN) architecture. In each block, a CNN module with a local block and a global context (GC) block is established to adaptively learn the sparse transform domain. SMD-Net models the interferometric phase filtering process. Instead of completely relying on data fitting like most networks, its network structure is simple. It has a large computational amount, a high training cost, and a high requirement for the number of training sets.

[0005] The prior art CN113920025A discloses an InSAR phase filtering method based on DWT. First, a three-level wavelet transform is performed on the complex interference phase. A mask image is generated for the wavelet sub-domains after the three-level wavelet transform using a set threshold. The real and imaginary parts of the coefficients detected as signals in the mask image are each multiplied by 2, and the noise coefficients remain unchanged. Then, an inverse discrete wavelet transform of the first level is performed, and a mask growth rule is designed to update the mask image. Finally, by calculating the phase value of the complex data, the filtered phase value is obtained. It only performs phase filtering on the adaptive mask image extracted by coherence during the inverse wavelet transform process, only retains the details of the provisions well, and has poor processing effects on large-area phase data. Summary of the Invention

[0006] In view of the above problems, the present invention proposes a deep learning-based InSAR phase filtering method and model for coal mining subsidence. The method is based on the UNet model and combines various advanced neural network structures such as a convolutional block attention module, a multi-head self-attention mechanism, and a residual convolutional block. Through a custom periodic phase loss function, precise suppression of phase noise in the coal mining subsidence area and high-precision restoration of phase information are achieved. The present invention aims to solve the problem of insufficient accuracy of existing InSAR filtering methods in complex subsidence environments and provides an innovative solution for the efficient monitoring of surface subsidence in coal mining areas.

[0007] To achieve the above technical objectives, the present invention discloses a deep learning-based InSAR phase filtering method for coal mining subsidence, including the following steps:

[0008] By processing the digital elevation model data of the coal mining area surface to generate various phases, and then generating noiseless wrapped phase data and noisy wrapped phase data in a simulation method;

[0009] In the way of pairing a noisy wrapped phase data with a corresponding noiseless wrapped phase data to construct a set of training data samples, all the noisy wrapped phase data and the corresponding noiseless wrapped phase data are used to construct a training data set;

[0010] Based on the U-Net network, a residual attention phase filtering network MCR-PFNet is optimally designed for the InSAR phase filtering task.

[0011] The MCR-PFNet network includes a downsampling module, a multi-head self-attention module, and an upsampling module arranged in sequence. The model structures of the downsampling module and the upsampling module are the same, and the downsampling module and the upsampling module are connected by an improved skip connection module.

[0012] Input the training dataset into the MCR-PFNet network and complete the training. Use multiple groups of noisy wrapped phase image data and corresponding noise-free wrapped phase image data as the input data for the training of the MCR-PFNet network to complete the training of the MCR-PFNet model;

[0013] Use the trained MCR-PFNet model to perform InSAR phase filtering on the noisy phase image of the coal mining area to be filtered, and obtain the corresponding noise-free wrapped phase image result.

[0014] Furthermore, process the digital elevation model data of the coal mining area surface: Obtain the digital elevation model data DEM of the coal mining area surface, crop the DEM data of the coal mining subsidence area into multiple small blocks of 128×128 pixel size, analyze the simulated terrain phase of the DEM data of the coal mining subsidence area, and obtain the influence of the terrain of the coal mining subsidence area on the phase by converting parameters such as the incident angle and wavelength of the DEM data. Generate the orbital phase based on the grid generation method to simulate the influence of satellite orbital errors on the phase; Use the Perlin noise generation technology to generate the atmospheric phase to simulate the interference of atmospheric turbulence on the atmospheric phase over the coal mining subsidence area; Simulate the surface deformation phase caused by coal mining subsidence by using a two-dimensional Gaussian distribution and random parameters to control the position, amplitude, width, and rotation angle of the subsidence deformation, and then generate noise-free wrapped phase data and noisy wrapped phase data: Randomly superimpose the generated surface deformation phase, atmospheric phase, terrain phase, and satellite orbital phase to obtain a large amount of noise-free wrapped phase data; Add phase jumps to 30% of the total number of noise-free wrapped phase data, and add Gaussian noise to all noise-free wrapped phase data to obtain the corresponding noisy wrapped phase data.

[0015] Furthermore, the specific structures of each module in the MCR-PFNet network are as follows:

[0016] The downsampling module includes an input convolutional block, the first downsampling layer, the second downsampling layer, the third downsampling layer, and the fourth downsampling layer connected in sequence; the structures of the first downsampling layer, the second downsampling layer, the third downsampling layer, and the fourth downsampling layer are the same, and each includes max pooling, a residual convolutional module, a convolutional block attention module CBAM, and an output arranged in sequence. Among them, each downsampling layer contains a residual convolutional module, and each residual convolutional module includes two layers of convolution, and each layer of convolution is followed by a batch normalization layer BN. Between the BN layer of the previous layer of convolution and the second layer of convolution, a ReLU activation function and a Dropout layer are sequentially provided; in addition, in the residual convolutional module, after the input features are processed by two layers of convolution, they are then added to the initial input features whose number of channels is adjusted by 1×1 convolution through a skip connection to complete feature fusion; the CBAM module contains channel attention and spatial attention mechanisms, which respectively adaptively adjust the weights of each channel and spatial position in the feature map; by introducing the residual convolutional module and the CBAM module in the downsampling stage, the residual convolutional module is used to capture local features, and combined with the CBAM module to adjust the weights of channel and spatial features, the attention ability of the MCR-PFNet network to important features is enhanced, and the parsing ability of complex phase signals is significantly improved;

[0017] The upsampling module has the same symmetrical structure as the downsampling module. It includes the first upsampling layer, the second upsampling layer, the third upsampling layer, and the fourth upsampling layer connected in sequence. Each of them includes transposed convolution, a residual convolutional module, a CBAM attention module, and an output arranged in sequence; each upsampling layer first performs upsampling through transposed convolution, and then passes through a residual convolutional module, which contains two layers of convolution, a BN layer, and a ReLU activation function; the residual convolutional module is connected to a CBAM module, which includes two parts: channel attention and spatial attention, and finally outputs the feature map processed by the CBAM module; in the upsampling stage, the MCR-PFNet network fuses the multi-scale features extracted in the downsampling stage with the upsampling features through skip connections, and at the same time embeds a residual convolutional module and a CBAM module in each upsampling block to ensure that more phase edge information can be retained during the process of restoring the resolution, and further reduce the loss of phase information;

[0018] The improved skip connection module includes max pooling, a residual convolutional module, a CBAM module, and a multi-head self-attention module connected in sequence; the max pooling of the improved skip connection module is connected to the CBAM module of the fourth downsampling layer, and the multi-head self-attention module of the improved skip connection module is connected to the transposed convolution of the first upsampling layer; through the multi-head self-attention mechanism, the long-range dependence relationship between different phase signals can be modeled globally, effectively improving the filtering performance in high-noise and complex terrain environments;

[0019] Furthermore, the downsampling module specifically includes: an input convolutional block: the input is the phase image data with a size of 128×128. First, it passes through a convolutional layer and batch normalization processing, and outputs a feature map with a size of 128×128; the first layer of downsampling: the size of the feature map is reduced to 64×64 through a max-pooling layer, and then passes through a residual block and a CBAM module to output a new feature map; the second layer, the third layer, and the fourth layer of downsampling all go through the same operations in sequence, with the size halved each time and the number of channels increasing layer by layer, and finally a feature map of 8×8 is obtained.

[0020] 5. The InSAR phase filtering method for coal mining subsidence based on deep learning according to claim 3, wherein the working process of the multi-head self-attention module is as follows: in the downsampling module, the processed feature map data is fed into the multi-head self-attention module to capture the global features and long-range dependencies of the input noisy wrapped phase; the multi-head self-attention module decomposes the information in the feature map into multiple self-attention heads, and each head adjusts the weights of each position in the feature map by calculating the similarity between features, and then globally processes the input feature map; after passing through the multi-head self-attention module, the global features of the feature map are extracted and further used in the upsampling stage; furthermore, a dynamic scaling strategy based on phase gradient is introduced for the long-range dependence characteristics of the fringe structure in InSAR data, and the attention weight allocation mechanism is adjusted, so that the module can more accurately focus on the fringe area while suppressing the interference of high-noise areas; the number of multi-heads is adjusted to 4, reducing redundant calculations while ensuring that the features with long-range correlations in the wrapped phase can be captured; the input embedding dimension of the self-attention module is optimized, and combined with the distribution characteristics of the wrapped phase, the global feature modeling ability is improved.

[0021] Furthermore, the specific steps of the upsampling module and the skip connection are as follows: during the process of passing the downsampled low-resolution features to the upsampling through the skip connection, the downsampled low-resolution features are first processed by a residual module and a CBAM module, and then passed to the upsampling. The residual module ensures the continuity of the phase features and the retention of phase edge information, and the CBAM module adaptively strengthens the key features and suppresses redundant information through channel and spatial attention mechanisms, significantly improving the adaptability of the model in the InSAR phase filtering task for coal mining subsidence and effectively solving the problem of information redundancy in traditional skip connections;

[0022] The first layer of upsampling: the size of the feature map is restored to 16×16 through a transposed convolution, and then processed by a residual block and a CBAM module; a skip connection is made with the feature map in the downsampling process; the second layer of upsampling, the third layer of upsampling, and the fourth layer of upsampling perform similar operations in sequence, and finally the size is restored to the input size of 128×128;

[0023] The skip connection plays a role in maintaining information flow in the network, enabling deep features to be directly combined with shallow features and enhancing the feature expression ability; during the above upsampling process, it can be observed that the feature maps obtained from the previous downsampling are concatenated with the upsampled feature maps; these connection processes are the realization of the skip connection;

[0024] Output convolution: After completing all upsampling steps, a 1×1 convolution kernel is used to reduce the number of channels to the target number of channels, and finally a filtered phase diagram with a size of 128×128 is output.

[0025] Furthermore, the constructed MCR-PFNet model is trained using training data, and the trained MCR-PFNet model is used for phase filtering of simulated data and real data in coal mining subsidence areas;

[0026] The noisy wrapped phase image data is used as the input data of the MCR-PFNet model, and the corresponding noise-free wrapped phase image data is used as the output data of the MCR-PFNet model, that is, the label data, to complete the training of the MCR-PFNet model; where the training uses a custom periodic loss function, and its calculation formula is as follows:

[0027]

[0028] where, L phase is the loss term, N is the number of samples in each batch participating in model training, Y i is the target label of the i-th training sample, is the output of the model for the i-th sample; this loss function calculates the cosine value of the phase difference, maintains the periodic characteristics of the phase, and avoids errors caused by phase jumps.

[0029] A model for phase filtering of InSAR images, the network is a residual attention phase filtering network MCR-PFNet, and the MCR-PFNet network includes a sequentially arranged downsampling module, a multi-head self-attention module, and an upsampling module. The model structures of the downsampling module and the upsampling module are the same, and the downsampling module and the upsampling module are connected by an improved skip connection module;

[0030] The downsampling module includes an input convolution block, the first downsampling layer, the second downsampling layer, the third downsampling layer, and the fourth downsampling layer connected in sequence; the structures of the first downsampling layer, the second downsampling layer, the third downsampling layer, and the fourth downsampling layer are the same, and each includes max pooling, a residual convolution module, a convolutional block attention module CBAM, and an output arranged in sequence. Among them, each downsampling layer contains a residual convolution module, and each residual convolution module includes two layers of convolution, and each layer of convolution is followed by a batch normalization layer BN. Between the BN layer of the previous layer of convolution and the second layer of convolution, a ReLU activation function and a Dropout layer are sequentially provided; in addition, in the residual convolution module, after the input features are processed by two layers of convolution, they are then added to the initial input features whose number of channels is adjusted by 1×1 convolution through a skip connection to complete feature fusion; the CBAM module contains channel attention and spatial attention mechanisms, which adaptively adjust the weights of each channel and spatial position in the feature map respectively; by introducing the residual convolution module and the CBAM module in the downsampling stage, the residual convolution module is used to capture local features, and combined with the CBAM module to adjust the weights of channel and spatial features, the attention ability of the MCR-PFNet network to important features is enhanced, and the parsing ability of complex phase signals is significantly improved;

[0031] The upsampling module has the same symmetric structure as the downsampling module, and includes the first upsampling layer, the second upsampling layer, the third upsampling layer, and the fourth upsampling layer connected in sequence. Each includes transposed convolution, a residual convolution module, a CBAM attention module, and an output arranged in sequence; each upsampling layer first performs upsampling through transposed convolution, and then passes through a residual convolution module, which contains two layers of convolution, a BN layer, and a ReLU activation function; the residual convolution module is connected to the CBAM module, which contains two parts: channel attention and spatial attention, and finally outputs the feature map processed by the CBAM module; in the upsampling stage, the MCR-PFNet network fuses the multi-scale features extracted in the downsampling stage with the upsampling features through skip connections, and at the same time embeds a residual convolution module and a CBAM module in each upsampling block to ensure that more phase edge information can be retained during the process of restoring the resolution, and further reduce the loss of phase information;

[0032] The improved skip connection module includes max pooling, a residual convolution module, a CBAM module, and a multi-head self-attention module connected in sequence; the max pooling of the improved skip connection module is connected to the CBAM module of the fourth downsampling layer, and the multi-head self-attention module of the improved skip connection module is connected to the transposed convolution of the first upsampling layer; through the multi-head self-attention mechanism, the long-range dependence relationship between different phase signals can be modeled globally, effectively improving the filtering performance in high-noise and complex terrain environments.

[0033] Furthermore, in the downsampling module: Input convolutional block: The input is the phase image data with a size of 128×128. First, it passes through a convolutional layer and batch normalization, and outputs a feature map with a size of 128×128. First-layer downsampling: The size of the feature map is reduced to 64×64 through a max-pooling layer, and then through a residual block and a CBAM module, a new feature map is output. The second, third, and fourth-layer downsamplings all go through the same operations in sequence, with the size halved each time and the number of channels increasing layer by layer, and finally a feature map of 8×8 is obtained.

[0034] A computer device includes a processor and a memory. The processor is electrically connected to the memory. The memory is used to store instructions and data, and the processor is used to execute the InSAR phase filtering method for coal mining subsidence based on deep learning.

[0035] Beneficial effects: First, through the MCR-PFNet model, the ability to capture complex deformation and noise features in the InSAR phase diagram is enhanced. Compared with the traditional U-Net model, the MCR-PFNet introduces a multi-head self-attention mechanism (Multi-Head Self Attention) in the encoding and decoding paths, improving the model's ability to capture long-range dependencies and subtle features in the phase diagram, and effectively improving the accuracy of InSAR phase filtering in coal mining subsidence areas. Second, the MCR-PFNet model optimizes the model structure by introducing a residual convolutional block (Residual Block) and a CBAM module. The residual module improves the training stability of the model. The CBAM module enables the model to better select and extract key features through channel and spatial attention mechanisms, effectively enhancing the phase filtering effect. In addition, by customizing a periodic loss function, the model fully considers the periodic characteristics of phase data, further improving the accuracy and robustness of the denoising results. Third, the MCR-PFNet model provided by the present invention can adapt to different phase noise scenarios and exhibits excellent performance in deformation monitoring of coal mining subsidence areas, and can effectively remove noise and restore surface deformation information. Through the above improvements, the MCR-PFNet model provided by the present invention performs excellently in the phase filtering task and can meet the processing requirements of InSAR phase data under complex terrains and different noise conditions.

[0036] This method comprehensively considers various complex factors, including the effects of additive Gaussian noise, local phase jumps, orbital errors, and atmospheric turbulence, and at the same time combines a two-dimensional Gaussian distribution to simulate the surface deformation characteristics caused by coal mining subsidence. This dataset generation strategy significantly enhances the generalization ability of the model, enabling it to maintain a high filtering accuracy even in an environment with complex stripes and strong noise interference. Description of the Drawings

[0037] Figure 1Schematic flow chart of the InSAR phase filtering method for coal mining subsidence based on deep learning according to the present invention;

[0038] Figure 2 Network architecture diagram of MCR-PFNet recorded in the embodiments of the present invention;

[0039] Figure 3 Schematic diagrams of each module of the MCR-PFNet network recorded in the embodiments of the present invention;

[0040] Figure 4 Schematic diagram of the training data set of the MCR-PFNet model in the embodiments of the present invention;

[0041] Figure 5 Schematic diagram of the filtering result and accuracy analysis of the MCR-PFNet model under simulated data in the embodiments of the present invention;

[0042] Figure 6 Result diagrams of filtering simulated data using Lee, Goldstein, MSFF-DCNN, and MCR-PFNet respectively in the embodiments of the present invention;

[0043] Figure 7 Result diagram of the accuracy analysis of the filtering of simulated data by Lee, Goldstein, MSFF-DCNN, and MCR-PFNet in the embodiments of the present invention;

[0044] Figure 8 Result diagram of phase filtering of a mining face in a certain mining area of Datong Coalfield, Shanxi Province, China by the Lee, Goldstein, MSFF-DCNN, and MCR-PFNet models in the embodiments of the present invention. Among them, (a) is a schematic diagram of the filtering results of the four methods and the result diagram of unwrapping using the SNAPHU algorithm; (b) is a schematic diagram of the change of phase values at each point of the profile line AA'. Detailed implementation manners

[0045] The present invention will be further described below in conjunction with the accompanying drawings and specific implementation cases.

[0046] As Figure 1 shown, an InSAR phase filtering method for coal mining subsidence based on deep learning obtains the information of the surface of the coal mining area through phase filtering, so as to obtain the noise-free surface information of the coal mining area, providing a basis for accurately analyzing whether there is coal mining subsidence caused by coal mining in the working face. The method includes the following steps:

[0047] Use InSAR technology to obtain DEM data of the coal mining area, crop it to a size of 128×128, simulate the topographic phase based on the selected DEM data, and obtain the influence of topography on the phase. Generate the orbital phase based on the grid generation method to simulate the influence of satellite orbital errors on the phase. Use the Perlin noise generation technique to generate the atmospheric phase and simulate the interference of atmospheric turbulence on the phase. Generate complex deformation phases by using two-dimensional Gaussian distribution and random parameters to control the position, amplitude, width, and rotation angle of the subsidence deformation, and simulate the surface deformation caused by coal mining subsidence;

[0048] Superimpose the deformation phase, atmospheric phase, topographic phase, and orbital phase to obtain noiseless wrapped phase data; add phase jumps to 30% of the total number of noiseless wrapped phase data, and add Gaussian noise to all noiseless wrapped phase data to obtain the corresponding noisy wrapped phase data;

[0049] Construct a training dataset from the noisy wrapped phase data and the corresponding noiseless wrapped phase data;

[0050] Based on the U-Net network, design the InSAR phase filtering model MCR-PFNet for coal mining subsidence; the MCR-PFNet model includes a downsampling module, a multi-head self-attention module, and an upsampling module arranged in sequence. The downsampling module and the upsampling module are connected by an improved skip connection module; the downsampling module includes an input convolutional block, the first layer of downsampling, the second layer of downsampling, the third layer of downsampling, and the fourth layer of downsampling connected in sequence. The structures of the four layers of downsampling are the same, and each includes a max pooling, a residual convolutional module, a CBAM module, and an output arranged in sequence. Among them, each residual convolutional module consists of two convolutional layers, and each convolution is followed by batch normalization (BatchNorm2d, BN) and the ReLU activation function. In the residual convolutional module, after the input features are processed by two convolutional layers, they are added to the initial input through a skip connection. The CBAM attention module contains channel attention and spatial attention mechanisms, which adaptively adjust the weights of each channel and spatial position in the feature map respectively;

[0051] The upsampling module includes the first layer of upsampling, the second layer of upsampling, the third layer of upsampling, and the fourth layer of upsampling connected in sequence. The structures of the four layers of upsampling are the same, and each includes a transposed convolution, a residual convolutional module, a CBAM module, and an output arranged in sequence. Among them, each layer of upsampling first performs upsampling through a transposed convolution, and then passes through a residual convolutional module, which contains two convolutional layers, BN, and the ReLU activation function. The residual convolutional module is followed by a CBAM module, which contains two parts: channel attention and spatial attention. Finally, the output is the feature map processed by the CBAM attention module;

[0052] The downsampling module includes the following: Input convolution block: The input is the phase image data with a size of 128×128. First, it passes through a convolutional layer and batch normalization, and outputs a feature map with a size of 128×128; First - layer downsampling: The size of the feature map is reduced to 64×64 through a max - pooling layer, and then passes through a residual block and a CBAM module, outputting a new feature map; The second - layer, third - layer, and fourth - layer downsamplings all go through the same operations in sequence. The size is halved each time, and the number of channels increases layer by layer, finally obtaining a feature map of 8×8.

[0053] The described multi - head self - attention module is as follows: After downsampling is completed, the feature map enters the multi - head self - attention module. This module is used to capture global features and long - range dependencies. The multi - head self - attention module decomposes the information in the feature map into multiple self - attention heads. Each head adjusts the weights of each position in the feature map by calculating the similarity between features, and then globally processes the input feature map. After passing through the multi - head self - attention module, the global features of the feature map are extracted and further used in the upsampling stage.

[0054] The specific steps of upsampling and skip connection are as follows: First - layer upsampling: The size of the feature map is restored to 16×16 through transposed convolution, and then processed through a residual block and a CBAM module; It is skip - connected with the feature map in the downsampling process; The second - layer, third - layer, and fourth - layer upsamplings perform similar operations in sequence, and finally restore to the input size of 128×128.

[0055] The skip connection plays a role in maintaining information flow in the network, enabling deep features to be directly combined with shallow features, enhancing the feature expression ability; In the above upsampling process, it can be observed that the feature maps obtained from the previous downsampling are concatenated with the upsampled feature maps; These connection processes are the realization of the skip connection;

[0056] Output convolution: After completing all upsampling steps, it passes through a 1×1 convolutional kernel, which is used to reduce the number of channels to the target number of channels, and finally outputs a filtered phase map with a size of 128×128;

[0057] The MCR - PFNet model draws on the basic architecture of the UNet model, adopts the design of upsampling, downsampling, and skip connection, but it has made original improvements for the InSAR phase filtering task in coal - mining subsidence areas, enabling the model to more efficiently cope with the challenges of complex noise environments and significant phase gradient changes:

[0058] In the downsampling module, each layer integrates a Residual Block and a Convolutional Block Attention Module (CBAM). The Residual Block effectively alleviates the vanishing gradient problem in deep networks through skip connections, while enhancing the ability to extract local features, retaining more phase detail information. Especially in the complex noise coal mining subsidence scenario, it can significantly improve the model's processing effect on stripe details. The CBAM module adaptively adjusts the weights of each channel and each position in the feature map through channel attention and spatial attention mechanisms, thereby enhancing the attention to the target area and suppressing background noise. The embedding of this module enables the model to more accurately retain the phase information of the target area when processing high-noise interference phases. In the bottleneck layer of the network (where downsampling ends and upsampling begins), MCR-PFNet introduces a multi-head self-attention mechanism. By using the multi-head self-attention mechanism to dynamically model global features, it further enhances the global perception ability of complex phase features. The multi-head self-attention mechanism can capture long-range dependencies in phase features, which is particularly important for processing scenarios with dense stripes and sudden gradient changes in coal mining subsidence areas. In such scenarios, the model not only needs to suppress noise but also ensure the continuity of the stripe structure, and the introduction of the multi-head self-attention mechanism precisely meets this requirement.

[0059] MCR-PFNet also optimizes the feature fusion strategy in the upsampling stage. During each upsampling process, the MCR-PFNet model concatenates the high-resolution features extracted in the downsampling stage with the current decoding layer features through skip connections. This feature fusion method can effectively combine local details with global features, avoiding the problem of detail loss that may occur in the pure upsampling process. Especially in areas with large gradient changes, the feature fusion of skip connections significantly improves the phase filtering accuracy after filtering.

[0060] The constructed training dataset is input into the MCR-PFNet model for training. During the training process of the MCR-PFNet model, the loss function is a key factor guiding the optimization of model parameters, directly affecting the learning efficiency and final performance of the model. Especially in the InSAR phase filtering task, the selection of the loss function needs to fully consider the periodic characteristics and noise complexity of the phase data to avoid the deficiencies of traditional loss functions in dealing with phase jumps and wrapping problems. This application proposes a custom periodic loss function, which is specifically optimized for the InSAR phase filtering task;

[0061] Specifically, the noisy wrapped phase image data is used as the input data of the MCR-PFNet model, and the corresponding noiseless wrapped phase image data is used as the output data of the MCR-PFNet model, that is, the label data, to train the network. The model training uses a custom periodic loss function, and its calculation formula is as follows:

[0062]

[0063] where L phase is the loss term, N is the number of samples in each batch participating in model training, Y i is the target label of the i-th training sample, is the output of the i-th sample model. This loss function calculates the cosine value of the phase difference, maintains the periodic characteristics of the phase, and avoids errors caused by phase jumps.

[0064] Use the trained MCR-PFNet model to perform phase filtering on the simulated data and the real data in the coal mining subsidence area;

[0065] Input the noisy wrapped phase data to be filtered into the trained MCR-PFNet model, and according to the prediction of the trained MCR-PFNet model, obtain the corresponding noiseless wrapped phase result, that is, complete the phase filtering process.

[0066] Example 1

[0067] 1. Construction of the training dataset

[0068] Randomly obtain some DEM data in China (the acquisition website is: https: / / earthexplorer.usgs.gov / ), crop it to a size of 128×128, simulate the terrain phase based on the selected DEM data, and obtain the influence of the terrain on the phase. Generate the orbital phase based on the grid generation method to simulate the influence of satellite orbital errors on the phase. Use the Perlin noise generation technology to generate the atmospheric phase to simulate the interference of atmospheric turbulence on the phase. Generate complex deformation phases by using a two-dimensional Gaussian distribution and random parameters to control the position, amplitude, width, and rotation angle of the settlement deformation, and simulate the surface deformation caused by coal mining subsidence;

[0069] Superimpose the deformation phase, atmospheric phase, terrain phase, and orbital phase to obtain noiseless wrapped phase data; Add phase jumps to 30% of the total number of noiseless wrapped phase data, and add Gaussian noise to the noiseless wrapped phase data to obtain the corresponding noisy wrapped phase data; Combine the noisy wrapped phase data and the corresponding noiseless wrapped phase data into a sample, and repeat the operation to obtain 30,000 training samples, thereby obtaining the training dataset for the MCR-UNet model, as Figure 4 shown;

[0070] 2. Construct the InSAR phase filtering model for coal mining subsidence, MCR-PFNet;

[0071] As Figure 2 and Figure 3 shown, the MCR-PFNet model includes downsampling, a multi-head self-attention module, and upsampling arranged in sequence. The downsampling and upsampling are connected by an improved skip connection module;

[0072] The downsampling module includes an input convolution block, the first layer of downsampling, the second layer of downsampling, the third layer of downsampling, and the fourth layer of downsampling connected in sequence. The structures of the four layers of downsampling are the same, and each includes max pooling, a residual convolution module, a CBAM attention module, and an output arranged in sequence. Among them, each residual convolution module consists of two convolutional layers, and after each convolution, batch normalization (BatchNorm2d, BN) and the ReLU activation function are connected. In the residual convolution module, after the input features are processed by two convolutional layers, they are added to the initial input through a skip connection. The CBAM attention module contains channel attention and spatial attention mechanisms, which adaptively adjust the weights of each channel and spatial position in the feature map respectively;

[0073] The upsampling includes the first layer of upsampling, the second layer of upsampling, the third layer of upsampling, and the fourth layer of upsampling connected in sequence. The structures of the four layers of upsampling are the same, and each includes a transposed convolution, a residual convolution module, a CBAM attention module, and an output arranged in sequence. Among them, each layer of upsampling first performs upsampling through a transposed convolution, and then passes through a residual convolution module, which contains two convolutional layers, BN, and the ReLU activation function. The residual convolution module is connected to a CBAM attention module, which contains two parts: channel attention and spatial attention. Finally, the output is the feature map processed by the CBAM attention module;

[0074] The downsampling module includes the following steps:

[0075] Input convolution block: The input phase image data with a size of 128×128 first passes through a convolutional layer and batch normalization, and outputs a feature map with a size of 128×128;

[0076] The first layer of downsampling: Reduces the size of the feature map to 64×64 through a max pooling layer, and then passes through a residual block and a CBAM module to output a new feature map. The second layer, the third layer, and the fourth layer of downsampling all go through the same operations in sequence, with the size halved each time and the number of channels increasing layer by layer, and finally a feature map of 8×8 is obtained.

[0077] Furthermore, the multi-head self-attention module is as follows: After downsampling is completed, the feature map enters the multi-head self-attention module. This module is used to capture global features and long-range dependencies. The multi-head self-attention module decomposes the information in the feature map into multiple self-attention heads. Each head adjusts the weights of each position in the feature map by calculating the similarity between features, and then globally processes the input feature map. After passing through the multi-head self-attention module, the global features of the feature map are extracted and further used in the upsampling stage.

[0078] The specific steps of upsampling and skip connection are as follows:

[0079] The first layer of upsampling: The size of the feature map is restored to 16×16 through transposed convolution, and then processed by a residual block and a CBAM module; a skip connection is made with the feature map in the downsampling process; similar operations are sequentially performed on the second, third, and fourth layers of upsampling, and finally restored to the input size of 128×128.

[0080] The skip connection plays a role in maintaining information flow in the network, enabling deep features to be directly combined with shallow features, enhancing the feature expression ability; in the above upsampling process, it can be observed that the feature map obtained from the previous downsampling is concatenated with the upsampled feature map; these connection processes are the realization of the skip connection;

[0081] Output convolution: After completing all upsampling steps, a 1×1 convolution kernel is used to reduce the number of channels to the target number of channels, and finally a filtered phase map with a size of 128×128 is output;

[0082] 3. Train the constructed MCR-PFNet model using training data

[0083] Use the mainstream deep learning framework pytorch 1.12.1 to train the above MCR-PFNet network. The computer software and hardware configuration are shown in Table 1. As attached Figure 2 As shown, the Adam optimizer is used for network training, a custom periodic loss function is selected, the initial learning rate is 0.00001, the batch size batch is 32, and the number of training epochs Epoch is 100. The noisy wrapped phase image data is used as the input data of the MCR-PFNet model, and the corresponding noise-free wrapped phase image data is used as the output data of the MCR-PFNet model, that is, the label data, to train the network; among them, the custom periodic loss function is used for model training, and its calculation formula is as follows:

[0084]

[0085] Among them, L phaseis the loss term, N is the number of samples in each batch participating in model training, and Y i is the target label of the i-th training sample, is the output of the i-th sample model. This loss function calculates the cosine value of the phase difference, maintains the periodic characteristics of the phase, and avoids errors caused by phase jumps.

[0086] Table 1 Computer software and hardware configuration

[0087]

[0088] 4. Input the noisy phase image to be filtered into the trained MCR-PFNet model to complete the phase filtering process

[0089] The simulated data phase filtering performed in this embodiment is as Figure 5 shown. Through Figure 6 and Figure 7 shown in the comparison diagram of simulated data phase filtering, the phase filtering and accuracy analysis of the surface subsidence image of a mining area in a certain coal mine field in Datong, Shanxi Province, China are as Figure 8 shown;

[0090] As Figure 5 shown, four groups of noisy wrapped phases are respectively selected and input into the MCR-PFNet model, and the corresponding phase filtering results can be obtained to complete the phase filtering task. It can be seen from Figure 5 that for sample data 1, the PSNR between the wrapped phase result predicted by MCR-PFNet and the noiseless wrapped phase data is 64.004dB, the SSIM is 0.983, and the MSE is 0.026; for sample data 2, the PSNR between the wrapped phase result predicted by MCR-PFNet and the noiseless wrapped phase data is 60.023dB, the SSIM is 0.964, and the MSE is 0.062; for sample data 3, the PSNR between the wrapped phase result predicted by MCR-PFNet and the noiseless wrapped phase data is 58.714dB, the SSIM is 0.964, and the MSE is 0.087; for sample data 4, the PSNR between the wrapped phase result predicted by MCR-PFNet and the noiseless wrapped phase data is 70.528dB, the SSIM is 0.988, and the MSE is 0.006. This accuracy can meet the needs of the vast majority of scientific research and practical applications.

[0091] As Figure 6 and Figure 7As shown in the figure, the Lee, Goldstein, MSFF-DCNN, and MCR-PFNet models were used to perform phase filtering on the simulated data, and the differences between the filtering results of each method and the true results were statistically calculated. It can be seen that the PSNR between the filtering results of Lee and the true data was 40.932 - 63.291 dB, and the SSIM value was 0.158 - 0.961; the PSNR between the filtering results of Goldstein and the true data was 46.825 - 59.372 dB, and the SSIM value was 0.586 - 0.892; the PSNR between the filtering results of MSFF-DCNN and the true data was 46.629 - 57.571 dB, and the SSIM value was 0.831 - 0.959; the PSNR between the filtering results of MCR-PFNet and the true data was 55.820 - 64.017 dB, and the SSIM value was 0.965 - 0.990; MCR-PFNet had a higher phase filtering accuracy;

[0092] As Figure 8 shown, the InSAR interference results of the surface of a mining area in the Datong Coalfield, Shanxi Province, China, which involve subsidence topography, were respectively phase-filtered using Lee, Goldstein, MSFF-DCNN, and MCR-PFNet, and SNAPHU was used to unwrap the filtering results of the four methods, and the profile diagram of AA' was drawn. From Figure 8 (a), it can be seen that the filtering result of MCR-PFNet contains less noise and has the best unwrapping effect; from Figure 8 (b), it can be seen that there are fluctuations of varying degrees in the phase values on the unwrapping result profiles after filtering by Lee, Goldstein, and MSFF-DCNN, while the phase changes on the unwrapping result profile after filtering by MCR-PFNet are stable and conform to the law of mining subsidence. Based on the above analysis, the filtering results of MCR-PFNet can meet the needs of the vast majority of scientific research and practical applications.

[0093] The present invention provides a deep learning-based InSAR phase filtering method for coal mining subsidence, which uses the original MCR-PFNet model for prediction. The MCR-PFNet model introduces a multi-head self-attention module and a CBAM attention module in the encoding and decoding paths, enhancing the ability to capture spatial and deformation features in InSAR phase data. At the same time, a residual block is introduced into the network structure to optimize the design of the convolutional layer, ensuring the high accuracy and robustness of the filtering results.

[0094] After multiple verifications and comparisons, the MCR-PFNet model has shown excellent performance in the prediction of InSAR phase diagrams in simulated data and coal mining subsidence areas. The accurate phase filtering results are of great value in fields such as surface deformation monitoring, disaster prevention, and resource management. The present invention not only improves the accuracy and robustness of InSAR phase filtering but also ensures the periodic characteristics of phase data through a customized periodic loss function, avoiding errors caused by phase jumps in traditional methods and further enhancing the stability of the filtering results.

[0095] The method of the present invention is easy to operate and has high computational efficiency. It is applicable to InSAR phase filtering tasks in different scenarios and can provide accurate surface deformation analysis for related fields. By evaluating the phase filtering results of the model, the surface deformation information after noise elimination can be clearly seen, such as Figure 8 (a) shows that the filtered phase image can accurately recover the unwrapped phase information. The present invention combines deep learning and InSAR technology and is applicable to phase filtering tasks in coal mining subsidence areas of various scales. Compared with traditional methods, it is more efficient and accurate. The attention mechanism and residual structure design of the MCR-PFNet model not only ensure sufficient feature extraction but also effectively reduce the computational complexity of the model, making it suitable for large-scale phase data processing.

Claims

1. A deep learning-based InSAR phase filtering method for coal mining subsidence, characterized in that The steps are as follows: By processing the digital elevation model data of the surface of the coal mining area, multiple phases are generated, and then the multiple phases are used to generate noiseless wrapped phase data and noisy wrapped phase data by simulation; In the way of constructing a set of training data samples by pairing a noisy wrapped phase data with a corresponding noiseless wrapped phase data, all the noisy wrapped phase data and the corresponding noiseless wrapped phase data are used to construct a training data set; Based on the U-Net network, a residual attention phase filtering network MCR-PFNet is optimally designed for the InSAR phase filtering task; The MCR-PFNet network includes a downsampling module, a multi-head self-attention module and an upsampling module which are arranged in sequence. The model structures of the downsampling module and the upsampling module are the same, and the downsampling module and the upsampling module are connected by an improved skip connection module; The training data set is input into the MCR-PFNet network and the training is completed. Multiple groups of noisy wrapped phase image data and the corresponding noiseless wrapped phase image data are used as the input data for the training of the MCR-PFNet network to complete the training of the MCR-PFNet model; The trained MCR-PFNet model is used to perform InSAR phase filtering on the noisy phase image of the coal mining area to be filtered, and the corresponding noiseless wrapped phase image result is obtained.

2. The InSAR phase filtering method for coal mining subsidence based on deep learning according to claim 1, wherein Processing the digital elevation model data of the surface of the coal mining area: Obtain the digital elevation model data DEM of the surface of the coal mining area, crop the DEM data of the coal mining subsidence area into multiple small blocks of 128×128 pixel size, analyze the simulated terrain phase of the DEM data of the coal mining subsidence area, and obtain the influence of the terrain of the coal mining subsidence area on the phase by converting parameters such as the incident angle and wavelength of the DEM data. Generate the orbital phase based on the grid generation method to simulate the influence of satellite orbital error on the phase; Use the Perlin noise generation technology to generate the atmospheric phase to simulate the interference of atmospheric turbulence on the atmospheric phase over the coal mining subsidence area; By using the two-dimensional Gaussian distribution and random parameters to control the position, amplitude, width and rotation angle of the subsidence deformation, simulate the surface deformation phase caused by coal mining subsidence, and then generate noiseless wrapped phase data and noisy wrapped phase data: Randomly superimpose the generated surface deformation phase, atmospheric phase, terrain phase and satellite orbital phase to obtain a large number of noiseless wrapped phase data; Add phase jumps to 30% of the total number of noiseless wrapped phase data, and add Gaussian noise to all the noiseless wrapped phase data to obtain the corresponding noisy wrapped phase data.

3. The InSAR phase filtering method for coal mining subsidence based on deep learning according to claim 1, characterized in that, The specific structures of each module in the MCR-PFNet network are as follows: The downsampling module includes an input convolution block, the first-layer downsampling, the second-layer downsampling, the third-layer downsampling, and the fourth-layer downsampling connected in sequence; the structures of the first-layer downsampling, the second-layer downsampling, the third-layer downsampling, and the fourth-layer downsampling are the same, and each includes max pooling, a residual convolution module, a convolutional block attention module CBAM, and an output arranged in sequence. Among them, each layer of downsampling contains a residual convolution module, and each residual convolution module includes two layers of convolution. After each layer of convolution, a batch normalization layer BN is connected. Between the BN layer of the previous layer of convolution and the second layer of convolution, a ReLU activation function and a Dropout layer are sequentially arranged; in addition, in the residual convolution module, after the input features are processed by two layers of convolution, they are then added to the initial input features whose number of channels is adjusted by 1×1 convolution through a skip connection to complete feature fusion; the CBAM module contains channel attention and spatial attention mechanisms, which adaptively adjust the weights of each channel and spatial position in the feature map respectively; by introducing the residual convolution module and the CBAM module in the downsampling stage, the local features are captured by the residual convolution module, and the weights of the channel and spatial features are adjusted in combination with the CBAM module, which enhances the MCR-PFNet network's ability to focus on important features and significantly improves the ability to analyze complex phase signals; The upsampling module has the same symmetric structure as the downsampling module. It includes the first-layer upsampling, the second-layer upsampling, the third-layer upsampling, and the fourth-layer upsampling connected in sequence, and each includes transposed convolution, a residual convolution module, a CBAM attention module, and an output arranged in sequence; each layer of upsampling first performs upsampling through transposed convolution, and then passes through a residual convolution module, which contains two layers of convolution, a BN layer, and a ReLU activation function; the residual convolution module is connected to a CBAM module, which contains two parts: channel attention and spatial attention. Finally, the output is the feature map processed by the CBAM module; in the upsampling stage, the MCR-PFNet network fuses the multi-scale features extracted in the downsampling stage with the upsampling features through skip connections, and at the same time embeds a residual convolution module and a CBAM module in each upsampling block to ensure that more phase edge information can be retained during the process of restoring the resolution, and further reduce the loss of phase information; The improved skip connection module includes max pooling, a residual convolution module, a CBAM module, and a multi-head self-attention module connected in sequence; the max pooling of the improved skip connection module is connected to the CBAM module of the fourth-layer downsampling, and the multi-head self-attention module of the improved skip connection module is connected to the transposed convolution of the first-layer upsampling; through the multi-head self-attention mechanism, the long-range dependencies between different phase signals can be modeled globally, effectively improving the filtering performance in high-noise and complex terrain environments.

4. The InSAR phase filtering method for coal mining subsidence based on deep learning according to claim 3, characterized in that The downsampling module specifically includes: Input convolution block: The input is phase image data with a size of 128×128. First, it passes through a convolution layer and batch normalization processing, and outputs a feature map with a size of 128×128; The first layer of downsampling: The size of the feature map is reduced to 64×64 through the max pooling layer, and then a new feature map is output through the residual block and the CBAM module. The second, third, and fourth layers of downsampling all go through the same operations in sequence. The size is halved each time, and the number of channels increases layer by layer, finally obtaining an 8×8 feature map.

5. The InSAR phase filtering method for coal mining subsidence based on deep learning according to claim 3, characterized in that, The working process of the multi-head self-attention module is as follows: In the downsampling module, the processed feature map data is fed into the multi-head self-attention module to capture the noisy wrapped phase global features and long-range dependencies in the input; the multi-head self-attention module decomposes the information in the feature map into multiple self-attention heads. Each head adjusts the weights of each position in the feature map by calculating the similarity between features, and then globally processes the input feature map; after passing through the multi-head self-attention module, the global features of the feature map are extracted and further used in the upsampling stage; then, a dynamic scaling strategy based on the phase gradient is introduced for the long-range dependence characteristics of the fringe structure in InSAR data, adjusting the attention weight allocation mechanism, enabling the module to more accurately focus on the fringe area while suppressing the interference in the high-noise area; the number of multi-heads is adjusted to 4, reducing redundant calculations while ensuring the ability to capture features with long-range correlations in the wrapped phase; the input embedding dimension of the self-attention module is optimized, and combined with the distribution characteristics of the wrapped phase, the global feature modeling ability is improved.

6. The InSAR phase filtering method for coal mining subsidence based on deep learning according to claim 3, wherein, The specific steps of the upsampling module and the skip connection are as follows: During the process of passing the low-resolution features of downsampling to upsampling through the skip connection, the low-resolution features of downsampling are first processed by the residual module and the CBAM module, and then passed to upsampling. The residual module ensures the continuity of the phase features and the retention of phase edge information. The CBAM module adaptively strengthens key features and suppresses redundant information through channel and spatial attention mechanisms, significantly improving the adaptability of the model in the InSAR phase filtering task of coal mining subsidence and effectively solving the problem of information redundancy in traditional skip connections; The first layer of upsampling: The size of the feature map is restored to 16×16 through transposed convolution, and then processed by the residual block and the CBAM module; It is skip-connected with the feature map in the downsampling process. Similar operations are performed successively for the second, third, and fourth layers of upsampling, and finally restored to the input size of 128×128; The skip connection plays a role in maintaining information flow in the network, enabling deep features to be directly combined with shallow features and enhancing the feature expression ability; during the above upsampling process, it can be observed that the feature map obtained from the previous downsampling is concatenated with the upsampled feature map; these connection processes are the realization of the skip connection; Output convolution: After completing all upsampling steps, it passes through a 1×1 convolution kernel to reduce the number of channels to the target number of channels, and finally outputs a filtered phase map with a size of 128×128.

7. The InSAR phase filtering method for coal mining subsidence based on deep learning according to claim 3, characterized in that The constructed MCR-PFNet model is trained using the training data, and the trained MCR-PFNet model is used for phase filtering of simulated data and real data in the coal mining subsidence area; The noisy wrapped phase image data is used as the input data of the MCR-PFNet model, and the corresponding noise-free wrapped phase image data is used as the output data of the MCR-PFNet model, that is, the label data, to complete the training of the MCR-PFNet model; among them, the training uses a custom periodic loss function, and its calculation formula is as follows: Among them, L phase is the loss term, N is the number of samples in each batch participating in model training, and Y i is the target label of the i-th training sample, is the output of the i-th sample model; this loss function calculates the cosine value of the phase difference, maintains the periodic characteristics of the phase, and avoids errors caused by phase jumps.

8. A model used for phase filtering of InSAR images, characterized in that: The network is the Residual Attention Phase Filtering Network MCR-PFNet. The MCR-PFNet network includes a downsampling module, a multi-head self-attention module, and an upsampling module arranged in sequence. The model structures of the downsampling module and the upsampling module are the same, and the downsampling module and the upsampling module are connected by an improved skip connection module; The downsampling module includes an input convolution block, the first layer of downsampling, the second layer of downsampling, the third layer of downsampling, and the fourth layer of downsampling connected in sequence; the structures of the first layer of downsampling, the second layer of downsampling, the third layer of downsampling, and the fourth layer of downsampling are the same, and all include max pooling, a residual convolution module, a Convolutional Block Attention Module CBAM, and an output arranged in sequence. Among them, each layer of downsampling contains a residual convolution module, and each residual convolution module includes two layers of convolution. After each layer of convolution, a batch normalization layer BN is connected. Between the BN layer of the previous layer of convolution and the second layer of convolution, a ReLU activation function and a Dropout layer are sequentially arranged; in addition, in the residual convolution module, after the input features are processed by two layers of convolution, they are then added to the initial input features whose number of channels is adjusted by a 1×1 convolution through a skip connection to complete feature fusion; the CBAM module contains channel attention and spatial attention mechanisms, which adaptively adjust the weights of each channel and spatial position in the feature map; by introducing the residual convolution module and the CBAM module in the downsampling stage, the residual convolution module is used to capture local features, and combined with the CBAM module to adjust the weights of channel and spatial features, the attention ability of the MCR-PFNet network to important features is enhanced, and the parsing ability of complex phase signals is significantly improved; The upsampling module has the same symmetric structure as the downsampling module. The first layer of upsampling, the second layer of upsampling, the third layer of upsampling, and the fourth layer of upsampling connected in sequence all include transposed convolution, a residual convolution module, a CBAM attention module, and an output arranged in sequence; each layer of upsampling first performs upsampling through transposed convolution, and then passes through a residual convolution module, which contains two layers of convolution, a BN layer, and a ReLU activation function; the residual convolution module is connected to a CBAM module, which contains two parts: channel attention and spatial attention, and finally outputs the feature map processed by the CBAM module; in the upsampling stage, the MCR-PFNet network fuses the multi-scale features extracted in the downsampling stage with the upsampling features through a skip connection, and at the same time embeds a residual convolution module and a CBAM module in each upsampling block to ensure that more phase edge information can be retained during the process of restoring the resolution, and further reduce the loss of phase information; The improved skip connection module includes a max pooling, a residual convolution module, a CBAM module, and a multi-head self-attention module connected in sequence; the max pooling of the improved skip connection module is connected to the CBAM module of the fourth-layer downsampling, and the multi-head self-attention module of the improved skip connection module is connected to the transposed convolution of the first-layer upsampling; through the multi-head self-attention mechanism, the long-range dependencies between different-phase signals can be modeled globally, effectively improving the filtering performance in high-noise and complex terrain environments.

9. The model used for InSAR phase filtering according to claim 8, wherein, In the downsampling module: Input convolution block: The input is the phase image data with a size of 128×128. First, it passes through a convolutional layer and batch normalization, and outputs a feature map with a size of 128×128. First-layer downsampling: The size of the feature map is reduced to 64×64 through a max pooling layer, and then passes through a residual block and a CBAM module to output a new feature map; the second-layer, third-layer, and fourth-layer downsamplings all go through the same operations in sequence, with the size halved each time and the number of channels increasing layer by layer, finally obtaining a feature map of 8×8.

10. A computer device, characterized in that, It includes a processor and a memory. The processor is electrically connected to the memory. The memory is used to store instructions and data, and the processor is used to execute the InSAR phase filtering method for coal mining subsidence based on deep learning according to any one of claims 1-7.

Images