A ground subsidence monitoring method and system fusing InSAR and neural network

Abstract

The application provides a ground collapse monitoring method and system fusing InSAR and a neural network, comprising the following steps: acquiring multi-time SAR images of a monitoring area; generating initial deformation fields and coherence coefficient maps of each time through differential interference processing; calculating local neighborhood structure tensors of each pixel according to the initial deformation fields; extracting the maximum eigenvalue as a deformation gradient value to generate a deformation gradient map; determining a coherence feature statistical space window size and calculating a kurtosis of the coherence coefficient; combining the kurtosis and the coherence coefficient value to form a multi-dimensional coherence feature; correcting the initial deformation field; stacking the multi-time corrected deformation field, the multi-dimensional coherence feature and the deformation gradient map along a time dimension to form a space-time feature tensor; inputting the space-time feature tensor into a convolutional neural network containing a parallel double branch and a channel attention module; and outputting a ground collapse probability of each pixel.

Description

Technical Field

[0001] This application belongs to the field of monitoring, and in particular relates to a ground subsidence monitoring method and system that integrates InSAR and neural networks. Background Technology

[0002] Synthetic Aperture Radar Interferometry (InSAR) technology, especially temporal InSAR, can extract millimeter-scale long-term deformation fields from multi-temporal SAR images. In temporal InSAR processing, the coherence coefficient is an indicator for evaluating the quality of interferograms and the reliability of deformation results, used to select high-quality pixels to ensure the accuracy of deformation monitoring. However, existing InSAR-based ground subsidence monitoring methods are prone to decoherence in areas where subsidence has occurred or is about to occur, where the surface deformation gradient is large. Traditional methods analyze based on deformation rate or cumulative deformation, ignoring the spatiotemporal information of the coherence coefficient. Furthermore, in identifying subsidence areas, methods based on setting deformation thresholds or manual visual interpretation are highly subjective and inefficient. Existing models based on machine learning or deep learning methods treat deformation and coherence as different features, failing to distinguish and weight the contributions of different features to subsidence identification, and failing to fully integrate the spatiotemporal variation characteristics of the subsidence process, thus limiting the reliability of monitoring. Summary of the Invention

[0003] This invention proposes a ground subsidence monitoring method that integrates InSAR and neural networks to address the problem of inaccurate ground subsidence monitoring in existing technologies, including: Acquire multi-temporal SAR images of the monitoring area, perform differential interferometry processing, and generate initial deformation field and coherence coefficient maps for each temporal sequence; based on the initial deformation field of each temporal sequence, calculate the structure tensor in the local neighborhood of each pixel, and extract the maximum eigenvalue as the deformation gradient value to generate deformation gradient maps for each temporal sequence. The size of the spatial window used for statistical coherence features is determined based on the deformation gradient value of each pixel; within the spatial window, the kurtosis of the coherence coefficient is calculated, and the kurtosis is combined with the coherence coefficient value to form the multidimensional coherence features of each time series. The initial deformation field is modified to obtain a modified deformation field. When the deformation gradient value and kurtosis value of a pixel are greater than a preset first threshold and a second threshold, respectively, the deformation value of the pixel is replaced by spatial interpolation using the deformation values ​​of other pixels in the neighborhood of the pixel that do not exceed the corresponding threshold. The modified deformation field, the multidimensional coherence feature, and the deformation gradient map of multiple time series are stacked along the time dimension to form a spatiotemporal feature tensor. The spatiotemporal feature tensor is input into a convolutional neural network, which outputs the ground collapse probability of each pixel. The convolutional neural network includes a parallel dual-branch and a channel attention module. The output of the first convolutional branch is used to generate channel attention weights to weight the feature map output by the second convolutional branch.

[0004] Furthermore, this invention also relates to a ground subsidence monitoring system that integrates InSAR and neural networks, comprising the following modules: The generation module is used to acquire multi-temporal SAR images of the monitoring area, perform differential interferometry processing, and generate initial deformation field and coherence coefficient maps for each temporal sequence; based on the initial deformation field of each temporal sequence, the structure tensor in the local neighborhood of each pixel is calculated, and the maximum eigenvalue is extracted as the deformation gradient value to generate deformation gradient maps for each temporal sequence. The first component module is used to determine the size of the spatial window for statistical coherence features based on the deformation gradient value of each pixel; within the spatial window, the kurtosis of the coherence coefficient is calculated, and the kurtosis is combined with the coherence coefficient value to form the multidimensional coherence features of each time series. The second component module is used to correct the initial deformation field to obtain a corrected deformation field. When the deformation gradient value and kurtosis value of a pixel are greater than a preset first threshold and a second threshold, respectively, the deformation values ​​of other pixels in the neighborhood of the pixel that do not exceed the corresponding thresholds are used to replace the deformation value of the pixel through spatial interpolation. The corrected deformation field, the multidimensional coherence feature, and the deformation gradient map of multiple time series are stacked along the time dimension to form a spatiotemporal feature tensor. The weighting module is used to input the spatiotemporal feature tensor into the convolutional neural network and output the ground collapse probability of each pixel. The convolutional neural network includes a parallel dual-branch and a channel attention module. The output of the first convolutional branch is used to generate channel attention weights to weight the feature map output by the second convolutional branch.

[0005] This invention constructs a multidimensional coherence feature capable of revealing ground instability anomalies by utilizing physical quantities such as deformation gradient and coherence kurtosis depth mining. By establishing the correlation between deformation gradient and analysis scale, coherence features of different deformation regions can be statistically analyzed. The initial deformation field is corrected by combining deformation gradient and kurtosis indices, suppressing noise and artifacts in high-risk collapse areas and obtaining reliable deformation input. The parallel convolutional neural network employing a channel attention mechanism can analyze spatiotemporal feature tensors that integrate multi-source information, learn and enhance the weights of feature channels, thereby improving the reliability of identification and monitoring. Attached Figure Description

[0006] Figure 1 A flowchart of the first embodiment; Figure 2 A schematic diagram illustrating the determination of window size based on deformation gradient; Figure 3 This is a schematic diagram of a parallel bi-branch convolutional network with an attention mechanism. Detailed Implementation

[0007] The features and exemplary embodiments of various aspects of this application will be described in detail below. To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only intended to explain this application and not to limit it. For those skilled in the art, this application can be implemented without some of these specific details. The following description of the embodiments is merely to provide a better understanding of this application by illustrating examples.

[0008] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.

[0009] In the first embodiment, the present invention proposes a ground subsidence monitoring method that integrates InSAR and neural networks, such as... Figure 1 ,include: S1. Acquire multi-temporal SAR images of the monitoring area, perform differential interferometry processing, and generate initial deformation field and coherence coefficient maps for each temporal sequence; based on the initial deformation field of each temporal sequence, calculate the structure tensor in the local neighborhood of each pixel, and extract the maximum eigenvalue as the deformation gradient value to generate deformation gradient maps for each temporal sequence. At least 20 Sentinel-1A single-look complex image data points from the same orbit and imaging mode covering the monitoring area were selected and processed using short baseline set temporal InSAR technology. The image with the center of the spatiotemporal baseline was selected as the master image, and the remaining images were registered to the master image coordinate system. Spatiotemporal baseline thresholds were set, for example, a temporal baseline less than 90 days and a spatial baseline less than 40% of the critical baseline, to generate an interferogram network. Topographic phase was removed using an external digital elevation model such as SRTM DEM to generate differential interferograms. The differential interferograms were filtered, and phase unwrapping was performed using the minimum cost flow method to obtain the unwrapped phase of each interferogram. The surface deformation of each time series relative to the initial time was obtained through singular value decomposition, which is the initial deformation field. Simultaneously, the coherence coefficient map corresponding to each time series was calculated during the processing.

[0010] For an initial deformation field at a specific time series, a local neighborhood is defined around each pixel, for example, a 5×5 pixel window. The first-order partial derivatives of the deformation values ​​of all pixels within this neighborhood are calculated in the horizontal and vertical directions, denoted as... and Based on the partial derivative values, a 2×2 structure tensor matrix J is constructed for the center pixel, where the matrix elements... For all within the neighborhood The sum of squares For all within the neighborhood The sum of squares and For all within the neighborhood and The sum of the products. Calculate the two eigenvalues ​​of the structure tensor matrix and select the largest eigenvalue as the deformation gradient value of the center pixel. Repeat this process for each pixel in the initial deformation field to generate the deformation gradient map for this time series.

[0011] In an optional embodiment, calculating the structure tensor in the local neighborhood of each pixel based on the initial deformation field at each time sequence includes: Obtain the partial derivatives of deformation for each pixel in the x and y directions; Within a neighborhood window centered on the pixel, the structure tensor is calculated by summing the second moments of the partial derivatives of the deformation.

[0012] Specifically, the input deformation field data is processed to obtain the deformation rate of each pixel in the horizontal x-direction and vertical y-direction, i.e., the deformation partial derivative. This can be achieved by applying operators such as the Sobel operator or similar gradient operators. For example, for the value of a pixel in the deformation field, the partial derivative value of that pixel is obtained by performing a convolution operation with the values ​​of its neighboring pixels. and The scalar deformation field is transformed into a gradient field with two channels, representing the deformation gradients in the x and y directions, respectively.

[0013] Define a local neighborhood window for each pixel, such as a 5×5 square region. Calculate the structure tensor S within this window. The structure tensor is a 2×2 symmetric matrix constructed by integrating or summing the second moments of the partial derivatives of the deformations of all pixels within the window. The formula for calculating this tensor is as follows: ; The W symbol under the summation symbol represents the operation within the entire neighborhood window. The diagonal elements of the matrix... and These represent the energies of the horizontal and vertical gradients within the window, respectively; off-diagonal elements. This reflects the correlation between the two directional gradients. The eigenvalues ​​of this tensor reflect the consistency and directionality of the deformation pattern in the local region.

[0014] S2, determine the size of the spatial window used for statistical coherence features based on the deformation gradient value of each pixel; within the spatial window, calculate the kurtosis of the coherence coefficient, and combine the kurtosis with the coherence coefficient value to form the multidimensional coherence features of each time series. Set a maximum window size, such as 15×15 pixels, and a minimum window size, such as 3×3 pixels. Normalize all gradient values ​​in the deformation gradient map so that the gradient values ​​fall between 0 and 1. For any pixel, calculate the normalized deformation gradient value for that pixel. The window size W is determined using a linear inverse proportional relationship. For example, ,in and These are the side lengths of the maximum and minimum windows, respectively. The calculation results must be rounded to an odd number. For pixels with larger deformation gradient values, the corresponding calculation window size is smaller; conversely, for pixels with smaller deformation gradient values, the corresponding window size is larger.

[0015] For a given pixel, a spatial window defined above is used to extract the coherence coefficient values ​​of all pixels within that window. These coherence coefficient values ​​are then used as a sample set, and the kurtosis value of this sample set is calculated. The kurtosis is calculated as the fourth central moment of the sample set divided by the square of the variance. After calculation, the pixel's own coherence coefficient value and the calculated kurtosis value are combined into a two-dimensional vector, i.e., [coherence coefficient, kurtosis]. This process is repeated for all pixels to generate a two-channel feature map, which represents the multidimensional coherence features of this time series.

[0016] To adjust the size of the spatial window used for calculating coherence based on the complexity of surface deformation features, in an optional embodiment, determining the size of the spatial window used for statistical coherence features based on the deformation gradient values ​​of each pixel includes: Through formula Calculate the side length W of the spatial window, where The deformation gradient value of the pixel is given by K, which is a preset scaling factor; and the minimum and maximum values ​​of the side length of the spatial window are defined.

[0017] The minimum side length of the spatial window is defined as 3, and the maximum value is 15; in regions of severe deformation, the deformation gradient value... A larger window indicates rich detail; in this case, a smaller window is needed to detect local changes and avoid over-smoothing. In areas of gentle deformation, When the value is small, using a larger window can include more samples, suppress noise, and improve the stability of the estimate.

[0018] This process is performed on each pixel i in the image, calculating the deformation gradient value at that point. Substitute the values ​​into the specified formula to calculate the window side length W. Here, K is a preset scaling factor used to adjust the window size's sensitivity to gradient values; for example, K can be set to 20. Assume the deformation gradient value of a pixel... If the value is 8, then the window side length is calculated as W=5. For a pixel in another region with gentle deformation, the pixel gradient value... If the value is 1.5, then W = 27. Boundary constraints are applied to the calculated W value. Since 27 exceeds the maximum value of 15, the window side length at that point is set to 15. If the calculated result is less than 3, then the minimum value of 3 is taken. Figure 2 .

[0019] In an optional embodiment, calculating the kurtosis of the coherence coefficient within the spatial window includes: If a pixel's adaptively calculated spatial window has a side length of 5 (i.e., a 5×5 pixel window) and contains 25 coherence coefficient samples, the first step is to extract the coherence coefficient sample set of all pixels within that window. First, through the formula Calculate the sample mean; then calculate the fourth central moment. and variance Through the kurtosis formula The kurtosis value of the coherence coefficient within the window is calculated. If the window is the smallest 3×3, then the sample set is first subjected to 3×3. The above calculations are performed after outliers are removed to reduce noise interference. Kurtosis indicates the steepness of the coherence coefficient distribution within the window.

[0020] In an optional embodiment, combining the kurtosis and coherence coefficient values ​​to form multidimensional coherence features for each time series includes: The coherence coefficient and kurtosis value of each pixel are combined to form a two-dimensional feature vector.

[0021] Specifically, after calculating the coherence coefficient and kurtosis of a specific time-series image, two two-dimensional data maps are obtained: a coherence coefficient map and a kurtosis map. These two maps have the same spatial dimension. Each pixel value on the coherence coefficient map represents the consistency of the deformation pattern at that point, typically ranging from 0 to 1; while each pixel value on the kurtosis map represents the sharpness of the coherence coefficient distribution within the neighborhood of that point.

[0022] To integrate the two complementary information sets, the data from the two images are combined pixel by pixel. For any coordinate position (i,j) in the image, the coherence coefficient value corresponding to that coordinate is extracted. and kurtosis value For example, a pixel might have a coherence coefficient of 0.85 and a kurtosis of 4.2. These two scalar values ​​are combined into a two-dimensional feature vector [0.85, 4.2]. This process is repeated for all pixels, fusing two H×W single-channel feature maps into a single H×W×2 multidimensional feature tensor. In this tensor, each spatial location is represented by a two-dimensional vector, comprehensively representing the surface stability and deformation characteristics of that point.

[0023] S3, the initial deformation field is corrected to obtain a corrected deformation field, wherein when the deformation gradient value and kurtosis value of a pixel are greater than a preset first threshold and a second threshold, respectively, the deformation value of the pixel is replaced by spatial interpolation using the deformation values ​​of other pixels in the neighborhood of the pixel that do not exceed the corresponding threshold; the corrected deformation field, the multidimensional coherence feature, and the deformation gradient map of multiple time series are stacked along the time dimension to form a spatiotemporal feature tensor; The first threshold is set based on the statistical histogram of the deformation gradient map, for example, taking the 90th percentile of all gradient values. The second threshold is set based on the physical meaning of kurtosis, for example, taking a value of 3 to identify regions with a distribution sharper than the normal distribution. All pixels are traversed; if a pixel's deformation gradient value is greater than the first threshold and its kurtosis value is greater than the second threshold, then that pixel is marked as a point to be corrected. For each point to be corrected, within a neighborhood of, for example, 7×7, stable pixels whose deformation gradient values ​​and kurtosis values ​​do not exceed the corresponding thresholds are searched. Using the deformation values ​​of these stable pixels, a new deformation value is calculated using inverse distance weighted interpolation, and this new value replaces the original deformation value of the point to be corrected. After all points to be corrected have been processed, the corrected deformation field is obtained.

[0024] Assume the monitoring time series comprises T time points, and the monitored area image size is H rows and W columns. After completing the aforementioned steps, T corrected deformation fields, T multidimensional coherence feature maps (including coherence and kurtosis channels), and T deformation gradient maps are obtained. For any pixel (h, w) in space, the corrected deformation value, coherence coefficient value, kurtosis value, and deformation gradient value of the pixel at all T time points are sequentially arranged to form a feature sequence of length 4×T. The feature sequences of all pixels are integrated to form a four-dimensional spatiotemporal feature tensor with dimensions H×W×T×4, where H and W are spatial dimensions, T is the temporal dimension, and 4 is the feature channel dimension.

[0025] In an optional embodiment, replacing the deformation value of the pixel by spatial interpolation using the deformation values ​​of other pixels in the neighborhood of the pixel that do not exceed the corresponding threshold includes: The deformation values ​​of other pixels in the neighborhood of the pixel that do not exceed the corresponding threshold are weighted and averaged to obtain the replacement deformation value.

[0026] If the deformation gradient value of a pixel is greater than a first threshold and the kurtosis value is greater than a second threshold, the deformation observation value of that pixel is considered unreliable or potentially noise. To correct these unreliable values, spatial interpolation is performed using inverse distance weighted interpolation to obtain the pixel value to be corrected. Define a neighborhood search window centered on a given area, such as a 9x9 region. Iterate through all neighboring pixels within this window.

[0027] During the traversal, for each neighboring pixel A filtering process is performed, selecting only reliable pixels whose deformation gradient value is less than a first threshold and whose kurtosis value is less than a second threshold as source data points for interpolation calculations. For each selected reliable neighbor point... Calculate the distance between the point and the center point. Euclidean distance between The square of the inverse of this distance is used as the weight. ,Right now Reliable points that are closer in distance will be assigned higher weights. This is determined by the deformation values ​​of all reliable neighboring points. Perform a weighted average and calculate the center point. new deformation value The calculation formula is: ; By replacing unreliable deformation values ​​with the deformation trends of reliable neighbors, the spatial continuity of the deformation field is maintained while suppressing noise.

[0028] To construct a high-dimensional data structure that a spatiotemporal convolutional neural network can process, in an optional embodiment, stacking the multi-temporal modified deformation field, the multi-dimensional coherence features, and the deformation gradient map along the time dimension to form a spatiotemporal feature tensor includes: The modified deformation field, the deformation gradient map, and the coherence coefficient map and kurtosis map that constitute the multidimensional coherence feature are stacked as feature channels.

[0029] Specifically, assuming the processed time series contains T time steps, for each time step t, four types of feature maps have been generated, each with a spatial size of H×W. The four types of feature maps are: 1. a modified deformation field map; 2. a deformation gradient map; 3. a coherence coefficient map; and 4. a kurtosis map.

[0030] For a single time step t, the four H×W two-dimensional feature maps are stacked along the channel dimension to form a three-dimensional tensor of size H×W×4. Each pixel in this tensor contains a four-dimensional feature vector, comprehensively representing the deformation state and statistical characteristics of that point at that moment. The three-dimensional tensors generated at all T time steps are then stacked along a new time dimension to form a four-dimensional spatiotemporal feature tensor of size T×H×W×4.

[0031] S4, the spatiotemporal feature tensor is input into the convolutional neural network, and the ground collapse probability of each pixel is output. The convolutional neural network includes a parallel dual-branch and a channel attention module. The output of the first convolutional branch is used to generate channel attention weights to weight the feature map output by the second convolutional branch.

[0032] The constructed H×W×T×4 spatiotemporal feature tensor is input into the planned network model. This tensor simultaneously enters two parallel 3D convolutional branches. The first branch, serving as the attention weight generation branch, compresses the input feature tensor into a channel feature vector of length C through several 3D convolutional layers and a global average pooling layer, where C is the number of channels in the feature map output by the second branch. This vector then passes through two fully connected layers and a sigmoid activation function to generate channel attention weights ranging from 0 to 1. The second branch is the backbone feature extraction network, composed of multiple 3D convolutional layers, batch normalization layers, and a ReLU activation function, used for deep extraction of spatiotemporal features, outputting a feature map of length H'×W'×T'×C. The channel attention weights of length C are broadcast to the H'×W'×T' dimension and then multiplied channel-by-channel with the feature map output by the second branch, achieving weighting of the feature map and enhancing the representation of the feature channels. The weighted feature map is then passed through convolutional and upsampling layers, followed by a 1×1 convolutional layer and a sigmoid activation function. The output is an H×W×1 probability map with the same spatial dimensions as the input image. Each pixel value in the map represents the probability of ground collapse at that point. Figure 3 .

[0033] In an optional embodiment, the convolutional neural network includes parallel dual branches, comprising: The first convolutional branch consists of three consecutive 3D convolutional layers, with each 3D convolutional layer having a stride of 1 and a padding mode of "same", used to detect deep spatiotemporal features. The second convolutional branch consists of a single 3D convolutional layer, used to detect shallow spatiotemporal features while preserving the original feature information.

[0034] The input to this neural network is the spatiotemporal feature tensor generated above, for example, with dimensions T×H×W×4. The input data is simultaneously fed into two parallel convolutional processing paths. The first branch, the deep feature extraction branch, consists of three cascaded 3D convolutional layers. For example, the first layer uses 16 3×3×3 convolutional kernels, the second layer uses 32 3×3×3 convolutional kernels, and the third layer uses 64 3×3×3 convolutional kernels. Each convolutional layer is typically followed by an activation function such as ReLU. Through the stacking of multiple convolutional layers, the receptive field of this branch gradually increases, enabling the detection of spatiotemporal correlation patterns and abstract features over a wider range and longer time span.

[0035] The second branch, the shallow feature preservation branch, has a simpler structure, containing only a single 3D convolutional layer. For example, this layer uses 64 1×1×1 convolutional kernels. The 1×1×1 kernels do not change the range of information aggregation in the spatiotemporal dimension, performing linear transformations and information integration in the channel dimension while preserving the spatiotemporal structure of the original feature map. This branch acts as a shortcut for feature extraction, ensuring that some basic and shallow details in the input features, such as deformation boundaries or abrupt changes, are not over-smoothed or lost during multiple convolutions and pooling operations in the deep network. The outputs of the two branches are combined to achieve complementarity between deep semantic information and shallow detailed information.

[0036] The complete structure of this dual-branch convolutional neural network, after fusing features from both branches, typically connects one or more additional convolutional layers for prediction. For example, a weighted shallow feature map is passed through a 1×1×1 3D convolutional layer, adjusting the number of channels to match the number of target classes, and then activated by a Softmax or Sigmoid activation function to generate a pixel-wise classification probability map. The network's input is the generated T×H×W×4 spatiotemporal feature tensor. The network's output is a probability map of spatial size H×W, where each pixel's value represents the probability that the point belongs to an unstable region. Training this network requires a labeled training set consisting of a large number of spatiotemporal feature tensor samples and corresponding ground truth labels. The ground truth labels are manually labeled by experts or generated from measurement data, indicating the stability of each pixel. The training process is end-to-end, optimized using a backpropagation algorithm. Specifically, a loss function such as cross-entropy is used to measure the difference between the network output and the ground truth labels, and the Adam optimizer is selected to iteratively update all trainable parameters of the network, such as convolutional kernel weights. During training, an initial learning rate is set, such as 0.001, and a mini-batch gradient descent method is used, for example, processing 8 spatiotemporal data cubes per batch, and iterating the entire training set for multiple epochs until the model performance converges.

[0037] In an optional embodiment, the output of the first convolutional branch is used to generate channel attention weights to weight the feature map output by the second convolutional branch, including: Global average pooling is performed on the output feature map of the first convolutional branch to obtain the channel representation; The channel representation is input into a bottleneck structure consisting of two fully connected layers, and attention weights for each channel are generated by an activation function. The attention weights are then multiplied element-wise by channel with the feature map output by the second convolutional branch.

[0038] Specifically, the output feature tensor of the first convolutional branch, i.e., the deep branch, is obtained, assuming its dimensions are C×T×H×W, where C is the number of channels, for example, 64. Global average pooling is applied to this tensor to compress the spatial and temporal dimensions T, H, W, i.e., averaging all elements of each channel. This results in a vector of length C, called the channel representative, which summarizes the global response of each feature channel.

[0039] The C-dimensional channel representation is input into a bottleneck module consisting of two fully connected layers. The first fully connected layer compresses the dimension from C to a smaller value, such as C / r, where r is the compression ratio, using the ReLU activation function. The second fully connected layer restores the dimension to C and uses a Sigmoid activation function. The Sigmoid function maps each element of the output to a value between 0 and 1; these C values ​​constitute the attention weights for each channel, representing the importance of each feature channel. The C-dimensional weight vector is weighted with the C×T×H×W feature tensor output by the second convolutional branch, i.e., the shallow branch. The weight vector is then broadcast to the same dimension as the feature tensor and multiplied element-wise by channel. Feature channels deemed important by deeper information are amplified, while less important ones are suppressed.

[0040] In the second embodiment, the present invention also proposes a ground subsidence monitoring system integrating InSAR and neural networks, comprising the following modules: The generation module is used to acquire multi-temporal SAR images of the monitoring area, perform differential interferometry processing, and generate initial deformation field and coherence coefficient maps for each temporal sequence; based on the initial deformation field of each temporal sequence, the structure tensor in the local neighborhood of each pixel is calculated, and the maximum eigenvalue is extracted as the deformation gradient value to generate deformation gradient maps for each temporal sequence. The first component module is used to determine the size of the spatial window for statistical coherence features based on the deformation gradient value of each pixel; within the spatial window, the kurtosis of the coherence coefficient is calculated, and the kurtosis is combined with the coherence coefficient value to form the multidimensional coherence features of each time series. The second component module is used to correct the initial deformation field to obtain a corrected deformation field. When the deformation gradient value and kurtosis value of a pixel are greater than a preset first threshold and a second threshold, respectively, the deformation values ​​of other pixels in the neighborhood of the pixel that do not exceed the corresponding thresholds are used to replace the deformation value of the pixel through spatial interpolation. The corrected deformation field, the multidimensional coherence feature, and the deformation gradient map of multiple time series are stacked along the time dimension to form a spatiotemporal feature tensor. The weighting module is used to input the spatiotemporal feature tensor into the convolutional neural network and output the ground collapse probability of each pixel. The convolutional neural network includes a parallel dual-branch and a channel attention module. The output of the first convolutional branch is used to generate channel attention weights to weight the feature map output by the second convolutional branch.

[0041] In an optional embodiment, calculating the structure tensor in the local neighborhood of each pixel based on the initial deformation field at each time sequence includes: Obtain the partial derivatives of deformation for each pixel in the x and y directions; Within a neighborhood window centered on the pixel, the structure tensor is calculated by summing the second moments of the partial derivatives of the deformation.

[0042] In an optional embodiment, determining the spatial window size for statistical coherence features based on the deformation gradient values ​​of each pixel includes: Through formula Calculate the side length W of the spatial window, where The deformation gradient value of the pixel is given by K, which is a preset scaling factor; and the minimum and maximum values ​​of the side length of the spatial window are defined.

[0043] In an optional embodiment, combining the kurtosis and coherence coefficient values ​​to form multidimensional coherence features for each time series includes: The coherence coefficient and kurtosis value of each pixel are combined to form a two-dimensional feature vector.

[0044] In an optional embodiment, replacing the deformation value of the pixel by spatial interpolation using the deformation values ​​of other pixels in the neighborhood of the pixel that do not exceed the corresponding threshold includes: The deformation values ​​of other pixels in the neighborhood of the pixel that do not exceed the corresponding threshold are weighted and averaged to obtain the replacement deformation value.

[0045] In an optional embodiment, stacking the modified deformation field, the multidimensional coherence features, and the deformation gradient map along the time dimension to form a spatiotemporal feature tensor includes: The modified deformation field, the deformation gradient map, and the coherence coefficient map and kurtosis map that constitute the multidimensional coherence feature are stacked as feature channels.

[0046] In an optional embodiment, the convolutional neural network includes parallel dual branches, comprising: The first convolutional branch consists of three consecutive 3D convolutional layers, used to detect deep spatiotemporal features; The second convolutional branch consists of a single 3D convolutional layer, used to detect shallow spatiotemporal features while preserving the original feature information.

[0047] In an optional embodiment, the output of the first convolutional branch is used to generate channel attention weights to weight the feature map output by the second convolutional branch, including: Global average pooling is performed on the output feature map of the first convolutional branch to obtain the channel representation; The channel representation is input into a bottleneck structure consisting of two fully connected layers, and attention weights for each channel are generated by an activation function. The attention weights are then multiplied element-wise by channel with the feature map output by the second convolutional branch.

[0048] It should be clarified that this application is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of this application is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of this application.

[0049] The functional modules shown in the above-described block diagram can be implemented as hardware, software, firmware, or a combination thereof. When implemented in hardware, they can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this application are programs or code segments used to perform the required tasks. Programs or code segments can be stored on a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried on a carrier wave. "Machine-readable medium" can include any medium capable of storing or transmitting information. Examples of machine-readable media include electronic circuits, semiconductor memory devices, ROM, flash memory, erasable ROM (EROM), floppy disks, CD-ROMs, optical disks, hard disks, fiber optic media, radio frequency (RF) links, etc. Code segments can be downloaded via computer networks such as the Internet, intranets, etc.

[0050] It should also be noted that the exemplary embodiments mentioned in this application describe methods or systems based on a series of steps or apparatus. However, this application is not limited to the order of the above steps; that is, the steps can be performed in the order mentioned in the embodiments, or in a different order, or several steps can be performed simultaneously.

[0051] The aspects of this application have been described above with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It should be understood that each block in the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that these instructions, executable via the processor of the computer or other programmable data processing apparatus, enable the implementation of the functions / actions specified in one or more blocks of the flowchart illustrations and / or block diagrams. Such a processor can be, but is not limited to, a general-purpose processor, a special-purpose processor, a special application processor, or a field-programmable logic circuit. It is also understood that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can also be implemented by dedicated hardware performing the specified functions or actions, or can be implemented by a combination of dedicated hardware and computer instructions.

[0052] The above description is merely a specific implementation of this application. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, modules, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. It should be understood that the protection scope of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the protection scope of this application.

Claims

1. A ground subsidence monitoring method integrating InSAR and neural networks, characterized in that, include: Acquire multi-temporal SAR images of the monitoring area, perform differential interferometry processing, and generate initial deformation field and coherence coefficient maps for each temporal sequence; Based on the initial deformation field of each time series, the structure tensor in the local neighborhood of each pixel is calculated, and the maximum eigenvalue is extracted as the deformation gradient value to generate the deformation gradient map of each time series. The size of the spatial window used for statistical coherence features is determined based on the deformation gradient value of each pixel; within the spatial window, the kurtosis of the coherence coefficient is calculated, and the kurtosis is combined with the coherence coefficient value to form the multidimensional coherence features of each time series. The initial deformation field is modified to obtain a modified deformation field. When the deformation gradient value and kurtosis value of a pixel are greater than a preset first threshold and a second threshold, respectively, the deformation value of the pixel is replaced by spatial interpolation using the deformation values ​​of other pixels in the neighborhood of the pixel that do not exceed the corresponding threshold. The modified deformation field, the multidimensional coherence feature, and the deformation gradient map of multiple time series are stacked along the time dimension to form a spatiotemporal feature tensor. The spatiotemporal feature tensor is input into a convolutional neural network, which outputs the ground collapse probability of each pixel. The convolutional neural network includes a parallel dual-branch and a channel attention module. The output of the first convolutional branch is used to generate channel attention weights to weight the feature map output by the second convolutional branch.

2. The method according to claim 1, characterized in that, The initial deformation field based on each time sequence is used to calculate the structure tensor in the local neighborhood of each pixel, including: Obtain the partial derivatives of deformation for each pixel in the x and y directions; Within a neighborhood window centered on the pixel, the structure tensor is calculated by summing the second moments of the partial derivatives of the deformation.

3. The method according to claim 1, characterized in that, The step of determining the spatial window size for statistical coherence features based on the deformation gradient values ​​of each pixel includes: Through formula Calculate the side length W of the spatial window, where The deformation gradient value of the pixel is given by K, which is a preset scaling factor; and the minimum and maximum values ​​of the side length of the spatial window are defined.

4. The method according to claim 1, characterized in that, The combination of the kurtosis and coherence coefficient values ​​to form the multidimensional coherence characteristics of each time series includes: The coherence coefficient and kurtosis value of each pixel are combined to form a two-dimensional feature vector.

5. The method according to claim 1, characterized in that, The step of replacing the deformation value of the pixel by spatial interpolation using the deformation values ​​of other pixels in the neighborhood of the pixel that do not exceed the corresponding threshold includes: The deformation values ​​of other pixels in the neighborhood of the pixel that do not exceed the corresponding threshold are weighted and averaged to obtain the replacement deformation value.

6. The method according to claim 1, characterized in that, The step of stacking the modified deformation field, the multidimensional coherence features, and the deformation gradient map along the time dimension to form a spatiotemporal feature tensor includes: The modified deformation field, the deformation gradient map, and the coherence coefficient map and kurtosis map that constitute the multidimensional coherence feature are stacked as feature channels.

7. The method according to claim 1, characterized in that, The convolutional neural network comprises parallel dual branches, including: The first convolutional branch consists of three consecutive 3D convolutional layers, used to detect deep spatiotemporal features; The second convolutional branch consists of a single 3D convolutional layer, used to detect shallow spatiotemporal features while preserving the original feature information.

8. The method according to claim 7, characterized in that, The output of the first convolutional branch is used to generate channel attention weights to weight the feature map output by the second convolutional branch, including: Global average pooling is performed on the output feature map of the first convolutional branch to obtain the channel representation; The channel representation is input into a bottleneck structure consisting of two fully connected layers, and attention weights for each channel are generated by an activation function. The attention weights are then multiplied element-wise by channel with the feature map output by the second convolutional branch.

9. A ground subsidence monitoring system integrating InSAR and neural networks, characterized in that, Includes the following modules: The generation module is used to acquire multi-temporal SAR images of the monitoring area, perform differential interferometry processing, and generate initial deformation field and coherence coefficient maps for each temporal sequence; based on the initial deformation field of each temporal sequence, the structure tensor in the local neighborhood of each pixel is calculated, and the maximum eigenvalue is extracted as the deformation gradient value to generate deformation gradient maps for each temporal sequence. The first component module is used to determine the size of the spatial window for statistical coherence features based on the deformation gradient value of each pixel; within the spatial window, the kurtosis of the coherence coefficient is calculated, and the kurtosis is combined with the coherence coefficient value to form the multidimensional coherence features of each time series. The second component module is used to correct the initial deformation field to obtain a corrected deformation field. When the deformation gradient value and kurtosis value of a pixel are greater than a preset first threshold and a second threshold, respectively, the deformation values ​​of other pixels in the neighborhood of the pixel that do not exceed the corresponding thresholds are used to replace the deformation value of the pixel through spatial interpolation. The corrected deformation field, the multidimensional coherence feature, and the deformation gradient map of multiple time series are stacked along the time dimension to form a spatiotemporal feature tensor. The weighting module is used to input the spatiotemporal feature tensor into the convolutional neural network and output the ground collapse probability of each pixel. The convolutional neural network includes a parallel dual-branch and a channel attention module. The output of the first convolutional branch is used to generate channel attention weights to weight the feature map output by the second convolutional branch.

10. The system according to claim 9, characterized in that, The initial deformation field based on each time sequence is used to calculate the structure tensor in the local neighborhood of each pixel, including: Obtain the partial derivatives of deformation for each pixel in the x and y directions; Within a neighborhood window centered on the pixel, the structure tensor is calculated by summing the second moments of the partial derivatives of the deformation.

Images