Coherence maximization method for InSAR mining area monitoring

Abstract

The invention provides a coherence maximization algorithm to improve the reliability of coherence estimation of a mining area, aiming at the problems of low manual interpretation efficiency of a settlement funnel area and influence of high phase gradient on subsequent data processing in a large-gradient deformation area of the mining area in an InSAR (Interferometric Synthetic Aperture Radar) technology. According to the method, a slice-assisted hyper-inference framework is introduced into a YOLOv11 network to extract mining area funnels, a two-dimensional Gaussian mixture model and a hybrid simulated annealing genetic algorithm are used for independently modeling each detected settlement funnel from a wound interference phase, and the simulated funnels are used for removing fringes, so that underestimation of coherence caused by fringes is eliminated. After the coherence maximization algorithm is applied, the unwrapping error is obviously reduced, and the average RMSE of 208 interference blocks is reduced from 3.90 to 2.64. The positive influence of phase unwrapping proves the value of the method.

Description

Technical Field

[0001] This invention belongs to the field of synthetic aperture radar interferometry (InSAR) technology, specifically relating to a method for maximizing coherence in InSAR mining area monitoring. Background Technology

[0002] Mineral resources are one of the main sources of energy supply in modern society, but frequent mining activities can cause rapid local subsidence of the earth's surface, leading to geological disasters such as collapses, surface deformation, landslides, and debris flows. Interferometric Synthetic Aperture Radar (InSAR) provides strong support for the refined monitoring of subsidence in mining areas due to its advantages of high precision, wide coverage, and non-contact operation. The large-gradient deformation caused by high-intensity mining exceeds the phase gradient range measurable by InSAR. The dense fringes resulting from the high phase gradient can cause deviations in phase unwrapping results and also lead to an underestimation of the region's coherence. Especially in densely vegetated mountainous areas, phase aliasing and decoherence are prone to occur, ultimately making it impossible to obtain accurate and reliable surface deformation information. Furthermore, methods for identifying subsidence funnel areas based on manual visual interpretation are labor-intensive and inefficient, making it difficult to meet the needs of large-scale monitoring. To address the issues of low manual interpretation efficiency in the subsidence funnel region and the impact of high phase gradient on subsequent data processing in large gradient deformation areas of mining areas using InSAR technology, this invention proposes a coherence maximization algorithm to improve the reliability of coherence estimation in mining areas.

[0003] The non-stationarity and phase gradient of interferometric signals lead to an underestimation of coherence, a problem that destriating algorithms can alleviate. Frequency domain methods, limited by kernel function selection, struggle to balance noise and fringe frequencies; spatial domain methods often rely on phase unwrapping or low-gradient assumptions, making them difficult to handle high-gradient phases. Deep learning offers new opportunities for InSAR target detection; for example, data augmentation methods based on TinyYOLOv2 can effectively identify settlement funnels. However, small targets are easily lost due to network downsampling, requiring the introduction of complex architectures (such as R-CNN and PUNet), which often face challenges such as high technical barriers and difficulty in utilizing pre-trained weights.

[0004] In summary, while existing advanced methods provide a wealth of technical means for target identification, signal filtering, and phase unwrapping in mining scenarios, they still have significant shortcomings in coherence estimation. The underestimation of coherence caused by high phase gradients not only reduces the number of available measurement points but also exacerbates discontinuities in the phase unwrapping process. Therefore, the settlement results obtained from InSAR inversion often have limited accuracy and reliability. Summary of the Invention

[0005] This invention addresses the problem that existing methods in mining scenarios suffer from severe performance degradation in target recognition, signal filtering, and phase unwrapping due to the underestimation of coherence caused by ignoring high phase gradients. This invention introduces the Slicing-Aided Hyper Inference (SAHI) framework into the YOLOv11 network to extract sinkholes in the mining area. A two-dimensional Gaussian mixture model and a Mixed Simulated Annealing-Genetic Algorithm (MSAGA) are used to model each detected sinkhole separately from the entangled interference phase. The simulated sinkhole is then used for stripe removal, thereby eliminating the underestimation of coherence caused by the presence of stripes.

[0006] A Gaussian mixture model was used to simulate surface deformation caused by mining activities, and synthetic samples were generated by overlaying real InSAR interferometric data. Theoretically, the Gaussian mixture model can approximate any shape, representing an irregular funnel as a weighted sum of k Gaussian components, each with independent weights, mean, and covariance matrices. Adjusting the model parameters allows the model to adapt to different shapes and data distributions, as shown in the following expression:

[0007] (1),

[0008] In the formula, It is a two-dimensional variable. As weight and , and These are the mean vector and covariance matrix of the i-th Gaussian distribution, respectively. For the i-th single Gaussian component:

[0009] (2),

[0010] Assuming the model value is the surface subsidence along the radar satellite's line of sight, the position is randomly adjusted using the mean vector, and the shape and area are modified according to the covariance. Before scaling the funnel amplitude, the model is normalized to [0, 1].

[0011] Using only a two-dimensional Gaussian mixture model for settlement simulation limits the generalization ability of the dataset. To enhance the diversity of settlement structures, the Mogi model, the Okada dislocation model, and the Probabilistic Integral Method (PIM) are employed. The parameters used are shown in Table 1. These three models generate three-dimensional displacement components (i.e., vertical displacement components). ,thing North and South (Direction), but the deformation detected by InSAR technology is along the radar line of sight, so it is necessary to synthesize the three-dimensional surface deformation into a deformation along the radar line of sight. :

[0012] (3),

[0013] In the formula, Indicates the angle of incidence of the radar wave. Indicates the satellite azimuth angle.

[0014] Table 1 - Range of Model Parameter Values

[0015]

[0016] Leveraging the latest version of YOLOv11 released by Ultralytics, which introduces innovative architectures and optimization strategies, such as C3K2 convolutional blocks (cross-stage partial blocks with 3×3 and 2×2 convolutions) which use smaller, stacked convolutional kernels to more effectively capture multi-scale features at different network depths; SPPF (Fast Spatial Pyramid Pooling) which expands the receptive field and reduces feature redundancy by performing feature concatenation after successive pooling operations; and C2PSA (Convolutional Blocks with Parallel Spatial Attention) which enhances the detection capability for small-sized and partially occluded targets by utilizing cross-scale spatial attention and fine-grained pixel-level optimization, making it particularly suitable for scenes with complex textures.

[0017] The SAHI framework is used to divide a large-scale interferogram into overlapping slices. The slice length, width, and overlap ratio are preset, and the size of each slice is adjusted while maintaining the aspect ratio. Target detection is performed on each slice. The detectability of small targets is improved by increasing the relative size of small targets within each slice and reducing background complexity, thereby reducing the false negative rate. The predictions of all slices are mapped back to their original coordinates, and overlapping predictions are merged using non-maximum suppression (NMS). The target locations are then reconstructed in the original image.

[0018] Coherence is used to quantify the correlation between two radar echoes. The coherence estimate between the first and second observations can be defined as:

[0019] (4),

[0020] In the formula, , This represents a single-view complex image, where L is the number of pixels in the sliding window. This represents the complex conjugate operator. Mining-induced subsidence introduces nonlinear signals with high phase gradients into the local window, leading to an underestimation of coherence. Assuming coherence is independent of local amplitude, this is addressed by compensating for nonlinear deformation phase within the coherence estimation window. Correcting the underestimation bias of any pixel p:

[0021] (5),

[0022] In the formula, and respectively represent the observed phase values of pixel p in two single-look complex (SLC) images. L is the number of pixels in the sliding window, is the simulated deformation phase value at pixel p. The closer the simulated phase is to the actual settlement near the mining area, the more local fringe rates can be removed, and thus a larger coherence amplitude can be obtained in the local area. Therefore, the previous formula can be expressed as an objective function:

[0023] (6),

[0024] (7),

[0025] In the formula, can be modeled using a Gaussian mixture function with k components. λ is the wavelength, represents the magnitude of the k-th Gaussian component. The covariance matrix defines the shape and orientation of the model.

[0026] Using the rotation matrix to decompose the covariance matrix to solve the numerical instability problem that may occur when constructing the covariance matrix using the correlation relationship in the inverse problem. This instability may lead to convergence failure and large numerical errors in the optimization process.

[0027] (8),

[0028] (9),

[0029] In the formula, is an orthogonal matrix that can keep the Euclidean norm of the vector unchanged and determines the orientation of the model in space. The variances and control the size of the model. In addition, a small positive lower limit is added to the variances to ensure that the covariance matrix remains in good condition and the optimization remains stable.

[0030] After using the YOLOv11 model to detect the settlement funnel in the mining area, obtain its position bounding box (including the center coordinates, width w, and height h) in the interferogram; according to this bounding box information, set the search range of each parameter of the Gaussian mixture model: the mean parameter is constrained within the intervals [-w / 2, w / 2] and [-h / 2, h / 2], the standard deviation parameter satisfies < w / 6, < h / 6, and the rotation angle is constrained within the range [-π, π]; at the same time, according to the relationship between the maximum possible number of fringes of the settlement funnel and the radar wavelength λ (a single fringe corresponds to a deformation of λ / 2), set the scaling factor The value range of is [-5λ / 2, 0]; an initial population is randomly generated within the above constraints for subsequent optimization.

[0031] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0032] (1) Compared with the original YOLO11 model, the method of the present invention reduces the overall false negative rate and provides accurate location information for early identification of mining area subsidence and subsequent InSAR processing;

[0033] (2) After weakening the nonlinear phase gradient, the method of the present invention can improve the average coherence of the settling funnel area and significantly improve the coherence point density under the same coherence threshold, providing support for point selection and phase unwrapping of InSAR technology in mining deformation monitoring.

[0034] (3) The method of the present invention reduces the average number of residual points before and after stripe removal. Compared with the traditional phase unwrapping method, it reduces the average mean square error in the synthetic data and reduces the number of unclosed pixels in the real data. This verifies that the joint application of target detection and phase modeling can effectively alleviate the impact of nonlinear high phase gradient on unwrapping accuracy and provides support for improving the deformation monitoring accuracy of large gradient mining areas. Attached Figure Description

[0035] Figure 1 This is a flowchart of the steps of the method of the present invention;

[0036] Figure 2 Here is an example of the test results;

[0037] Figure 2 (a) shows the YOLO11 results of the synthetic data;

[0038] Figure 2 (b) YOLO11 results incorporating the SAHI framework;

[0039] Figure 2 (c) shows the actual YOLO11 results;

[0040] Figure 2 (d) YOLO11 results that incorporate the SAHI framework;

[0041] Figure 3 Example of stripe removal results;

[0042] Figure 3 (a) shows the test results;

[0043] Figure 3 (b) shows the result after stripe removal;

[0044] Figure 3(c) is the coherence diagram. Detailed Implementation

[0045] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0046] Example: A coherence maximization algorithm technique for InSAR mining area monitoring, applicable to glacial regions, such as... Figure 1 As shown, the specific steps include the following:

[0047] Step S1: Data Collection. Based on C-band InSAR data acquired by the Sentinel-1 satellite and L-band data acquired by the Lutan-1 satellite, systematic experimental verification was carried out.

[0048] Step S2: Data Preprocessing. This invention uses SNAP software developed by the European Space Agency to generate the wrapped interferogram. The preprocessing steps mainly include cropping, importing precise orbital data for correction, image registration, stitching gaps, sub-block creation, and wrapped interferogram generation. The original interferogram is a binary grayscale image, which is converted into an RGB color image to facilitate better recognition by the deep learning network. To reduce the impact of noise, multi-view processing and filtering techniques are used to improve the quality of the real dataset.

[0049] Step S3: Simulate surface deformation caused by mining activities using a Gaussian mixture distribution and overlay real InSAR interferometric data to generate a synthetic sample. Theoretically, a Gaussian mixture model can approximate any shape, representing an irregular funnel as a weighted sum of k Gaussian components, each with independent weights, mean, and covariance matrices. Adjusting the model parameters allows the model to adapt to different shapes and data distributions, as shown in the following expression:

[0050] (1),

[0051] In the formula, It is a two-dimensional variable. As weight and , and These are the mean vector and covariance matrix of the i-th Gaussian distribution, respectively. For the i-th single Gaussian component:

[0052] (2),

[0053] Assuming the model value is the surface subsidence along the radar satellite's line of sight, the position is randomly adjusted using the mean vector, and the shape and area are modified according to the covariance. Before scaling the funnel amplitude, the model is normalized to [0, 1].

[0054] Using only a two-dimensional Gaussian mixture model for settlement simulation limits the generalization ability of the dataset. The Mogi model, Okada dislocation model, and Probabilistic Integral Method (PIM) are employed to enhance the diversity of settlement structures. The parameters used are shown in the table below. These three models generate three-dimensional displacement components (i.e., vertical displacement). ,thing North and South (Direction), but the deformation detected by InSAR technology is along the radar line of sight, so it is necessary to synthesize the three-dimensional surface deformation into a deformation along the radar line of sight. :

[0055] (3),

[0056] In the formula, Indicates the angle of incidence of the radar wave. Indicates the satellite azimuth angle.

[0057] Table 1 - Range of Model Parameter Values

[0058]

[0059] Step S4: Utilize the latest version of YOLOv11 released by Ultralytics, which introduces innovative architectures and optimization strategies. For example, C3K2 convolutional blocks (cross-stage partial blocks with 3×3 and 2×2 convolutions) use smaller, stacked convolutional kernels, which can more effectively capture multi-scale features at different network depths; SPPF (Fast Spatial Pyramid Pooling) expands the receptive field and reduces feature redundancy by performing feature concatenation after successive pooling operations; C2PSA (convolutional blocks with parallel spatial attention) enhances the detection capability for small-sized and partially occluded targets by utilizing cross-scale spatial attention and fine-grained pixel-level optimization, which is especially suitable for scenes with complex textures.

[0060] Step S5: Using the SAHI framework, the large-scale interferogram is divided into overlapping slices. The length, width, and overlap ratio of each slice are preset. While maintaining the aspect ratio, the size of each slice is adjusted, and target detection is performed on each slice. The detectability of small targets is improved by increasing the relative size of small targets within each slice and reducing background complexity, thereby reducing the false negative rate. All slice predictions are mapped back to their original coordinates. Overlapping predictions are merged using non-maximum suppression (NMS), and the target location is reconstructed in the original image.

[0061] Step S6: Coherence is used to quantify the correlation between two radar echoes. The coherence estimate between the first and second observations can be defined as:

[0062] (4),

[0063] In the formula, , This represents a single-view complex image, where L is the number of pixels in the sliding window. This represents the complex conjugate operator. Mining-induced subsidence introduces nonlinear signals with high phase gradients into the local window, leading to an underestimation of coherence. Assuming coherence is independent of local amplitude, this is addressed by compensating for nonlinear deformation phase within the coherence estimation window. Correcting the underestimation bias of any pixel p:

[0064] (5),

[0065] In the formula, and Let represent the observed phase values ​​of pixel p in the two single-view complex images (SLC), and L be the number of pixels in the sliding window. Let be the simulated deformation phase value at pixel p. The closer the simulated phase matches the actual subsidence near the mining area, the better the local fringe rate can be removed, thus obtaining a larger coherence amplitude in the local region. Therefore, the previous formula can be expressed as an objective function:

[0066] (6),

[0067] (7),

[0068] In the formula, This can be modeled using a Gaussian mixture function with k components, where λ is the wavelength. This represents the magnitude of the k-th Gaussian component. Covariance matrix. The shape and orientation of the model were defined.

[0069] Solving inverse problems by decomposing the covariance matrix using rotation matrix decomposition may lead to numerical instability when constructing the covariance matrix using correlation relationships. This instability may result in convergence failure and large numerical errors during the optimization process.

[0070] (8),

[0071] (9),

[0072] In the formula, An orthogonal matrix that preserves the Euclidean norm of vectors determines the orientation of the model in space. Variance and Control the size of the model. Additionally, add a small positive lower bound on the variance to ensure the covariance matrix remains in good condition and the optimization remains stable.

[0073] Step S7: After detecting the subsidence funnel in the mining area using the YOLOv11 model, obtain its position bounding box (including the center coordinates, width w, and height h) in the interferogram; according to this bounding box information, set the search range for each parameter of the Gaussian mixture model: the mean parameter is constrained within the intervals [-w / 2, w / 2] and [-h / 2, h / 2], the standard deviation parameter satisfies < w / 6, < h / 6, and the rotation angle is constrained within the range [-π, π]; at the same time, according to the relationship between the maximum possible number of fringes of the subsidence funnel and the radar wavelength λ (a single fringe corresponds to a deformation of λ / 2), set the value range of the scaling factor to be [-5λ / 2, 0]; randomly generate an initial population within the above constraints for subsequent optimization.

[0074] Step S8: After detecting the subsidence funnel in the mining area using the YOLOv11 model, obtain its position bounding box (including the center coordinates, width w, and height h) in the interferogram; according to this bounding box information, set the search range for each parameter of the Gaussian mixture model: the mean parameter is constrained within the intervals [-w / 2, w / 2] and [-h / 2, h / 2], the standard deviation parameter satisfies <w / 6, <h / 6, and the rotation angle is constrained within the range [-π, π]; at the same time, according to the relationship between the maximum possible number of fringes of the subsidence funnel and the radar wavelength λ (a single fringe corresponds to a deformation of λ / 2), set the value range of the scaling factor to be [-5λ / 2, 0]; randomly generate an initial population within the above constraints for subsequent optimization.

[0075] Test case

[0076] To verify the feasibility and effectiveness of the "Coherence Maximization Algorithm for InSAR Mining Area Monitoring" proposed in this invention, this invention selects the Datong Mining Area and Changzhi Mining Area in Shanxi Province, China as the experimental areas, and conducts systematic experimental verification. The specific steps are as follows:

[0077] Step S1: Based on the C-band InSAR data obtained from real Sentinel-1 satellites (time span from November 2022 to January 2023) and the L-band data obtained from Lutan-1 satellites (time span from June 12, 2023 to June 16, 2023), select the Datong Mining Area and Changzhi Mining Area in Shanxi Province, China as the experimental areas, and conduct systematic experimental verification. Due to the relatively limited Lutan-1 data, it is mainly used for network model testing and not added to the training set.

[0078] All networks were trained and tested on computers equipped with an Intel® i7-12700 CPU, 32GB of memory, and an NVIDIA Ge Force RTX 3070 GPU. The PyTorch 1.8.0 deep learning framework based on Python 3.8 was used, with Ultralytics YOLOv11 and GAMMA software used for InSAR processing.

[0079] Step S2: Construction of the synthetic dataset. A Gaussian mixture model was used in conjunction with three physical models: Mogi, Okada, and PIM. Settlement amplitudes were randomly generated within the range [−5λ / 2, 0] to simulate settlement funnels of different shapes, scales, and depths. Ten 2000×2000 and ten 1000×1000 synthetic interferograms were synthesized, overlaid with real noise and terrain phase, to construct the dataset for YOLOv11 training. The background interferograms for these images were all derived from real interferograms in the Sentinel-1 dataset. Different numbers and shapes of simulated funnels were randomly distributed within the synthetic interferograms. The 1000×1000 dataset contained 47 simulated funnels, with widths ranging from 34 to 87 pixels and lengths from 31 to 96 pixels; the 2000×2000 dataset contained 161 simulated funnels, with widths ranging from 28 to 179 pixels and lengths from 27 to 159 pixels.

[0080] Step S3: Constructing the Real Dataset. After generating the interferogram in the preprocessing step, the regions containing settlement funnels are visually cropped into sample data suitable for network input. In the YOLO network model, the size of the input image must be predefined. When the input sample image size is large, the image contains multiple settlement funnels, which can improve the network's detection accuracy for small targets. However, it requires more memory and computing resources. Small-sized samples train faster, but it is necessary to ensure that a complete settlement funnel region is covered to guarantee the validity of the data and the accuracy of detection. If a 320×320 pixel sample is cropped from a 20m×20m resolution Sentinel-1 multiview interferogram of a mining area in Shanxi Province, this sample can contain 1 to 4 settlement funnels. Therefore, the size of all samples in the dataset is uniformly 320×320 pixels. The funnels identified visually are manually labeled using the LabelImg annotation tool. In the real dataset, there are a total of 567 320×320 pixel interferogram patches, of which 738 settlement funnels are labeled.

[0081] Step S4: Settlement funnel detection. A sliding window detection using the SAHI framework is employed, with a slice overlap rate of 20%. The YOLOv11 model released by Ultralytics is used for settlement funnel detection. YOLOv11 employs C3K2 convolutional blocks, SPPF pooling modules, and a C2PSA attention mechanism to enhance multi-scale feature extraction capabilities. Training is performed for 100 epochs with a learning rate of 1e−3 and a batch size of 4. The detection output is the bounding box (center coordinates, width w, height h) for each settlement funnel. Detection results for simulated data are shown below. Figure 2 As shown in (a) and 2(b), in the 1000×1000 interferogram dataset, the YOLO11+SAHI model with the integrated framework achieved an accuracy of 89.6%, significantly higher than YOLO11 (80.8%), with a false negative rate reduced to 8.5%. In the 2000×2000 interferogram dataset, the YOLO11 network achieved a detection accuracy of 44.7% and a false negative rate of 55.3%. The YOLO11 model with the integrated SAHI framework achieved an accuracy of 88.8% and a false negative rate of 6.8%. The detection results for real data are as follows... Figure 2 As shown in (c) and 2(d), the false negative rate of YOLOv11 exceeds 50%, but when SAHI is integrated, the false negative rate drops to less than 8%.

[0082] Step S5: Gaussian mixture modeling and parameter optimization. For each detected funnel region, a two-dimensional Gaussian mixture model based on two Gaussian components is constructed to represent the deformation phase model of the mining area.

[0083] (10)

[0084] Where λ = 5.66cm (Sentinel-1 C band), amplitude factor [-5λ / 2, 0]mm controls the number of stripes. [0,1] represents the weights. All parameter boundaries are automatically generated by the detection box: mean... [-w / 2, w / 2], [-h / 2,h / 2], standard deviation <w / 6, <h / 6。

[0085] Step S6: To avoid numerical instability caused by ill-conditioned covariance matrix during optimization, this embodiment uses orthogonal rotation matrix decomposition to reconstruct the inverse covariance matrix:

[0086] (11),

[0087] Among them Rotation matrix. This parameterization method transforms the optimization of the covariance matrix into optimization of the independent variables. The search significantly improves the convergence stability of the optimization process. No iteration failures due to matrix singularities were observed in the experiments, verifying the effectiveness of the proposed method.

[0088] Step S6: Inversion is performed using the MSAGA method. The population size is 200, crossover rate is 0.8, mutation rate is 0.1, and the initial simulated annealing temperature is adjusted to 100 to improve convergence speed, while maintaining a minimum temperature of 0.01 and an annealing factor of 0.95. In data testing, the average coherence estimate of the region remained essentially unchanged after 30 iterations.

[0089] Step S7: Coherence Estimation and Phase Compensation. Reconstruct the simulated phase using optimal parameters. Substitute into the formula:

[0090] (12),

[0091] In the formula, and Let represent the observed phase values ​​of pixel p in the two single-view complex images (SLC), and L be the number of pixels in the sliding window. Let be the simulated deformation phase value at pixel p. Calculate the compensated local coherence and use it for subsequent phase unwrapping. Figure 3 (a) and 3(b) show the experimental results of the synthetic interferometric data. The original average coherence value of the five simulated funnel identification regions is 0.42, and the average coherence value after stripe removal is 0.60. Figure 3 (c) and (d) show the experimental results of the Sentinel-1 interferometric data. The average coherence of all selected funnel regions is 0.42, and the average coherence after stripe removal is 0.58.

[0092] It should be noted that the above content merely illustrates the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. For those skilled in the art, various improvements and modifications can be made without departing from the principle of the present invention, and all such improvements and modifications fall within the scope of protection of the claims of the present invention.

Claims

1. A method for maximizing coherence in InSAR mining area monitoring, characterized in that, The method includes the following steps: S1, Data Collection: Based on C-band InSAR data acquired by the real Sentinel-1 satellite and L-band data acquired by the Lutan-1 satellite, a mining area was selected as the experimental area, and systematic experimental verification was carried out. S2, Data preprocessing: Generate a tangled interferogram from the data samples collected in step S1; S3, Construction of synthetic dataset: Using a Gaussian mixture model combined with three physical models, Mogi, Okada and PIM, a settling funnel of different shapes, scales and depths is simulated to construct a dataset for YOLOv11 training. S4, Construction of the real dataset: For the interferogram generated in step S2, the size of all samples is uniform, and the funnels identified by visual inspection are manually labeled using the LabelImg annotation tool; S5, Settlement funnel detection: The SAHI framework is used for sliding window detection, and the YOLOv11 model released by Ultralytics is used for settlement funnel detection. The detection output is the bounding box of each settlement funnel. S6, Calculate the phase: For each detected funnel region, construct a two-dimensional Gaussian mixture model based on the dual Gaussian components to construct the deformation phase model of the mining area, calculate the corresponding two-dimensional Gaussian mixture model phase, and subtract it from the original entanglement phase to obtain the residual phase; S7, Local Fine-tuning: Genetic iteration is performed through selection, crossover and mutation operations to generate a new generation of population, and the newly generated solution is locally fine-tuned using the simulated annealing algorithm. S8, Settlement funnel phase inversion: Repeat steps S6 and S7 until the preset convergence condition is met or the maximum number of iterations is reached. Output the parameter combination with the highest fitness as the optimal solution to complete the inversion modeling of the settlement funnel phase.

2. The coherence maximization method for InSAR mining area monitoring according to claim 1, characterized in that, In step S2, the preprocessing process must at least convert the original interferogram from a binary grayscale image to an RGB color image.

3. The coherence maximization method for InSAR mining area monitoring according to claim 1, characterized in that, The data preprocessing step S2 further includes processing the real dataset using multi-view processing and filtering techniques. Both multi-view processing and filtering are performed in SNAP software, with a range view count of 6 and a directional view count of 1. The Lee Filter is selected for filtering, and the filter window size is [size missing]. .

4. The coherence maximization method for InSAR mining area monitoring according to claim 1, characterized in that, The expression for the deformation phase model of the mining area based on the two-dimensional Gaussian mixture model with two Gaussian components in step S6 is as follows: （10）， Where λ = 5.66cm, Sentinel-1 C band, amplitude factor [-5λ / 2, 0]mm controls the number of stripes. [0,1] represents the weights, and all parameter boundaries are automatically generated by the detection box: mean. [-w / 2, w / 2], [-h / 2, h / 2], standard deviation < w / 6, < h / 6.

5. The coherence maximization method for InSAR mining area monitoring according to claim 1, characterized in that, The detailed parameters for simulated annealing in step S7 are as follows: population size 200, crossover rate 0.8, mutation rate 0.1, initial simulated annealing temperature adjusted to 100 to improve convergence speed, while maintaining minimum temperature at 0.01, annealing factor at 0.

95. In data testing, the average coherence estimate of the region remains basically unchanged after 30 iterations.

6. The coherence maximization method for InSAR mining area monitoring according to claim 1, characterized in that, The preprocessing step S2 includes trimming, importing and correcting precise orbital data, image registration, splicing gaps, sub-block creation, and generating a winding interferogram.

Images