A method for identifying mining subsidence basin in mining area combining deformable DE TR and InSAR technology

Abstract

The present application belongs to the technical field of mining subsidence monitoring, and particularly relates to a mining subsidence basin identification method combining Deformable DETR and InSAR technology. The method effectively solves the problem of insufficient training data by constructing a mixed sample set, combining real deformation data with simulated deformation samples generated based on a dynamic probability integral method. An improved Deformable DETR model is used to realize high-precision detection of subsidence basins in InSAR images by using a deformable attention mechanism and multi-scale feature fusion technology. The dynamic probability integral method dynamically calculates the subsidence rate through the Boltzmann function, and generates simulated data with geological authenticity by combining elastic transformation and noise synthesis technology. The Deformable DETR model significantly improves the identification ability of subsidence basins of different scales through sparse sampling and the Transformer architecture.

Description

Technical Field

[0001] This invention belongs to the field of mining subsidence monitoring technology, specifically relating to a method for identifying mining subsidence basins that combines Deformable DETR and InSAR technologies. Background Technology

[0002] Currently, monitoring of mining subsidence basins mainly relies on traditional measurement methods and Synthetic Aperture Radar Interferometry (InSAR). Traditional methods, such as ground leveling and GPS monitoring, while highly accurate, are limited by their small monitoring range, are time-consuming and labor-intensive, and costly. InSAR technology, with its advantages of being available 24 / 7 and in all weather conditions with wide coverage, has become an important tool for monitoring surface deformation in mining areas. However, traditional InSAR methods struggle to quickly and accurately identify coal mining subsidence basins when processing wide-swath SAR imagery, especially under complex geological conditions where deformation signals are easily interfered with by noise, thus limiting monitoring accuracy.

[0003] Traditional monitoring methods are limited by their efficiency and scope, making it difficult to meet the needs of real-time monitoring in large-scale mining areas. While InSAR technology possesses wide-area monitoring capabilities, it is susceptible to atmospheric delay, orbital errors, and decoherent noise when processing complex surface deformations, leading to artifacts or errors in the deformation results. Furthermore, existing methods often rely on threshold segmentation or simple machine learning models, which are insufficient for extracting complex deformation features. This makes it difficult to accurately identify coal mining subsidence basins in wide-swath SAR imagery, especially given the low detection accuracy of small-scale deformation targets, thus limiting its widespread application in monitoring geological hazards in mining areas. Summary of the Invention

[0004] The purpose of this invention is to provide a method for identifying mining subsidence basins that combines Deformable DETR (Deformable Detection Transformer) and InSAR technology, in order to solve the technical problem that existing technologies are unable to quickly, accurately and robustly identify mining subsidence basins in wide-swath SAR images, especially the low accuracy of small-scale deformed target detection under the influence of noise interference and complex geological conditions.

[0005] The present invention achieves the above objectives through the following technical solutions:

[0006] This invention proposes a method for identifying mining subsidence basins by combining Deformable DETR and InSAR technologies. The method includes:

[0007] Acquire InSAR imagery of the area to be measured;

[0008] A deep learning model is constructed based on the Deformable DETR model architecture and a hybrid sample set of sample regions. The hybrid sample set is obtained by combining real deformation samples and simulated deformation samples. The simulated deformation samples are synthesized by generating simulated deformation maps using a dynamic probability integral method based on parameter changes, followed by elastic transformation and the addition of noise.

[0009] The InSAR image of the area to be tested is input into the trained deep learning model, which outputs the location and boundary information of the coal mining subsidence basin.

[0010] Furthermore, the Deformable DETR model architecture includes:

[0011] The feature extraction module uses a ResNet50 backbone network to extract multi-scale feature maps from the input InSAR image.

[0012] The deformable attention module includes a multi-scale deformable attention layer for computing attention features;

[0013] Each attention head samples K points, given an input feature map. C represents the number of color channels in the image, H represents the height of the image, and W represents the width of the image, which has content features z. q and two-dimensional reference point p q The attention features of the query elements satisfy the following formula:

[0014] ;

[0015] In the formula: m is the attention head index; M is the total number of attention heads; k is the sampling point; K is the total number of sampling points, J< <H×W;W m and W' m For feature weights; A mqk ∆p represents the attention weight of the k-th sampling point in the m-th attention head. mqk This is the corresponding sampling offset;

[0016] The Transformer encoder-decoder module includes an encoder and a decoder. The encoder uses multi-scale deformable attention instead of standard multi-head attention to process feature maps and positional encoding. The decoder matches the target query with image features through cross attention while retaining the self-attention mechanism.

[0017] Prediction output module: Outputs the bounding box coordinates and confidence score of the subsidence basin, where the bounding box is represented by normalized coordinates.

[0018] Furthermore, the steps for obtaining the real deformation sample include:

[0019] Deformation maps of the mining area were obtained based on D-InSAR technology.

[0020] The deformed image is processed into image blocks of a preset size of pixels, and threshold classification is performed using the natural discontinuity grading method. Zero values ​​are added as discontinuities, and the image is converted into a color RGB image.

[0021] Mark the mining subsidence basin area on RGB imagery to form a true deformation sample. The annotation information includes bounding boxes and category labels.

[0022] Furthermore, the generation of the simulated deformation map using the dynamic probability integral method based on parameter changes includes the following steps:

[0023] The integral parameters include: subsidence coefficient q, horizontal movement coefficient b, tangent of the main influence angle tanβ, mining influence propagation angle θ0, inflection point offset distance S1~S4, and Boltzmann coefficients A3 and A4.

[0024] Based on the working face strike length D and mining depth H, the subsidence rate q' under fully exploited conditions is calculated using the improved Boltzmann function, as shown in the following formula:

[0025] ;

[0026] Substituting the dynamic subsidence rate q' into the probability integral method prediction model, the subsidence value W of any point on the ground surface during the mining period from t1 to t1+T is calculated. z As shown in the following formula:

[0027] ;

[0028] ;

[0029] Where m is the coal seam thickness. The dip angle of the coal seam;

[0030] At any time t1, the length of the working face is D. t1 Given the predicted parameters, the subsidence value W at any point on the Earth's surface can be obtained using the probability integral method prediction model. t1 After a satellite undergoes one revisit period T, its trajectory length is D. t1+T Let the advancing speed of the working face be V m / d, and calculate D. t1+T As shown in the following formula:

[0031] ;

[0032] Ignoring the effect of horizontal movement, the vertical sinking value W z Converted to radar line-of-sight deformation d LOS As shown in the following formula:

[0033] ;

[0034] in, This is the angle of incidence for the satellite.

[0035] Furthermore, the elastic transformation includes the following steps:

[0036] For each pixel (x,y) in the simulated deformation map, generate random translation amounts Δx(x,y) and Δy(x,y) in the range of -1 to 1;

[0037] Use a Gaussian convolution kernel with a mean of 0 and a standard deviation of σ. The smoothed random translation after convolution of Δx and Δy is as follows:

[0038] .

[0039] Furthermore, the application of noise synthesis includes the following steps:

[0040] Berlin noise and Gaussian noise of different frequencies and amplitudes were applied to the deformation map after elastic transformation to simulate the decorrelation effect of turbulent atmosphere and SAR imagery, where the noise intensity was related to the land cover type.

[0041] The processed deformation diagram is used as a simulated deformation sample.

[0042] Furthermore, the training steps of the deep learning model include:

[0043] Input the mixed sample set into the Deformable DETR model;

[0044] Image features are extracted using a backbone network to generate feature maps at different scales;

[0045] The feature map and the position encoding are fed into the Transformer encoder module for processing to obtain a new feature vector.

[0046] The feature vectors are matched by the Transformer decoder, and the final target category and bounding box are obtained by the prediction module.

[0047] Adjust the model parameters based on the difference between the predicted results and the true labels, and perform iterative training until the model converges.

[0048] Furthermore, the method also includes:

[0049] Non-maximum suppression is applied to the predicted bounding boxes output by the model;

[0050] The segmented prediction results are stitched together to form a complete regional deformation distribution map.

[0051] The beneficial effects of this invention are as follows:

[0052] This invention proposes an intelligent identification method for mining subsidence basins combining Deformable DETR and InSAR technologies. Specifically, by fusing real deformation samples with simulated samples based on dynamic probability integral method, elastic transformation, and noise synthesis, a diverse hybrid training set is constructed. This significantly enhances the model's adaptability to complex surface deformation and reduces its dependence on real data. The deformable attention mechanism of the Deformable DETR model can accurately capture the spatial distribution of deformation features, especially performing well when dealing with small-scale or irregularly deformed targets, effectively improving detection sensitivity. Attached Figure Description

[0053] Figure 1 This is a schematic flowchart of a method for identifying mining subsidence basins according to a specific embodiment of the present invention.

[0054] Figure 2 This is another flowchart illustrating a specific embodiment of the method for identifying subsidence basins in mining areas proposed in this invention.

[0055] Figure 3 This is a schematic diagram of the geographical location of the study area in the experimental case section of a specific embodiment of the present invention;

[0056] Figure 4 This is a schematic diagram illustrating the change process of AP values ​​in some experimental cases in a specific embodiment of the present invention;

[0057] Figure 5 This is a schematic diagram of the location of the concentrated mining subsidence basin in the experimental case section of a specific embodiment of the present invention;

[0058] Figure 6 This is a schematic diagram of the detection results of identifying six subsidence areas using three models in the experimental case section of a specific embodiment of the present invention;

[0059] Figure 7 This is a schematic diagram of the detection results of the Deformable DETR model in the experimental case section of a specific embodiment of the present invention;

[0060] Figure 8 The experimental case examples in this invention show the number of samples detected in the mining subsidence area at different time points and the AP (Average Per Second) values. 50 Value diagram;

[0061] Figure 9 This is a schematic diagram illustrating the distribution characteristics of subsidence funnels in the Huainan mining area, as part of an experimental case study in a specific embodiment of the present invention. Detailed Implementation

[0062] The present application will now be described in further detail with reference to the accompanying drawings. It should be noted that the following specific embodiments are only used to further illustrate the present application and should not be construed as limiting the scope of protection of the present application. Those skilled in the art can make some non-essential improvements and adjustments to the present application based on the above application content.

[0063] Example 1

[0064] like Figure 1-2 As shown in the figure, this embodiment proposes a method for identifying mining subsidence basins by combining Deformable DETR and InSAR technologies. The method includes: acquiring InSAR image data of the mining area to be tested through Sentinel-1A satellite, processing it with D-InSAR technology to generate a deformation interferogram, and cropping it into image patches of a set pixel size; constructing a deep learning model trained on a hybrid sample set based on the Deformable DETR model architecture and the sample area, wherein the hybrid sample set is obtained by combining real deformation samples and simulated deformation samples, and the simulated deformation samples are obtained by generating simulated deformation maps by a dynamic probability integral method based on parameter changes, and then applying noise after elastic transformation; inputting the InSAR image of the area to be tested into the trained deep learning model, and outputting the location and boundary information of the mining subsidence basin.

[0065] In this embodiment, the dynamic probability integral method introduces a variation mechanism for geological parameters such as subsidence coefficient and mining depth, enabling the simulated deformation map to reflect the deformation distribution pattern under different mining conditions. Elastic transformation and noise synthesis are applied to the simulated deformation map to simulate the geometric distortion and atmospheric interference present in real SAR images, enhancing the model's adaptability to complex noise environments.

[0066] It should be noted that the basic structure of the Deformable DETR model in this embodiment mainly includes a backbone network, a Transformer encoder, a decoder, and a prediction module. The specific process is as follows: First, sample data is input into the backbone network to extract image features and generate feature maps at different scales. Then, these feature maps, along with positional encodings, are fed into the Transformer encoder module for processing, resulting in new feature vectors. Finally, the Transformer decoder matches the feature vectors, and the prediction module then determines the final target category and bounding box. Compared to DETR (DEtection Transformer is an end-to-end object detection model based on the Transformer architecture), Deformable DETR replaces the multi-head attention module in the Transformer encoder with a multi-scale deformable attention module, and replaces the cross-attention module in the Transformer decoder with a multi-scale deformable cross-attention module, while the self-attention module in the Transformer decoder remains unchanged.

[0067] In a preferred embodiment, the Deformable DETR model architecture includes a feature extraction module, a deformable attention module, a Transformer encoding / decoding module, and a prediction output module. The feature extraction module uses a ResNet50 backbone network to extract multi-scale feature maps from the input InSAR image. The deformable attention module includes multi-scale deformable attention layers for calculating attention features. Each attention head samples K sampling points, given an input feature map. C represents the number of color channels in the image, H represents the height of the image, and W represents the width of the image, which has content features z. q and two-dimensional reference point p q The attention features of the query elements satisfy the following formula:

[0068] ;

[0069] In the formula: m is the attention head index; M is the total number of attention heads; k is the sampling point; K is the total number of sampling points, J< <H×W;W m and W' m For feature weights; A mqk ∆p represents the attention weight of the k-th sampling point in the m-th attention head. mqk This is the corresponding sampling offset.

[0070] Optionally, the deformable attention module performs weighted fusion of multi-scale feature maps, where feature maps of different scales are aggregated through adaptive weights; given the input multi-feature mapping ,in ,set up Using the normalized coordinates of the reference point, the multi-scale deformable attention features can be calculated using the following equation:

[0071] ;

[0072] In the formula: is the input feature level; L is the total number of feature layers; is the normalized coordinate of the reference point; is the sampling offset between the k-th sampling point of the l-th feature level and the m-th attention head; is the sampling offset between the k-th sampling point of the l-th feature level and the m-th attention head. The attention weight of the k-th sampling point of the m-th attention head in each feature level; and the input feature map of the l-th layer, which is the normalized coordinates rescaled to the l-th layer.

[0073] The Transformer encoding / decoding module includes an encoder and a decoder. The encoder uses multi-scale deformable attention instead of standard multi-head attention to process feature maps and positional encoding. The decoder matches the target query with image features through cross-attention while retaining the self-attention mechanism. The prediction output module outputs the bounding box coordinates and confidence scores of the subsidence basin, where the bounding box is represented by normalized coordinates.

[0074] In the preferred embodiment described above, the Deformable DETR model architecture is implemented as follows: The feature extraction module uses a pre-trained ResNet50 as the backbone network, and extracts multi-scale feature maps from the input InSAR image through progressive downsampling to form a multi-level feature pyramid. To enhance adaptability to irregular depression boundaries, deformable convolutional layers are introduced in the deeper stages of the backbone network.

[0075] The deformable attention module employs a multi-layer, multi-scale deformable attention structure, with each attention head sampling several key points. For each query element, its attention weight is calculated through a fully connected network, while a sampling position offset is generated through an independent offset prediction network. During multi-scale feature fusion, an adaptive weighting mechanism is used to balance the contributions of features at different scales.

[0076] The Transformer encoder-decoder module consists of stacked encoder and decoder layers. The encoder uses multi-scale deformable attention instead of the standard attention mechanism to process feature maps with positional encoding. The decoder achieves dynamic matching of the target query with image features through a combination of self-attention and cross-attention. The initial target query is initialized with learnable parameters and gradually optimized during training.

[0077] The prediction output module comprises multiple parallel prediction branches, each handling tasks such as bounding box regression and class confidence estimation. During training, an improved matching strategy and a composite loss function are employed to ensure the model comprehensively learns the features of the subsidence basin. In the inference phase, an optimized post-processing workflow effectively filters out low-quality predictions and eliminates redundant detection boxes.

[0078] This architecture incorporates several optimizations to address the specific needs of subsidence detection in mining areas: an enhanced feature extraction network improves the ability to capture complex deformation patterns; an improved multi-scale attention mechanism strengthens the detection performance of small targets; and a newly added auxiliary prediction branch enhances the accuracy of quantitative deformation analysis. These optimizations collectively improve the model's detection performance in complex mining environments.

[0079] In a preferred embodiment, the steps for obtaining real deformation samples include: obtaining a mining area deformation map of the sample area based on D-InSAR technology; processing the deformation map into image blocks of a preset size of pixels, performing threshold classification using the natural discontinuity grading method, adding zero values ​​as discontinuities, and converting it into a color RGB image; marking the mining subsidence basin area on the RGB image to form a real deformation sample, with the marking information including bounding boxes and category labels.

[0080] In the preferred embodiment described above, the acquisition of real deformation samples specifically employs D-InSAR technology to process SAR image data from the Sentinel-1A satellite. To enhance the visualization and feature distinguishability of deformation information, the original deformation map is divided into standardized image blocks of a preset size, and intelligent classification is performed using a natural discontinuity grading method. Zero values ​​are specifically introduced as key discontinuities in this process to effectively distinguish deformed areas from the stable background. Subsequently, the classification results are converted into pseudo-color RGB images, and different levels of subsidence characteristics are visually presented through color gradients.

[0081] During the sample annotation phase, professional surveyors, combining geological exploration data and field measurement data, accurately marked the spatial extent of the mining subsidence basins on RGB imagery. The annotation information was stored in a standardized vector bounding box format, while also recording the category attributes (such as active subsidence area, stable subsidence area, etc.) and deformation characteristic parameters of each subsidence basin.

[0082] In a preferred embodiment, the simulated deformation diagram is generated based on the dynamic probability integral method of parameter variation, including the following steps: setting the integral parameters including: subsidence coefficient q, horizontal movement coefficient b, tangent of the main influence angle tanβ, mining influence propagation angle θ0, inflection point offset distance S1~S4, and Boltzmann coefficients A3 and A4; calculating the subsidence rate q' under fully exploited conditions using the improved Boltzmann function based on the working face strike length D and mining depth H, as shown in the following formula:

[0083] ;

[0084] Substituting the dynamic subsidence rate q' into the probability integral method prediction model, the subsidence value W of any point on the ground surface during the mining period from t1 to t1+T is calculated. z As shown in the following formula:

[0085] ;

[0086] ;

[0087] Where m is the coal seam thickness. The dip angle of the coal seam;

[0088] At any time t1, the length of the working face is D. t1 Given the predicted parameters, the subsidence value W at any point on the Earth's surface can be obtained using the probability integral method prediction model. t1 After a satellite undergoes one revisit period T, its trajectory length is D. t1+T Let the advancing speed of the working face be V m / d, and calculate D. t1+T As shown in the following formula:

[0089] ;

[0090] Ignoring the effect of horizontal movement, the vertical sinking value W z Converted to radar line-of-sight deformation d LOS As shown in the following formula:

[0091] ;

[0092] in, This is the satellite's angle of incidence.

[0093] In a preferred embodiment, the specific implementation process of generating simulated deformation maps based on the dynamic probability integral method of parameter changes is as follows: First, a complete parameter system including geological mining parameters and satellite observation parameters is established, wherein the geological parameters include coal seam thickness (m) and dip angle. Inherent properties, mining parameters include dynamic variables such as working face length and advance speed, while satellite parameters include incident angle. Observation conditions such as revisit period T were used. A dynamic calculation model for subsidence rate was established using an improved Boltzmann function. This model, by introducing a mining degree correction coefficient (D / H represents the ratio of strike length to mining depth), can accurately characterize the nonlinear subsidence characteristics of the entire process from initial mining to full mining operation.

[0094] In the deformation simulation calculation stage, a time discretization method is used to decompose the continuous mining process into multiple calculation time steps. Each time step calculates the three-dimensional subsidence field based on the current working face location and mining parameters using the probability integral method, focusing on mining boundary effects and inflection point offset characteristics. To adapt to InSAR observation characteristics, a geometric projection model is used to convert vertical subsidence into radar line-of-sight deformation, and a Gaussian random field is introduced to simulate minor perturbations during the mining process. By adjusting parameter combinations, diverse deformation patterns covering different geological conditions, mining stages, and observation geometries can be generated, effectively expanding the coverage of the training samples.

[0095] In the post-processing stage of the simulated data, an elastic deformation algorithm is used to enhance the spatial variability of the deformation field. By controlling the correlation length and amplitude parameters of the displacement field, an irregular deformation pattern that conforms to the actual geomechanical characteristics is generated. Finally, a composite noise field that conforms to the characteristics of SAR imaging is superimposed, including spatially correlated fractal noise and speckle noise that conforms to radar scattering characteristics, so that the simulated data is consistent with the statistical characteristics of real InSAR observations.

[0096] In a preferred embodiment, the elastic transformation includes the following steps: generating random translation amounts Δx(x,y) and Δy(x,y) within the range of -1 to 1 for each pixel (x,y) in the simulated deformation image; and employing a Gaussian convolution kernel with a mean of 0 and a standard deviation of σ. The smoothed random translation after convolution of Δx and Δy is as follows:

[0097] .

[0098] The noise synthesis process includes the following steps: applying Berlin noise and Gaussian noise of different frequencies and amplitudes to the elastically transformed deformation map to simulate the decorrelation effect of turbulent atmosphere and SAR imagery, wherein the noise intensity is related to the land cover type; and using the processed deformation map as a sample for simulated deformation.

[0099] Regarding elastic transformation, as an example, displacement field generation: For each pixel (x, y) of the simulated deformation map, two independent random displacement fields Δx(x, y) and Δy(x, y) are first generated, with their values ​​uniformly distributed in the interval [-1, 1]. Displacement field smoothing: A two-dimensional Gaussian convolution kernel is used for smoothing, where the standard deviation σ is set to a range of 3-7 pixels according to the target deformation scale, and the convolution kernel size is (6σ+1)×(6σ+1). Deformation field application: The smoothed displacement field is applied to the original deformation map through bilinear interpolation, where the displacement amplitude coefficient α is controlled between 0.1 and 0.3 to ensure that the deformation pattern maintains both geological rationality and natural variation characteristics.

[0100] Regarding noise synthesis, as an example, Perlin noise generation: A fractal noise synthesis method is used, superimposing five layers of Perlin noise with different frequencies (wavelengths from 5 pixels to 50 pixels) and amplitudes (decreasing with frequency) to simulate phase disturbances caused by atmospheric turbulence. Gaussian noise addition: Differentiated noise levels are set according to the land cover type, where: vegetated areas: SNR=10-15dB; bare soil areas: SNR=15-20dB; water areas: SNR=5-10dB; Composite noise fusion: Perlin noise and Gaussian noise are mixed with a 3:1 weight, and morphological filtering is used to enhance the spatial correlation of the noise, making it more consistent with the noise characteristics of real SAR imagery.

[0101] In a preferred embodiment, the training steps of the deep learning model include: inputting a mixed sample set into the Deformable DETR model; extracting image features using a backbone network to generate feature maps at different scales; feeding the feature maps and location encodings together into a Transformer encoder module for processing to obtain new feature vectors; matching the feature vectors through a Transformer decoder and obtaining the final target category and bounding box through a prediction module; adjusting the model parameters based on the difference between the prediction results and the true labels, and performing iterative training until the model converges.

[0102] In the preferred embodiment described above, the training process of the deep learning model employs an end-to-end optimization framework. First, the mixed sample set undergoes standardization and spatial augmentation preprocessing to ensure the uniformity and diversity of data distribution. The backbone network, based on a pre-trained convolutional neural network architecture, generates a feature pyramid rich in semantic information through multi-level feature extraction. These feature maps, combined with learnable location encodings, are input into an improved Transformer encoder module, which uses a multi-scale deformable attention mechanism to capture the correlation of deformation features across different ranges. The decoder part gradually optimizes the target prediction results through the dynamic interaction between learnable query vectors and encoded features. A phased optimization strategy is adopted during training: initial training with fixed backbone network parameters, followed by overall fine-tuning of model parameters. The loss function design comprehensively considers the multi-task requirements of bounding box accuracy, classification confidence, and deformation estimation, guiding the model optimization direction through weighted combination. The optimizer uses an adaptive learning rate algorithm combined with warm-up and annealing scheduling strategies to balance the stability and convergence efficiency of the training process. The entire training process iteratively optimizes the model to gradually master the key features for identifying subsidence basins from complex InSAR imagery, ultimately achieving stable detection performance. To improve the detection performance of small targets, the loss weight for small-scale depression samples was strengthened in the later stages of training, and the recognition boundary of the model was improved through a hard sample mining strategy.

[0103] In a preferred embodiment, the method further includes: performing non-maximum suppression on the predicted bounding boxes output by the model; and stitching the block prediction results together to form a complete regional deformation distribution map. The output regional deformation distribution map contains three types of information: the spatial location and boundaries of the subsidence basin, the deformation intensity level classification, and the detection confidence assessment.

[0104] According to the above embodiments, the working principle of the present invention is as follows: intelligent identification of mining subsidence basins is achieved through a multi-stage processing flow. First, satellite SAR images are processed using D-InSAR technology to obtain surface deformation interferograms and construct a hybrid training set containing real and simulated samples. The simulated samples are generated using a parameterized dynamic probability integral method, combined with elastic transformation and composite noise synthesis techniques to enhance data authenticity. Subsequently, an improved Deformable DETR model is used for feature learning and target detection: the backbone network extracts multi-scale deformation features, the deformable attention module focuses on key deformation areas through sparse sampling and dynamic offset prediction, and the Transformer encoding and decoding structure realizes global context modeling and local feature matching. Model training adopts multi-task loss function optimization, combined with a staged training strategy to improve convergence efficiency. Finally, the prediction results output by the model are intelligently post-processed, including non-maximum suppression based on geometric and semantic information, block prediction stitching, and deformation field reconstruction, generating a complete subsidence distribution map containing location boundaries, deformation intensity, and confidence assessment.

[0105] To more clearly illustrate the present invention and its advantages, the method provided by the present invention will be further explained below with reference to specific experimental examples and related partial figures.

[0106] Experimental environment and evaluation indicators

[0107] 1.1 Experimental Environment

[0108] PyTorch (GPU version) was used as the deep learning framework to build the experimental environment for training and testing. The CPU was a 12th Gen Intel(R) Core(TM) i5-12400F, the GPU was an NVIDIA GeForce RTX 4060 Ti with 16GB of VRAM, the deep learning framework was PyTorch 2.00, and the programming language was Python 3.9. The batch size was set to 2, and the training epochs were 300. The initial learning rate was 0.0001, which decreased continuously with subsequent training. The Adam optimizer was used to optimize and adjust the network parameters, and the validation set performance was evaluated every 10 epochs. To verify the performance of the Deformable DETR network model in wide-area sedimentation recognition, the widely used Faster-RCNN and DETR model were selected for comparison.

[0109] 1.2 Evaluation Indicators

[0110] To evaluate the model's performance, mean average precision (mAP), precision (P), recall (R), F1 score, and mean intersection over union (MIoU) were used as evaluation metrics. Since the mining area has only one category of target, "mining subsidence basin", mAP = AP in this experimental case.

[0111] Experimental Results and Analysis

[0112] 2.1 Overview of the Study Area

[0113] The Huainan mining area is located between 115°50' and 117°45' east longitude and 32°25' and 33°10' north latitude, mainly distributed in the area north of 27°40' north latitude. It extends from the Tanlu Fault Zone in the east to eastern Fuyang in the west, from Minglong Mountain and Shangyao in the north to Shungeng Mountain and Bagong Mountain in the south, spanning Huainan, eastern Fuyang, and Bozhou counties and cities. The mining area is approximately 180 km long from east to west and 15-25 km wide from north to south, with a total area of ​​approximately 3600 km². The Huainan mining area is situated on the Huaihe River alluvial plain; except for some hilly and ridged areas, the rest is alluvial plain, with a surface elevation generally between +20 and +30 m. The Huainan mining area has a long mining history, abundant mineral resources, and numerous and significant coal mining subsidence areas. Therefore, this experiment selected the Huainan mining area as the experimental zone. The distribution of mines in the Huainan mining area is as follows: Figure 3 As shown.

[0114] 2.2.1 Experimental Results and Analysis

[0115] This paper selects eight “Sentinel-1A” IW mode images covering the Huainan mining area from November 9, 2023 to February 1, 2024, forming 18 interferometric pairs for the construction of the test set. Detailed parameters are shown in Table 1.

[0116] Table 1 Sentine-1A Image Parameters

[0117] ;

[0118] The experiment used the method constructed in this application to automatically identify mining subsidence basins in the converted RGB images. Before model prediction, each image needs to be segmented into 512×512 pixels with zero overlap. The edges are not processed. A strategy of block prediction and mosaicking to restore the whole image is adopted.

[0119] Faster R-CNN (Faster Region-based Convolutional Neural Network) is a deep learning model for object detection that achieves a good balance between accuracy and speed, and is a significant improvement on the R-CNN series (including Fast R-CNN). This network has a deep structure, making it suitable for various application scenarios. Therefore, to verify the effectiveness of Deformable DETR in identifying subsidence and mobile basins, this experiment uses Deformable DETR, DETR, and Faster R-CNN networks to identify mining subsidence basins based on training samples. The model training results are compared in Table 2. Figure 4 The curves show the AP value as a function of the number of iterations during the training process for the three models.

[0120] Table 2. Identification results of the three networks in the subsidence basin dataset.

[0121] ;

[0122] As shown in Table 2, Deformable DETR outperforms both DETR and Faster-RCNN in overall detection performance, achieving an accuracy of 92%. This model not only demonstrates a balanced performance in precision and recall but also surpasses the other two models in MIoU, showcasing its superior performance in locating and accurately segmenting targets of varying sizes. Figure 6 The model training process shows that Deformable DETR exhibits faster convergence speed and smaller AP value fluctuations, reflecting that the model has better convergence speed and generalization performance, and is more suitable for complex and variable target detection tasks.

[0123] Figure 5 The map shows the distribution of mining subsidence areas, including six subsidence zones (A, B, C, D, E, and F). Color coding is used to visually represent the differences in subsidence among these zones. Three models were used to identify the six subsidence zones, and the results are as follows. Figure 6As shown in the figure, Faster-RCNN has high accuracy in detecting individual targets, but its overall recognition accuracy is low, with three targets being falsely detected. DETR's candidate box fitting is unstable, resulting in two false detections, and its accuracy is low for targets E with small subsidence areas. Overall, Faster-RCNN and DETR identify more candidate boxes than the true value, leading to a high false detection rate. Deformable DETR, on the other hand, delineates boundaries closer to the true value, with a lower false detection rate than the other two models. It performs well in recognizing regularly shaped mining subsidence targets, achieving a confidence level of over 90%. The confidence level for recognizing deformed targets E with subsidence amounts ranging from 1mm to 9mm reaches 88%. It exhibits strong anti-interference capabilities and can more accurately identify the location information of mining subsidence areas from complex backgrounds.

[0124] The experiment conducted detailed statistical analysis on all detected subsidence basins with different subsidence ranges, revealing that the model could detect deformations ranging from less than 1 mm to over 100 mm. The performance of the three models in different deformation zones was also compared. In areas of significant deformation, the Deformable DETR model achieved higher accuracy than DETR, with slight lower accuracy for individual targets compared to Faster-RCNN. When deformation was small and visual recognition was not obvious, Deformable DETR successfully identified targets with high confidence and higher accuracy than the DETR model; in contrast, Faster R-CNN failed to detect any targets. This result further demonstrates that Deformable DETR can detect a wide range of deformations and exhibits excellent adaptability when handling complex image features. This advantage may stem from its flexible feature extraction and dynamic adjustment capabilities, maintaining high detection accuracy even in mining subsidence basins of varying shapes and sizes.

[0125] Overall, from November 2023 to February 1, 2024, a total of 451 mining subsidence areas were detected. These mining subsidence basins are mainly distributed in the northeastern part of Yingshang County, the central part of Fengtai County, and Panji District. Ground subsidence caused by mining mainly occurred at Liuzhuang Coal Mine, Xieqiao Coal Mine, Zhangji Coal Mine, and Guqiao Coal Mine. The mining subsidence areas detected by the model are consistent with the distribution of the main mines. The test identification results are as follows... Figure 7 As shown.

[0126] 2.2.2 Transferability Validation of the Model

[0127] To verify the transferability of Deformable DETR in identifying mining subsidence at a wide-scale, the optimal model trained on a hybrid dataset was applied to a dynamically updated dataset, and the model's real-time performance was analyzed. Based on data from November 9, 2023 to February 1, 2024, 22 Sentinel-1 satellite images from February 13, 2024 to October 22, 2024 were supplemented. Interferometric results with a 24-day time baseline were used for continuous monitoring of the Huainan mining area, acquiring a total of 20 deformation maps. When newly acquired data became available, the experiment performed the same operations on the newly acquired data as on historical images.

[0128] 454 settlement zones were identified through manual visual interpretation, and the model detected a total of 402. The accuracy evaluation of Deformable DETR is shown in Table 3.

[0129] Table 3. Accuracy Evaluation Table of Dynamically Updated Data Results

[0130] ;

[0131] As shown in Table 3, Deformable DETR achieved an accuracy of 88.39% on the dynamically updated dataset, indicating that the model has excellent recognition capabilities and few false positives. The F1 score reached 82%, reflecting a good balance between precision and recall, effectively identifying most subsidence areas. In terms of inference efficiency, the average inference time per image was 0.2 seconds, meeting the rapid response requirements for real-time subsidence identification. These results demonstrate that Deformable DETR has a certain transferability to time-varying data, quickly adapting to new datasets and providing accurate recognition results.

[0132] Although the Deformable DETR model performed well on the initial test set, its recall and MIoU were lower than its test precision on new datasets. This indicates that there are more missed detections on dynamically updated datasets, and the overlap between the predicted results and the true labels is reduced. This change may be related to differences in data quality. To further evaluate the performance changes of Deformable DETR on dynamic datasets, this experiment analyzed the number of mining subsidence areas identified by the model in images from different time periods and their corresponding AP values. 50 The values ​​were compared and analyzed, and the results are as follows: Figure 8 As shown in the figure, the number of mining subsidence areas detected by the model remained relatively stable from February to May 2024, while AP 50 The values ​​also remained at a high level, indicating high detection accuracy during this period. However, from June to July, the number of subsidence areas and AP values... 50The values ​​all decreased significantly, indicating an increased false negative rate and decreased detection performance during this period. This phenomenon may be related to the dense vegetation cover in summer. The dense vegetation cover may cause signal absorption or scattering, making the concentric circles or elliptical fringes in the interference results blurred or missing, thus increasing the difficulty for the model to identify and locate mining subsidence. In contrast, in winter, due to less vegetation cover, the deformation fringes are clearer, and the model can more accurately identify mining subsidence moving basins.

[0133] The spatial distribution of mining subsidence basins detected by Deformable DETR from February 13, 2024 to October 22, 2024, is as follows: Figure 9 As shown in the figure, the yellow area represents the mining area, and the red dots represent mining subsidence areas detected by the model. Without considering the influence of summer vegetation, the number of subsidence areas is relatively uniform, without sudden increases or decreases. The subsidence areas are mainly concentrated in the central and eastern parts, which is basically consistent with the distribution of the main mining areas.

[0134] Furthermore, this experiment also observed a small number of false detection areas in the model recognition results located near water bodies. These areas exhibit morphological characteristics similar to those of mining areas in the deformation map. For example, in the shallow waters near a tributary of the Huai River, the areas showed subsidence with distinct boundaries during imaging, making it difficult for the model to accurately interpret the data.

[0135] Comparative analysis of dataset selection

[0136] 3.1 Optimize dataset comparison

[0137] To test the advantages of the proposed hybrid sample set in monitoring mining subsidence and mobile basins, the model was trained using a real dataset and then retrained using the hybrid dataset. The detection accuracy of the three models was compared across different datasets. During training, all algorithms used the same parameter settings, and the comparison results are shown in Table 4. (Table shows AP...) L For high-precision large target recognition (targets with an area greater than 96×96 pixels); AP M For medium target average accuracy (targets with an area between 32×32 and 96×96 pixels); AP S The average accuracy for small targets (targets with an area smaller than 32×32 pixels) is calculated; and the average detection accuracy at Intersection over Union (IoU) of 50 and 75 is introduced as an evaluation metric (AP). 50 and AP 75 ).

[0138] Table 4. Detection accuracy of the three models trained on different datasets.

[0139] ;

[0140] Experimental results show that all three models improve mAP on the mixed dataset compared to training only on the real dataset. Specifically, Faster-RCNN improves mAP by 0.7%, DETR by 1.5%, and DeformableDETR shows the most significant improvement, reaching 1.6%.

[0141] At different IoU thresholds (AP) 50 and AP 75 Under these conditions, Faster-RCNN improved across all metrics on the mixed dataset, but the improvements were small. DETR and Deformable DETR showed a slight decrease in AP50, but still maintained high detection accuracy. For AP... 75 This more stringent standard resulted in improved performance for all models. Specifically, the Deformable DETR achieved higher AP on the mixed dataset. 75 The most significant improvement was achieved, reaching 6.3%. Although performance declined at lower IoU thresholds, it still yielded significant performance gains in high-precision detection tasks. Furthermore, performance analysis at different target scales revealed that the Deformable DETR exhibited higher accuracy in small target detection tasks, with its AP... S It increased from 0.550 to 0.600, an increase of 0.5%.

[0142] In conclusion, introducing mixed datasets has a positive impact on model performance, effectively enhancing the model's learning ability and improving its recognition performance in mining subsidence areas at different scales, thereby improving the overall efficiency of the detection task.

[0143] This invention, through systematic experimental verification, fully demonstrates the superiority and practicality of the mining subsidence identification method integrating InSAR technology and Deformable DETR. Experimental results show that the hybrid training set, composed of simulated data generated using the dynamic probability integral method and real InSAR deformation maps, significantly improves the detection performance of the deep learning model, enabling Deformable DETR to exhibit excellent identification capabilities in complex mining environments. In practical applications in the Huainan mining area, this method successfully detected hundreds of mining subsidence areas with an accuracy rate of 92.6%, verifying its reliability in wide-area monitoring. Particularly noteworthy is that the model maintains an accuracy rate of 88.39% even with dynamically updated data, demonstrating good temporal adaptability and engineering practical value. Compared to traditional detection methods, this scheme, through the innovative combination of physical models and deep learning, not only solves the problem of small-scale deformation detection but also significantly improves the efficiency of automated processing. These achievements provide a new technological paradigm for geological disaster monitoring in mining areas. Its core innovations—the hybrid sample construction strategy and the deformable attention mechanism—provide a generalizable solution for similar surface deformation detection tasks.

[0144] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0145] In addition, the functional modules in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0146] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for identifying mining subsidence basins by combining Deformable DETR and InSAR technologies, characterized in that, The method includes: Acquire InSAR imagery of the area to be measured; A deep learning model is constructed based on the Deformable DETR model architecture and a hybrid sample set of sample regions. The hybrid sample set is obtained by combining real deformation samples and simulated deformation samples. The simulated deformation samples are synthesized by generating simulated deformation maps using a dynamic probability integral method based on parameter changes, followed by elastic transformation and the addition of noise. Input the InSAR image of the area to be tested into the trained deep learning model, and output the location and boundary information of the coal mining subsidence basin; The method for generating simulated deformation diagrams based on the dynamic probability integral method of parameter changes includes the following steps: The integral parameters include: subsidence coefficient q, horizontal displacement coefficient b, and the tangent of the principal influence angle. The angle of propagation of the impact of mining Inflection point offset S1~S4, Boltzmann coefficients A3 and A4; Based on the working face strike length D and mining depth H, the subsidence rate under fully exploited conditions is calculated using the improved Boltzmann function. As shown in the following formula: ； Substituting the dynamic subsidence rate q' into the probability integral method prediction model, the subsidence value W of any point on the ground surface during the mining period from t1 to t1+T is calculated. z As shown in the following formula: ； Where m is the coal seam thickness. The dip angle of the coal seam; At any time t1, the length of the corresponding working face is Given the predicted parameters, the subsidence value at any point on the Earth's surface can be obtained using a probability integral method prediction model. After a satellite completes one revisit period T, its trajectory length is... Let the advancing speed of the working face be V m / d, and calculate... As shown in the following formula: ； Ignoring the effect of horizontal movement, the vertical sinking value W z Converted to radar line-of-sight deformation As shown in the following formula: ； in, This is the satellite's angle of incidence.

2. The method for identifying mining subsidence basins combining Deformable DETR and InSAR technologies according to claim 1, characterized in that, The Deformable DETR model architecture includes: The feature extraction module uses a ResNet50 backbone network to extract multi-scale feature maps from the input InSAR image. The deformable attention module includes a multi-scale deformable attention layer for computing attention features; Each attention head samples K points, given an input feature map. C represents the number of color channels in the image, H represents the height of the image, and W represents the width of the image, which has content features z. q and two-dimensional reference point p q The attention features of the query elements satisfy the following formula: ； In the formula: Note the header index; This represents the total number of attention heads. For sampling points; For the total sampling points, ; W m and For feature weights; For the first The first one to pay attention to Attention weights for each sampling point; This is the corresponding sampling offset; The Transformer encoder-decoder module includes an encoder and a decoder. The encoder uses multi-scale deformable attention instead of standard multi-head attention to process feature maps and positional encoding. The decoder matches the target query with image features through cross attention while retaining the self-attention mechanism. Prediction output module: Outputs the bounding box coordinates and confidence score of the subsidence basin, where the bounding box is represented by normalized coordinates.

3. The method for identifying mining subsidence basins combining Deformable DETR and InSAR technologies according to claim 1, characterized in that, The steps for obtaining the real deformation sample include: Deformation maps of the mining area were obtained based on D-InSAR technology. The deformed image is processed into image blocks of a preset size of pixels, and threshold classification is performed using the natural discontinuity grading method. Zero values ​​are added as discontinuities, and the image is converted into a color RGB image. Mark the mining subsidence basin area on RGB imagery to form a true deformation sample. The annotation information includes bounding boxes and category labels.

4. The method for identifying mining subsidence basins combining Deformable DETR and InSAR technologies according to claim 1, characterized in that, The elastic transformation includes the following steps: Generate a random translation amount in the range of -1 to 1 for each pixel (x,y) in the simulated deformation map. and ; Use a Gaussian convolution kernel with a mean of 0 and a standard deviation of σ. The smoothed random translation after convolution of Δx and Δy is as follows: 。 5. The method for identifying mining subsidence basins combining Deformable DETR and InSAR technologies according to claim 1, characterized in that, The applied noise synthesis includes the following steps: Berlin noise and Gaussian noise of different frequencies and amplitudes were applied to the deformation map after elastic transformation to simulate the decorrelation effect of turbulent atmosphere and SAR imagery, where the noise intensity was related to the land cover type. The processed deformation diagram is used as a simulated deformation sample.

6. The method for identifying mining subsidence basins combining Deformable DETR and InSAR technologies according to claim 1, characterized in that, The training steps of the deep learning model include: Input the mixed sample set into the Deformable DETR model; Image features are extracted using a backbone network to generate feature maps at different scales; The feature map and the position encoding are fed into the Transformer encoder module for processing to obtain a new feature vector. The feature vectors are matched by the Transformer decoder, and the final target category and bounding box are obtained by the prediction module. Adjust the model parameters based on the difference between the predicted results and the true labels, and perform iterative training until the model converges.

7. The method for identifying mining subsidence basins combining Deformable DETR and InSAR technologies according to claim 1, characterized in that, The method further includes: Non-maximum suppression is applied to the predicted bounding boxes output by the model; The segmented prediction results are stitched together to form a complete regional deformation distribution map.

Images