Mine subsidence identification and extraction method based on multi-source monitoring data fusion processing

Abstract

The application discloses a mine subsidence identification and extraction method based on multi-source monitoring data fusion processing, and relates to the field of mine subsidence monitoring and segmentation. The method comprises the following steps: a mine subsidence identification model comprising a convolutional neural network and a spatial adaptive regression model is constructed, multi-source monitoring data of a research mine area is acquired, and the multi-source monitoring data is input into the mine subsidence identification model to extract a multi-dimensional feature data set; the convolutional neural network outputs subsidence prediction probabilities of pixels in the research mine area; the spatial adaptive regression model performs spatial weighted regression processing on the subsidence prediction probabilities of the pixels in the research mine area to obtain corrected subsidence prediction probabilities, and then obtains fusion probabilities of the pixels. The application can quickly screen a research area to a complete cascade extraction process of fine segmentation, spatial correction and multi-source and multi-temporal fusion from a large range, and realizes high-precision, automatic and dynamic updating extraction and interpretation processing of a subsidence area of the research mine area.

Description

Technical Field

[0001] This invention relates to the field of mining area subsidence monitoring and segmentation, and in particular to a mining area subsidence identification and extraction method based on multi-source monitoring data fusion processing. Background Technology

[0002] Subsidence zones caused by underground mining operations are surface depressions formed when the overlying strata lose support and shift, deform, or collapse after the underground coal seam is extracted. These subsidence zones typically manifest as slow subsidence, surface cracks, and sinkholes, with blurred boundaries and dynamic evolution over time. They are important targets for geological hazard monitoring and ecological environment assessment in mining areas.

[0003] Traditional methods for identifying the extent of subsidence areas primarily rely on manual field measurements and ground subsidence monitoring networks. While these methods can acquire high-precision point data, they suffer from high costs, low efficiency, limited coverage, and difficulty in achieving continuous monitoring over large areas. With the development of remote sensing technology, using remote sensing to invert and monitor the extent of subsidence areas has gradually become a development direction. Currently, the main development directions include the following two categories: First, using satellite remote sensing to monitor underground mining areas and acquire satellite remote sensing data. Based on machine learning methods (such as support vector machines and random forests), pixel classification is performed using spectral and textural features. While this method has some ability to identify obvious subsidence areas, its generalization ability is insufficient in areas with blurred subsidence boundaries and slow subsidence. Second, using differential interferometric synthetic aperture radar (D-InSAR) technology to detect surface deformation through phase information in underground mining areas. Although this method can reflect subsidence trends, it is susceptible to interference from atmospheric delay and decoherence, resulting in limited boundary positioning accuracy. Existing technologies generally suffer from the following shortcomings: First, the boundaries of subsidence areas are blurred and the deformation is weak, making it difficult for traditional methods to capture them accurately; Second, machine learning methods and deep learning models have limitations in capturing weak signals and complex scenes, have poor adaptability, and the lack of an attention mechanism leads to insufficient response to weak features; Third, existing technologies ignore the dynamics and cumulative effects of the subsidence process, and cannot meet the needs of time-series monitoring. Summary of the Invention

[0004] The purpose of this invention is to solve the problems of poor adaptability and insufficient dynamic monitoring capability of existing technologies, and to provide a mining area subsidence identification and extraction method based on multi-source monitoring data fusion processing. This method can quickly screen large areas of the research area to fine segmentation, spatial correction and multi-source, multi-temporal fusion, and complete cascade extraction process, realizing high-precision, automated and dynamically updated extraction and interpretation of subsidence areas in underground mine research areas.

[0005] The objective of this invention is achieved through the following technical solution: A method for identifying and extracting subsidence in mining areas based on multi-source monitoring data fusion processing, the method comprising: S1. Construct a mining area subsidence identification model that includes a convolutional neural network and a spatial adaptive regression model. Obtain multi-source monitoring data of the research mining area and input it into the mining area subsidence identification model to extract a multi-dimensional feature dataset including texture, spectrum, deformation, and shape. S2. The convolutional neural network includes an encoder, a decoder, and an output module. The encoder weights and encodes the multidimensional feature dataset along the channel dimension to obtain a fused feature A. The decoder weights and decodes the fused feature A along the spatial dimension to obtain a fused feature B. The encoder and decoder have one-to-one correspondences and skip connections between their layers. The output module uses the fused feature B to map and output the subsidence prediction probability of pixel i in the mining area. ; S3. The spatial adaptive regression model uses the subsidence prediction probability of pixel i in the research mining area to perform spatial weighted regression processing to obtain the corrected subsidence prediction probability. The subsidence identification model for mining areas obtains the fusion probability of pixel i according to the following formula. : ,in , These are the weighting coefficients.

[0006] To better achieve the present invention, the present invention also includes the following methods: S4. The spatial adaptive regression model uses the fusion probability of all pixels in the study area to obtain the fusion probability map of the study area. Clustering is performed on the fusion probability map to obtain the initial subsidence area. Canny edge detection is used to detect and identify the boundary of the initial subsidence area based on a preset threshold. Then, morphological operations are used to optimize the boundary to obtain the subsidence area segmentation result. The subsidence area segmentation result includes the internal area of ​​the subsidence area and the boundary of the subsidence area.

[0007] Furthermore, the present invention also includes the following method: S5. The spatial adaptive regression model obtains the subsidence area segmentation results of the study area in the preceding and following time phases according to methods S1 to S4, and extracts the boundary vectors. The subsidence change detection results are output using the boundary vectors of the preceding and following time phases. The subsidence change detection results include subsidence expansion, static stability of subsidence, dynamic stability of subsidence, and subsidence shrinkage. Subsidence expansion indicates that the boundary vector of the following time phase is greater than that of the preceding time phase. Subsidence shrinkage indicates that the boundary vector of the following time phase is less than that of the preceding time phase. Static stability of subsidence indicates that the boundary vectors of the preceding and following time phases are the same. Dynamic stability of subsidence indicates that the boundary vector area of ​​the preceding and following time phases remains dynamically equal under the subsidence restoration treatment.

[0008] Furthermore, the present invention also includes the following method: S6. The spatial adaptive regression model obtains the subsidence zone segmentation results of the study area in the preceding and following time phases according to methods S1 to S4, and extracts the boundary vectors. The boundary update response is then calculated according to the following expression. : , , Let be the boundary vectors for the previous time phase t and the next time phase t+1, respectively. Set a boundary update threshold. If the boundary update response... If the value exceeds the boundary update threshold, a boundary update is triggered, and the spatial adaptive regression model outputs the updated subsidence area segmentation result.

[0009] Preferably, in method S1, the multi-source monitoring data includes multi-source remote sensing image monitoring data and digital elevation model data covering the research mining area. The multi-source remote sensing image monitoring data includes Landsat-8 multispectral optical image, Sentinel-2 multispectral optical image and SAR image data. The multi-source remote sensing image monitoring data undergoes preprocessing including geometric correction, radiometric correction, atmospheric correction, resampling and coordinate registration.

[0010] Preferably, in method S1, the method for the mining area subsidence identification model to obtain a multi-dimensional feature dataset is as follows: The LBP texture feature vector and HOG texture feature vector were extracted using the RotoHog texture feature extraction method and then concatenated to obtain multidimensional texture features. Multidimensional spectral features were obtained by calculating the band surface reflectance data and spectral index data of each pixel using multi-source monitoring data. The spectral index data included Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), Normalized Difference Building Index (NDBI), and Enhanced Bare Soil Index (EBSI). Deformation features were obtained from SAR image data in the multi-source monitoring data based on the differential synthetic aperture radar interferometry method.

[0011] Preferably, in method S2, the encoder of the convolutional neural network includes five layers, and the decoder includes four layers. The first to fourth layers of the encoder correspond one-to-one with the first to fourth layers of the decoder and are connected in a skip connection. The first to fifth layers of the encoder each include a convolutional module, which includes two convolution operations and ReLU activation. There is a channel attention submodule between adjacent layers of the encoder. The channel attention submodule is used to obtain channel attention weights in the channel dimension. The first to fifth layers of the encoder sequentially perform downsampling feature extraction and channel attention weight weighting. The fifth layer of the encoder outputs fused feature A. Some or all layers of the decoder include a spatial attention submodule, which is used to obtain spatial attention weights in the spatial dimension. The fourth to first layers of the decoder sequentially perform upsampling feature extraction and spatial attention weight weighting. The first layer of the decoder outputs fused feature B.

[0012] Preferably, a multi-source image sample dataset containing subsidence labels is constructed. The mining area subsidence identification model is trained using the multi-source image sample dataset for subsidence identification and prediction. The output module obtains the subsidence prediction probability of pixel i through a 1×1 convolution and sigmoid activation. .

[0013] Preferably, in method S3, the multi-source monitoring data includes elevation data, and the spatial adaptive regression model performs spatial weighted regression processing based on the location, elevation, and subsidence prediction probability of pixel i. The weight function of the spatial adaptive regression model adopts a Gaussian kernel, and the expression is as follows: , Let i be the weight of pixel i to pixel j. Spatial distance between pixel i and pixel j This is the bandwidth parameter.

[0014] Preferably, in method S3, the F1 scores corresponding to the convolutional neural network and the spatial adaptive regression model are obtained respectively, and the weight coefficients of the mining area subsidence identification model are also obtained. , The F1 score of the convolutional neural network and spatial adaptive regression model is dynamically and adaptively adjusted, and the weight coefficients are... , All expressions are as follows: , This indicates a convolutional neural network DGMCN or a spatial adaptive regression model SARM, when For the convolutional neural network DGMCN, Corresponding to ;when For the spatial adaptive regression model SARM, Corresponding to ; The F1 score corresponding to the DGMCN convolutional neural network. This represents the F1 score of the spatial adaptive regression model (SARM).

[0015] Compared with the prior art, the present invention has the following advantages and beneficial effects: (1) The mining area subsidence identification model of the present invention consists of multi-level feature extraction, convolutional neural network, spatial adaptive regression model and fusion probability processing to form multi-level feature processing. Attention enhancement is performed in the multi-level feature processing to make the distinction between the target in the subsidence area and the complex background clearer. The present invention monitors and acquires multi-source monitoring data of the research mining area and extracts multi-dimensional feature data including texture, spectrum, deformation and shape. Multi-level feature processing and attention enhancement have strengthened the network’s ability to pay attention to key spectral, texture and deformation features, which greatly improves the accuracy of semantic segmentation, reduces the risk of false detection and false detection, and realizes the precise and complete identification and segmentation of the mining area subsidence range.

[0016] (2) The present invention also adopts a multi-temporal cascade detection and subsidence update process, which can capture the subsidence change detection results of the subsidence area in a timely manner within the monitoring period. The subsidence change detection results include subsidence expansion, subsidence static stability, subsidence dynamic stability and subsidence shrinkage. From the subsidence change detection results, stable subsidence areas, newly added subsidence areas, continuous subsidence areas and subsidence recovery areas can be analyzed. The subsidence detection result update delay is short, which meets the high-frequency monitoring requirements.

[0017] (3) This invention can rapidly screen large areas of the study area to fine segmentation, spatial correction and multi-source and multi-temporal fusion of complete cascade extraction process, realize high-precision, automated and dynamic update extraction and interpretation of the subsidence area of ​​the mining area; this invention realizes high-precision, automated and dynamic update intelligent extraction of the subsidence area of ​​the mining area, providing technical support for the monitoring of geological disasters and ecological restoration management in the mining area. Attached Figure Description

[0018] Figure 1 This is a flowchart of the method for identifying and extracting subsidence in mining areas according to the present invention; Figure 2 This is a schematic diagram of the network structure of the convolutional neural network DGMCN in the embodiment; Figure 3 This is a schematic diagram illustrating the principle of the RotoHog texture feature extraction method used in this embodiment; Figure 4 This is a schematic diagram illustrating the principle of the channel attention submodule in the embodiment; Figure 5 This is a schematic diagram illustrating the principle of the spatial attention submodule in the embodiment. Detailed Implementation

[0019] The present invention will be further described in detail below with reference to embodiments: Example like Figure 1 As shown, a method for identifying and extracting subsidence in mining areas based on multi-source monitoring data fusion processing is described, the method comprising: S1. Construct a mining area subsidence identification model incorporating a convolutional neural network and a spatial adaptive regression model. Obtain multi-source monitoring data of the research mining area and input it into the subsidence identification model to extract a multi-dimensional feature dataset including texture, spectrum, deformation, and shape. The multi-source monitoring data refers to various monitoring data corresponding to the research mining area, including multi-source remote sensing image monitoring data and digital elevation model data (DEM data) covering the research mining area. Preferably, the multi-source monitoring data also includes a vector boundary file of the mining area. The multi-source remote sensing image monitoring data includes Landsat-8 multispectral optical images (obtained by monitoring the mining area using the Landsat-8 satellite, which carries the Land Imager (OLI) and Thermal Infrared Sensor (TIRS) to monitor and obtain multispectral optical images in multiple bands), Sentinel-2 multispectral optical images (obtained by monitoring the mining area using the Sentinel-2 satellite, which carries a multispectral imager to monitor and obtain multispectral optical images in multiple bands, including visible light, near-infrared, red-edge, and shortwave infrared), and SAR image data (preferably Sentinel-1 C-band interferometric wide-swath mode data). The multi-source remote sensing image monitoring data undergoes preprocessing including geometric correction, radiometric correction, atmospheric correction, resampling, and coordinate registration. ENVI software is used for Landsat image processing. -8. Geometric fine correction is performed on Sentinel-2 multispectral optical images. Atmospheric correction is performed on Sentinel-2 multispectral optical images using SNAP software to convert the original DN values ​​into surface reflectance, unifying the reflectance benchmark for all images. Orbit correction, thermal noise removal, and radiometric calibration are applied to SAR image data (preferably Sentinel-1 SAR data) to generate backscattering coefficient images. Then, various types of data from multi-source monitoring are precisely registered, and bilinear interpolation is used to resample various types of data from multi-source monitoring to a unified spatial resolution and project them onto a unified coordinate system (e.g., the WGS_1984_UTM_Zone_49N coordinate system). This invention constructs a multi-source image sample dataset containing subsidence labels. The multi-source image sample dataset is divided into training, validation, and test sets in a 70%:15%:15% ratio. The mining area subsidence identification model uses the multi-source image sample dataset for subsidence identification and prediction training.

[0020] The method for obtaining a multi-dimensional feature dataset based on multi-source monitoring data in the mining area subsidence identification model of this invention is as follows: The RotoHog texture feature extraction method is used to extract LBP texture feature vectors and HOG texture feature vectors, which are then concatenated to obtain multidimensional texture features; specifically, such as Figure 3As shown, firstly, rotation-invariant local binary mode (LBP) values ​​are calculated for each pixel (a pixel represents a region of the mining area under study, and a pixel includes several pixels). Then, the LBP histogram for each pixel is calculated, and the LBP texture feature vector is extracted. Specifically, a threshold function s(v) is used to segment neighboring pixels (… Figure 3 The grayscale difference between the 8 neighboring pixels (i.e., the number of neighboring points) and the center pixel is converted into binary code, and the threshold function expression is as follows: Then, the pixels are divided into cells, and the gradient direction histogram of each cell is calculated (gradient components are calculated using horizontal and vertical gradient operators to obtain the gradient magnitude G and gradient direction θ), generating a gradient direction histogram (HOG) and extracting the HOG texture feature vector. Multi-source monitoring data is used to calculate and obtain the band surface reflectance data and spectral index data of each pixel as multi-dimensional spectral features. The spectral index data includes Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), Normalized Difference Building Index (NDBI), and Enhanced Bare Soil Index (EBSI). In this embodiment, the spectral index data of the pixels uses the mean. The various spectral indices and calculation formulas are shown in Table 1.

[0021] Deformation features (including average deformation rate and cumulative deformation) are obtained from SAR image data in multi-source monitoring data based on differential synthetic aperture radar interferometry (DSAR Interferometry adopts existing technology, including the extraction of surface deformation rate and cumulative deformation). An example is as follows: Preprocessed Sentinel-1 SAR data is acquired, and the time series data of the study year is processed using Small Baseline Set Interferometry (SBAS-InSAR) to extract surface deformation features. First, interferometric pairs are generated based on spatiotemporal baseline threshold combinations, followed by differential interferometric processing: the registered master and slave images are conjugately multiplied to generate an interferogram, and the flatland effect and topographic phase are removed sequentially to obtain a differential interferometric phase containing deformation, atmospheric, and noise signals. The differential interferogram is unwrapped to obtain the absolute phase difference, and then the deformation time series is solved using methods such as singular value decomposition (SVD). The average deformation rate and cumulative deformation of each pixel are obtained by inversion. The deformation rate is obtained from the cumulative deformation time series through least squares fitting, and the average deformation rate and cumulative deformation are used as deformation features.

[0022] In this embodiment, the multidimensional feature dataset also includes shape features, which are the geometric shapes and attributes presented in the image within a pixel. Geometric attributes include area (number of pixels), aspect ratio (the ratio of the longer side to the shorter side of the circumscribed rectangle), and rectangularity, where rectangularity = area / (width × height). The multidimensional feature dataset stitches together texture (95 dimensions), spectral (28 dimensions), deformation (2 dimensions), and shape (3 dimensions) features at the pixel level to form a multi-channel feature matrix of size H×W×N (1024×1024×128).

[0023] S2. The Deep Global Attention Mechanism and Convolutional Neural Network (DGMCN) consists of an encoder, a decoder, and an output module. The encoder weights and encodes the multidimensional feature dataset along the channel dimension to obtain a fused feature A. The decoder weights and decodes the fused feature A along the spatial dimension to obtain a fused feature B. The encoder and decoder have one-to-one correspondences and skip connections. The output module uses the fused feature B to map and output the subsidence prediction probability of pixel i in the mining area. .

[0024] In some embodiments, such as Figure 2 As shown, the encoder of a convolutional neural network consists of five layers (the number of channels gradually increases from 128 channels in the input, forming 128, 256, 512, 1024, and 1024 channels), and the decoder consists of four layers, the first to fourth layers of the encoder (see...). Figure 2 (arranged from top to bottom) respectively correspond to the first to fourth layers of the decoder (see...) Figure 2 The encoder layers are arranged from bottom to top, with the number of channels decreasing sequentially to form 1024, 512, 256, 128, and 64, forming a one-to-one correspondence with corresponding skip connections. Layers one through five of the encoder each include a convolutional module. Each convolutional module includes two convolution operations (preferably two 3×3 convolution operations) and ReLU activation (preferably, ReLU activation is performed after each convolution operation). Adjacent layers of the encoder have channel attention submodules (CA submodules), which are used to obtain channel attention weights along the channel dimension, such as... Figure 4As shown, the channel attention submodule first performs global average pooling, then passes through two fully connected layers (with ReLU activation added in between), and then outputs a channel weight vector of length C via a sigmoid function. This channel weight vector is then multiplied back into the original feature map channel by channel to achieve adaptive recalibration of the channel dimension. The encoder's first to fifth layers sequentially perform downsampling feature extraction and channel attention weighting (i.e., multiplying the channel attention weights with the extracted features to achieve weighted feature processing along the channel dimension). The fifth layer of the encoder outputs the fused feature A.

[0025] The first to fourth layers of the decoder correspond one-to-one with the first to fourth layers of the encoder, that is, the first layer of the decoder corresponds to the first layer of the encoder, the second layer of the decoder corresponds to the second layer of the encoder, the third layer of the decoder corresponds to the third layer of the encoder, and the fourth layer of the decoder corresponds to the fourth layer of the encoder (in this example, the fourth layer of the decoder skips to the features output by the fourth layer of the encoder during upsampling decoding and performs feature concatenation). Some or all layers of the decoder include a spatial attention submodule, preferably, as follows: Figure 2 As shown, the fourth and second layers of the decoder each include an attention submodule (SA submodule). The spatial attention submodule is used to obtain the spatial attention weights in the spatial dimension, such as... Figure 5 As shown, the attention submodule first performs average pooling and max pooling along the channel dimension to obtain two H×W×1 projection images, which are then concatenated along the channel dimension to form H×W×2 images. These are then input into a 7×7 convolutional layer and activated by a sigmoid function to generate H×W×1 spatial weights. These spatial weights are then multiplied element-wise with the original features along the spatial dimension to highlight important regions. The decoder's fourth to first layers sequentially perform upsampling feature extraction and spatial attention weighting (i.e., multiplying the spatial attention weights with the extracted features to achieve spatial weighting). The decoder's first layer outputs the fused feature B.

[0026] A multi-source image sample dataset containing subsidence labels is constructed. The mining area subsidence identification model is trained using the multi-source image sample dataset for subsidence identification and prediction. The output module obtains the subsidence prediction probability of pixel i through a 1×1 convolution and sigmoid activation. The mining area subsidence identification model uses the following loss function during subsidence identification and prediction training: , ; ,in This represents the total loss (the optimization objective is to minimize the total loss). Represents cross-entropy loss, This represents the Dice coefficient loss (the Dice coefficient is a commonly used similarity metric in semantic segmentation). Here, M represents the weighting coefficient, and M represents the total number of pixels. and denoted as the ground truth value of the sinking label and the sinking prediction probability of pixel i, respectively.

[0027] S3. The spatial adaptive regression model uses the subsidence prediction probability of pixel i in the research mining area to perform spatial weighted regression processing to obtain the corrected subsidence prediction probability. The multi-source monitoring data includes elevation data. The spatial adaptive regression model performs spatial weighted regression processing based on the location, elevation, and subsidence prediction probability of pixel i. The weight function of the spatial adaptive regression model uses a Gaussian kernel, and the expression is as follows: , Let i be the weight of pixel i to pixel j. Spatial distance between pixel i and pixel j This is a bandwidth parameter (controlling the influence range of the neighborhood); The larger it is, the wider its influence. The smaller the value, the stronger the locality.

[0028] The subsidence identification model for mining areas in this invention obtains the fusion probability of pixel i according to the following formula. : ,in , These are the weighting coefficients. This embodiment employs a multi-index comprehensive evaluation system to verify and evaluate the mine subsidence identification model of this invention from two aspects: pixel-level classification accuracy and boundary space consistency. First, using on-site measured subsidence area range data and visual interpretation results of high-resolution UAV orthophotos as ground truth, the range results extracted by this invention are spatially registered and rasterized with the reference data to generate a binary mask image. Based on this, a confusion matrix is ​​constructed, and four classification performance indicators—accuracy, precision, recall, and F1 score—are calculated to comprehensively evaluate the accuracy of pixel-level prediction. Accuracy reflects the overall classification correctness, precision measures the proportion of real subsidence areas in the extracted range, recall assesses the proportion of real subsidence areas correctly extracted, and the F1 score is the harmonic mean of precision and recall, used to balance the two types of errors. Preferably, the weighting coefficients of the mine subsidence identification model... , The dynamic adaptive adjustment is adopted, and the method is as follows: obtain the F1 scores corresponding to the convolutional neural network and the spatial adaptive regression model, respectively, and the weight coefficients of the mining area subsidence identification model. , The F1 score of the convolutional neural network and spatial adaptive regression model is dynamically and adaptively adjusted, and the weight coefficients are... , All expressions are as follows: , This indicates a convolutional neural network DGMCN or a spatial adaptive regression model SARM, when For the convolutional neural network DGMCN, Corresponding to ;when For the spatial adaptive regression model SARM, Corresponding to . The F1 score corresponding to the DGMCN convolutional neural network. This represents the F1 score of the spatial adaptive regression model (SARM).

[0029] To further evaluate the consistency of boundary spatial locations, two indicators, Intersection over Union (IoU) and Boundary Displacement Error (BDE), are used to evaluate the subsidence identification model in the mining area. IoU reflects the segmentation quality by calculating the degree of overlap between the predicted and actual boundary regions, and is defined as the ratio of the intersection to the union of the two regions: IoU = |A∩B| / |A∪B|, where A is the predicted region and B is the actual region. BDE evaluates the boundary positioning accuracy by calculating the average Euclidean distance between the predicted and actual boundary lines. Specifically, a two-way distance mapping method is used to calculate the average distance from the predicted boundary point to the nearest actual boundary point and the average distance from the actual boundary point to the nearest predicted boundary point, respectively, and the average value is taken as the final BDE value.

[0030] By employing a cross-validation strategy, this invention constructs a multi-source image sample dataset containing sinking labels. The dataset is divided into training, validation, and test sets in a 70%:15%:15% ratio. The experiment is repeated three times, and the average is used as the final accuracy metric to ensure the statistical robustness of the evaluation results. Experimental results show that the proposed method achieves an F1 score of over 0.89 on the test set, an Intersection over Union (IoU) exceeding 0.82, and a Boundary Depth Error (BDE) controlled within 2.0 meters, significantly outperforming traditional machine learning and semantic segmentation methods.

[0031] S4. The spatial adaptive regression model uses the fusion probability of all pixels in the study area to obtain the fusion probability map of the study area. Clustering is performed on the fusion probability map to obtain the initial depression area. Canny edge detection is used to detect and identify the boundary of the initial depression area based on a preset threshold. Then, morphological operations are used to optimize the boundary (specifically, to remove noise and small-area pseudo-boundaries) to obtain the depression area segmentation result. The depression area segmentation result includes the internal area of ​​the depression area and the boundary of the depression area.

[0032] S5. The spatial adaptive regression model obtains the subsidence zone segmentation results of the study area in the preceding and following time phases according to methods S1 to S4, and extracts the boundary vectors. Using the boundary vectors of the preceding and following time phases (denoted as t for the preceding phase and t+1 for the following phase), the subsidence change detection results are output. These results include subsidence expansion, static stability, dynamic stability, and shrinkage. Subsidence expansion indicates that the boundary vector of the following phase is greater than that of the preceding phase; subsidence shrinkage indicates that the boundary vector of the following phase is less than that of the preceding phase; static stability indicates that the boundary vectors of the preceding and following time phases are the same; and dynamic stability indicates that the boundary vector areas of the preceding and following time phases remain dynamically equal under subsidence recovery treatment. From the subsidence change detection results, stable subsidence areas, newly added subsidence areas, continuously subsiding areas, and subsidence recovery areas can be analyzed. Then, targeted field measurements are conducted on these areas to facilitate the verification and evaluation of the updated subsidence zone segmentation results, ensuring accuracy. Finally, the updated subsidence zone segmentation results of the study area are output and stored in vector format using geographic information system software. This invention enables a complete cascaded extraction process, from rapid screening of a large area of ​​study region to fine segmentation, spatial correction, and fusion of multiple sources and time phases, achieving high-precision, automated, and dynamically updated extraction and interpretation of subsidence areas in underground mining research areas.

[0033] S6. The spatial adaptive regression model obtains the subsidence zone segmentation results of the study area in the preceding and following time phases according to methods S1 to S4, and extracts the boundary vectors. The boundary update response is then calculated according to the following expression. : , , Let be the boundary vectors for the previous time phase t and the next time phase t+1, respectively. Set a boundary update threshold. If the boundary update response... If the value exceeds the boundary update threshold, a boundary update is triggered, and the spatial adaptive regression model outputs the updated subsidence area segmentation result.

[0034] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for identifying and extracting subsidence in mining areas based on multi-source monitoring data fusion processing, characterized in that: The methods include: S1. Construct a mining area subsidence identification model that includes a convolutional neural network and a spatial adaptive regression model. Obtain multi-source monitoring data of the research mining area and input it into the mining area subsidence identification model to extract a multi-dimensional feature dataset including texture, spectrum, deformation, and shape. S2. The convolutional neural network includes an encoder, a decoder, and an output module. The encoder weights and encodes the multidimensional feature dataset along the channel dimension to obtain a fused feature A. The decoder weights and decodes the fused feature A along the spatial dimension to obtain a fused feature B. The encoder and decoder have one-to-one correspondences and skip connections between their layers. The output module uses the fused feature B to map and output the subsidence prediction probability of pixel i in the mining area. ; S3. The spatial adaptive regression model uses the subsidence prediction probability of pixel i in the research mining area to perform spatial weighted regression processing to obtain the corrected subsidence prediction probability. The fusion probability of pixel i is obtained according to the following formula. : ,in , These are the weight coefficients; the F1 scores corresponding to the convolutional neural network and the spatial adaptive regression model are obtained respectively, and these are the weight coefficients of the mining area subsidence identification model. , The F1 score of the convolutional neural network and spatial adaptive regression model is dynamically and adaptively adjusted, and the weight coefficients are... , All expressions are as follows: , This indicates a convolutional neural network DGMCN or a spatial adaptive regression model SARM, when For the convolutional neural network DGMCN, Corresponding to ;when For the spatial adaptive regression model SARM, Corresponding to ; The F1 score corresponding to the DGMCN convolutional neural network. This represents the F1 score of the spatial adaptive regression model (SARM).

2. The method for identifying and extracting mining area subsidence based on multi-source monitoring data fusion processing according to claim 1, characterized in that: It also includes the following methods: S4. The spatial adaptive regression model uses the fusion probability of all pixels in the study area to obtain the fusion probability map of the study area. Clustering is performed on the fusion probability map to obtain the initial subsidence area. Canny edge detection is used to detect and identify the boundary of the initial subsidence area based on a preset threshold. Then, morphological operations are used to optimize the boundary to obtain the subsidence area segmentation result. The subsidence area segmentation result includes the internal area of ​​the subsidence area and the boundary of the subsidence area.

3. The method for identifying and extracting subsidence in mining areas based on multi-source monitoring data fusion processing according to claim 2, characterized in that: It also includes the following methods: S5. The spatial adaptive regression model obtains the subsidence area segmentation results of the study area in the preceding and following time phases according to methods S1 to S4, and extracts the boundary vectors. The subsidence change detection results are output using the boundary vectors of the preceding and following time phases. The subsidence change detection results include subsidence expansion, static stability of subsidence, dynamic stability of subsidence, and subsidence shrinkage. Subsidence expansion indicates that the boundary vector of the following time phase is greater than that of the preceding time phase. Subsidence shrinkage indicates that the boundary vector of the following time phase is less than that of the preceding time phase. Static stability of subsidence indicates that the boundary vectors of the preceding and following time phases are the same. Dynamic stability of subsidence indicates that the boundary vector area of ​​the preceding and following time phases remains dynamically equal under the subsidence restoration treatment.

4. The method for identifying and extracting subsidence in mining areas based on multi-source monitoring data fusion processing according to claim 2, characterized in that: It also includes the following methods: S6. The spatial adaptive regression model obtains the subsidence zone segmentation results of the study area in the preceding and following time phases according to methods S1 to S4, and extracts the boundary vectors. The boundary update response is then calculated according to the following expression. : , , Let be the boundary vectors for the previous time phase t and the next time phase t+1, respectively. Set a boundary update threshold. If the boundary update response... If the value exceeds the boundary update threshold, a boundary update is triggered, and the spatial adaptive regression model outputs the updated subsidence area segmentation result.

5. The method for identifying and extracting mining area subsidence based on multi-source monitoring data fusion processing according to claim 1, characterized in that: In method S1, the multi-source monitoring data includes multi-source remote sensing image monitoring data and digital elevation model data covering the research mining area. The multi-source remote sensing image monitoring data includes Landsat-8 multispectral optical image, Sentinel-2 multispectral optical image and SAR image data. The multi-source remote sensing image monitoring data undergoes preprocessing including geometric correction, radiometric correction, atmospheric correction, resampling and coordinate registration.

6. The method for identifying and extracting mining area subsidence based on multi-source monitoring data fusion processing according to claim 5, characterized in that: In method S1, the method for obtaining the multidimensional feature dataset by the mining area subsidence identification model is as follows: The LBP texture feature vector and HOG texture feature vector were extracted using the RotoHog texture feature extraction method and then concatenated to obtain multidimensional texture features. Multidimensional spectral features were obtained by calculating the band surface reflectance data and spectral index data of each pixel using multi-source monitoring data. The spectral index data included Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), Normalized Difference Building Index (NDBI), and Enhanced Bare Soil Index (EBSI). Deformation features were obtained from SAR image data in the multi-source monitoring data based on the differential synthetic aperture radar interferometry method.

7. The method for identifying and extracting mining area subsidence based on multi-source monitoring data fusion processing according to claim 1, characterized in that: In method S2, the encoder of the convolutional neural network includes five layers, and the decoder includes four layers. The first to fourth layers of the encoder correspond one-to-one with the first to fourth layers of the decoder and are connected in a skip connection. The first to fifth layers of the encoder each include a convolutional module, which includes two convolution operations and ReLU activation. There is a channel attention submodule between adjacent layers of the encoder. The channel attention submodule is used to obtain channel attention weights in the channel dimension. The first to fifth layers of the encoder sequentially perform downsampling feature extraction and channel attention weight weighting. The fifth layer of the encoder outputs fused feature A. Some or all layers of the decoder include a spatial attention submodule, which is used to obtain spatial attention weights in the spatial dimension. The fourth to first layers of the decoder sequentially perform upsampling feature extraction and spatial attention weight weighting. The first layer of the decoder outputs fused feature B.

8. The method for identifying and extracting subsidence in mining areas based on multi-source monitoring data fusion processing according to claim 1, characterized in that: A multi-source image sample dataset containing subsidence labels is constructed. The mining area subsidence identification model is trained using the multi-source image sample dataset for subsidence identification and prediction. The output module obtains the subsidence prediction probability of pixel i through a 1×1 convolution and sigmoid activation. .

9. The method for identifying and extracting mining area subsidence based on multi-source monitoring data fusion processing according to claim 5, characterized in that: In method S3, the multi-source monitoring data includes elevation data. The spatial adaptive regression model performs spatial weighted regression processing based on the location, elevation, and subsidence prediction probability of pixel i. The weight function of the spatial adaptive regression model uses a Gaussian kernel, and the expression is as follows: , For pixel i to pixel j right Heavy, Spatial distance between pixel i and pixel j This is the bandwidth parameter.

Images