A mine area disturbance space identification method and system, computer device and medium

Abstract

The application provides a mining area disturbance space identification method, system, computer device and medium, and belongs to the image recognition field, comprising: acquiring remote sensing images of a mining area; performing feature extraction on the remote sensing images to obtain a plurality of coding feature maps of different scales; performing weighted fusion on the plurality of coding feature maps of different scales to obtain a multi-scale fusion feature map; calculating edge gradients of the remote sensing images and the multi-scale fusion feature map respectively, and fusing to obtain a normalized edge feature map; determining an edge attention weight according to the normalized edge feature map, and performing weighted fusion on the edge attention weight and the coding feature map of the corresponding scale to obtain an edge guided feature map; and performing resolution reconstruction on the edge guided feature map to obtain a disturbance space classification image of the mining area. The method improves the accuracy of the mining area disturbance space identification, and meets the fine identification requirement of the disturbance space.

Description

Technical Field

[0001] This invention belongs to the field of image recognition, specifically relating to a method, system, computer equipment, and medium for spatial recognition of disturbances in mining areas. Background Technology

[0002] Open-pit mining has been widely used in mining areas due to its advantages such as short construction period, high production efficiency, and high resource recovery rate. However, its "excavation-style" mining method causes far greater disturbance to the regional ecological environment than underground mining, often resulting in a series of ecological and environmental problems such as land excavation and occupation, vegetation degradation, and soil erosion. Along the entire mining-restoration process, open-pit mines can generally be divided into stripping zones, excavation zones, spoil heaps, and restoration zones. Different zones undertake differentiated production functions and face their own prominent ecological problems. While traditional methods relying on manual field surveys can achieve high accuracy, the process is labor-intensive and inefficient, making it difficult to meet the dynamic characteristics of rapid evolution in open-pit mines and the practical needs of ecological supervision.

[0003] In recent years, with the rapid development of remote sensing technology and deep learning, the automatic identification of the spatial extent of open-pit mines has gradually become a research hotspot, improving monitoring efficiency and automation. However, existing solutions mostly focus on extracting the overall outline of the mine, which is prone to classification errors not only in spectrally similar areas such as stripping and spoil heaps, but also in depicting complex boundaries, leading to blurred, broken, or adhered outlines and low identification accuracy. Summary of the Invention

[0004] To address the problem of low accuracy in identifying disturbance space in existing open-pit mines during the stripping-mining-drainage-reconstruction process, this invention provides a method, system, computer equipment, and medium for identifying disturbance space in mining areas.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A method for spatial identification of mining area disturbances includes: Acquire remote sensing images of the mining area; Feature extraction is performed on remote sensing images to obtain coded feature maps at multiple different scales; Weighted fusion of multiple encoded feature maps at different scales yields a multi-scale fused feature map. Edge gradients are calculated for remote sensing images and multi-scale fused feature maps respectively, and the edge gradients of remote sensing images and multi-scale fused feature maps are fused to obtain normalized edge feature maps. The edge attention weights are determined based on the normalized edge feature map, and the edge attention weights are then weighted and fused with the corresponding scale of the encoded feature map to obtain the edge guidance feature map. Resolution reconstruction is performed on the edge-guided feature map to obtain a perturbation spatial classification image of the mining area.

[0006] Optionally, in the mining area disturbance space identification method provided by the present invention, remote sensing images are processed by an enhanced U-net network to obtain a disturbance space classification image; The enhanced U-net network consists of an encoder, a multi-scale feature fusion module, an edge feature extraction module, an edge semantic fusion module, and a decoder connected in sequence; the edge feature extraction module includes a parallel remote sensing image edge feature extraction branch and a multi-scale fused feature map edge feature extraction branch; The remote sensing images are convolutionally encoded and downsampled by an encoder to obtain multiple encoded feature maps at different scales; A multi-scale fused feature map is obtained by fusing multiple encoded feature maps at different scales through a multi-scale feature fusion module; a normalized edge feature map is obtained by processing remote sensing images and multi-scale fused feature maps through an edge feature extraction module; and an edge-guided feature map is obtained by processing the normalized edge feature map and encoded feature map through an edge semantic fusion module. Based on the skip connection mechanism, the edge-guided feature map is upsampled layer by layer by the decoder to obtain the perturbation space classification image.

[0007] Optionally, the method for spatial identification of mining area disturbance provided by the present invention further includes: Convolution of multiple encoded feature maps at different scales yields a compressed channel of feature encoded maps; Upsampling of the feature coding map yields a feature coding map of uniform size, wherein the feature coding map of uniform size is consistent with the spatial size of the remote sensing image; Multiple feature-encoded maps of the same size are summed at the pixel level to obtain an aggregated feature map; Convolve the aggregated feature map to obtain a multi-scale fused feature map.

[0008] Optionally, the method for spatial identification of mining area disturbance provided by the present invention further includes: By applying the Sobel operator to the remote sensing image through the remote sensing image edge feature extraction branch, a pixel-level edge feature map is obtained; By applying the Sobel operator to the multi-scale fused feature map through the multi-scale fused feature map edge feature extraction branch, a semantic-level edge feature map is obtained. The pixel-level edge feature map and the semantic-level edge feature map are added and fused to obtain the edge gradient fusion result; Batch normalization and L2 normalization are performed on the edge gradient fusion results to obtain normalized edge feature maps.

[0009] Optionally, the method for spatial identification of mining area disturbance provided by the present invention further includes: Global max pooling and downsampling are performed on the normalized edge feature map to obtain a downsampling result that matches the spatial size of the encoded feature map; The downsampling results are processed using the Sigmoid activation function to obtain the edge attention weights; The edge attention weights are multiplied channel by channel with the encoded feature map of the corresponding scale to obtain the semantic response of the boundary enhancement. The semantic response is added element-wise to the encoded feature map at the corresponding scale to obtain the edge-guided feature map.

[0010] Optionally, the method for spatial identification of mining area disturbance provided by the present invention further includes: The edge guiding feature map of the first resolution is upsampled to the second resolution and concatenated with the edge guiding feature map of the second resolution to obtain the decoding result of the second resolution, wherein the first resolution is smaller than the second resolution; The decoding result at the second resolution is upsampled to the third resolution and then concatenated with the edge-guided feature map at the third resolution to obtain the decoding result at the third resolution. The third resolution is greater than the second resolution and less than or equal to the resolution of the remote sensing image. The decoding result of the third resolution is output as the class probability of each pixel through Softmax, and the perturbation space classification image is obtained based on the class probability of the pixels.

[0011] Optionally, the method for spatial identification of mining area disturbance provided by the present invention further includes: Based on the spatial topological relationships between the mining area and the stripping area, and between the mining area and the spoil disposal area, the perturbation spatial classification image of the mining area is optimized to obtain a perturbation spatial classification image with optimized spatial topology.

[0012] The present invention also provides a spatial identification system for mining area disturbance, comprising: Image acquisition module, used to acquire remote sensing images of the mining area; The image coding module is used to extract features from remote sensing images and obtain coded feature maps at multiple different scales. The feature fusion module is used to weight and fuse multiple encoded feature maps at different scales to obtain a multi-scale fused feature map. The edge gradient fusion module is used to calculate the edge gradients of the remote sensing image and the multi-scale fused feature map respectively, and fuse the edge gradients of the remote sensing image and the multi-scale fused feature map to obtain a normalized edge feature map. The edge attention weighting module is used to determine the edge attention weights based on the normalized edge feature map, and then weight and fuse the edge attention weights with the encoded feature map of the corresponding scale to obtain the edge guidance feature map. The spatial classification module is used to reconstruct the resolution of the edge-guided feature map to obtain a perturbation spatial classification image of the mining area.

[0013] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement any of the steps in a method for identifying spatial disturbances in a mining area.

[0014] The present invention also provides a computer-readable storage medium storing a computer program that, when loaded by a processor, can execute any step of a method for identifying spatial disturbances in a mining area.

[0015] The method for spatial identification of mining area disturbance provided by this invention has the following beneficial effects: Because the mining area disturbance space identification method provided by this invention aggregates encoded feature maps of different scales across all scales, it achieves the collaborative expression of shallow texture and deep semantics, significantly improving the ability to distinguish similar regions such as stripping areas and spoil heaps in open-pit mines, thus meeting the needs for refined identification of disturbance spaces. Furthermore, by extracting dual-source edges from original images and multi-scale fused features, a normalized edge attention map is generated. This map guides channel-wise learning and residual enhancement of semantic features at each scale, allowing edge features to deeply participate in the decoding process. This effectively solves the problems of blurred, broken, or adhered boundaries in complex areas, further improving the accuracy and robustness of mining area disturbance space identification. Attached Figure Description

[0016] To more clearly illustrate the embodiments and design schemes of the present invention, the accompanying drawings required for this embodiment will be briefly described below. The drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 A schematic diagram of the mining area disturbance space identification method provided in an embodiment of the present invention; Figure 2 This is an example of a multi-scale feature fusion module provided in an embodiment of the present invention; Figure 3 This is an example of an edge feature extraction module provided in an embodiment of the present invention; Figure 4 This is an example of an edge semantic fusion module provided in an embodiment of the present invention; Figure 5 This is an example of a labeled sample set of spatial disturbances caused by stripping and drainage in open-pit mines, provided in an embodiment of the present invention. Figure 6 This is an example of a network architecture for precise spatial identification of open-pit coal mine stripping-mining-drainage-re-disturbance provided in an embodiment of the present invention; Figure 7 An example of the spatial identification results of "stripping-mining-drainage-recovery" disturbance in open-pit mines provided in an embodiment of the present invention. Detailed Implementation

[0018] To enable those skilled in the art to better understand and implement the technical solutions of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be construed as limiting the scope of protection of the present invention.

[0019] While the rapid development and application of remote sensing technology and deep learning can improve the monitoring efficiency and automation of open-pit mines, the extraction of the overall outline of the mining area lacks fine-grained identification of disturbance units such as stripping zones, mining zones, spoil heaps, and remediation zones. Furthermore, in fine-grained classification, confusion often arises due to the high similarity in spectral and textural features between stripping zones and spoil heaps. In addition, models relying on single-scale or shallow features often fail to accurately represent complex boundaries, easily leading to blurred, broken, or adhered outlines, resulting in low overall identification accuracy. In short, existing technologies still have significant shortcomings in the accurate identification of disturbance spaces in open-pit mines involving stripping, mining, spoil heaps, and remediation, making it difficult to meet the dual needs of refined management and ecological protection in open-pit mines.

[0020] To address the limitations of existing technologies that only identify disturbances within the overall mining area, such as binary classification tasks focusing on "mining area / non-mining area" and "changed / unchanged," the mining disturbance spatial identification method provided by this invention, under single-temporal high-resolution remote sensing imagery, improves the precision of identification results by finely identifying disturbance units such as stripping zones, mining areas, spoil heaps or dumps, and remediation zones, thereby meeting regulatory requirements. Furthermore, this method constructs a multi-scale feature fusion mechanism to effectively coordinate shallow texture features and deep semantic features, enhancing the ability to distinguish similar regions such as stripping zones and spoil heaps, thus improving the accuracy of fine-grained classification. In addition, this method introduces an edge-guided mechanism to deeply fuse edge information with semantic features, enhancing the ability to delineate complex area boundaries, reducing blurry, broken, or adhered contours in the identification results, and improving the overall accuracy and reliability of the identification.

[0021] Specifically, in terms of multi-scale feature fusion, the mining area disturbance spatial recognition method provided by this invention uniformly upsamples features of different scales to the original image size and performs full-scale aggregation at the pixel level, realizing the collaborative expression of shallow texture and deep semantics. This significantly improves the ability to distinguish similar regions such as stripping areas and dumping areas, meeting the needs of refined recognition of disturbance units. In terms of edge information processing, the mining area disturbance spatial recognition method provided by this invention generates a normalized edge attention map through dual-source edge extraction of the original image and multi-scale fused features. It guides channel-wise learning and residual enhancement of semantic features at each scale, enabling edge features to deeply participate in the decoding process. This effectively solves the problems of blurred, broken, or adhered boundaries in complex areas, significantly improving the overall recognition accuracy and robustness.

[0022] Example 1 This invention provides a method for spatial identification of disturbances in mining areas. It processes remote sensing images using an enhanced U-net network to obtain a disturbance spatial classification image. The enhanced U-net network consists of an encoder, a multi-scale feature fusion module (MFM), an edge feature extraction module (EM), an edge semantic fusion module (FAM), and a decoder, connected sequentially. The edge feature extraction module includes parallel branches for extracting edge features from remote sensing images and multi-scale fused feature maps. Specifically... Figure 1 As shown, it includes the following steps: Step 11: Acquire remote sensing images of the mining area; extract features from the remote sensing images to obtain multiple coded feature maps at different scales. For example, by performing convolutional coding and downsampling on the remote sensing images using an encoder, multiple coded feature maps at different scales can be obtained.

[0023] Step 12: Weighted fusion of multiple encoded feature maps at different scales yields a multi-scale fused feature map; edge gradients are calculated for both the remote sensing image and the multi-scale fused feature map, and these gradients are fused to obtain a normalized edge feature map; edge attention weights are determined based on the normalized edge feature map, and these weights are weighted and fused with the corresponding encoded feature maps to obtain an edge-guided feature map. For example, a multi-scale fused feature map can be obtained by fusing multiple encoded feature maps at different scales using a multi-scale feature fusion module; a normalized edge feature map can be obtained by processing the remote sensing image and the multi-scale fused feature map using an edge feature extraction module; and an edge-guided feature map can be obtained by processing the normalized edge feature map and the encoded feature map using an edge semantic fusion module.

[0024] In the multi-scale feature fusion module, the multi-scale fused feature map can be extracted through the following steps: Step 121: Convolve multiple coded feature maps of different scales to obtain a compressed channel feature coded map; upsample the feature coded map to obtain a feature coded map of uniform size, wherein the feature coded map of uniform size is consistent with the spatial size of the remote sensing image.

[0025] Step 122: Sum the multiple feature encoding maps of the same size at the pixel level to obtain the aggregated feature map; convolve the aggregated feature map to obtain the multi-scale fused feature map.

[0026] Specifically, firstly, single-temporal high-resolution remote sensing images of the open-pit mine area are acquired as input images for the enhanced U-net network. Secondly, multiple convolutional coding units in the encoder of the enhanced U-net network perform convolution and downsampling operations to extract coding feature maps of different scales layer by layer. Each encoded feature map .

[0027] Next, for each encoded feature map, a 1×1 convolution is applied to compress the channel dimension to 64, generating a compressed channel feature encoding map. and will Upsampled to the same spatial size as the remote sensing image, resulting in n coded feature maps of uniform size. Then, all the uniformly sized encoded feature maps are summed pixel-by-pixel to obtain the aggregated feature map. As shown in formula (1): (1) Finally, the aggregated feature map By sequentially applying 3×3 convolutions and 1×1 convolutions, a multi-scale fused feature map is obtained. .

[0028] Specifically, such as Figure 2 As shown, firstly, the multi-level features output by the backbone network are subjected to channel normalization. Then, 1×1 convolutions are used to compress the feature channels of each layer—512×512×64 high-resolution features, 256×256×128 medium-resolution features, 128×128×256 medium-low-resolution features, and 64×64×512 low-resolution features—to 64 dimensions, thus preserving key information while reducing computational complexity. Subsequently, bilinear interpolation upsampling is performed on the deep features to restore them to the original input size. For example, upsampling is performed by 8x (8×up), 4x (4×up), and 2x (2×up) to control the size to 512×512, thereby eliminating scale differences between different levels of features and ensuring feature alignment.

[0029] Subsequently, an element-wise addition method is used to complementarily fuse the high-resolution features of the shallow layer with the upsampled deep features, enabling the model to capture the subtle textures of the target region while preserving its overall structure. The fused feature map is further enhanced by 3×3 convolution to improve local information interaction, and channel relationships are adjusted by 1×1 convolution to extract key information related to the target region, thereby improving the model's ability to distinguish easily confused categories.

[0030] In summary, to address the identification confusion between the spoil heap and stripping areas in open-pit mines due to similar texture and spectral features, the mining area disturbance spatial identification method provided by this invention enhances the target discrimination capability by fusing shallow spatial details and deep semantic information in the multi-scale feature fusion module.

[0031] In the edge feature extraction module, the normalized edge feature map can be determined through the following steps: Step 123: Apply the Sobel operator to the remote sensing image through the remote sensing image edge feature extraction branch to obtain a pixel-level edge feature map; apply the Sobel operator to the multi-scale fusion feature map through the multi-scale fusion feature map edge feature extraction branch to obtain a semantic-level edge feature map.

[0032] Step 124: Add and fuse the pixel-level edge feature map and the semantic-level edge feature map to obtain the edge gradient fusion result; perform batch normalization and L2 normalization on the edge gradient fusion result to obtain the normalized edge feature map.

[0033] Specifically, such as Figure 3 As shown, in the spatial identification method for mining area disturbances provided by this invention, the edge feature extraction module (EM) adopts a dual-branch parallel processing strategy to extract edge information from different sources and fuse them to enhance the clarity and stability of the target boundary. The first branch applies Sobel operators in the horizontal and vertical directions to the original high-resolution remote sensing image to obtain pixel-level edge features that are rich in detail but easily affected by background noise, such as fine-grained texture information like step cracks and soil dumping area fissures. The second branch applies Sobel processing to the multi-scale fused feature map to obtain semantic-level edge features with strong semantic constraints that can effectively suppress noise, such as the outer contour of the mining area and the boundary line of the reclamation area. Subsequently, the edge feature maps extracted by the two branches are fused by element-wise addition to effectively suppress background noise and highlight the key edge features of the target. Finally, the fused edge feature map is sequentially subjected to batch normalization (BatchNorm) to balance the differences in feature distribution, and L2 normalization (L2Norm) is used to globally constrain the gradient magnitude to ensure the stability and consistency of edge information at different scales.

[0034] For example, as shown in Equation (2), the Sobel convolution kernel is applied in the edge feature extraction branch of the multi-scale fusion feature map. and Calculate multi-scale fused feature maps Edge gradient : (2) Furthermore, as shown in Equation (3), in the remote sensing image edge feature extraction branch, a Sobel operation is performed on remote sensing image I to calculate its edge gradient. : (3) Then, as shown in formula (4), the edge feature maps obtained from the two edge gradient calculations are added together to obtain the edge gradient fusion result. : (4) Finally, the edge gradient fusion results Batch normalization and L2 normalization are performed to obtain normalized edge feature maps. .

[0035] In summary, addressing the challenges of blurred feature boundaries and significant noise interference in open-pit mines, the spatial identification method for mine disturbances provided by this invention integrates pixel-level spatial details from the original image with semantic guidance information from multi-scale feature maps in the edge feature extraction module, thereby constructing a more accurate edge detection algorithm. Specifically, while the original high-resolution remote sensing image contains rich spatial details, the complex terrain, shadows, and noise interference in the mining area may lead to unstable edge information. The feature maps extracted by the edge feature extraction module of this invention, however, contain deep semantic information and can characterize the overall structural features of targets such as stripping zones and spoil heaps.

[0036] In the edge semantic fusion module, the edge-guided feature map can be determined through the following steps: Step 125: Perform global max pooling and downsampling on the normalized edge feature map to obtain a downsampling result that matches the spatial size of the encoded feature map; process the downsampling result through the Sigmoid activation function to obtain the edge attention weights.

[0037] Step 126: Multiply the edge attention weights channel-by-channel with the corresponding scale's encoded feature map to obtain the semantic response for edge enhancement; add the semantic response element-by-element with the corresponding scale's encoded feature map to obtain the edge guidance feature map.

[0038] Specifically, such as Figure 4As shown, to achieve accurate identification of the disturbance space in open-pit mines during the stripping-mining-drainage-reconstruction process, the disturbance space identification method provided by this invention achieves feature fusion of edges and semantics through an edge semantic fusion module. First, the high-precision edge features output by the edge feature extraction module (EM) are downsampled to the spatial dimension of the corresponding semantic feature map using global max pooling. Next, an edge attention weight map is generated through sigmoid activation, which quantifies the spatial confidence distribution of the boundary region. Subsequently, through a channel-wise dot product operation between the edge attention weights and semantic features, the semantic response of the boundary transition region is dynamically strengthened, while the interference of redundant features within homogeneous materials is weakened. This module, through explicit guided boundary learning, effectively improves the representation of edge details and can effectively alleviate the boundary blurring and internal hole problems commonly found in traditional segmentation methods.

[0039] For example, first, normalize the edge feature map. Perform global max pooling to downsample the data to each encoded feature map. Matching spatial dimensions yields downsampling results. .

[0040] Then, as shown in formula (5), the downsampling results are... Input the sigmoid activation function to generate edge attention weights. : (5) in, The Sigmoid activation function is used. Then, as shown in Equation (6), the encoded feature maps at each scale are... With edge attention map weight map Perform channel-by-channel multiplication and then combine it with the encoded feature maps at each scale. Element-wise addition is performed, and feature information is further extracted through 1×1 convolution to obtain the edge-guided feature map. : (6) in, This is a channel-by-channel element-wise multiplication.

[0041] Step 13: Reconstruct the edge-guided feature map to obtain a perturbation spatial classification image of the mining area. For example, based on a skip connection mechanism, the edge-guided feature map is upsampled layer by layer by the decoder to obtain a perturbation spatial classification image.

[0042] The decoder can obtain the perturbation spatial classification image of the mining area through the following steps: Step 131: Upsample the edge guiding feature map of the first resolution to the second resolution, and concatenate it with the edge guiding feature map of the second resolution to obtain the decoding result of the second resolution, wherein the first resolution is smaller than the second resolution.

[0043] Step 132: Upsample the decoding result of the second resolution to the third resolution, and stitch it with the edge-guided feature map of the third resolution to obtain the decoding result of the third resolution. The third resolution is greater than the second resolution and less than or equal to the resolution of the remote sensing image.

[0044] Step 133: The decoding result of the third resolution is output as the class probability of each pixel through Softmax, and the perturbation space classification image is obtained based on the class probability of the pixels.

[0045] Specifically, in the decoder, as shown in formula (7), it first starts with the feature map of the lowest resolution and then sequentially upsamples to higher spatial resolutions, where the input of the i-th layer decoder is... Output from the previous layer Upsampling and corresponding The result obtained by piecing together: (7) in, For upsampling operation, For feature stitching. It should be emphasized that the number of decoder layers in this embodiment can be set based on the resolution, for example, set to four or more resolutions, and upsampling is performed through the corresponding layers in the decoder until the resolution of the remote sensing image is upsampled from the lowest resolution. The specific number of layers is not limited. Then, as shown in formula (8), for Perform convolution, activation, and normalization to obtain the output of the i-th layer decoder. : (8) Here, BN is the batch normalization layer, and ReLU is the activation function. The operation is repeated until the output feature map is obtained. The dimensions are consistent with the original image. Finally, as shown in formula (9), the dimensions are... The input is processed through a 1×1 convolutional layer and a softmax layer to obtain the final perturbation spatial classification image. : (9) Step 14: Based on the spatial topological relationship between the mining area and the stripping area and between the mining area and the spoil disposal area, optimize the disturbance spatial classification image of the mining area to obtain the disturbance spatial classification image after spatial topology optimization.

[0046] Specifically, considering the confusion between mining areas and stripping areas, and between mining areas and spoil heaps, in existing algorithm recognition results, the mining area disturbance space recognition method provided by this invention further optimizes the recognition results through spatial topological relationships during open-pit mining. The spatial topological relationships during open-pit mining can be described as follows: the line connecting the geometric center of the repair area and the geometric center of the mining area is defined as the main direction of coal mining; the line connecting the geometric center of the stripping area and the geometric center of the mining area forms an obtuse angle relative to the mining direction, while the line connecting the geometric center of the spoil heap area and the geometric center of the mining area forms an acute angle with the mining direction. By combining these spatial topological relationships to limit and optimize the disturbance space classification image, the confusion between stripping areas and spoil heaps can be effectively reduced, further improving the model's recognition accuracy under complex geological conditions.

[0047] Example 2 Based on Example 1, this invention also provides a complete example of a mining area disturbance space identification process, further demonstrating the accurate identification effect on the disturbance space of open-pit mining stripping and drainage.

[0048] Specifically, the coordinated advancement of open-pit coal mining and ecological restoration faces severe challenges, urgently requiring the construction of a dynamic identification and control system for the disturbed space of "stripping-mining-dumping-restoration". Throughout the entire life cycle of open-pit mining, the spatial evolution of various functional zones overlaps, and the ecological disturbance effects vary significantly. Traditional extensive management models are no longer sufficient to address the complex ecological risks caused by high-intensity mining, such as topsoil resource loss, geomorphological imbalance, and the spread of water and soil pollution. There is an urgent need for systematic governance through precise zoning: In the stripping zone, the key issue is the spatiotemporal misalignment between topsoil resource protection and soil and rock migration in the mining area. For example, premature stripping may lead to excessively large mining faces, increasing the pressure on topsoil resource storage and reuse, while delayed stripping may affect mining efficiency. Similarly, there is the challenge of coordinating stability control with the spatiotemporal reconstruction of the restoration zone. For instance, the dumping construction in the dumping zone requires a large amount of space, while soil reconstruction in the restoration zone depends on a specific time window, creating potential conflicts in space utilization and construction progress.

[0049] However, the boundaries between the stripping zone, mining zone, spoil heap, and restoration zone in open-pit coal mines are blurred, and the spatial transformation of different areas exhibits complex nonlinear dynamic changes. Influenced by factors such as topographic relief, lighting conditions, soil moisture, and vegetation cover, traditional spatial recognition algorithms have limitations in feature extraction and region classification. They struggle to effectively distinguish similar categories, easily leading to large boundary recognition errors, severe inter-class confusion, and loss of local details, thus reducing classification accuracy and failing to meet the needs of refined dynamic monitoring and ecological restoration assessment in mining areas.

[0050] To address the aforementioned challenges, this invention innovatively proposes a semantic segmentation network combining multi-scale feature fusion and edge feature-guided decoding, based on high-resolution remote sensing imagery from Jilin-1. By introducing a multi-scale feature fusion module into the encoder, the semantic representation capability of complex terrain and inter-class differences in mining areas is enhanced. Simultaneously, the edge feature extraction module and the feature adaptive fusion module explicitly guide boundary learning, improving the ability to depict edge details. This effectively alleviates the jagged edges and local voids caused by feature hierarchy fragmentation in traditional methods, thereby significantly improving the classification accuracy of open-pit mines and providing more reliable technical support for dynamic monitoring and ecological restoration of mining areas.

[0051] The open-pit mine stripping, drainage, and remediation disturbance spatial sample set is based on downloaded Jilin-1 L3D remote sensing imagery. Open-pit mines with typical "stripping-mining-drainage-remediation" characteristics were selected through visual interpretation to construct a high-resolution sample dataset. For example... Figure 5 As shown, firstly, within the selected open-pit mining area, vector boundaries for each disturbance type of "stripping-mining-drainage-recovery" are manually delineated based on high-resolution imagery, and the vector data are precisely classified and labeled. Secondly, the labeled vector data is converted into a raster data format that matches the imagery to ensure consistency between the imagery and the label data. Finally, the generated raster data is subjected to sample partitioning and data augmentation processing, including operations such as rotation, flipping, and cropping, to expand the sample size.

[0052] Ultimately, this study constructed a dataset containing 2385 high-resolution image samples with a size of 512 pixels × 512 pixels. This dataset was randomly divided into a training set and a validation set in a 7:3 ratio, with the training set containing 1670 image samples and the validation set containing 715 image samples. To further enhance the robustness and generalization of the model, future research can incorporate remote sensing imagery from more regions and time series to continuously expand the sample dataset, providing more comprehensive and richer data support for model optimization and performance improvement.

[0053] The specific network architecture is as follows: Figure 6As shown, based on the classic U-net framework, a multi-scale feature fusion module, an edge feature extraction module, and a feature fusion module are introduced. In the encoder stage, a hierarchical feature extraction strategy is adopted. First, multi-scale feature maps are extracted layer by layer through downsampling of the U-net backbone network. Then, the MFM module realizes the interaction and fusion of cross-layer features, strengthening the complementarity between low-level detail features and high-level semantic features, thereby effectively improving the representation ability of complex features, especially inter-class similar regions. After fusion, the original high-resolution image and the multi-scale feature map are input into the EM module. A dual-branch edge detection network is constructed using the Sobel operator: the original image branch focuses on extracting pixel-level spatial details, while the multi-scale feature branch provides semantic guidance, suppressing background noise and enhancing the target boundary response, thus generating a highly discriminative edge feature map. In the decoding stage, the FAM module is designed as the feature integration hub to dynamically fuse semantic features and edge features, further optimizing the classification confidence of boundary regions. Finally, the spatial resolution is restored by progressive upsampling of the decoder, generating accurate stripping-sampling-sorting-reconstruction spatial classification prediction results.

[0054] The performance evaluation metrics of the model include accuracy and mean intersection over union (mIoU), and the calculation formulas are shown in formulas (10) and (11): (10) (11) Where TP represents true positives, FP represents false positives, TN represents true negatives, FN represents false negatives, and C represents the total number of categories. Used to measure the proportion of correctly predicted pixels out of all pixels predicted by the model. This represents the ratio of the intersection to the union between the predicted and actual regions, and is then averaged over all categories. A higher value indicates that the model is more accurate in its predictions.

[0055] The proposed algorithm and the traditional U-net network model were applied to the accurate identification of the "stripping-mining-drainage-reclamation" disturbance space in open-pit mines to evaluate the classification performance of the proposed algorithm. Experimental results show that the overall classification accuracy of the proposed algorithm on high-resolution remote sensing datasets is significantly better than that of the U-net network, with an mIoU of 77.44% and an accuracy of 93.58%, representing improvements of 7.58% and 2.6% respectively compared to U-net. This indicates that the proposed algorithm has higher stability and robustness in classification tasks involving complex terrain features. Figure 7As shown, the U-net model exhibits certain limitations in classification, primarily manifested in noticeable jagged edges and patchy formation in some areas, resulting in fragmented and empty classification results and discontinuous and uneven classification boundaries. In contrast, the algorithm proposed in this invention performs better in edge detail processing, effectively reducing jagged edges, filling local gaps, improving the completeness and spatial consistency of classification results, and making the boundaries of disturbed areas clearer. Furthermore, in distinguishing land cover categories, the U-net model is prone to misclassifying some stripping areas as dumping areas, leading to significant category confusion. This algorithm, through a multi-scale feature fusion mechanism, fully integrates local details and global semantic information, improving the distinguishability of different disturbance types, effectively reducing misclassification of feature-similar regions, and further enhancing the accuracy and reliability of classification. Overall, this algorithm improves classification accuracy while optimizing the boundary representation of disturbed spaces, accurately depicting the "stripping-mining-dumping-reclamation" spatial pattern of open-pit mines, and providing more precise data support for refined monitoring and management of the ecological environment of mining areas.

[0056] In summary, the present invention significantly outperforms traditional schemes in overall classification accuracy on high-resolution remote sensing datasets, improving mIoU by approximately 7.58% and accuracy by approximately 2.6%, demonstrating higher accuracy, stability, and robustness in identifying complex perturbation units. Furthermore, traditional schemes often exhibit jagged edges and patchy phenomena in boundary regions, resulting in fragmented and discontinuous boundaries, and are prone to confusion in areas with similar features, such as stripping zones and spoil heaps. The present invention, through multi-scale feature fusion and edge guidance mechanisms, effectively reduces jagged edges, fills local voids, and makes boundaries clear and smooth, while improving the distinguishability of perturbation types and significantly reducing misclassification. In addition, the present invention reduces reliance on manual field surveys, achieving automated and dynamic identification of perturbation spaces. This not only reduces monitoring costs and improves regulatory efficiency but also assists management departments in timely understanding the mining and restoration status of mining areas, bringing significant economic and social benefits to the coordinated development of resource development and ecological environmental protection.

[0057] The multi-scale feature fusion module upsamples features at different scales to the original image size and performs full-scale aggregation at the pixel level, achieving collaborative expression of shallow texture and deep semantics. This significantly improves the ability to distinguish similar regions such as stripping areas and dumping areas, meeting the needs of refined identification of disturbance units. Meanwhile, the dual-source edge extraction of the original image and multi-scale fused features generates a normalized edge attention map, which guides channel-by-channel learning and residual enhancement of semantic features at each scale. This allows edge features to participate deeply in the decoding process, effectively solving the problems of blurred, broken, or adhered boundaries in complex areas. This significantly improves the overall accuracy and robustness of identification, and comprehensively enhances the accuracy of identification of the disturbance space in open-pit mine stripping-dumping-recovery.

[0058] Example 3 The present invention also provides a spatial identification system for mining area disturbance, comprising: Image acquisition module, used to acquire remote sensing images of the mining area; The image coding module is used to extract features from remote sensing images and obtain coded feature maps at multiple different scales. The feature fusion module is used to weight and fuse multiple encoded feature maps at different scales to obtain a multi-scale fused feature map. The edge gradient fusion module is used to calculate the edge gradients of the remote sensing image and the multi-scale fused feature map respectively, and fuse the edge gradients of the remote sensing image and the multi-scale fused feature map to obtain a normalized edge feature map. The edge attention weighting module is used to determine the edge attention weights based on the normalized edge feature map, and then weight and fuse the edge attention weights with the encoded feature map of the corresponding scale to obtain the edge guidance feature map. The spatial classification module is used to reconstruct the resolution of the edge-guided feature map to obtain a perturbation spatial classification image of the mining area.

[0059] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory. The processor executes the computer program to implement the steps in an embodiment of a method for spatial identification of disturbance in a mining area. Specific implementation methods can be found in the method embodiments, and will not be repeated here.

[0060] Furthermore, the present invention also provides a non-transitory computer-readable storage medium containing instructions, on which a computer program is stored. For example, a memory containing instructions that can be executed by a processor of a computer device to perform the above-described method. For example, the non-transitory computer-readable storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device, etc. When the computer program is executed by the processor, it can implement the steps in an embodiment of a method for identifying spatial disturbances in a mining area. Specific implementation methods can be found in the method embodiments, which will not be repeated here.

[0061] Those skilled in the art will understand that embodiments of the present invention can provide methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0062] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0063] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0064] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0065] It should be noted that the specific embodiments described above enable those skilled in the art to more fully understand the present invention, but do not limit the present invention in any way. Therefore, although the present invention has been described in detail in this specification and embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the present invention; and all technical solutions and improvements that do not depart from the spirit and scope of the present invention are covered within the protection scope of the present invention patent. No reference numerals in the claims should be construed as limiting the scope of the claims. Any simple variations or equivalent substitutions of technical solutions that can be readily obtained by those skilled in the art within the scope of the technology disclosed in the present invention are within the protection scope of the present invention.

Claims

1. A method for spatial identification of disturbances in a mining area, characterized in that, include: Acquire remote sensing images of the mining area; Feature extraction is performed on the remote sensing image to obtain multiple coded feature maps at different scales; The weighted fusion of the multiple encoded feature maps at different scales yields a multi-scale fused feature map. Edge gradients are calculated for the remote sensing image and the multi-scale fused feature map respectively, and the edge gradients of the remote sensing image and the multi-scale fused feature map are fused to obtain a normalized edge feature map; The edge attention weights are determined based on the normalized edge feature map, and the edge attention weights are weighted and fused with the encoded feature map at the corresponding scale to obtain the edge guidance feature map; The edge-guided feature map is reconstructed at a higher resolution to obtain a perturbation spatial classification image of the mining area.

2. The method for spatial identification of mining area disturbance according to claim 1, characterized in that, The remote sensing image is processed by an enhanced U-net network to obtain the perturbation spatial classification image; The enhanced U-net network consists of an encoder, a multi-scale feature fusion module, an edge feature extraction module, an edge semantic fusion module, and a decoder connected in sequence; the edge feature extraction module includes a parallel remote sensing image edge feature extraction branch and a multi-scale fused feature map edge feature extraction branch; The remote sensing image is convolutionally encoded and downsampled using the encoder to obtain multiple encoded feature maps at different scales; The multi-scale fusion feature map is obtained by fusing the multiple encoded feature maps at different scales through the multi-scale feature fusion module; the normalized edge feature map is obtained by processing the remote sensing image and the multi-scale fusion feature map through the edge feature extraction module; and the edge-guided feature map is obtained by processing the normalized edge feature map and the encoded feature map through the edge semantic fusion module. Based on the skip connection mechanism, the edge-guided feature map is upsampled layer by layer by the decoder to obtain the perturbation space classification image.

3. The method for spatial identification of mining area disturbance according to claim 2, characterized in that, The multi-scale fused feature map is obtained by fusing the multiple encoded feature maps of different scales through the multi-scale feature fusion module, including: Convolution of the multiple encoded feature maps at different scales yields a compressed channel of feature encoded maps; The feature coding map is upsampled to obtain a feature coding map of uniform size, wherein the feature coding map of uniform size is consistent with the spatial size of the remote sensing image; The pixel-level summation of multiple feature-encoded maps of the same size yields an aggregated feature map. The aggregated feature map is convolved to obtain the multi-scale fused feature map.

4. The method for spatial identification of mining area disturbance according to claim 2, characterized in that, The normalized edge feature map is obtained by processing the remote sensing image and the multi-scale fused feature map through the edge feature extraction module, including: The Sobel operator is applied to the remote sensing image through the remote sensing image edge feature extraction branch to obtain a pixel-level edge feature map; The Sobel operator is applied to the multi-scale fused feature map by the multi-scale fused feature map edge feature extraction branch to obtain a semantic-level edge feature map; The pixel-level edge feature map and the semantic-level edge feature map are added and fused to obtain the edge gradient fusion result; The edge gradient fusion result is subjected to batch normalization and L2 normalization to obtain the normalized edge feature map.

5. The method for spatial identification of mining area disturbance according to claim 2, characterized in that, The edge guidance feature map is obtained by processing the normalized edge feature map and the encoded feature map through the edge semantic fusion module, including: The normalized edge feature map is subjected to global max pooling and downsampling to obtain a downsampling result that matches the spatial size of the encoded feature map; The downsampling results are processed by the Sigmoid activation function to obtain the edge attention weights; The edge attention weights are multiplied channel-by-channel with the encoded feature map of the corresponding scale to obtain the semantic response for edge enhancement. The semantic response is added element-by-element to the encoded feature map of the corresponding scale to obtain the edge-guided feature map.

6. The method for spatial identification of mining area disturbance according to claim 2, characterized in that, Based on the skip connection mechanism, the edge-guided feature map is upsampled layer by layer by the decoder to obtain the perturbation space classification image, including: The edge guiding feature map of the first resolution is upsampled to the second resolution and concatenated with the edge guiding feature map of the second resolution to obtain the decoding result of the second resolution, wherein the first resolution is smaller than the second resolution; The decoding result at the second resolution is upsampled to the third resolution and then concatenated with the edge-guided feature map at the third resolution to obtain the decoding result at the third resolution, wherein the third resolution is greater than the second resolution and less than or equal to the resolution of the remote sensing image. The decoding result of the third resolution is output as the class probability of each pixel through Softmax, and the perturbation space classification image is obtained based on the class probability of the pixels.

7. The method for spatial identification of mining area disturbance according to claim 1, characterized in that, After obtaining the perturbation spatial classification image of the mining area, the method further includes: Based on the spatial topological relationships between the mining area and the stripping area, and between the mining area and the spoil disposal area, the disturbance spatial classification image of the mining area is optimized to obtain a disturbance spatial classification image with optimized spatial topology.

8. A spatial identification system for mining area disturbance, characterized in that, include: Image acquisition module, used to acquire remote sensing images of the mining area; The image coding module is used to extract features from the remote sensing image to obtain multiple coded feature maps at different scales; The feature fusion module is used to weight and fuse the multiple encoded feature maps at different scales to obtain a multi-scale fused feature map; The edge gradient fusion module is used to calculate the edge gradients of the remote sensing image and the multi-scale fused feature map respectively, and fuse the edge gradients of the remote sensing image and the multi-scale fused feature map to obtain a normalized edge feature map. An edge attention weighting module is used to determine edge attention weights based on the normalized edge feature map, and to weight and fuse the edge attention weights with the encoded feature map at the corresponding scale to obtain an edge guidance feature map; The spatial classification module is used to reconstruct the resolution of the edge-guided feature map to obtain a perturbation spatial classification image of the mining area.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the mining area disturbance spatial identification method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is loaded by the processor, it is able to execute the steps of the mining area disturbance spatial identification method according to any one of claims 1 to 7.

Images