Tailing pond extraction method, device and equipment based on wide-area remote sensing image

Abstract

The invention provides a tailing pond extraction method, device and equipment based on a wide-area remote sensing image, and belongs to the technical field of tailing pond management.The method comprises the steps that the wide-area remote sensing image containing a tailing pond is obtained; inputting the wide-area remote sensing image into a pre-trained target detection model for processing, and outputting a candidate bounding box containing tailing pond position information; on the basis of the candidate bounding box, cutting out a local image area containing the tailing pond from the original wide-area remote sensing image; and inputting the local image area into a pre-trained semantic segmentation model for processing, and outputting a tailing pond extraction result. Therefore, when the remote sensing image of a complex scene is processed, the feature extraction limitation of a single model can be effectively overcome through double-model complementation, and the extraction accuracy of the tailings pond is improved.

Description

Technical Field

[0001] This invention belongs to the field of tailings dam management technology, specifically relating to a tailings dam extraction method, apparatus, and equipment based on wide-area remote sensing imagery. Background Technology

[0002] Tailings dams are sites constructed by sealing off valleys or enclosing areas to store tailings or other industrial waste generated during the beneficiation of metallic or non-metallic ores. With the continuous expansion of mineral resource extraction and the rapid growth of mineral consumption, the number and scale of tailings dams, as a major site for centralized waste storage, have shown a significant upward trend. However, due to historical issues and cost control factors, many tailings dams were constructed to low standards. This makes them highly susceptible to dam imbalance and serious safety accidents if they lack effective management and regular maintenance during operation and are exposed to adverse factors such as floods, geological disasters, or overloading. In recent years, the frequent occurrence of tailings dam accidents has further highlighted the severity and urgency of tailings dam safety management. Against this backdrop, how to achieve efficient and accurate identification and dynamic monitoring of tailings dams has become a critical issue that urgently needs to be addressed. Compared to traditional monitoring methods that primarily rely on manual inspections, ground exploration, and on-site testing, the approach based on remote sensing imagery and deep learning models can automatically identify and extract information from tailings ponds over large areas, offering advantages such as low cost, high efficiency, and higher prediction accuracy. This not only improves the efficiency and accuracy of tailings pond identification and monitoring but also provides technical support for large-scale, periodic, and non-contact risk early warning. It can ensure engineering safety while providing a data foundation and decision support for ecological environmental protection and the optimization of tailings pond management systems, thus having significant practical implications for the sustainable development of tailings ponds.

[0003] Currently, existing research on tailings dam identification in remote sensing imagery largely focuses on method validation on well-preprocessed sample datasets. However, when applied to large-scale real-world remote sensing imagery scenarios, these methods often suffer from weak generalization ability and poor adaptability. Taking semantic segmentation algorithms as an example, although they perform excellently in target boundary extraction and shape reconstruction, achieving pixel-level fine-grained recognition, their accuracy is easily compromised in areas with complex backgrounds or small target scales. Especially in remote sensing images of tailings dams, due to the varying shapes and significant area differences of tailings dams, coupled with cluttered backgrounds, semantic segmentation models are prone to missed or false detections, thus affecting overall recognition accuracy. Summary of the Invention

[0004] To address the aforementioned problems, this invention provides a method, apparatus, and equipment for tailings dam extraction based on wide-area remote sensing imagery.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A method for tailings dam extraction based on wide-area remote sensing imagery, the method comprising: Acquire wide-area remote sensing imagery including tailings ponds; The wide-area remote sensing image is input into a pre-trained target detection model for processing, and candidate bounding boxes containing tailings dam location information are output. Based on candidate bounding boxes, a local image region containing the tailings pond is cropped from the original wide-area remote sensing image; the local image region is input into a pre-trained semantic segmentation model for processing, and the tailings pond extraction result is output.

[0006] Optionally, the target detection model is an improved YOLOv8-obb ​​network model, the improvement including: The efficient convolutional unit SCConv is embedded in the C2f module of the YOLOv8-obb ​​network, and the coordinate attention module CA is introduced at the connection between the Neck and Head.

[0007] Optionally, the wide-area remote sensing image is input into a pre-trained target detection model for processing, and the output candidate bounding boxes containing tailings dam location information include: Image features are extracted using a backbone network embedded with SCConv high-efficiency convolutional units; Multi-scale fused features are generated using a multi-scale feature fusion network with a Neck structure; Spatial-channel weighted enhancement of multi-scale fused features is performed using the coordinate attention module (CA). Candidate bounding boxes for tailings ponds are generated based on weighted features.

[0008] Optionally, the semantic segmentation model is an improved DeepLabv3+ network model, the improvements including: MobileNetV3 is introduced as the backbone feature extraction network, and strip pooling (SP) is introduced to replace the global average pooling of the ASPP module in the DeepLabv3+ model. CBAM modules are added to the feature fusion stages at both the encoder and decoder ends of the DeepLabv3+ model.

[0009] Optionally, the local image region is input into a pre-trained semantic segmentation model for processing, and the output tailings pond extraction results include: Multi-level features of local image regions are extracted using MobileNetV3; Contextual information is extracted using a spatial pyramid pooling structure that includes strip pooling SPs; By fusing the multi-level features and contextual information, tailings dam extraction results are generated based on the fused features.

[0010] An apparatus for extracting equipment fault features, the apparatus comprising: The acquisition module is used to acquire wide-area remote sensing images containing tailings ponds; The determination module is used to input the wide-area remote sensing image into a pre-trained target detection model for processing and output candidate bounding boxes containing tailings dam location information. The segmentation module is used to crop out local image regions containing tailings ponds from the original wide-area remote sensing image based on candidate bounding boxes. The extraction module is used to input the local image region into a pre-trained semantic segmentation model for processing and output the tailings pond extraction results.

[0011] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned tailings dam extraction method based on wide-area remote sensing imagery.

[0012] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the aforementioned tailings dam extraction method based on wide-area remote sensing imagery.

[0013] The tailings dam extraction method based on wide-area remote sensing imagery provided by this invention has the following beneficial effects: This invention combines object detection with semantic segmentation models, effectively mitigating the issues of missed detections and false detections in tailings dam identification using a single semantic segmentation model, and significantly improving overall recognition accuracy. First, the object detection model quickly locates potential tailings dams using candidate bounding boxes, greatly narrowing the processing range of the semantic segmentation model and avoiding interference from complex backgrounds in wide-area imagery on segmentation accuracy. Second, cropping local image regions allows the semantic segmentation model to focus on high-value areas, compensating for the coarse-grained nature of object detection through pixel-level fine segmentation. The two models work together through a "detect first, segment later" approach, forming a "coarse localization - fine characterization" collaborative mechanism. This cascaded strategy fully integrates the technical advantages of object detection and semantic segmentation, improving the accuracy of tailings dam boundary extraction while reducing the missed detection rate of small tailings dams. Especially when processing remote sensing images with complex scenes, the complementary nature of the two models effectively overcomes the limitations of feature extraction from a single model, reducing the false positive rate of similar features. Thus, this method, through dual verification in spatial and semantic dimensions, provides a more robust technical path for tailings dam extraction from remote sensing images. Attached Figure Description

[0014] 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.

[0015] Figure 1 This is a schematic flowchart of a tailings dam extraction method based on wide-area remote sensing imagery provided by the present invention according to an exemplary embodiment.

[0016] Figure 2 This is a schematic flowchart of another tailings dam extraction method based on wide-area remote sensing imagery provided by the present invention according to an exemplary embodiment.

[0017] Figure 3 This is a schematic diagram of the structure of an optimized target detection model provided by the present invention according to an exemplary embodiment.

[0018] Figure 4 This is an accuracy curve of a target detection stage provided by the present invention according to an exemplary embodiment.

[0019] Figure 5 This is a schematic diagram of the structure of an optimized semantic segmentation model provided by the present invention according to an exemplary embodiment.

[0020] Figure 6 This is a graph showing the average intersection-union ratio (OCR) of a semantic segmentation model provided by the present invention according to an exemplary embodiment.

[0021] Figure 7 This is a comparison diagram of extraction results from various semantic segmentation models provided by the present invention according to an exemplary embodiment.

[0022] Figure 8 This is a block diagram of a tailings dam extraction device based on wide-area remote sensing imagery, provided by the present invention according to an exemplary embodiment. Detailed Implementation

[0023] 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.

[0024] like Figure 1As shown, this invention, based on existing publicly available tailings dam target detection datasets, creates a corresponding tailings dam semantic segmentation dataset and performs data augmentation on both datasets. Addressing the issue of large differences in target size and sparse distribution of tailings dams in remote sensing images, a spatial context feature fusion method is used in the detection stage to enhance the model's perception of targets at different scales. A coordinate attention mechanism is introduced to strengthen the spatial information representation of key areas, achieving accurate localization of tailings dam targets in wide-area remote sensing images and generating remote sensing image slices containing tailings dam areas and their spatial location information. Addressing the issue that tailings dams often present as long strip structures and have unclear boundaries in remote sensing images, strip pooling is introduced in the segmentation stage to enhance the extraction capability of structural features. Combined with channel attention and local context information perception, the overall recognition effect of the semantic segmentation model is optimized, achieving refined extraction of tailings dam areas. The remote sensing image slices containing tailings dam areas obtained in the detection stage are input into the optimized semantic segmentation model to obtain the tailings dam extraction results from wide-area remote sensing images.

[0025] The technical solutions provided by the various embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0026] First, this invention provides a tailings dam extraction method based on wide-area remote sensing imagery, specifically as follows: Figure 2 As shown, it includes the following steps: S101. Acquire wide-area remote sensing images including tailings ponds.

[0027] For example, medium-resolution satellite data such as Landsat-9 (30m resolution) or Sentinel-2 (10-60m resolution) can be prioritized, as their multispectral characteristics can effectively identify the spectral features of tailings ponds. For key areas, Gaofen-7 (sub-meter resolution) data can be supplemented to enhance detail. Based on the distribution density of tailings ponds, image coverage should be planned according to a 1:50,000 scale standard, ensuring that each image covers an area ≥100km². A 30% overlap between adjacent images should be used to ensure stitching integrity. Dry season (cloud cover <10%) images should be selected, with a focus on configuring SWIR (shortwave infrared) and NIR (near-infrared) band combinations to enhance the spectral differences between tailings and surrounding features. Simultaneous acquisition of concurrent DEM data will assist in terrain correction.

[0028] S102. Input the wide-area remote sensing image into the pre-trained target detection model for processing, and output candidate bounding boxes containing tailings dam location information.

[0029] In this step, the untrained target detection model first needs to be trained. For the target detection model, this invention addresses the problem of large differences in target size and sparse distribution of tailings ponds in remote sensing images. In the detection stage, a spatial context feature fusion method is used to improve the model's ability to perceive targets at different scales, and a coordinate attention mechanism is introduced to strengthen the spatial information representation of key areas, thereby achieving accurate localization of tailings pond targets in wide-area remote sensing images.

[0030] In one embodiment, a sample dataset is first obtained, and then the object detection model is trained based on the sample dataset.

[0031] For example, when acquiring a sample dataset, existing publicly available historical tailings dam detection datasets can be obtained. Based on the correspondence between the coordinate information of the original labels in these historical tailings dam detection datasets and the images, the annotation information of the original datasets can be restored onto the images, resulting in a fusion result of the original image and the historical detection labels. Based on this fusion result, the Labelme tool can be used to re-perform semantic segmentation and annotation, creating a tailings dam semantic segmentation dataset. To improve training performance, data augmentation can also be performed. Four data augmentation methods can be used to expand the samples: rotation (rotating the original image and corresponding labels counterclockwise by 90°), flipping (flipping the original image and corresponding labels horizontally), color enhancement (adjusting the color display changes of the original image), and Gaussian blur processing (setting the blur value to 7.0 and applying Gaussian blur to the original image), resulting in a data-augmented sample dataset.

[0032] Before training an object detection model, the model structure can be optimized and improved to enhance its performance. For example... Figure 3As shown, the object detection model can be an improved YOLOv8-obb ​​network model. Efficient convolutional units SCConv are embedded into the C2f module in the Backbone and Neck of the original YOLOv8-obb ​​network, resulting in the reconstructed C2f_SCConv module. An attention mechanism module CA is introduced at the connection between the Neck and Head. In the backbone network, CONV is a convolutional layer with different parameters, including kernel size, stride, and padding. The backbone network performs initial feature extraction through the first convolutional layer and then adjusts the number of channels and integrates information through subsequent convolutional layers, progressively extracting higher-level and more abstract features in a hierarchical manner to increase the receptive field. SPPF is a fast spatial pyramid pooling module that fuses multi-scale features through pooling operations at different scales, enhancing the model's adaptability to inputs of different sizes while reducing computational redundancy to improve efficiency. Finally, the Detect module in the detection head is used to generate the predicted oriented bounding boxes. The model was then trained, with experimental parameters set. The optimizer was SGD, the total training batches were set to 300 epochs, and experimental results were recorded every 3 training epochs. The batch size was set to 16, the number of workers to 4, and the initial learning rate to 0.01. The augmented tailings dam target detection dataset was divided into training, testing, and validation sets in an 8:1:1 ratio for training, resulting in a pre-trained target detection model. By improving the YOLOv8-obb ​​network model and embedding SCConv, its channel shuffling and other operations reduced computation and parameter requirements, decreasing resource demands and accelerating inference speed when processing wide-area remote sensing imagery, thus improving the efficiency of local image region cropping. Simultaneously, it better captures local features and spatial information, accurately extracts features related to the tailings dam, reduces irrelevant interference, and improves cropping accuracy. Introducing the CA module enhances spatial location information, allowing the model to focus more on the specific location of the tailings dam, resulting in more accurate detection and reducing bounding box offsets and misjudgments. Moreover, the CA module can dynamically adjust according to the importance of location features, highlight key location features, enhance the model's ability to distinguish tailings ponds in complex backgrounds, and thus generate more accurate and compliant local image regions.

[0033] In addition, evaluation indicators can be selected to assess the model's detection results, and comparative and ablation experiments can be set up to verify the reliability and accuracy of the optimized target detection model. Specific steps include: Precision, recall, mean average precision (mAP), and F1 score were selected to quantitatively evaluate the reliability of the optimized object detection model. The results are as follows: Figure 4 As shown; ; ; ; ; In the formula, TP indicates that a positive sample is correctly identified, FN indicates that a positive sample is missed, TN indicates that a negative sample is correctly excluded, FP indicates that a negative sample is falsely detected, m represents the total number of categories of the target to be detected, n represents the number of targets within a certain category, and P ij (r) represents the precision given a recall rate of R.

[0034] After training the target detection model, candidate bounding boxes containing tailings dam location information are extracted using the target detection model. The process is as follows: Based on the pre-trained target detection model, wide-area remote sensing images containing tailings dams are processed, and image features are extracted through a backbone network embedding SCConv high-efficiency convolutional units; multi-scale fusion features are generated through a Neck structure multi-scale feature fusion network; spatial-channel weighted enhancement is performed on the multi-scale fusion features using the coordinate attention module CA; and candidate bounding boxes for tailings dams are generated based on the weighted features.

[0035] In this step, YOLOv5_obb, YOLOv7_obb, and YOLOv8_obb models can be selected for comparative experiments with the optimized object detection model. Four ablation experiments are designed to verify the model's performance by sequentially adding the SCConv convolution module and the CA attention mechanism module to the YOLOv8_obb model: YOLOv8_obb model, YOLOv8_obb model (C2f_SCConv convolution module), YOLOv8_obb model (CA module), and the optimized object detection model. Based on these comparative experiments, the superiority of the object detection model in this invention is clearly demonstrated.

[0036] S103. Based on candidate bounding boxes, crop out the local image region containing the tailings pond from the original wide-area remote sensing image.

[0037] In this step, the wide-area remote sensing image is selected and cropped based on the range of the candidate bounding box to obtain a local image region containing the tailings pond.

[0038] For example, the candidate bounding box coordinates can first be preprocessed to ensure they correspond to the pixel positions in the wide-area remote sensing image. Then, using the cropping function of an image processing library (such as OpenCV or GDAL), rectangular regions can be extracted from the wide-area image based on the pixel coordinates of the candidate bounding boxes. During cropping, the original image's metadata and band information should be preserved. Additionally, the bounding boxes can be appropriately expanded (e.g., by 5-10 pixels) to avoid truncating the tailings dam edge information and to facilitate subsequent analysis.

[0039] S104. Input the local image region into the pre-trained semantic segmentation model for processing, and output the tailings pond extraction results.

[0040] In this step, the untrained semantic segmentation model first needs to be trained. For the semantic segmentation model, this invention addresses the problem that tailings ponds often present as long, strip-shaped structures with unclear boundaries in remote sensing images. It introduces strip pooling during the segmentation stage to enhance the extraction of structural features. Furthermore, by combining channel attention mechanisms and local contextual information perception, the overall recognition performance of the semantic segmentation model is optimized, achieving refined extraction of tailings pond areas.

[0041] For example, the semantic segmentation model can be trained based on the sample dataset obtained in the aforementioned steps.

[0042] Among them, such as Figure 5 As shown, this semantic segmentation model can be an improved DeepLabv3+ network model. MobileNetV3 is introduced as the backbone feature extraction network, and strip pooling (SP) replaces the global average pooling of the ASPP module in the DeepLabv3+ model. CBAM modules are added to the feature fusion stages at both the encoder and decoder ends of the DeepLabv3+ model. Introducing MobileNetV3 as the backbone feature extraction network reduces computational complexity and the number of parameters due to its lightweight nature, accelerates feature extraction, meets real-time requirements, and maintains a certain feature extraction capability through innovative design, providing an effective feature foundation for segmentation tasks. Replacing the global average pooling of the ASPP module with strip pooling (SP) enhances the ability to model long-distance dependencies, better captures slender or directional structural information, enriches feature representation, and improves the segmentation accuracy of tailings ponds. The CBAM module is added to the feature fusion stage at both the encoder and decoder ends. The encoder end can enhance feature selection capabilities, highlighting feature channels and key areas related to the tailings dam. The decoder end can improve the feature fusion effect, optimize the fused features, make them more focused on the tailings dam area, reduce background interference, and improve the accuracy and boundary clarity of the extraction results.

[0043] For training the improved DeepLabv3+ network model, the model experiment parameters can be set, with Adam selected as the optimizer, a total training batch size of 100 epochs, experimental results recorded every 3 training batches, a batch size of 16, 4 workers, and a maximum learning rate of 5×10-3. The augmented sample dataset obtained in the previous steps is divided into training, testing, and validation sets in an 8:1:1 ratio for training, resulting in a pre-trained semantic segmentation model.

[0044] In addition, in this step, evaluation metrics can be selected to assess the model's detection results, and comparative and ablation experiments can be set up to verify the reliability and accuracy of the optimized semantic segmentation model in step five: Five evaluation metrics were selected: mean Intersection over Union (mIoU), mean pixel classification precision (mPA), accuracy, complexity (FLOPs), and number of parameters (Params) to quantitatively evaluate the reliability of the optimized semantic segmentation model in step five. The results are as follows: Figure 6 As shown; ; ; ; ; ; In the formula, TP i FP represents the actual number of pixels in category i. i FN represents the number of false positive pixels of category i. i IoU represents the number of false negative pixels in category i. i The intersection-union ratio (IU) of category i is represented by C, where C is the number of categories, Cinput represents the number of input channels, Coutput represents the number of output channels, K is the kernel size, and H×W represents the size of the input feature map.

[0045] Comparative experiments were conducted using U-Net, HPNet, PSPNet, and DeepLabv3+ models, along with the optimized semantic segmentation model from step five. The magnitudes of the five evaluation metrics in the results of the five sets of experiments were compared. Four ablation experiments were designed to verify the application effect of the model: the DeepLabv3+ model, the DeepLabv3+ model with only the backbone network replaced by MobileNetV3, the DeepLabv3+ model with the backbone network replaced by MobileNetV3 and the introduction of a strip pooling module, and the optimized semantic segmentation model from step five. The magnitudes of the five evaluation metrics in the results of the four sets of experiments were compared.

[0046] The local image regions obtained by cropping in the previous steps are processed based on a pre-trained semantic segmentation model. Multi-level features of the local image regions are extracted using MobileNetV3. Contextual information is extracted using a spatial pyramid pooling structure containing strip pooling (SP). The multi-level features and contextual information are fused together, and the tailings pond extraction result is generated based on the fused features.

[0047] By combining object detection with semantic segmentation, the aforementioned method effectively alleviates the issues of missed detections and false detections in tailings dam identification using a single semantic segmentation model, significantly improving overall recognition accuracy. First, the object detection model quickly locates potential tailings dams using candidate bounding boxes, greatly narrowing the processing range of the semantic segmentation model and avoiding interference from complex backgrounds in wide-area imagery. Second, cropping local image regions allows the semantic segmentation model to focus on high-value areas, compensating for the coarse-grained nature of object detection through pixel-level fine segmentation. The two models work together through a "detect first, segment later" approach, forming a "coarse localization - fine characterization" collaborative mechanism. This cascaded strategy fully integrates the technical advantages of object detection and semantic segmentation, improving the accuracy of tailings dam boundary extraction while reducing the missed detection rate of small tailings dams. Especially when processing remote sensing images with complex scenes, the complementary nature of the two models effectively overcomes the limitations of feature extraction from a single model, reducing the false positive rate of similar features. Thus, this method, through dual verification in both spatial and semantic dimensions, provides a more robust technical path for tailings dam extraction from remote sensing images.

[0048] Furthermore, based on the above steps, the present invention also provides a tailings dam extraction method tailored to specific circumstances, comprising the following steps: Step 1: Overview of the study area and experimental data: The remote sensing image data released by the historical water-based map of a certain year was selected as the data source. The data covers the entire target city. The image originated from the Ziyuan-2 and Ziyuan-3 satellites, and its spatial resolution is 2.39m.

[0049] Step Two: Image Data Preprocessing: Due to the massive amount of remote sensing image data, to improve the efficiency and accuracy of the target detection model for tailings ponds, this invention utilizes Python combined with the GDAL library to perform block slicing preprocessing on the original remote sensing images of the target city, dividing them into image slices of 1024×1024 pixels to ensure that the slice size is consistent with the training dataset used in this invention. Simultaneously, the geographic latitude and longitude coordinates of the image slices are preserved during the slicing process to ensure that after the model prediction is completed, the slice stitching can accurately restore the spatial location in the original remote sensing image. A total of 3683 remote sensing image slices were obtained through image slicing.

[0050] Step 3: Obtain remote sensing image slices containing tailings ponds: Input the image slices obtained in Step 2 into the optimized target detection model for identification and detection, and a total of 288 candidate target instances are identified.

[0051] Step 4: Through visual interpretation, further screen the candidate target instances of tailings ponds from Step 3 to obtain remote sensing image slices of the target city containing tailings ponds.

[0052] Step 5: Input the remote sensing image slices containing tailings ponds of the target city from Step 4 into the optimized semantic segmentation model to extract the tailings ponds, and obtain the outline boundary information and spatial location information of the tailings ponds in the remote sensing image of the target city.

[0053] Step Six: To verify the effectiveness of the proposed method in tailings dam extraction in wide-area remote sensing scenarios, this invention selected two areas with relatively uniform tailings dam distribution in the target city, each with an area of ​​44.9 square kilometers. Three sets of control experiments were designed. Scheme 1 directly used a semantic segmentation model to extract tailings dams from the entire remote sensing image; Scheme 2 directly segmented the original remote sensing image and extracted tailings dams based on the slice data using a semantic segmentation model; Scheme 3 adopted the proposed method, employing a two-stage extraction strategy of "detection first, segmentation later," first locating the tailings dam area, and then performing refined segmentation on the detection results. The model extraction results of the three schemes in the two demonstration areas were compared, and the results are as follows: Figure 7 As can be clearly seen from the figure, the method of the present invention utilizes a two-stage extraction strategy of "detection first, segmentation later". First, it detects and identifies the spatial location of the tailings pond, which simplifies the range that needs to be segmented later, eliminates a large amount of useless background information, effectively avoids missegmentation when segmenting the tailings pond in a complex background, and achieves higher overall segmentation accuracy.

[0054] Secondly, the present invention also provides a tailings dam extraction device based on wide-area remote sensing imagery, such as... Figure 8 As shown, it includes: The acquisition module 201 is used to acquire wide-area remote sensing images containing tailings ponds.

[0055] The determination module 202 is used to process the wide-area remote sensing image into a pre-trained target detection model and output candidate bounding boxes containing tailings dam location information.

[0056] The segmentation module 203 is used to crop out a local image region containing a tailings pond from the original wide-area remote sensing image based on candidate bounding boxes.

[0057] The extraction module 204 is used to input the local image region into a pre-trained semantic segmentation model for processing and output the tailings pond extraction results.

[0058] By employing the aforementioned device and combining object detection with semantic segmentation models, the problems of missed detections and false detections in tailings dam identification using a single semantic segmentation model can be effectively mitigated, significantly improving overall recognition accuracy. First, the object detection model quickly locates potential tailings dams using candidate bounding boxes, greatly narrowing the processing range of the semantic segmentation model and avoiding interference from complex backgrounds in wide-area images on segmentation accuracy. Second, cropping local image regions allows the semantic segmentation model to focus on high-value areas, compensating for the coarse-grained defects of object detection through pixel-level fine segmentation. The two models form a collaborative mechanism of "coarse localization - fine characterization" through "detection first, segmentation later." This cascaded strategy fully integrates the technical advantages of object detection and semantic segmentation, improving the accuracy of tailings dam boundary extraction while reducing the missed detection rate of small tailings dams. Especially when processing remote sensing images with complex scenes, the complementary nature of the two models effectively overcomes the limitations of feature extraction by a single model, reducing the false positive rate of similar features. Thus, this method, through dual verification in spatial and semantic dimensions, provides a more robust technical path for tailings dam extraction from remote sensing images.

[0059] The present invention also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described... Figure 2 The steps of the tailings dam extraction method based on wide-area remote sensing imagery are provided.

[0060] This invention also provides a computer device. At the hardware level, the computer device includes a processor, an internal bus, a network interface, memory, and non-volatile memory, and may also include other hardware required for various operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then executes it to achieve the above-mentioned functions. Figure 2 The steps of the tailings dam extraction method based on wide-area remote sensing imagery are provided.

[0061] Those skilled in the art will understand that embodiments of the present invention can be provided as 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, 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 patent of the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A method for extracting tailing ponds based on wide-area remote sensing images, characterized in that, The method comprises: acquiring a wide-area remote sensing image containing a tailings pond; inputting the wide-area remote sensing image into a pre-trained target detection model for processing to output a candidate bounding box containing tailings pond location information; based on the candidate bounding box, cutting a local image region containing the tailings pond from the original wide-area remote sensing image; inputting the local image region into a pre-trained semantic segmentation model for processing to output a tailings pond extraction result.

2. The method of claim 1, wherein, The target detection model is an improved YOLOv8-obb network model, and the improvement comprises: embedding an efficient convolution unit SCConv in the C2f module of the YOLOv8-obb network, and introducing a coordinate attention module CA at the connection between the Neck and the Head.

3. The method of claim 2, wherein, Inputting the wide-area remote sensing image into the pre-trained target detection model for processing to output a candidate bounding box containing tailings pond location information comprises: extracting image features through the backbone network embedded with the SCConv efficient convolution unit; generating multi-scale fusion features through the multi-scale feature fusion network of the Neck structure; spatial-channel weighted enhancing the multi-scale fusion features through the coordinate attention module CA; generating the tailings pond candidate bounding box based on the weighted features.

4. The method of claim 1, wherein, The semantic segmentation model is an improved DeepLabv3+ network model, and the improvement comprises: introducing MobileNetV3 as the backbone feature extraction network, introducing strip pooling SP to replace the global average pooling of the ASPP module in the DeepLabv3+ model, and adding CBAM modules at the feature fusion stage of the encoder end and the decoder end of the DeepLabv3+ model, respectively.

5. The method of claim 4, wherein, Inputting the local image region into the pre-trained semantic segmentation model for processing to output a tailings pond extraction result comprises: extracting multi-level features of the local image region through MobileNetV3; extracting context information using a spatial pyramid pooling structure containing strip pooling SP; generating the tailings pond extraction result based on the fused features by fusing the multi-level features and the context information.

6. An apparatus for device failure feature extraction, comprising: The device comprises: an acquisition module configured to acquire a wide-area remote sensing image containing a tailings pond; a determination module configured to input the wide-area remote sensing image into a pre-trained target detection model for processing to output a candidate bounding box containing tailings pond location information; a segmentation module configured to cut a local image region containing the tailings pond from the original wide-area remote sensing image based on the candidate bounding box; an extraction module configured to input the local image region into a pre-trained semantic segmentation model for processing to output a tailings pond extraction result.

7. A computer readable storage medium characterized in that, The storage medium stores a computer program, and the computer program is executed by the processor to implement the method of any one of claims 1-5.

8. A computer device, comprising: The device comprises a memory, a processor, and a computer program stored on the memory and executable on the processor, and the processor implements the method of any one of claims 1-5 when executing the program.

Images