A solid waste yard remote sensing automatic identification method and system

Abstract

The present application belongs to the technical field of remote sensing, and discloses a solid waste yard remote sensing automatic identification method and system, wherein the method comprises: obtaining original remote sensing images of a target area and performing pretreatment; constructing a solid waste yard multi-element remote sensing image database according to in-situ solid waste yard image data sets, the pretreated remote sensing images and public remote sensing image data sets; introducing RFB modules and ECA modules into a YOLOv8 model; training the improved YOLOv8 model using the solid waste yard multi-element remote sensing image database, and using a SIoU loss function as a target frame regression loss in the training stage; inputting the pretreated remote sensing images into the trained improved YOLOv8 model, and outputting the predicted frame position, range and confidence of the solid waste yard. The present application improves the authenticity and reliability of the training samples, has stronger adaptability, better feature discrimination ability, and improves the positioning and boundary regression accuracy of the solid waste yard target.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing technology, and specifically relates to a remote sensing automatic identification method and system for solid waste dumps. Background Technology

[0002] With the continuous advancement of urbanization and industrial activities, the accumulation of solid waste such as construction waste, industrial solid waste, and domestic waste in areas surrounding cities, construction sites, riverbanks, and wastelands is becoming increasingly prominent. If solid waste dumps are not effectively regulated for a long period, they will not only waste land resources but may also cause soil pollution, water pollution, and ecological damage, posing potential risks to regional ecological security and public health. Therefore, timely and accurate monitoring of the spatial distribution, accumulation range, and changes in solid waste dumps is a crucial task in ecological environment supervision and urban governance.

[0003] Traditional solid waste monitoring primarily relies on manual on-site inspections, a method that is not only time-consuming, labor-intensive, and costly, but also struggles to achieve large-scale, high-frequency coverage. Remote sensing technology, with its advantages of wide coverage, high acquisition efficiency, and relatively low cost, is widely used in environmental monitoring and resource surveys. High-resolution remote sensing imagery allows for the identification and location of solid waste dumps on a large regional scale, providing decision-making support for environmental regulatory departments. Early solid waste identification relied mainly on manual visual interpretation or rule-based image processing methods; however, with the improvement in remote sensing data resolution and coverage, traditional methods have become insufficient in terms of efficiency and accuracy to meet practical needs.

[0004] With the development of deep learning technology, target detection methods based on convolutional neural networks (CNN) have been gradually introduced into the field of remote sensing image interpretation, providing a new technical path for the automatic identification of solid waste dumps.

[0005] For example, patent CN120544041A discloses a method for identifying urban-scale solid waste landfill sites based on artificial intelligence target recognition algorithms. The core idea is to construct a multi-source remote sensing image database and improve the structure of the YOLOv8 model to train a target recognition model suitable for identifying urban-scale solid waste landfill sites. Then, the model is applied to remote sensing images of the target area to achieve automatic identification and on-site verification of solid waste landfill sites.

[0006] While the aforementioned patents can achieve automatic identification of solid waste landfill sites on an urban scale, improving identification efficiency compared to manual survey methods, the following technical problems still exist: 1. Solid waste dumps exhibit significant differences in target scale, and remote sensing imagery fails to adequately represent their multi-scale features. Solid waste dumps in remote sensing imagery span a wide range of scales, including both large construction waste dumps and scattered, small-scale solid waste accumulation points. Existing remote sensing identification methods and general target detection models have limited ability to represent multi-scale features, making it difficult to simultaneously account for targets of different scales, and leading to frequent missed detections.

[0007] 2. Solid waste features are unstable in complex backgrounds, resulting in low recognition accuracy. Solid waste dumps are often mixed with bare land, vegetation, roads, and piles of soil. Their spectral and textural features are easily affected by lighting conditions, shadow changes, and differences in pile morphology. Existing methods are not good at focusing on key target features and are easily affected by background interference, leading to false detections and misdetections.

[0008] 3. Insufficient accuracy in target localization and boundary regression: Existing solid waste dump identification methods typically only focus on the overlap between the predicted and actual bounding boxes during the target bounding box regression process, without adequately considering the spatial positional relationship, shape differences, and orientation information of the target. This results in low target localization accuracy in complex remote sensing scenarios, affecting subsequent regulatory and area assessment applications. Summary of the Invention

[0009] To address the above problems, this invention provides a remote sensing automatic identification method for solid waste dumps, comprising the following steps: According to the monitoring task requirements, acquire the original remote sensing images of the target area and preprocess the original remote sensing images. Based on the acquired on-site solid waste dump image dataset, preprocessed remote sensing images, and publicly available remote sensing image datasets, a multi-dimensional remote sensing image database of solid waste dumps is constructed. An RFB module is introduced into the SPPF structure of the YOLOv8 model to enhance the YOLOv8 model's ability to model the receptive field of solid waste dumps at different scales. An ECA module is also introduced into the YOLOv8 model to adaptively weight the feature channels, thereby enhancing the YOLOv8 model's ability to focus on key features of solid waste dumps, thus obtaining an improved YOLOv8 model. The improved YOLOv8 model was trained using a multivariate remote sensing image database of solid waste dumps as a sample set. During the training phase, the SIoU loss function was used as the target box regression loss to constrain the difference between the predicted box and the true box. The preprocessed remote sensing image is input into the trained improved YOLOv8 model, which outputs the predicted bounding box location, extent, and confidence level of the solid waste dump.

[0010] Furthermore, based on the monitoring task requirements, the original remote sensing images of the target area are acquired, and the original remote sensing images are preprocessed, including the following steps: Based on the monitoring task requirements, select raw visible light remote sensing images that cover the target area and meet the spatial resolution and temporal requirements; The original remote sensing image is sequentially subjected to radiometric correction, atmospheric correction and geometric correction to obtain the corrected remote sensing image; The panchromatic and multispectral images in the corrected remote sensing image are fused, and the fused remote sensing image is used as the preprocessed remote sensing image.

[0011] Furthermore, based on the acquired on-site solid waste dump image dataset, preprocessed remote sensing images, and publicly available remote sensing image datasets, a multivariate remote sensing image database of solid waste dumps is constructed, including the following steps: The vector data of the solid waste dump that has been verified on-site is spatially overlaid with the preprocessed remote sensing image, and the corresponding image area is extracted to obtain the on-site solid waste dump image dataset. By combining publicly available remote sensing image datasets, we have unified and organized on-site solid waste dump image datasets from different sources, at different resolutions, and at different times. The solid waste dump targets in the processed solid waste dump image dataset were manually labeled to construct a multi-source remote sensing image database of solid waste dumps containing multi-source, multi-scale, and multi-background features.

[0012] Furthermore, the ability of the YOLOv8 model to model the receptive field of solid waste disposal sites at different scales is enhanced through the RFB module, including the following steps: In the RFB module, multiple parallel convolutional branches are set up, each using a convolutional kernel with a different dilation rate to simulate receptive field ranges of different sizes, including: The input feature map is simultaneously fed into multiple convolutional branches with different dilation rates to extract feature information at different scales. The feature information output by each branch is concatenated and fused through convolution. The fused features are then added element-wise to the original input features to form a residual connection structure. Through the residual connection structure, the improved YOLOv8 model can simultaneously perceive the overall outline features of large-scale solid waste dumps and the local detail features of small-scale solid waste targets.

[0013] Furthermore, the YOLOv8 model's ability to focus on key features of solid waste stockpiles is enhanced by adaptively weighting the feature channels using the ECA module, including the following steps: The input feature map is subjected to global average pooling through the ECA module to obtain global description information for each feature channel. Local interaction between feature channels is achieved through one-dimensional convolution. Feature channel weights are generated by the Sigmoid activation function and then multiplied with the original feature map channel by channel to obtain the recalibrated feature map.

[0014] Furthermore, during the training phase, the SIoU loss function is used as the regression loss for the target box to constrain the difference between the predicted and ground truth boxes, including: The SIoU loss function takes into account the following factors during training: the spatial distance between the predicted bounding box and the ground truth bounding box, the angular deviation between the predicted bounding box and the ground truth bounding box, and the shape difference between the predicted bounding box and the ground truth bounding box.

[0015] Furthermore, the predicted frame location, extent, and confidence level of the solid waste disposal site are output in vector file format.

[0016] The present invention also provides a remote sensing automatic identification system for solid waste dumps, comprising: The image preprocessing module is used to acquire the original remote sensing images of the target area according to the monitoring task requirements, and to preprocess the original remote sensing images. The database construction module is used to construct a multi-dimensional remote sensing image database of solid waste dumps based on the acquired on-site solid waste dump image dataset, preprocessed remote sensing images, and publicly available remote sensing image datasets. The model building module is used to introduce the RFB module into the SPPF structure of the YOLOv8 model. The RFB module enhances the YOLOv8 model's ability to model the receptive field of solid waste dumps at different scales. The ECA module is introduced into the YOLOv8 model. The ECA module performs adaptive weighting on the feature channels, which enhances the YOLOv8 model's ability to focus on the key features of solid waste dumps, thus obtaining an improved YOLOv8 model. The model training module is used to train the improved YOLOv8 model using a multivariate remote sensing image database of solid waste dumps as a sample set. During the training phase, the SIoU loss function is used as the target box regression loss to constrain the difference between the predicted box and the true box. The model prediction module is used to input preprocessed remote sensing images into a trained, improved YOLOv8 model and output the predicted bounding box location, extent, and confidence level of the solid waste dump.

[0017] Furthermore, the model building module is specifically used for: In the RFB module, multiple parallel convolutional branches are set up, each using a convolutional kernel with a different dilation rate to simulate receptive field ranges of different sizes, including: The input feature map is simultaneously fed into multiple convolutional branches with different dilation rates to extract feature information at different scales. The feature information output by each branch is concatenated and fused through convolution. The fused features are then added element-wise to the original input features to form a residual connection structure. Through the residual connection structure, the improved YOLOv8 model can simultaneously perceive the overall outline features of large-scale solid waste dumps and the local detail features of small-scale solid waste targets.

[0018] Furthermore, the model building module is specifically used for: The input feature map is subjected to global average pooling through the ECA module to obtain global description information for each feature channel. Local interaction between feature channels is achieved through one-dimensional convolution. Feature channel weights are generated by the Sigmoid activation function and then multiplied with the original feature map channel by channel to obtain the recalibrated feature map.

[0019] The beneficial effects of this invention are: 1. At the sample construction level, this invention improves the authenticity and reliability of training samples. Existing technologies mainly rely on public datasets or indirectly compiled sample data. However, this invention constructs a sample library containing ground truth constraints by spatially overlaying solid waste dump vector data obtained from field verification with remote sensing images. This sample library is then fused with multi-source remote sensing image data for training, which helps to reduce sample labeling errors and improve the credibility and stability of model training data.

[0020] 2. This invention has stronger adaptability in the identification of solid waste dumps at multiple scales. In view of the large scale difference of solid waste dumps in remote sensing images, this invention introduces the RFB module into the SPPF structure of the YOLOv8 model. By enhancing the model's feature perception ability of solid waste dump targets of different sizes through multi-scale receptive fields, it has better recognition adaptability in scenarios where multiple scale targets exist at the same time.

[0021] 3. The present invention has better feature discrimination ability under complex background conditions. Compared with the existing technology that focuses on cross-scale feature fusion, the present invention introduces the ECA channel attention mechanism in the neck of the model to adaptively weight the feature channels, which helps to enhance the relevant features of solid waste dumps and suppress background interference such as bare soil and cultivated land, thereby improving the recognition stability in complex remote sensing scenarios.

[0022] 4. This invention has advantages in the accuracy of target positioning in solid waste dumps. In response to the problem of irregular shape and complex boundaries of solid waste dumps, this invention uses the SIoU loss function as the target box regression constraint during the model training stage. By comprehensively considering the differences in position, direction and shape, it guides the predicted box to align with the real box faster and more accurately, thereby improving the positioning and boundary regression accuracy of solid waste dump targets.

[0023] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description and the drawings. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 A schematic diagram of the main process of a remote sensing automatic identification method for solid waste dumps according to an embodiment of the present invention is shown. Figure 2 A detailed flowchart of a remote sensing automatic identification method for solid waste dumps according to an embodiment of the present invention is shown. Figure 3 A schematic diagram of a remote sensing automatic identification system for solid waste dumps according to an embodiment of the present invention is shown. Figure 4 The diagram illustrates the identification method according to an embodiment of the present invention for on-site inspection of solid waste storage site locations. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] It should be noted that the terms "first," "second," etc., used in this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate for the embodiments of this application described herein.

[0028] This invention provides a remote sensing automatic identification method and system for solid waste dumps. Based on deep learning and remote sensing technology, it achieves automated identification and location of solid waste dumps, thereby reducing manual intervention, improving the efficiency and timeliness of large-scale solid waste dump monitoring, enhancing the overall identification accuracy of multi-scale solid waste dumps, reducing missed detections, minimizing background interference, improving the model's identification stability and accuracy in complex remote sensing scenarios, and enhancing the model's accuracy in predicting the location and boundaries of solid waste dump targets in remote sensing images. This makes the identification results more accurate in spatial positioning and shape representation, meeting the needs of subsequent regulatory and analytical applications.

[0029] like Figure 1 and Figure 2 As shown, a remote sensing automatic identification method for solid waste dumps includes the following steps: S1. Based on the monitoring task requirements, acquire the original remote sensing images of the target area and preprocess the original remote sensing images, including: S11. Based on the monitoring task requirements, select raw visible light remote sensing images that cover the target area and meet the spatial resolution and temporal requirements.

[0030] S12. Perform radiometric correction, atmospheric correction and geometric correction on the original remote sensing image in sequence to eliminate sensor differences, atmospheric effects and geometric distortion, and obtain the corrected remote sensing image.

[0031] S13. When conditions permit, perform image fusion on the panchromatic image and multispectral image in the corrected remote sensing image to obtain a high-quality input image that takes into account both spatial resolution and spectral information. Use the fused remote sensing image as the preprocessed remote sensing image. The processed remote sensing image is used as the unified input data for subsequent database construction and model inference.

[0032] In this step, the original remote sensing images of the target area are standardized to ensure the consistency of the quality of subsequent model input data.

[0033] S2. Based on the acquired on-site solid waste dump image dataset, preprocessed remote sensing images, and publicly available remote sensing image datasets, construct a multivariate remote sensing image database for solid waste dumps, including the following steps: S21. Spatially overlay the vector data of the solid waste dump after on-site verification with the preprocessed remote sensing image, extract the corresponding image area, and obtain the on-site solid waste dump image dataset.

[0034] S22. Combining publicly available remote sensing image datasets, uniformly organize on-site solid waste dump image datasets from different sources, resolutions, and time phases; manually label the solid waste dump targets in the organized solid waste dump image datasets to construct a multi-source remote sensing image database of solid waste dumps containing multi-source, multi-scale, and multi-background features.

[0035] In this embodiment of the invention, the multi-dimensional remote sensing image database of solid waste dumps is used as a training sample input to the model building and training module, and together with the model structure, it determines the recognition effect.

[0036] At the sample construction level, the embodiments of the present invention improve the authenticity and reliability of training samples. Existing technologies mainly rely on public datasets or indirectly compiled sample data. However, the embodiments of the present invention construct a sample library containing ground truth constraints by spatially overlaying vector data of solid waste dumps obtained from field verification with remote sensing images. This sample library is then fused with multi-source remote sensing image data for training, which helps to reduce sample labeling errors and improve the credibility and stability of model training data.

[0037] S3. Introduce the RFB (Receptive Field Block) module into the SPPF (Spatial Pyramid Pooling-Fast) structure of the YOLOv8 model, and introduce the ECA module into the YOLOv8 model to obtain an improved YOLOv8 model.

[0038] The improved YOLOv8 model in this embodiment of the invention is used to achieve effective extraction and accurate localization of multi-scale features of solid waste dumps.

[0039] The RFB module is located at the SPPF position in the YOLOv8 model. It enhances the YOLOv8 model's ability to model the receptive field of solid waste disposal sites at different scales. This includes setting up multiple parallel convolutional branches within the RFB module, each using a convolutional kernel with a different dilation rate to simulate receptive field ranges of varying sizes. Its operation involves: The input feature map is simultaneously fed into multiple convolutional branches with different dilation rates to extract feature information at different scales. The feature information output by each branch is concatenated and fused through convolution. The fused features are then added element-wise to the original input features to form a residual connection structure.

[0040] By using residual connection structures, the improved YOLOv8 model can simultaneously perceive the overall outline features of large-scale solid waste dumps and the local detailed features of small-scale solid waste targets, effectively alleviating the difficulty of multi-scale target identification.

[0041] The identification method of this invention has stronger adaptability in the identification of multi-scale solid waste dumps. In view of the large scale difference of solid waste dumps in remote sensing images, this invention introduces the RFB module into the SPPF structure of the YOLOv8 model. By enhancing the model's feature perception ability of solid waste dump targets of different sizes through multi-scale receptive fields, it has better identification adaptability in scenarios where multiple targets exist at the same time.

[0042] The ECA (Efficient Channel Attention) module is located in the neck of the YOLOv8 model. The ECA module adaptively weights the feature channels, thereby enhancing the YOLOv8 model's ability to focus on key features of solid waste dumps. This includes: using the ECA module to perform global average pooling on the input feature map to obtain global descriptive information for each feature channel. Local interaction between feature channels is achieved through one-dimensional convolution, avoiding parameter redundancy caused by fully connected layers. The Sigmoid activation function is used to generate feature channel weights, and then the feature channel weights are multiplied with the original feature map channel by channel to obtain the recalibrated feature map.

[0043] In this embodiment of the invention, the ECA module enables the model to automatically enhance the characteristic channel response related to solid waste dumps, suppress background ground object interference, and improve the recognition stability in complex remote sensing scenarios.

[0044] The identification method of this invention has better feature discrimination ability under complex background conditions. Compared with the existing technology that focuses on cross-scale feature fusion, this invention introduces the ECA channel attention mechanism in the neck of the YOLOv8 model to adaptively weight the feature channels, which helps to enhance the relevant features of solid waste dumps and suppress background interference such as bare soil and cultivated land, thereby improving the identification stability in complex remote sensing scenarios.

[0045] S4. The improved YOLOv8 model was trained using a multivariate remote sensing image database of solid waste dumps as a sample set.

[0046] In the training phase of the improved YOLOv8 model, the SIoU loss function is used as the target box regression loss to constrain the difference between the predicted box and the ground truth box.

[0047] Unlike traditional IoU loss functions, the SIoU loss function takes into account the spatial distance between the predicted bounding box and the ground truth bounding box, the angular deviation between the predicted bounding box and the ground truth bounding box, and the shape difference between the predicted bounding box and the ground truth bounding box during training.

[0048] In terms of target positioning accuracy in solid waste dumps, this invention addresses the problem of irregular shapes and complex boundaries in solid waste dumps by using the SIoU loss function as a target box regression constraint during the model training phase. By comprehensively considering differences in position, orientation, and shape, it guides the predicted box to align with the real box more quickly and accurately, thereby improving the positioning and boundary regression accuracy of solid waste dump targets.

[0049] S5. Input the preprocessed remote sensing image into the trained improved YOLOv8 model and output the predicted bounding box location, range, and confidence level of the solid waste dump.

[0050] The trained and improved YOLOv8 model in this embodiment of the invention is used to apply the trained model to actual remote sensing images to achieve automatic identification of solid waste dumps.

[0051] For example, the predicted location, extent, and confidence level of the solid waste disposal site can be output as a vector file for subsequent on-site verification, statistical analysis, and regulatory applications.

[0052] Based on the above-mentioned remote sensing automatic identification method for solid waste dumps, such as Figure 3As shown in the figure, this embodiment of the invention also provides a remote sensing automatic identification system for solid waste dumps, including an image preprocessing module, a database construction module, a model building module, a model training module, and a model prediction module.

[0053] The image preprocessing module is used to acquire the original remote sensing images of the target area according to the monitoring task requirements and to preprocess the original remote sensing images; the database construction module is used to construct a multi-dimensional remote sensing image database of solid waste dumps based on the acquired on-site solid waste dump image dataset, the preprocessed remote sensing images, and the publicly available remote sensing image dataset.

[0054] The model building module is used to introduce the RFB module into the SPPF structure of the YOLOv8 model. The RFB module enhances the YOLOv8 model's ability to model the receptive field of solid waste stockpiles at different scales. The ECA module is introduced into the YOLOv8 model. The ECA module adaptively weights the feature channels, enhancing the YOLOv8 model's ability to focus on key features of solid waste stockpiles, thus obtaining an improved YOLOv8 model.

[0055] The model training module is used to train the improved YOLOv8 model using a multivariate remote sensing image database of solid waste dumps as a sample set. During the training phase, the SIoU loss function is used as the target box regression loss to constrain the difference between the predicted box and the true box.

[0056] The model prediction module is used to input preprocessed remote sensing images into a trained, improved YOLOv8 model and output the predicted bounding box location, extent, and confidence level of the solid waste dump.

[0057] Compared with existing YOLOv8-based urban-scale solid waste landfill identification methods, the identification method and system of this invention have at least the following differences and advantages: like Figure 4 As shown, Figure 4 This diagram illustrates the location of a solid waste storage site during on-site inspection, based on an embodiment of the present invention. Figure 4 As can be seen from the above, at the feature extraction level, the embodiments of the present invention enhance the multi-scale receptive field modeling capability by introducing the RFB module, rather than relying solely on conventional convolution or other structures, thus making it more suitable for the characteristics of large scale differences in solid waste dumps in remote sensing images.

[0058] At the feature fusion level, this embodiment of the invention introduces a lightweight ECA channel attention mechanism to improve the model's ability to focus on key features without significantly increasing the number of model parameters. At the target regression level, this embodiment of the invention uses the SIoU loss function to improve the model's prediction accuracy of the location and shape of solid waste dumps and enhance the practicality of the identification results.

[0059] Through the synergistic effect of the above technical solutions, the embodiments of the present invention can achieve high-precision, stable and automated identification of solid waste dumps in complex remote sensing scenarios.

[0060] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A remote sensing automatic identification method for solid waste dumps, characterized in that, Includes the following steps: According to the monitoring task requirements, acquire the original remote sensing images of the target area and preprocess the original remote sensing images. Based on the acquired on-site solid waste dump image dataset, preprocessed remote sensing images, and publicly available remote sensing image datasets, a multi-dimensional remote sensing image database of solid waste dumps is constructed. An RFB module is introduced into the SPPF structure of the YOLOv8 model to enhance the YOLOv8 model's ability to model the receptive field of solid waste dumps at different scales. An ECA module is also introduced into the YOLOv8 model to adaptively weight the feature channels, thereby enhancing the YOLOv8 model's ability to focus on key features of solid waste dumps, thus obtaining an improved YOLOv8 model. The improved YOLOv8 model was trained using a multivariate remote sensing image database of solid waste dumps as a sample set. During the training phase, the SIoU loss function was used as the target box regression loss to constrain the difference between the predicted box and the true box. The preprocessed remote sensing image is input into the trained improved YOLOv8 model, which outputs the predicted bounding box location, extent, and confidence level of the solid waste dump.

2. The method for automatic remote sensing identification of solid waste dumps according to claim 1, characterized in that, Based on the monitoring task requirements, acquire raw remote sensing images of the target area and preprocess the raw remote sensing images, including the following steps: Based on the monitoring task requirements, select raw visible light remote sensing images that cover the target area and meet the spatial resolution and temporal requirements; The original remote sensing image is sequentially subjected to radiometric correction, atmospheric correction and geometric correction to obtain the corrected remote sensing image; The panchromatic and multispectral images in the corrected remote sensing image are fused, and the fused remote sensing image is used as the preprocessed remote sensing image.

3. The method for automatic remote sensing identification of solid waste dumps according to claim 1, characterized in that, Based on the acquired on-site solid waste disposal site image dataset, preprocessed remote sensing images, and publicly available remote sensing image datasets, a multivariate remote sensing image database of solid waste disposal sites is constructed, including the following steps: The vector data of the solid waste dump that has been verified on-site is spatially overlaid with the preprocessed remote sensing image, and the corresponding image area is extracted to obtain the on-site solid waste dump image dataset. By combining publicly available remote sensing image datasets, we have unified and organized on-site solid waste dump image datasets from different sources, at different resolutions, and at different times. The solid waste dump targets in the processed solid waste dump image dataset were manually labeled to construct a multi-source remote sensing image database of solid waste dumps containing multi-source, multi-scale, and multi-background features.

4. The method for automatic remote sensing identification of solid waste dumps according to claim 1, characterized in that, Enhancing the receptive field modeling capability of YOLOv8 models for solid waste disposal sites of different scales by using the RFB module includes the following steps: In the RFB module, multiple parallel convolutional branches are set up, each using a convolutional kernel with a different dilation rate to simulate receptive field ranges of different sizes, including: The input feature map is simultaneously fed into multiple convolutional branches with different dilation rates to extract feature information at different scales. The feature information output by each branch is concatenated and fused through convolution. The fused features are then added element-wise to the original input features to form a residual connection structure. Through the residual connection structure, the improved YOLOv8 model can simultaneously perceive the overall outline features of large-scale solid waste dumps and the local detail features of small-scale solid waste targets.

5. The method for automatic remote sensing identification of solid waste dumps according to claim 1, characterized in that, The YOLOv8 model enhances its ability to focus on key features of solid waste stockpiles by adaptively weighting the feature channels using the ECA module, including the following steps: The input feature map is subjected to global average pooling through the ECA module to obtain global description information for each feature channel. Local interaction between feature channels is achieved through one-dimensional convolution. Feature channel weights are generated by the Sigmoid activation function and then multiplied with the original feature map channel by channel to obtain the recalibrated feature map.

6. The method for automatic remote sensing identification of solid waste dumps according to claim 1, characterized in that, During the training phase, the SIoU loss function is used as the regression loss for the target box, constraining the difference between the predicted and ground truth boxes, including: The SIoU loss function takes into account the following factors during training: the spatial distance between the predicted bounding box and the ground truth bounding box, the angular deviation between the predicted bounding box and the ground truth bounding box, and the shape difference between the predicted bounding box and the ground truth bounding box.

7. The method for automatic remote sensing identification of solid waste dumps according to any one of claims 1-6, characterized in that, The predicted frame location, extent, and confidence level for solid waste disposal sites are output as vector files.

8. A remote sensing automatic identification system for solid waste dumps, characterized in that, include: The image preprocessing module is used to acquire the original remote sensing images of the target area according to the monitoring task requirements, and to preprocess the original remote sensing images. The database construction module is used to construct a multi-dimensional remote sensing image database of solid waste dumps based on the acquired on-site solid waste dump image dataset, preprocessed remote sensing images, and publicly available remote sensing image datasets. The model building module is used to introduce the RFB module into the SPPF structure of the YOLOv8 model. The RFB module enhances the YOLOv8 model's ability to model the receptive field of solid waste dumps at different scales. The ECA module is introduced into the YOLOv8 model. The ECA module performs adaptive weighting on the feature channels, which enhances the YOLOv8 model's ability to focus on the key features of solid waste dumps, thus obtaining an improved YOLOv8 model. The model training module is used to train the improved YOLOv8 model using a multivariate remote sensing image database of solid waste dumps as a sample set. During the training phase, the SIoU loss function is used as the target box regression loss to constrain the difference between the predicted box and the true box. The model prediction module is used to input preprocessed remote sensing images into a trained, improved YOLOv8 model and output the predicted bounding box location, extent, and confidence level of the solid waste dump.

9. The remote sensing automatic identification system for solid waste dumps according to claim 8, characterized in that, The model building module is specifically used for: In the RFB module, multiple parallel convolutional branches are set up, each using a convolutional kernel with a different dilation rate to simulate receptive field ranges of different sizes, including: The input feature map is simultaneously fed into multiple convolutional branches with different dilation rates to extract feature information at different scales. The feature information output by each branch is concatenated and fused through convolution. The fused features are then added element-wise to the original input features to form a residual connection structure. Through the residual connection structure, the improved YOLOv8 model can simultaneously perceive the overall outline features of large-scale solid waste dumps and the local detail features of small-scale solid waste targets.

10. The remote sensing automatic identification system for solid waste dumps according to claim 8, characterized in that, The model building module is specifically used for: The input feature map is subjected to global average pooling through the ECA module to obtain global description information for each feature channel. Local interaction between feature channels is achieved through one-dimensional convolution. Feature channel weights are generated by the Sigmoid activation function and then multiplied with the original feature map channel by channel to obtain the recalibrated feature map.

Images