Waste slag field intelligent identification method and system based on high-resolution remote sensing image

Abstract

The invention provides a waste slag field intelligent identification method and system based on a high-resolution remote sensing image, belongs to the technical field of remote sensing image processing, and aims to solve the problem that real-time monitoring of a waste slag field, especially an ultra-large waste slag field cannot be realized in the prior art. Comprising the following steps: S1, acquiring a remote sensing image of a waste slag field, and preprocessing the remote sensing image; s2, initializing GLI D-DeepLab model parameters, and performing feature extraction on the remote sensing image in the step S1 by adopting a deep convolutional neural network RMT and combining a bidirectional feature pyramid network BiFPN; s3, performing iterative regularization deconvolution processing on the features obtained in the step S2, and fusing shallow features output by the encoder backbone network; and S4, selecting a remote sensing image sample to be detected, and inputting the remote sensing image sample to the trained GLI D-DeepLab model. According to the method, the deep convolutional network RMT and the bidirectional feature pyramid network are combined, and the visual hybrid coding auxiliary detection head is prompted in an auxiliary manner, so that multi-scale features can be effectively extracted, and the identification of the waste slag field is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of remote sensing image processing, and particularly relates to an intelligent identification method and system for waste dumps based on high-resolution remote sensing images. Background Art

[0002] With the acceleration of industrialization and urbanization, the treatment and storage of solid waste have become important issues in environmental management and ecological protection. Waste dumps, as areas for centralized storage and treatment of solid waste, have a significant impact on the environment and human health in terms of their stability and safety. In China, due to complex and diverse terrains and uneven distribution of towns, many waste dumps are located in remote or inaccessible areas, which poses great challenges to traditional manual monitoring methods. Manual inspection is not only inefficient but also difficult to achieve real-time monitoring of these areas, making it difficult to detect and respond to potential environmental problems in a timely manner.

[0003] Along with large-scale infrastructure construction, such as highway, railway network expansion, and water conservancy projects, a large amount of solid waste has been generated, leading to a sharp increase in the number of waste dumps. While these construction projects promote economic development, they also exert great pressure on the ecological environment. Especially in areas with complex terrains, environmental risk management and stability monitoring of waste dumps have become a severe challenge. Traditional manual monitoring methods are not only time-consuming and labor-intensive but also difficult to cover all areas, especially in remote mountainous areas and uninhabited areas, which limits the ability to effectively supervise waste dumps. In addition, manual monitoring cannot achieve real-time tracking of the dynamic changes of waste dumps, resulting in obvious deficiencies in environmental risk assessment and disaster warning.

[0004] At present, there are some problems and weak links in the research on ecological restoration and comprehensive monitoring of waste dumps at home and abroad. Especially in the real-time and effective monitoring of super-large waste dumps, existing technologies such as low-altitude remote sensing by drones cannot achieve real-time monitoring of waste dumps, especially super-large waste dumps, due to limitations in endurance and flight conditions. Therefore, developing an intelligent identification method based on high-resolution remote sensing images is of great significance for improving the efficiency and accuracy of waste dump monitoring. Summary of the Invention

[0005] In view of this, the present invention provides an intelligent identification method and system for waste dumps based on high-resolution remote sensing images to solve the problem that existing technologies such as low-altitude remote sensing by drones cannot achieve real-time monitoring of waste dumps, especially super-large waste dumps, due to limitations in endurance and flight conditions.

[0006] The technical solution adopted by the present invention is as follows:

[0007] An intelligent identification method for waste dumps based on high-resolution remote sensing images includes:

[0008] Step S1: Obtain high-resolution remote sensing images of the spoil ground distribution, preprocess the obtained remote sensing images, and obtain the preprocessed remote sensing images;

[0009] The specific steps of step 1 include the following steps:

[0010] Step S11: Obtain cloudless high-resolution optical remote sensing images;

[0011] Step S12: Draw the boundary of the spoil ground based on the obtained remote sensing images, slice the spoil ground boundary and the remote sensing images, and obtain the sliced images and the corresponding boundary vector files;

[0012] Step S13: Modify the attribute values of the target area and non-target area of the boundary vector file.

[0013] Step S2: Initialize the GLID-DeepLab model parameters, use the deep convolutional neural network RMT as the backbone network Backbone and combine it with the bidirectional feature pyramid network BiFPN to extract features from the remote sensing images preprocessed in step S1;

[0014] The specific steps of step S2 include the following steps:

[0015] Step S21: Put the sliced images and the boundary vector files into the deep convolutional neural network to obtain multi-scale feature information, where the loss function of the deep convolutional neural network RMT is MALoss, as shown in the following formula:

[0016] L MALoss =L MAGICLoss +λL aux

[0017] In the formula, L MAGIC Loss is the MAGIC Loss, L aux is the auxiliary detection head loss function, and λ is the dynamically adjusted weight;

[0018] Step S22: Introduce multi-modal regularization, adaptive weighting, gradient penalty term, information entropy term, and C-Class-Balanced Cross-Entropy into L MAGICLoss to optimize the loss function, as shown in the following formula:

[0019] L MAGIC LOSS =L A +L M +L G +L I +L C

[0020] The specific steps of step S22 include:

[0021] Multimodal data is introduced to regularize the loss function, enabling the GLID-DeepLab model to learn richer feature representations from multiple data sources, as shown in the following equation:

[0022]

[0023] where f image is the image feature, f other_modality is the feature of other modalities, and λ M is the regularization coefficient;

[0024] Adaptive weighting is introduced to dynamically adjust the loss weight according to the complexity of remote sensing image samples or the difficulty of classes, enabling the GLID-DeepLab model to pay more attention to difficult-to-classify samples, as shown in the following equation:

[0025]

[0026] where α i is the adaptive weight, which is proportional to the classification difficulty of the sample, Z i is the label, and f(x i ) is the model output;

[0027] A gradient penalty term is added to the loss function to prevent gradient explosion or disappearance and enhance the generalization ability of the model, as shown in the following equation:

[0028]

[0029] where λ G is the penalty coefficient, is the gradient of the model at the input x i ;

[0030] An information entropy term is introduced to make the prediction results of the GLID-DeepLab model closer to a uniform distribution, thereby improving the model's ability to handle uncertainty, as shown in the following equation:

[0031]

[0032] where H(f(x i )) is the information entropy of the probability distribution of the model output, and λ I is the weight coefficient;

[0033] C-Class-Balanced Cross-Entropy is introduced to make the GLID-DeepLab model handle samples of different classes more fairly by dynamically adjusting the weights of each class, as shown in the following equation:

[0034]

[0035] Among them, β Mi is the weight of category M i and is inversely proportional to the number of samples in this category.

[0036] L aux The loss function of L is as follows:

[0037] L aux = αL Terrain + γL Veg

[0038] In the formula, L Terrain is the loss of the waste dump terrain information, and L Veg is the loss of vegetation coverage. α and γ are dynamically adjusted weights;

[0039]

[0040] In the formula, f terrain (x i ) is the terrain feature predicted by the model, is the real terrain feature, and λ terrain is the regularization coefficient;

[0041]

[0042] In the formula, f veg (x i ) is the vegetation coverage feature predicted by the model, is the real vegetation coverage feature, and λ veg is the regularization coefficient.

[0043] Step S23: The bidirectional feature pyramid network performs cross-layer feature fusion on the multi-scale feature information obtained in step S22 through the top-down and bottom-up bidirectional information flows, and combines the auxiliary detection head that prompts visual hybrid coding. The visual features and the prior prompt information are deeply fused through the layer-by-layer and multi-scale fusion modules to obtain preliminary features;

[0044] Step S24: The encoder improves the ability to capture details through self-supervised pre-training, and optimizes the preliminary features in step S22 through self-supervised learning tasks.

[0045] At the same time, an auxiliary detection head that prompts visual hybrid coding is introduced. The visual features and the prior prompt information are deeply fused through the layer-by-layer and multi-scale fusion modules, and the model training is optimized through the intermediate supervision mechanism to construct an efficient encoder;

[0046] Step S3: Use iterative regularization deconvolution to process the features obtained in step S2, and fuse the shallow features output by the encoder backbone network to finely restore the target boundary;

[0047] The specific steps of step S3 are as follows:

[0048] Step S31: Perform deconvolution processing on the features obtained in step S2 to achieve up-down connection matching. Through iterative regularized deconvolution operations, map the high-dimensional features of the encoder back to a lower dimension to prepare for subsequent feature fusion.

[0049] Step S32: Fuse the features after deconvolution processing with the shallow features extracted by RMT to achieve multi-scale information acquisition.

[0050] Step S4: Randomly select remote sensing image samples to be detected and input them into the trained GLID-DeepLab model to obtain the targets to be recognized.

[0051] The specific steps of step S4 are as follows:

[0052] Step S41: Select relatively typical waste dump image slices for processing, put them into the GLID-DeepLab model to test the recognition accuracy, and evaluate the recognition effect of the GLID-DeepLab model by comparing the output of the GLID-DeepLab model with the actual annotation.

[0053] Step S42: Considering the influence of geological conditions and climate characteristics in different regions on the recognition of waste dumps, multi-region samples are incorporated during the training of the GLID-DeepLab model to enhance the adaptability and robustness of the GLID-DeepLab model.

[0054] An intelligent waste dump recognition system based on high-resolution remote sensing images, comprising:

[0055] Preprocessing module: Obtain high-resolution remote sensing images of the waste dump distribution, preprocess the obtained remote sensing images to obtain preprocessed remote sensing images.

[0056] Feature extraction module: Initialize the parameters of the GLID-DeepLab model, use the deep convolutional neural network RMT as the backbone network Backbone and combine it with the bidirectional feature pyramid network BiFPN to extract features from the remote sensing images preprocessed in step S1.

[0057] Meanwhile, introduce an auxiliary detection head that prompts visual hybrid coding, deeply fuse visual features with prior prompt information through a layer-by-layer and multi-scale fusion module, and optimize model training through an intermediate supervision mechanism to construct an efficient encoder.

[0058] Fusion module: Use iterative regularized deconvolution to process the features obtained in step S2, and fuse the shallow features output by the encoder backbone network to finely restore the target boundary.

[0059] Recognition module: Randomly select remote sensing image samples to be detected and input them into the trained GLID-DeepLab model to obtain the targets to be recognized.

[0060] In summary, due to the adoption of the above technical solutions, the beneficial effects of the present invention are as follows:

[0061] In the present invention, by combining the deep convolutional network RMT, the bidirectional feature pyramid network, and the auxiliary prompt visual hybrid coding auxiliary detection head, the model can effectively extract multi-scale features and achieve high-precision recognition of waste dumps; by using iterative regularization deconvolution and adaptive feature fusion technologies, the model can quickly recover the target boundaries, improve the processing efficiency, and reduce the calculation of non-target background regions; through self-supervised pre-training and multi-region sample training, the model enhances its adaptability to different regional geological and climatic conditions and improves the generalization ability of the model. Brief Description of the Drawings

[0062] The present invention will be described by way of examples with reference to the accompanying drawings, where:

[0063] Figure 1 is the flowchart of the present invention;

[0064] Figure 2 is the technical implementation example of each step of the present invention;

[0065] Figure 3 is the comparison chart of the recognition results of typical waste dumps of the present invention. Detailed Embodiments

[0066] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Usually, the components of the embodiments of the present invention described and illustrated herein can be arranged and designed in various different configurations.

[0067] Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed present invention, but merely represents the selected embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0068] It should be noted that, without conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other.

[0069] It should be noted that like reference numerals and letters refer to like items in the following figures, and thus, once an item is defined in one figure, it need not be further defined or explained in subsequent figures.

[0070] In the present invention, unless otherwise clearly specified and limited, the first feature being "on" or "under" the second feature may include the first and second features being in direct contact, or may include the first and second features not being in direct contact but being in contact through additional features therebetween. Moreover, the first feature being "above", "over" and "on top of" the second feature includes the first feature being directly above and obliquely above the second feature, or merely indicating that the first feature has a higher horizontal height than the second feature. The first feature being "under", "beneath" and "underneath" the second feature includes the first feature being directly below and obliquely below the second feature, or merely indicating that the first feature has a lower horizontal height than the second feature.

[0071] It should be noted that, without conflict, the embodiments in the present invention and the features in the embodiments may be combined with each other.

[0072] Embodiment 1

[0073] Combined with Figures 1 - 3 A detailed description of the present invention will be given.

[0074] Figure 1 What is shown is the overall architecture design of the present invention. This architecture is divided into four parts, S1, S2, S3, and S4, in sequence and is carried out sequentially. Each part has a separate output and is the input for the next part.

[0075] Figure 2 What is shown is the convergence trend of the GLID-DeepLab training function and the verification function. Compared with the deeplabv3+ models with Xception and ResNet as the backbone networks, the curve shows good fitting effect and high test accuracy.

[0076] Figure 3 What is shown is the recognition effect (ablation experiment) of the sliced images of typical waste dumps in Luding County, western Sichuan. The results in the second row are the recognition results of GLID-DeepLab, and the results in the third and fourth rows are the recognition results of the deeplabv3+ models with Xception and ResNet. Visual interpretation can prove the high accuracy of the model in recognizing waste dumps.

[0077] In an embodiment of the present invention, an intelligent recognition method for waste dumps based on high-resolution remote sensing images is disclosed, including:

[0078] Step S1: Obtain high-resolution remote sensing images of the distribution of waste dumps, and preprocess the obtained remote sensing images to obtain preprocessed remote sensing images;

[0079] Step 1 specifically includes the following steps:

[0080] Step S11: Obtain cloudless high-resolution optical remote sensing images;

[0081] Step S12: Draw the boundary of the spoil ground based on the obtained remote sensing images, and perform slicing processing on the spoil ground boundary and the remote sensing images to obtain sliced images;

[0082] Step S13: Modify the attribute values of the target area and the non-target area.

[0083] Step S2: Initialize the GLID-DeepLab model parameters, and use the deep convolutional neural network RMT as the backbone network Backbone and combine it with the bidirectional feature pyramid network BiFPN to extract features from the remotely sensed images preprocessed in Step S1;

[0084] Step S2 specifically includes the following steps:

[0085] Step S21: Put the sliced images and the boundary vector files into the deep convolutional neural network to obtain multi-scale feature information. The loss function of the deep convolutional neural network RMT is MALoss, as shown in the following formula:

[0086] L MALoss =L MAGICLoss +λL aux

[0087] In the formula, L MAGIC Loss is the MAGIC Loss, L aux is the auxiliary detection head loss function, and λ is the dynamically adjusted weight;

[0088] Step S22: Introduce multi-modal regularization, adaptive weighting, gradient penalty term, information entropy term, and C-Class-Balanced Cross-Entropy to optimize the loss function, as shown in the following formula: MAGICLoss is introduced into L

[0089] L MAGIC LOSS =L A +L M +L G +L I +L C

[0090] Step S22 specifically includes:

[0091] Introduce multi-modal data to regularize the loss function, so that the GLID-DeepLab model can learn richer feature representations from multiple data sources, as shown in the following formula:

[0092]

[0093] Among them, f image is the image feature, f other_modality is the feature of other modalities, and λ M is the regularization coefficient;

[0094] Adaptive weighting is introduced to dynamically adjust the loss weight according to the complexity of remote sensing image samples or the difficulty of categories, so that the GLID-DeepLab model pays more attention to difficult-to-classify samples, as shown in the following formula:

[0095]

[0096] Among them, α i is the adaptive weight, which is proportional to the classification difficulty of the sample, and Z i is the label, and f(x i ) is the model output;

[0097] A gradient penalty term is added to the loss function to prevent gradient explosion or disappearance, and at the same time enhance the generalization ability of the model, as shown in the following formula:

[0098]

[0099] Among them, λ G is the penalty coefficient, is the gradient of the model at the input x i ;

[0100] An information entropy term is introduced to make the output of the GL ID-DeepLab model closer to the prediction result of a uniform distribution, thereby improving the uncertainty processing ability of the model, as shown in the following formula:

[0101]

[0102] Among them, H(f(x i )) is the information entropy of the probability distribution of the model output, and λ I is the weight coefficient;

[0103] C-Class-Balanced Cross-Entropy is introduced to make the GLID-DeepLab model handle samples of different categories more fairly by dynamically adjusting the weights of each category, as shown in the following formula:

[0104]

[0105] Among them, β Mi is the weight of category M i , which is inversely proportional to the number of samples in that category.

[0106] Laux The loss function is as follows:

[0107] L aux = αL Terrain + γL Veg

[0108] Wherein, L Terrain is the loss of the waste dump terrain information, and L Veg is the loss of vegetation cover. α and γ are dynamically adjusted weights;

[0109]

[0110] Wherein, f terrain (x i ) is the terrain feature predicted by the model, is the real terrain feature, and λ terrain is the regularization coefficient;

[0111]

[0112] Wherein, f veg (x i ) is the vegetation cover feature predicted by the model, is the real vegetation cover feature, and λ veg is the regularization coefficient.

[0113] Step S23: The bidirectional feature pyramid network performs cross-layer feature fusion on the multi-scale feature information obtained in step S22 through the top-down and bottom-up bidirectional information flows, and combines an auxiliary detection head that prompts visual hybrid coding to deeply fuse the visual features with the prior prompt information through the layer-by-layer and multi-scale fusion modules to obtain preliminary features;

[0114] Step S21: The encoder improves the ability to capture details through self-supervised pre-training, and optimizes the preliminary features in step S22 through self-supervised learning tasks.

[0115] Meanwhile, an auxiliary detection head that prompts visual hybrid coding is introduced to deeply fuse the visual features with the prior prompt information through the layer-by-layer and multi-scale fusion modules, and the model training is optimized through the intermediate supervision mechanism to construct an efficient encoder;

[0116] Step S3: Perform iterative regularization deconvolution on the features obtained in step S2, and fuse the shallow features output by the encoder backbone network to finely restore the target boundary;

[0117] The specific steps of step S3 are as follows:

[0118] Step S31: Perform deconvolution processing on the features obtained in step S2 to achieve up-down connection matching. Through iterative regularization deconvolution operations, map the high-dimensional features of the encoder back to lower dimensions to prepare for subsequent feature fusion;

[0119] Step S32: Fuse the features processed by deconvolution with the shallow features extracted by RMT to achieve multi-scale information acquisition.

[0120] Step S4: Randomly select remote sensing image samples to be detected and input them into the trained GLID-DeepLab model to obtain the targets to be recognized.

[0121] The specific steps of step S4 are as follows:

[0122] Step S41: Select relatively typical waste dump image slices for processing, put them into the GLID-DeepLab model to test the recognition accuracy, and evaluate the recognition effect of the GLID-DeepLab model by comparing the output of the GLID-DeepLab model with the actual annotation;

[0123] Step S42: Considering the influence of geological conditions and climate characteristics in different regions on the recognition of waste dumps, include multi-region samples during the training of the GLID-DeepLab model to enhance the adaptability and robustness of the GLID-DeepLab model.

[0124] Embodiment 2

[0125] This embodiment proposes an intelligent waste dump recognition system based on high-resolution remote sensing images, including:

[0126] Preprocessing module: Obtain high-resolution remote sensing images of waste dump distributions, preprocess the obtained remote sensing images to obtain preprocessed remote sensing images;

[0127] Feature extraction module: Initialize the parameters of the GLID-DeepLab model, use the deep convolutional neural network RMT as the backbone network Backbone and combine it with the bidirectional feature pyramid network BiFPN to extract features from the remote sensing images preprocessed in step S1;

[0128] At the same time, introduce an auxiliary detection head that prompts visual hybrid coding, deeply fuse visual features with prior prompt information through a layer-by-layer and multi-scale fusion module, and optimize model training through an intermediate supervision mechanism to construct an efficient encoder;

[0129] Fusion module: Use iterative regularization deconvolution to process the features obtained in step S2, and fuse the shallow features output by the encoder backbone network to finely restore the target boundary;

[0130] Recognition module: Randomly select remote sensing image samples to be detected and input them into the trained GLID-DeepLab model to obtain the targets to be recognized.

[0131] The circuits, electronic components and modules involved are all prior arts and can be fully implemented by those skilled in the art without further elaboration. The content protected by the present invention does not involve improvements to software and methods either.

[0132] In this specification, the various embodiments are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. For the same or similar parts among the various embodiments, reference can be made to each other.

[0133] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. The general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but will be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. An intelligent recognition method for waste dumps based on high-resolution remote sensing images, characterized in that, Including: Step S1: Obtain high-resolution remote sensing images of the spoil ground distribution, preprocess the obtained remote sensing images to obtain preprocessed remote sensing images; Step S2: Initialize the GLID-DeepLab model parameters, use the deep convolutional neural network RMT as the backbone network Backbone and combine the bidirectional feature pyramid network BiFPN to extract features from the remote sensing images preprocessed in Step S1; When extracting features, introduce an auxiliary detection head that prompts visual hybrid coding, deeply fuse visual features with prior prompt information through a layer-by-layer and multi-scale fusion module, and optimize model training through an intermediate supervision mechanism to construct an efficient encoder; Step S3: Use iterative regularization deconvolution to process the features obtained in Step S2, and fuse the shallow features output by the encoder backbone network to finely restore the target boundary and complete the training of the GLID-DeepLab model; Step S4: Randomly select remote sensing image samples to be detected and input them into the trained GLID-DeepLab model to obtain the targets to be recognized.

2. The intelligent recognition method for waste dumps based on high-resolution remote sensing images according to claim 1, characterized in that, The specific steps of Step 1 are as follows: Step S11: Obtain cloudless high-resolution optical remote sensing images; Step S12: Draw the boundary of the spoil ground based on the obtained remote sensing images, slice the spoil ground boundary and the remote sensing images to obtain sliced images and corresponding boundary vector files; Step S13: Modify the attribute values of the target area and non-target area of the boundary vector file.

3. The intelligent identification method of waste dump based on high-resolution remote sensing images according to claim 2, characterized in that, The specific steps of Step S2 are as follows: Step S21: Put the sliced images and the boundary vector files into the deep convolutional neural network RMT to obtain multi-scale feature information. The loss function of the deep convolutional neural network RMT is MALoss, as shown in the following formula: L MALoss = L MAGICLoss + λL aux where L MAGICLoss is the comprehensive loss function MAGIC Loss, L aux is the auxiliary detection head loss function, and λ is the dynamically adjusted weight; Step S22: Introduce multi-modal regularization, adaptive weighting, gradient penalty term, information entropy term, and class-balanced cross-entropy to optimize the loss function as shown in the following equation: MAGICLoss Introduce multi-modal regularization, adaptive weighting, gradient penalty term, information entropy term, and class-balanced cross-entropy to optimize the loss function as follows: L MAGICLOSS = L A + L M + L G + L I + L C Where, L MAGICLoss is the comprehensive loss function, L A is the adaptive weighting, L M is the multi-modal data, L G is the gradient penalty term, L I is the information entropy term, L c is the class balance cross entropy; Step S23: The bidirectional feature pyramid network performs cross-layer feature fusion on the multi-scale feature information obtained in Step S21 through top-down and bottom-up bidirectional information flows, and combines an auxiliary detection head that prompts visual hybrid coding. The visual features of the multi-scale feature information are deeply fused with the prior prompt information through a layer-by-layer and multi-scale fusion module to obtain preliminary features; Step S24: The encoder improves the ability to capture details through self-supervised pre-training, and optimizes the preliminary features in Step S23 through self-supervised learning tasks.

4. An intelligent identification method for waste dumps based on high-resolution remote sensing images according to claim 3, characterized in that The specific steps of Step S22 include: Introduce multi-modal data to regularize the loss function, so that the GLID-DeepLab model can learn richer feature representations from multiple data sources, as shown in the following formula: Among them, f image is the image feature, f other_modality is the feature of other modalities, λ M is the regularization coefficient, x i is the pixel value of the image, y i is the feature extracted by the neural network; Introduce adaptive weighting to dynamically adjust the loss weight according to the complexity or category difficulty of the remote sensing image samples, so that the GLID-DeepLab model pays more attention to difficult-to-classify samples, as shown in the following formula: Among them, α i is the adaptive weight, which is proportional to the classification difficulty of the sample, z i is the label, and f(x i ) is the model output; Add a gradient penalty term to the loss function, as shown in the following formula: Among them, λ G is the penalty coefficient, is the gradient of the model at the input x i ; Introduce an information entropy term to make the prediction results output by the GLID-DeepLab model closer to a uniform distribution, as shown in the following formula: Among them, H(f(x i )) is the information entropy of the probability distribution output by the model, and λ I is the weight coefficient; Introduce C-Class-Balanced Cross-Entropy, which makes the GLID-DeepLab model handle samples of different classes more fairly by dynamically adjusting the weights of each class, as shown in the following formula: Among them, β yi is the weight of category M i and is inversely proportional to the number of samples in this category.

5. The intelligent identification method for waste dumps based on high-resolution remote sensing images according to claim 4, characterized in that In step S21, the loss function of L aux is as follows: L aux = αL Terrain + γL Veg where, L Terrain is the loss of topographic information of the waste dumping site, and L Veg is the loss of vegetation cover. α and γ are the dynamic adjustment weights; where, f terrain (x i ) is the terrain feature predicted by the model, is the true terrain feature, and λ terrain is the regularization coefficient; where, f veg (x i ) is the vegetation cover feature predicted by the model, is the true vegetation cover feature, and λ veg is the regularization coefficient.

6. The intelligent identification method for waste dumps based on high-resolution remote sensing images according to claim 1, wherein The specific steps of step S3 are as follows: Step S31: Perform deconvolution on the features obtained in step S2 to achieve up-down connection matching. Through iterative regularized deconvolution operations, map the high-dimensional features of the encoder back to a lower dimension to prepare for subsequent feature fusion; Step S32: Fuse the features after deconvolution with the shallow features extracted by RMT to achieve multi-scale information acquisition.

7. The intelligent identification method of waste dump based on high-resolution remote sensing images according to claim 1, characterized in that The specific steps of step S4 are as follows: Step S41: Select the waste dump image slices for processing, put them into the GLID-DeepLab model to test the recognition accuracy, and evaluate the recognition effect of the GLID-DeepLab model by comparing the output of the GLID-DeepLab model with the actual annotation; Step S42: Considering the influence of geological conditions and climate characteristics in different regions on the recognition of waste dumps, multi-region samples are included in the training of the GLID-DeepLab model to enhance the adaptability and robustness of the GLID-DeepLab model.

8. An intelligent identification system for waste dumps based on high-resolution remote sensing images, which is used to implement the intelligent identification method for waste dumps based on high-resolution remote sensing images described in any one of claims 1-7, and is characterized in that, Include: Preprocessing module: Obtain high-resolution remote sensing images of the waste dump distribution, preprocess the obtained remote sensing images to obtain preprocessed remote sensing images; Feature extraction module: Initialize the parameters of the GLID-DeepLab model, use the deep convolutional neural network RMT as the backbone network Backbone and combine it with the bidirectional feature pyramid network BiFPN to extract features from the remote sensing images preprocessed in step S1; At the same time, introduce an auxiliary detection head that prompts visual hybrid coding, deeply fuse visual features with prior prompt information through a layer-by-layer and multi-scale fusion module, and optimize model training through an intermediate supervision mechanism to construct an efficient encoder; Fusion module: Use iterative regularized deconvolution to process the features obtained in step S2, and fuse the shallow features output by the encoder backbone network to finely restore the target boundary; Recognition module: Randomly select remote sensing image samples to be detected and input them into the trained GLID-DeepLab model to obtain the targets to be recognized.

Images