A remote sensing image solid waste intelligent identification method and system

Abstract

The embodiment of the application provides a kind of remote sensing image solid waste intelligent identification method and system, it is related to remote sensing image processing technical field.The method can be according to dynamic cutting parameters after obtaining multispectral remote sensing image data sliding window cutting, and the multiple image blocks obtained by cutting are sequentially input into remote sensing solid waste detection model to obtain complete inference parameters.Then effective boundary box is filtered out by traversing boundary box coordinates in complete inference parameters, and solid waste yard quantitative index is calculated, complete inference parameters and solid waste yard quantitative index are realized structured export, and solid waste monitoring achievement file is output.The method can dynamically adjust cutting parameters according to image resolution and ground object density, adapt to different solid waste targets, and improve feature expression capability by adding parallel spatial feature, spectral feature double extraction branch and feature fusion in remote sensing solid waste detection model, improve the accuracy of solid waste identification result.

Description

Technical Field

[0001] This application relates to the field of remote sensing image processing technology, and in particular to a method and system for intelligent identification of solid waste from remote sensing images. Background Technology

[0002] Solid waste, or SDF for short, refers to solid, semi-solid, and gaseous substances or materials in containers that are generated during production, daily life, and other activities, have lost their original utilization value, or have been discarded or abandoned even if they have not lost their utilization value. Illegal dumping and open landfilling of solid waste damage the local ecological environment and pollute water, soil, and air, making it a key focus of ecological and environmental supervision. The supervision of solid waste primarily relies on manual on-site inspections, but this is limited by terrain conditions, inspection range, and manpower, easily creating blind spots in supervision. Furthermore, the inspection cycle is long and the timeliness is weak, making it difficult to meet the needs of routine monitoring across the entire region.

[0003] To meet the needs of routine monitoring across the entire region, and with the development of high-resolution remote sensing technology, satellite imagery, with its advantages of wide coverage, rapid updates, and rich information, has gradually become an important data source for monitoring surface solid waste. Based on satellite remote sensing imagery combined with deep learning-based target detection, intelligent interpretation of remote sensing images can be performed to identify the solid waste accumulation situation in a specific area. For example, this can be achieved using convolutional networks, keypoint detection networks, and Transformer architectures, along with processes such as image cropping, model inference, and post-processing, to intelligently identify the solid waste accumulation situation in a specific area.

[0004] However, the feature utilization dimension of the deep learning-based intelligent recognition process is single, the image segmentation strategy is fixed, and the model has insufficient scene adaptability and weak anti-interference ability, resulting in low accuracy of solid waste recognition results. It also has shortcomings in feature mining, scene adaptation, efficiency balance, and engineering practicality. Summary of the Invention

[0005] In view of this, embodiments of this application provide a method and system for intelligent identification of solid waste in remote sensing images, in order to solve the problem of low accuracy in the identification results of solid waste in remote sensing images.

[0006] According to a first aspect of this application, a method for intelligent identification of solid waste from remote sensing images is provided, the method comprising: Acquire multispectral remote sensing image data, which includes remote sensing images of multiple spectral bands; The remote sensing image is sliced ​​into multiple image blocks using a sliding window according to dynamic cropping parameters, which are dynamically adjusted based on the image resolution and ground feature density of the remote sensing image. Multiple image blocks are sequentially input into the remote sensing solid waste detection model to obtain the complete inference parameters output by the model. The remote sensing solid waste detection model is an image detection model based on the YOLO network as its backbone network. Spatial feature branches and spectral feature branches are built in parallel at the front end of the backbone network, and a dilated spatial pyramid pooling module and an efficient channel attention mechanism are embedded in it. The dilated spatial pyramid pooling module uses multiple sets of dilated convolutions with different dilation rates in parallel. The complete inference parameters include the pixel-level bounding box coordinates and target confidence for the solid waste target. The process involves filtering valid bounding boxes by traversing the bounding box coordinates in the complete inference parameters, and calculating quantitative indicators for solid waste dumps based on the complete inference parameters corresponding to the valid bounding boxes. This includes: calculating the pixel width and pixel height of the valid bounding boxes in the complete inference parameters; obtaining the image spatial resolution of the remote sensing image; calculating the solid waste area based on the pixel width and pixel height, combined with the image spatial resolution; extracting the vertex geographic coordinates of the valid bounding boxes; and determining the actual boundary range of the solid waste dump based on the vertex geographic coordinates. The valid bounding boxes are the bounding box coordinates filtered based on a post-processing threshold. The post-processing threshold is a threshold dynamically set according to different landform types using a terrain-adaptive dynamic non-maximum suppression strategy. The complete inference parameters and the quantitative indicators of the solid waste storage site are exported in a structured manner to output a solid waste monitoring results document.

[0007] In some embodiments, acquiring multispectral remote sensing image data includes: Acquire raw input data, which includes multispectral remote sensing images of the area to be monitored collected by the remote sensing monitoring equipment and recorded metadata; the metadata includes at least one of raw spatial resolution, geographic coordinates, and projected coordinate system; The remote sensing images in the original input data are subjected to standardization correction processing to obtain the corrected images. The standardization correction processing includes atmospheric correction, orthorectification, and panchromatic multispectral fusion. The corrected image is subjected to remote sensing noise removal and vegetation masking to obtain the multispectral remote sensing image data.

[0008] In some embodiments, the remote sensing image is sliced ​​into multiple image blocks by a sliding window according to dynamic cropping parameters, including: Define a basic clipping window, whose basic clipping parameters include window size, minimum overlap rate, and maximum overlap rate; The image resolution and ground feature density of the remote sensing image are obtained, wherein the ground feature density is the prior ground feature distribution density obtained based on prior knowledge of the remote sensing image. The basic cropping window is dynamically adjusted based on the image resolution and the density of ground features. The adjustment parameters of the basic cropping window include the image size of a single image block and the overlap rate between blocks. The number of pixels in the single-sided overlapping area is calculated based on the image size and the inter-block overlap rate to generate the dynamic cropping parameters; The entire remote sensing image is cropped using the dynamic cropping parameters to obtain multiple image blocks, and the pixel coordinates of the marker points of each image block are recorded synchronously.

[0009] In some embodiments, multiple image blocks are sequentially input into a remote sensing solid waste detection model to obtain complete inference parameters output by the remote sensing solid waste detection model, including: The image block is input into the spatial feature branch to extract spatial features through the convolutional structure of the YOLO network; the spatial features include at least one of image texture features, image contour features, and image morphological features. The image block is input into the spectral feature branch to extract spectral features through a channel fusion convolution and a depthwise separable convolution cascade structure in the spectral feature branch; the spectral features include multi-band original spectral reflectance features and spectral response difference features; Learnable weight coefficients are obtained, and the spatial features and spectral features are adaptively weighted and fused based on the learnable weight coefficients to obtain fused features; The fused features are fed into the feature pyramid network and combined with the void space pyramid pooling module to output multi-scale perceptual features. According to the efficient channel attention mechanism, feature filtering is performed on the multi-scale perceptual features to obtain recalibrated features; By dividing the recalibrated features into a grid, the association between local detection features and surrounding landform features is established, and global context features are generated by combining the context information passed by the adjacency graph convolution. The detection head outputs the complete inference parameters based on the global context features and the local detection features.

[0010] In some embodiments, the method further includes: Annotated image blocks are obtained by adding sample annotations to the image blocks; the sample annotations include solid waste category and target bounding rectangle boundary; A sample dataset is constructed based on the labeled image blocks. The sample dataset includes a global general solid waste sample set and a complex terrain sample set. The initial model is trained using the global general solid waste sample set to obtain a general pre-trained model, wherein some network parameters are frozen during the training process of the initial model; The general pre-trained model is fine-tuned using the complex terrain sample set to obtain the remote sensing solid waste detection model. During the fine-tuning training process, all network parameters of the general pre-trained model are unfrozen.

[0011] In some embodiments, the method further includes: By adjusting the spectral band reflectance values ​​corresponding to the labeled image blocks within a preset physical range, a band-enhanced image block is obtained; A random scale transformation is performed on the labeled image block to obtain a scale-enhanced image block; the random scale transformation includes at least one of random scaling and random cropping. Disturbance elements are randomly superimposed onto the labeled image blocks to obtain terrain simulation enhanced image blocks; the interference elements include at least one of natural shadows, vegetation occlusion, and surface debris coverage; An expanded sample set is generated based on the band-enhanced image block, the scale-transformed enhanced image block, and the terrain simulation enhanced image block; The expanded sample set is input into the remote sensing solid waste detection model to obtain the training result data output by the remote sensing solid waste detection model; The combined training loss is calculated based on the training results data. The combined training loss is a weighted sum of the classification loss, bounding box regression loss, and confidence loss. The network parameters of the remote sensing solid waste detection model are adjusted based on the combined training loss.

[0012] In some embodiments, valid bounding boxes are filtered out by traversing the bounding box coordinates in the complete inference parameters, including: Extract the bounding box coordinates from the complete inference parameters; Calculate the intersection-union ratio between the two bounding boxes based on the bounding box coordinates; Extract the image slope and NDVI vegetation index of the remote sensing image, and determine the landform type of the remote sensing image based on the image slope and the NDVI vegetation index. The post-processing threshold is set according to the landform type, and the post-processing threshold includes an intersection-over-union threshold and a confidence threshold. The effective bounding boxes are obtained by filtering out bounding boxes from the complete inference parameters whose intersection-union ratio is greater than the intersection-union ratio threshold and whose target confidence is greater than the confidence threshold.

[0013] In some embodiments, the complete inference parameters and the quantitative indicators of the solid waste storage site are structured and exported to output a solid waste monitoring results document, including: Extract valid bounding box coordinates and block offset coordinates from the complete inference parameters; The effective bounding box coordinates are calculated back to the remote sensing image according to the block offset coordinates; Different types of valid bounding boxes are labeled using different display styles on the remote sensing image, and labeled result images are generated by synchronously labeling detection result information; the detection result information includes at least one of category name, target confidence level, and area. Based on the complete inference parameters and the quantitative indicators of the solid waste storage site, standardized detection reports and target vector information are generated in batches. According to the output format set by the regulatory business requirements information, the result image, the standardized test report, and the target vector information are output as solid waste monitoring result files.

[0014] According to a second aspect of this application, a remote sensing image-based intelligent identification system for solid waste is provided, the system comprising: The data acquisition module is used to acquire multispectral remote sensing image data, which includes remote sensing images of multiple spectral bands. The image cropping module is used to crop the remote sensing image into multiple image blocks by a sliding window according to dynamic cropping parameters, wherein the dynamic cropping parameters are dynamically adjusted based on the image resolution and ground feature density of the remote sensing image. The solid waste detection module is used to sequentially input multiple image blocks into the remote sensing solid waste detection model to obtain the complete inference parameters output by the remote sensing solid waste detection model. The remote sensing solid waste detection model is an image detection model formed by building spatial feature branches and spectral feature branches in parallel at the front end of the basic backbone network, and embedding a dilated spatial pyramid pooling module and an efficient channel attention mechanism. The dilated spatial pyramid pooling module uses multiple sets of dilated convolutions with different dilation rates in parallel. The complete inference parameters include the pixel-level bounding box coordinates and target confidence for the solid waste target. The post-processing module is used to filter out valid bounding boxes by traversing the bounding box coordinates in the complete inference parameters, and to calculate the quantitative indicators of the solid waste dump based on the complete inference parameters corresponding to the valid bounding boxes. This includes: calculating the pixel width and pixel height of the valid bounding boxes in the complete inference parameters; obtaining the image spatial resolution of the remote sensing image; calculating the solid waste area based on the pixel width and pixel height, combined with the image spatial resolution; extracting the vertex geographic coordinates of the valid bounding boxes; and determining the actual boundary range of the solid waste dump based on the vertex geographic coordinates. The valid bounding boxes are the bounding box coordinates filtered based on a post-processing threshold. The post-processing threshold is a threshold dynamically set according to different landform types using a terrain-adaptive dynamic non-maximum suppression strategy. The results output module is used to export the complete inference parameters and the quantitative indicators of the solid waste storage site in a structured manner to output a solid waste monitoring results file.

[0015] By employing the above technical solutions, this application provides a method and system for intelligent identification of solid waste from remote sensing images. After acquiring multispectral remote sensing image data, the method performs sliding window cropping according to dynamic cropping parameters and sequentially inputs multiple image blocks obtained from the cropping into a remote sensing solid waste detection model to obtain complete inference parameters. Then, by traversing the bounding box coordinates in the complete inference parameters, valid bounding boxes are selected, and quantitative indicators of solid waste dumps are calculated. This enables the structured export of the complete inference parameters and the quantitative indicators of solid waste dumps, outputting a solid waste monitoring results file. The method can dynamically adjust the cropping parameters according to image resolution and ground feature density to adapt to different solid waste targets. Furthermore, by adding parallel spatial and spectral feature dual extraction branches and feature fusion to the remote sensing solid waste detection model, the feature expression capability is improved, thereby increasing the accuracy of solid waste identification results.

[0016] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, specific embodiments of this application are given below. Attached Figure Description

[0017] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a schematic flowchart of the intelligent identification method for solid waste from remote sensing images provided in an embodiment of this application; Figure 2 This is a schematic diagram of the solid waste identification process provided in the embodiments of this application; Figure 3 This is a schematic diagram of the remote sensing solid waste detection model structure provided in the embodiments of this application; Figure 4 This is a schematic diagram of the training process for the remote sensing solid waste detection model provided in the embodiments of this application; Figure 5 This is a schematic diagram of the structure of the intelligent identification system for solid waste from remote sensing images provided in an embodiment of this application. Detailed Implementation

[0018] The present application will be described in detail below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the embodiments of the present application can be combined with each other.

[0019] In this embodiment, the intelligent identification method and system for solid waste from remote sensing images can be applied to a remote sensing monitoring system to identify the stacking areas of solid waste from remote sensing images using an image detection model. Solid waste, referred to as solid waste, refers to solid, semi-solid, and gaseous items or substances in containers that are generated during production, daily life, and other activities, have lost their original utilization value, or have not lost their utilization value but have been discarded or abandoned.

[0020] Because illegal dumping and open landfilling of solid waste damage the local ecological environment and pollute water, soil, and air, it is a key area of ​​ecological and environmental supervision. Therefore, it is necessary to monitor the dumping of solid waste. In some embodiments, the monitoring of solid waste can be conducted by analyzing and judging the dumping situation of solid waste in a specific area through methods such as visual interpretation, spectral index analysis, and object-oriented classification.

[0021] However, manual visual interpretation relies on personnel experience, and interpretation standards are prone to inconsistencies. It also has limited efficiency when processing massive amounts of imagery, and is constrained by terrain conditions, patrol range, and manpower, easily creating blind spots in supervision. Furthermore, patrol cycles are long and timeliness is weak, making it difficult to meet the needs of routine monitoring across the entire area. The single spectral index method has weak feature representation capabilities, easily leading to identification errors when local object textures and appearances are similar, and is easily affected by atmospheric and lighting factors. Object-oriented classification methods are highly sensitive to image resolution and ground object morphology, and their segmentation effect on scattered or contiguous solid waste is unstable, resulting in fluctuating recognition accuracy. Therefore, the above-mentioned regulatory methods are only suitable for small-scale, low-frequency auxiliary inspections and cannot support the needs of large-scale (comprehensive) routine monitoring.

[0022] To meet the needs of routine monitoring across the entire region, some embodiments can also leverage satellite remote sensing imagery combined with deep learning target detection. This utilizes the advantages of satellite imagery—wide coverage, rapid updates, and rich information—to intelligently interpret remote sensing images of solid waste accumulation in a given area, thereby monitoring surface solid waste conditions. For example, this can be implemented using convolutional networks, keypoint detection networks, and Transformer architectures, along with processes such as image cropping, model inference, and post-processing, to achieve intelligent identification of solid waste accumulation in specific areas.

[0023] However, the feature utilization dimension of the intelligent recognition process based on deep learning is singular. That is, the deep learning scheme processes remote sensing images as ordinary visible light images, focusing on extracting spatial texture and contour features, without fully exploring the unique information of multispectral bands. Solid waste and bare land, wasteland and other land features have similar appearances but significantly different spectral characteristics, and relying solely on visual features is prone to false detection; targets covered by vegetation or shadows are also prone to missed detection.

[0024] Furthermore, the image segmentation strategy of the deep learning-based intelligent recognition process is fixed, and the cropping method with fixed size and fixed overlap rate cannot be flexibly adjusted according to image resolution, terrain complexity and target scale. The adaptability to solid waste targets with large size differences is generally poor, which will indirectly affect the recognition accuracy and data processing efficiency.

[0025] The intelligent recognition process based on deep learning suffers from insufficient model scene adaptability. This means the network models used in the deep learning process are general-purpose networks, not customized and optimized for remote sensing scenarios with large fields of view and complex terrain. Furthermore, some solutions introduce global inference modules with complex structures and high computational overhead; while Transformer-type models have strong global modeling capabilities, they consume significant computational power, making it difficult to balance detection accuracy and running speed.

[0026] Furthermore, the anti-interference capability of deep learning-based intelligent recognition processes is limited. Due to the common interference factors in outdoor scenes such as vegetation obstruction, terrain shadows, and surface debris, and the lack of targeted optimization in these deep learning models, recognition performance in complex environments is easily affected. Moreover, post-processing and quantization capabilities are insufficient; that is, using fixed thresholds for detection box selection fails to adapt to different terrains such as plains, mountains, and villages. Most solutions only output target category and location information, lacking quantitative statistical functions such as storage area and boundary range, resulting in insufficient adaptability to actual regulatory operations.

[0027] To address the issue of low accuracy in identifying solid waste from remote sensing images, this application provides an intelligent method for identifying solid waste from remote sensing images in certain embodiments. This method dynamically adjusts cropping parameters based on image resolution and ground feature density to adapt to different solid waste targets. Furthermore, by adding parallel spatial and spectral feature extraction branches and feature fusion to the remote sensing solid waste detection model, the method improves feature representation capabilities and increases the accuracy of solid waste identification results.

[0028] The method can be applied to remote sensing monitoring systems or electronic devices that establish a communication connection with the remote sensing monitoring system and have data processing capabilities. These electronic devices include, but are not limited to, computers, servers, mobile terminals, smart wearable devices, and industrial control computers. For ease of description, this application embodiment uses a remote sensing monitoring system as the executing entity of the method. It should be understood that the method can also be applied to other types of executing entities, which are not illustrated in this application embodiment. Figure 1 As shown, the method includes: S101. Acquire multispectral remote sensing image data.

[0029] When identifying solid waste, multispectral remote sensing image data can be acquired first. This multispectral remote sensing image data includes remote sensing images across multiple spectral bands. During the acquisition of multispectral remote sensing image data, it is necessary to perform raw remote sensing image standardization and interference suppression.

[0030] In some embodiments, when acquiring multispectral remote sensing image data, raw input data can be acquired first. This raw input data includes multispectral remote sensing images of the area to be monitored collected by the remote sensing monitoring device and recorded metadata; the metadata includes at least one of the following: original spatial resolution, geographic coordinates, and projected coordinate system.

[0031] For example, a remote sensing monitoring system can first acquire raw multispectral remote sensing images, that is, collect commercial high-resolution satellite remote sensing images such as the Gaofen series of images of the area to be monitored. The remote sensing images contain multiple spectral bands such as red, green, blue, and near-infrared, and record metadata such as the original spatial resolution, geographic coordinates, and projected coordinate system of the images as raw input data.

[0032] Then, standardized correction processing is performed on the remote sensing images in the original input data to obtain corrected images. This standardized correction processing includes atmospheric correction, orthorectification, and panchromatic multispectral fusion. For example, after acquiring the original input data, the remote sensing monitoring system can perform standardized correction processing on the remote sensing images. By sequentially performing professional remote sensing correction operations on the original images, distortions and biases caused by imaging, atmosphere, and topography are eliminated. Atmospheric correction can use radiometric calibration combined with atmospheric scattering correction to reduce the interference of atmospheric aerosols and illumination differences on spectral reflectance values, restoring the true spectral characteristics of ground features. Orthorectification can be combined with a digital elevation model (DEM) to complete topographic geometric correction, correcting image geometric distortions caused by topographic undulations and satellite attitude, ensuring accurate location of ground features. Panchromatic multispectral fusion, also known as pan-sharpening, can perform pixel-level fusion of high-resolution panchromatic bands and multispectral bands, improving image spatial resolution while preserving multispectral information, resulting in a fused image that combines spectral and detail features.

[0033] For the corrected image obtained through standardized correction processing, remote sensing noise removal and vegetation masking can be applied to the corrected image to obtain multispectral remote sensing image data. For example, remote sensing monitoring systems can achieve specific suppression of interference terms through noise removal and vegetation masking. Therefore, when specifically processing the common problems of salt-and-pepper noise, stripe noise, and vegetation interference in remote sensing images, remote sensing noise removal and vegetation masking can be performed separately.

[0034] Among them, remote sensing noise removal can use an adaptive windowed mid-range filtering algorithm to denoise the fused image. While removing noise points, it retains effective features such as solid waste edges and textures, avoiding the detail blurring problem caused by ordinary mid-range filtering.

[0035] Vegetation masking can use the Normalized Difference Vegetation Index (NDVI) to distinguish vegetated areas and generate vegetation masks to weaken vegetation disturbance. The formula for calculating NDVI is as follows:

[0036] In the formula, NDVI represents the normalized vegetation index; For near-infrared reflectivity, This refers to the reflectivity in the red band.

[0037] By setting a vegetation determination threshold T NDVI When NDVI>T NDVI When a pixel is identified as a vegetation area, feature weakening is then applied to that area to reduce the impact of vegetation occlusion and large areas of forest on solid waste identification.

[0038] S102. Cut the remote sensing image into multiple image blocks using a sliding window according to the dynamic cropping parameters.

[0039] After acquiring multispectral remote sensing image data, adaptive scaling of the remote sensing images within the multispectral remote sensing image data can be performed, i.e., as... Figure 2 As shown, remote sensing images are sliced ​​into multiple image blocks using a sliding window according to dynamic cropping parameters. These dynamic cropping parameters are dynamically adjusted based on the image resolution and ground feature density of the remote sensing image. Sliding window cropping performed according to these dynamic parameters eliminates the need for fixed-size, fixed-overlap cropping methods, instead dynamically adjusting the cropping method based on image resolution and ground feature density, thus achieving full-scale solid waste target adaptation.

[0040] In some embodiments, when a remote sensing image is clipped into multiple image blocks by a sliding window according to dynamic clipping parameters, a basic clipping window can be defined first. The basic clipping parameters corresponding to the basic clipping window include the window size, minimum overlap rate, and maximum overlap rate. Then, the image resolution and ground feature density of the remote sensing image are obtained, wherein the ground feature density is the prior ground feature distribution density obtained statistically based on prior knowledge of the remote sensing image.

[0041] Then, the basic cropping window is dynamically adjusted based on image resolution and ground feature density. The adjustment parameters of the basic cropping window include the image size of a single image block and the inter-block overlap rate. The number of pixels in the single-sided overlapping area is then calculated based on the image size and inter-block overlap rate to generate dynamic cropping parameters. A sliding window cropping process is then performed on the entire remote sensing image according to the dynamic cropping parameters to obtain multiple image blocks, and the pixel coordinates of the marker points for each image block are recorded simultaneously.

[0042] For example, when performing adaptive scaling image cropping, you can first define the basic cropping window size W and the minimum overlap rate. r min Maximum overlap rate r max Then, the image resolution and ground feature density of the remote sensing image are acquired, and the size of a single image patch is dynamically adjusted based on the image resolution and prior ground feature distribution density. W i Inter-block overlap rate r i In this analysis, the density of ground features can be calculated using existing, coarse-grained data or prior knowledge through statistical or machine learning methods before the analysis begins, resulting in a more refined distribution density. Therefore, for areas with a high proportion of small, scattered solid waste, a small window and high overlap rate can be used; while for areas with a high proportion of large, contiguous solid waste dumps, a large window and low overlap rate can be used. Next, the number of pixels in the overlapping area on one side is calculated, that is:

[0043] In the formula, This refers to the number of overlapping pixels on one side of an image block, used to alleviate the problems of broken or missed detection of solid waste targets at the edge caused by cropping. W i The pixel dimensions of the dynamically adjusted image block side length; r i This is for real-time adjustment of inter-block overlap rate.

[0044] Then, the entire image is cropped using a sliding window according to dynamic parameters, and the pixel coordinates of the upper left corner of each image block are recorded simultaneously for subsequent detection box coordinate back calculation.

[0045] S103. Input multiple image blocks into the remote sensing solid waste detection model in sequence to obtain the complete inference parameters output by the remote sensing solid waste detection model.

[0046] After obtaining multiple image blocks through adaptive scaling image cropping, the remote sensing solid waste detection model can be applied to adaptively detect the multiple image blocks. Therefore, multiple image blocks can be input into the remote sensing solid waste detection model in sequence to obtain the complete inference parameters output by the remote sensing solid waste detection model.

[0047] Because the "You Only Look Once" (YOLO) model and mainstream remote sensing detection models mostly adopt a single-branch feature extraction structure, they only mine two-dimensional spatial texture information of images and rarely utilize the unique spectral dimension features of multispectral satellite imagery. This results in the loss of spectral discrimination information for ground objects, making it difficult to distinguish between targets with similar spatial morphology and significant differences in spectral features, such as bare vegetation and solid waste, or different types of solid waste. Therefore, adaptive detection can be achieved by constructing a remote sensing solid waste detection model. This model refers to the Spectral-Spatial Adaptive Waste Detection Network (SSAW-Net), which builds spatial feature branches and spectral feature branches in parallel at the front end of the YOLO network backbone. This achieves decoupled extraction and adaptive fusion of spatial morphological features and spectral attribute features, overcoming the technical shortcomings of traditional models that lack spectral perception.

[0048] The remote sensing solid waste detection model can also progressively embed a lightweight Atrous Spatial Pyramid Pooling (ASPP) multi-scale perception module, an Efficient Channel Attention (ECA) intelligent attention recalibration module, and a lightweight global inference unit. That is, the remote sensing solid waste detection model is an image detection model formed by building spatial feature branches and spectral feature branches in parallel at the front end of the basic backbone network, and embedding an atrous spatial pyramid pooling module and an efficient channel attention mechanism. By constructing the SSAW-Net remote sensing multispectral solid waste-specific detection network, it can balance spectral feature specificity, spatial detail integrity, multi-scale adaptability, and lightweight inference performance, solving the problems of poor remote sensing scene adaptability, missed detection of small solid waste, and false detection of complex terrain in traditional models from the network structure level.

[0049] like Figure 3 As shown, the basic network architecture of the remote sensing solid waste detection model follows the YOLO network architecture, that is, it retains the original YOLO backbone network, feature pyramid network and detection head structure. The backbone network is responsible for basic feature extraction, the feature pyramid completes multi-scale feature fusion, and the detection head outputs target category, bounding box and confidence score, which serves as the basic architecture carrier of the SSAW-Net model.

[0050] The spatial feature branch and the spectral feature branch are two feature extraction branches built in parallel at the front end of the network backbone, realizing the decoupled extraction and adaptive fusion of spatial morphological features and spectral attribute features. For the spatial feature branch, the convolutional structure of the YOLO network can be used to extract two-dimensional spatial features such as image texture, contour, and morphology.

[0051] The spectral feature branch is a dedicated lightweight convolutional group added to the SSAW-Net model. Unlike the standard convolutional structure of the native YOLO network, it is specifically adapted to the multi-band input characteristics of multispectral remote sensing images, enabling specialized in-depth mining of spectral features. For example, the lightweight convolutional group adopts a simplified structure of cascaded 1×1 channel fusion convolution and depthwise separable convolution. First, the 1×1 convolution completes the channel integration and dimensionality reduction of redundant information across multiple spectral bands, aggregating the original spectral reflectance features of red, green, blue, and near-infrared bands. Then, the depthwise separable convolution independently extracts spectral variation patterns band by band, accurately capturing the inherent spectral response differences of different ground features.

[0052] Compared to standard convolution operations, the fusion of convolution and depthwise separable convolution cascade structures can reduce computational redundancy, with the number of parameters and computational cost being only about 15% of that of conventional convolution. Without increasing the model's inference burden and ensuring lightweight performance, it is dedicated to learning the spectral distinguishing features of solid waste and background objects, making up for the technical deficiency that the model can only extract spatial texture features and cannot analyze spectral attributes, and providing an independent, high-quality spectral feature source for subsequent dual-branch adaptive fusion.

[0053] The Atrous Spatial Pyramid Pooling (ASPP) module can be used for multi-scale perception optimization, enabling efficient extraction of multi-scale features. Due to the large scale and irregular distribution of remotely sensed solid waste targets—ranging from scattered small targets (square meters) to contiguous waste dumps (tens of thousands of square meters)—the receptive field of the YOLO feature pyramid structure is fixed and cannot simultaneously adapt to both extremely small and extremely large targets, easily leading to missed detections of small targets and incomplete fitting of large target boundaries. Furthermore, the ASPP module has redundant structure and a large number of parameters, which significantly reduces the efficiency of batch inference of remote sensing images after embedding it into the detection network. Therefore, a lightweight, customized ASPP module is designed and embedded into the SSAW-Net model feature pyramid network to achieve efficient multi-scale feature extraction.

[0054] In some embodiments, the ASPP module can be customized with dilation rate parameter configuration, that is, multiple sets of dilated convolutions with different dilation rates can be used for parallel operation, such as dilation rates set to 3, 6 and 9 in sequence. This parameter combination is a dedicated adaptation parameter for remote sensing solid waste scenarios, which can accurately cover the multi-scale feature extraction needs of scattered small solid waste, medium-sized solid waste piles and large contiguous piles in remote sensing images, and solve the problem that a single receptive field cannot adapt to multi-scale targets.

[0055] The ASPP module can also be simplified with a lightweight structure, that is, the redundant ASPP architecture is abandoned, the global average pooling branch and redundant 1×1 convolutional layers in the ASPP architecture are removed, and the number of module parameters and computation is reduced by about 30% while ensuring that the core perception capability of multi-scale dilated convolution remains unchanged. This alleviates the defects of redundant computing power and slow inference speed of multi-scale modules, and allows the SSAW-Net model to achieve a two-way balance between accuracy and speed.

[0056] The ASPP module can also perform multi-scale feature fusion enhancement. By splicing and fusing the global multi-scale features output by ASPP with the original pyramid hierarchical features, it enriches the network feature expression dimension, makes up for the shortcomings of traditional networks where deep features lack details and shallow features lack global features, and significantly improves the accuracy and completeness of the SSAW-Net model in recognizing multi-scale solid waste targets.

[0057] Embedding the ECA channel attention mechanism can enhance the anti-interference capability of feature selection. Due to the characteristics of remote sensing imagery—multiple band dimensions, noisy channels, and redundant invalid features—a large number of vegetation, shadows, and bare land interference channels can mask the effective features of solid waste, leading to low feature utilization and high false detection rates in the model. Furthermore, without channel selection capabilities, the detection network cannot distinguish between effective and noisy features. Therefore, the proposed remote sensing solid waste detection model integrates the efficient ECA channel attention mechanism into the lightweight ASPP module backend, endowing the SSAW-Net model with adaptive feature channel selection and weight recalibration capabilities. This accurately strengthens the effective features of solid waste, suppresses invalid remote sensing noise, and significantly improves the model's anti-interference capability.

[0058] The ECA (Efficient Channel Attention) mechanism performs global average pooling on the input feature map, compressing the spatial dimension to obtain a one-dimensional channel feature vector. It then uses one-dimensional convolution to perform local cross-channel interactions and calculates the weight coefficients for each channel. Finally, these weight coefficients are multiplied channel-by-channel by the original feature map to complete feature recalibration.

[0059] In the formula, F eca For recalibration features; F in Input feature map; AvgPool ( ) represents the global average pooling operation; Conv1D ( ) represents a one-dimensional convolution operation; σ Use the Sigmoid activation function; This indicates a channel-by-channel multiplication operation.

[0060] In some embodiments, the remote sensing solid waste detection model can also embed a lightweight global inference unit for global context modeling. Because YOLO series models rely solely on local convolutions for feature extraction, their global context modeling capabilities are limited. They cannot combine surrounding terrain features to assist in identifying solid waste targets, easily misclassifying suspected targets such as mountain rocks and debris in open areas as solid waste. Existing global modeling modules are complex in structure and consume extremely high computational power, making them unsuitable for batch remote sensing detection scenarios. To address these issues, a lightweight global inference unit can be embedded at the end of the feature pyramid of the remote sensing solid waste detection model, enabling the SSAW-Net model to balance global context awareness with lightweight performance.

[0061] The lightweight global inference unit can perform grid partitioning on the features of the entire image, establish associations between local features and surrounding terrain features, and then optimize lightweight graph convolution. It abandons the full matrix operation of global graph convolution and uses lightweight adjacency graph convolution to complete the feature transfer between local and global features. It only retains the core computational logic of terrain association and target association, reducing computational complexity. Compared with the global inference module, the computational power requirement is reduced by more than 40%. After fusing global context features and local detection features, it is input into the detection head to complete inference, improving the global perception and local detection collaborative capabilities of the SSAW-Net model.

[0062] Using the remote sensing solid waste detection model described in the above embodiments, complete inference parameters can be obtained through adaptive detection. These complete inference parameters include the pixel-level bounding box coordinates and target confidence for the solid waste target. The bounding box refers to the marker symbol displayed on the remote sensing image using shapes such as rectangles to represent detected solid waste targets; the bounding box coordinates are the numerical values ​​used to define the position of this shape marker in the image. For example, the bounding box coordinates can be represented as (x... min y min x max y max ), representing the pixel coordinates of the top-left and bottom-right corners of the rectangle, respectively. Target confidence is a value between 0 and 1, used to represent the model's degree of confidence in the detection results.

[0063] In some embodiments, when multiple image patches are sequentially input into a remote sensing solid waste detection model to obtain the complete inference parameters output by the model, the image patches can be input into the spatial feature branch and the spectral feature branch respectively. Spatial features are extracted through the convolutional structure of the YOLO network, and spectral features are extracted through the channel fusion convolution and depthwise separable convolution cascade structure in the spectral feature branch. The spatial features include at least one of image texture features, image contour features, and image morphological features; the spectral features include multi-band raw spectral reflectance features and spectral response difference features.

[0064] Next, learnable weight coefficients are obtained, and spatial and spectral features are adaptively weighted and fused based on these coefficients to obtain fused features. These fused features are then fed into a feature pyramid network and combined with a hollow spatial pyramid pooling module to output multi-scale perceptual features. Following an efficient channel attention mechanism, feature filtering is performed on the multi-scale perceptual features to obtain recalibrated features. By dividing the recalibrated features into a grid, the correlation between local detection features and surrounding terrain features is established, and global contextual features are generated by combining contextual information passed through adjacency graph convolution. Thus, the detection head outputs complete inference parameters based on the global contextual features and local detection features.

[0065] For example, in the input stage of the SSAW-Net model, the SSAW-Net model can receive multispectral remote sensing images containing multiple bands such as red, green, blue, and near-infrared. Feature extraction is then performed in parallel through a spectral-spatial dual-branch approach. At the front end of the network backbone, the input remote sensing image is copied and fed into two parallel branches. In the spatial feature branch, the native YOLO convolutional structure can be used to extract the two-dimensional spatial morphological features of the image, such as texture, contour, and edge features. In the spectral feature branch, a dedicated lightweight convolutional group can process the multi-band input. The spectral feature branch structure consists of a cascade of 1×1 convolutions for channel integration and dimensionality reduction and depthwise separable convolutions for independently extracting spectral variation patterns band by band, specifically designed to extract the spectral reflectance differences of ground features.

[0066] After extracting spatial and spectral features in parallel through two branches, adaptive weighted fusion is then performed to combine the feature maps output by the two branches, i.e., the spatial feature F. spatial With spectral characteristics F spectral Deep stitching is performed, and learnable weight parameters α and β are introduced. The sum of these two weights is constrained to 1 using the Softmax function, allowing the network to dynamically learn the optimal weight ratio during training, achieving adaptive fusion of spatial and spectral features. The fusion formula is as follows:

[0067] In the formula, F spatial For spatial feature maps, F spectral The spectral feature map is represented by α and β, which are learnable weight parameters, i.e., the normalized weight coefficients for network autonomous learning.

[0068] The adaptive weighted fusion mechanism differs from fixed-weight concatenation and simple channel overlay fusion methods, enabling deep concatenation and fusion of feature maps from both branches. By introducing two sets of learnable weight coefficients α and β, and using the Softmax normalization algorithm to constrain the dynamic balance of weights in real time, forcing α+β=1, the SSAW-Net model autonomously learns the weight ratios of spatial and spectral features in different scenarios. It automatically strengthens the weights of spectral features in regions with complex terrain and strong texture interference, and automatically strengthens the weights of spatial features in regions with blurred targets and missing details, achieving adaptive complementarity between the two features.

[0069] The fused feature maps are then fed into the feature pyramid network, where a lightweight ASPP module embedded within it performs parallel computation using dilated convolutions with dilation rates of 3, 6, and 9 to adapt to the multi-scale characteristics of remote sensing solid waste. The extracted multi-scale features are then concatenated and fused with the original hierarchical features of the pyramid to output multi-scale perception features, thereby enhancing the network's ability to perceive multi-scale targets.

[0070] The multi-scale perceptual feature maps output by ASPP are then fed into the ECA module. The ECA module performs global average pooling on the multi-scale perceptual feature maps to compress the spatial dimension. Local cross-channel interaction is achieved through one-dimensional convolution to calculate the weight coefficients of each channel. Then, the weights are multiplied with the original feature map channel by channel to complete feature recalibration, resulting in recalibrated features that enhance effective features and suppress invalid noise.

[0071] The recalibrated feature map enters the lightweight global inference unit at the end of the feature pyramid. The lightweight global inference unit then performs grid partitioning of the entire image features, establishing the association between local features and surrounding terrain features. Lightweight adjacency graph convolution replaces the full matrix operation of traditional global graph convolution, passing contextual information. The global contextual features and local detection features are then fused and input into the detection head. Based on the enhanced feature map that fuses multi-scale and multi-dimensional information, the detection head completes the final solid waste target detection and localization, outputting the detection results and obtaining complete inference parameters.

[0072] S104. Valid bounding boxes are selected by traversing the bounding box coordinates in the complete inference parameters, and the quantitative indicators of solid waste dumps are calculated based on the complete inference parameters corresponding to the valid bounding boxes.

[0073] After obtaining the complete inference parameters, intelligent post-processing can be performed on the detection results output by the remote sensing solid waste detection model. Specifically, valid bounding boxes are selected by traversing the bounding box coordinates in the complete inference parameters. The valid bounding boxes are the bounding box coordinates selected based on the post-processing threshold. The post-processing threshold is a threshold dynamically set according to different landform types based on the landform adaptive dynamic non-maximum suppression strategy.

[0074] By constructing a terrain-adaptive intelligent post-processing system and matching it with the detection characteristics of the SSAW-Net model, we can achieve differentiated target selection, accurate pixel-to-geographic mapping, and intelligent quantification of solid waste indicators, thereby improving the accuracy and engineering applicability of the detection results.

[0075] In some embodiments, when filtering valid bounding boxes by traversing the bounding box coordinates in the complete inference parameters, the bounding box coordinates can be extracted from the complete inference parameters first, and the intersection-union ratio (IU) between two bounding boxes can be calculated based on the bounding box coordinates. Then, the image slope and NDVI vegetation index of the remote sensing image are extracted, and the landform type of the remote sensing image is determined based on the image slope and NDVI vegetation index. Then, post-processing thresholds are set according to the landform type, including an IU threshold and a confidence threshold. Finally, bounding boxes with an IU greater than the IU threshold and a target confidence score greater than the confidence threshold are filtered from the complete inference parameters to obtain valid bounding boxes.

[0076] For example, the preprocessed image blocks to be tested are sequentially input into the trained SSAW-Net model to perform forward inference calculations. The detection head of the SSAW-Net model can accurately output the complete inference parameters of each solid waste target, including the category number, pixel-level bounding box coordinates and target confidence, providing accurate raw data support for subsequent intelligent post-processing and quantization calculations.

[0077] Then, post-processing is performed using a terrain-adaptive non-maximum suppression strategy. Since the Non-Maximum Suppression (NMS) algorithm uses a globally fixed IOU and confidence threshold, it processes all terrain scenes uniformly, resulting in serious scene adaptation defects. For example, fragmented solid waste in mountainous areas is easily mistakenly removed, forest interference easily generates a large number of false detections, and regular targets in plains are prone to duplicate detections. Therefore, the globally fixed threshold mode can be abandoned, and a terrain-adaptive dynamic NMS strategy is proposed. This strategy matches the detection output characteristics of the SSAW-Net model, automatically classifies terrain types based on image topography and vegetation features, matches differentiated screening thresholds, accurately filters out duplicate detection boxes and low-confidence false detection boxes, and retains the maximum number of valid solid waste targets.

[0078] In post-processing, the Intersection over Union (IOU) can be calculated first, that is, the crossover ratio between each pair of detection boxes can be calculated:

[0079] In the formula, IOU is the intersection-union ratio; S i S is the pixel area of ​​the intersection of the two detection boxes; u The pixel area of ​​the union of the two detection boxes.

[0080] Then, intelligent landform classification and threshold adaptive matching are performed. Based on the dual-dimensional quantitative indicators of image slope and NDVI vegetation index, the image is automatically divided into four typical landforms. Then, for the solid waste distribution characteristics and interference characteristics of different landforms, exclusive differentiated IOU thresholds and confidence thresholds are configured to accurately reduce the false detection and false negative rates of the SSAW-Net model from the post-processing level.

[0081] The classification standards and adaptation thresholds for various landforms are as follows: For plains, the image slope is <5°, the target distribution is regular, and the IOU threshold is set to 0.5 and the confidence threshold is set to 0.4; for mountainous areas, the image slope is ≥25°, the target fragmentation is severe, and the IOU threshold is set to 0.3 and the confidence threshold is set to 0.4; for forest / vegetation areas, the NDVI is >0.6, there are many interfering factors, and the IOU threshold is set to 0.5 and the confidence threshold is set to 0.6 to strengthen false detection filtering; for urban and rural settlement areas, the image slope is between 5° and 25° and the NDVI is ≤0.6, and the IOU threshold is set to 0.4 and the confidence threshold is set to 0.4.

[0082] After intelligently identifying the corresponding terrain, the NMS algorithm is executed according to the configured threshold. All detection boxes are traversed, and valid bounding boxes whose intersection-over-union ratio and target confidence meet the post-processing threshold are retained. Overlapping redundant boxes and low-confidence false detection boxes are removed, thus completing the adaptive optimization and filtering of SSAW-Net detection results.

[0083] After filtering out valid bounding boxes by traversing the bounding box coordinates in the complete inference parameters, the quantitative indicators of the solid waste dump can be calculated based on the complete inference parameters corresponding to the valid bounding boxes. In some embodiments, when calculating the quantitative indicators of the solid waste dump based on the complete inference parameters corresponding to the valid bounding boxes, the pixel width and pixel height of the valid bounding boxes can be counted in the complete inference parameters first, and the image spatial resolution of the remote sensing image can be obtained. Then, based on the pixel width and pixel height, combined with the image spatial resolution, the solid waste area is calculated, and the vertex geographic coordinates of the valid bounding boxes are extracted. Finally, the actual boundary range of the solid waste dump is determined based on the vertex geographic coordinates.

[0084] For example, after selecting valid bounding boxes, pixel coordinates can be mapped to geographic coordinates. By establishing a precise mapping algorithm from pixel coordinates to geographic latitude and longitude, and combining it with geographic projection information from image metadata, the image pixel coordinates output by the SSAW-Net model are accurately converted into geographic latitude and longitude coordinates in the WGS-84 universal coordinate system, achieving sub-meter-level geographic positioning of solid waste targets.

[0085] Then, the quantitative indicators of solid waste dumps are calculated. By designing an automated solid waste quantitative calculation system, based on the SSAW-Net model, the system can accurately locate solid waste boundaries and accurately calculate the area occupied. This provides quantitative data support for ecological environment supervision and solid waste source tracing law enforcement. Specifically, the solid waste area is calculated by statistically analyzing the pixel width and pixel height of the detection box and combining it with the spatial resolution of the image. The geographical coordinates of the four vertices of the detection box are extracted to determine the actual boundary range of the solid waste dump.

[0086] S105. The complete inference parameters and quantitative indicators of solid waste storage sites are exported in a structured manner to output solid waste monitoring results documents.

[0087] After intelligent post-processing is completed, the results can be visualized and batch output by combining the intelligent post-processing results with the detection results output by the remote sensing solid waste detection model. That is, the complete inference parameters and the quantitative indicators of solid waste dumps are exported in a structured manner to output solid waste monitoring result files. This enables visualization rendering and structured data export of the detection results completed by the SSAW-Net model post-processing, outputting multi-format result files to connect with ecological and environmental supervision business.

[0088] In some embodiments, when structurally exporting complete inference parameters and solid waste site quantitative indicators to output solid waste monitoring results files, effective bounding box coordinates and block offset coordinates can be extracted from the complete inference parameters first, and the effective bounding box coordinates can be recalculated back to the remote sensing image according to the block offset coordinates. Then, different types of effective bounding boxes are labeled on the remote sensing image using different display styles, and labeled result images are generated by synchronously labeling detection result information. The detection result information includes at least one of category name, target confidence level, and land area.

[0089] For example, when outputting results, the detection results can be visualized first. This involves combining the block-level bounding box coordinates with the block offset coordinates recorded during adaptive scaling image cropping to calculate back to the entire original remote sensing image. Then, different colors are used to mark the bounding boxes for different categories of solid waste on the original image, along with information such as category name, confidence level, and area occupied, generating an labeled result image that visually presents the detection effect of the SSAW-Net model.

[0090] Then, based on the structured data, standardized detection reports and target vector information are generated in batches according to the complete inference parameters and the quantitative indicators of solid waste storage sites. In this way, the result images, standardized detection reports and target vector information are output as solid waste monitoring result files in accordance with the output format set by the regulatory business requirements.

[0091] For example, standardized detection reports can be generated in batches according to regulatory business needs. The report fields may include image number, solid waste category, target geographic coordinates, boundary range, land area, confidence level, and landform type, etc., and the quantitative data detected by the SSAW-Net model can be completely retained.

[0092] By enabling multi-format file storage and output, three types of output files can be uniformly stored and output: visualized and annotated images in common image formats, statistical reports in Excel or CSV table formats, and target vector information in Shapefile (SHP) vector format. All files are archived according to monitoring area and monitoring time, supporting batch export, online viewing, and ledger statistics. This meets the practical application needs of solid waste investigation, dynamic monitoring, and law enforcement tracing, fully leveraging the engineering application value of the SSAW-Net model.

[0093] By applying the technical solutions of the above embodiments, the intelligent identification method for solid waste from remote sensing images described in the above embodiments is based on the YOLO algorithm and customized and improved in combination with the business characteristics of remote sensing solid waste identification, building a multi-module collaborative integrated identification solution. The method can improve the feature expression capability through spectral-spatial dual-branch feature fusion. This application adds parallel spatial feature and spectral feature dual extraction branches and completes feature fusion. Unlike the single spatial feature extraction mode, the method utilizes multispectral band information and combines the spectral patterns of ground objects to distinguish solid waste with bare land and wasteland with similar appearances, reducing false detections; on the other hand, it relies on spatial features to identify target morphology, alleviating the problem of missed detection caused by vegetation occlusion. At the same time, the branch structure adopts a lightweight design, which does not significantly increase the model's computing power burden while enriching feature dimensions, and has better adaptability to remote sensing images from different sources and under different imaging conditions.

[0094] The proposed method can adapt to multi-scale targets through adaptive multi-scale cropping and feature enhancement. This scheme abandons the fixed parameter cropping mode and can dynamically adjust the cropping size and overlap rate according to image resolution, terrain, and target distribution, adapting to different targets such as small, scattered solid waste and large, contiguous stockpiles. Furthermore, it enhances multi-scale feature perception by combining a lightweight dilated convolution module and an attention mechanism, balancing recognition accuracy and image preprocessing efficiency.

[0095] The method can also balance accuracy and operational efficiency through a lightweight global inference module. By embedding a low-computational-cost global context inference unit into the network, feature associations between local targets and the overall terrain can be established. This assists the model in making comprehensive judgments based on the surrounding environment, reducing the probability of misjudgment in complex terrain. This streamlined module structure ensures efficient batch detection of large-scale remote sensing images, achieving a balance between accuracy and speed.

[0096] The method can also improve its practicality by combining terrain-adaptive post-processing with quantitative statistics. This solution dynamically adjusts the post-processing threshold according to different terrain types to optimize the deduplication and filtering effects of detection boxes. At the same time, by combining remote sensing geographic information, it automatically calculates quantitative indicators such as the boundary and area of ​​solid waste dumps, and can simultaneously output multi-dimensional results such as category, location, and statistical data, which can better connect with actual work such as environmental supervision and law enforcement statistics, making the solution more valuable for implementation.

[0097] In some embodiments, as a refinement and extension of the specific implementation of the above embodiments, and to fully illustrate the specific implementation process of this embodiment, some embodiments of this application also provide a method for intelligent identification of solid waste from remote sensing images. The difference between this method and the above embodiments is that a remote sensing solid waste detection model can also be trained using a hierarchical training mechanism. For example... Figure 4 As shown, the method includes: S201. Obtain labeled image blocks by adding sample annotations to image blocks; S202. Construct a sample dataset based on the labeled image patches; S203. Use a universal solid waste sample set to train the initial model to obtain a universal pre-trained model; S204. Use a complex terrain sample set to perform fine-tuning training on the general pre-trained model to obtain a remote sensing solid waste detection model.

[0098] When developing a remote sensing solid waste detection model, labeled image patches can be obtained by adding sample annotations to image blocks, and then a sample dataset can be constructed based on the labeled image patches. The sample annotations include the solid waste category and the bounding rectangle boundary of the target. The sample dataset includes a global general solid waste sample set and a complex terrain sample set.

[0099] For example, after cropping the entire image using a sliding window based on dynamic parameters, sample annotation and dataset partitioning can be performed. For sample annotation, the cropped image blocks can be labeled, including information such as solid waste category and the bounding rectangle of the target. Solid waste categories can include scattered solid waste, industrial solid waste dumps, and municipal solid waste landfills. During sample annotation, the target pixel area can also be labeled for subsequent SSAW-Net model small target loss weight determination; the annotation format is fully compatible with model training specifications.

[0100] For the dataset construction process, a complete sample library can be formed by aggregating samples from multiple regions, terrains, and scales. Based on the application scope of the sample data, the sample dataset can be divided into a universal solid waste sample set and a complex terrain sample set. Then, it can be randomly divided into a training set, a validation set, and a test set in a 7:2:1 ratio for training the SSAW-Net model, validating its performance, and testing its effectiveness, respectively.

[0101] After constructing the sample dataset, an initial model can be trained using a universal solid waste sample set to obtain a general pre-trained model. Then, a complex terrain sample set can be used to perform fine-tuning training on the general pre-trained model to obtain a remote sensing solid waste detection model. During the training of the initial model, some network parameters are frozen, while all network parameters of the general pre-trained model are unfrozen during the fine-tuning training.

[0102] For example, for remote sensing solid waste detection scenarios, a complete training system that is remote sensing-specific, hierarchically adapted, and dynamically optimized can be built to adapt to the structural characteristics of the SSAW-Net model. Through hierarchical transfer training, remote sensing multidimensional data augmentation, and adaptive combined loss function, the model's accuracy, robustness, and generalization ability can be improved in all aspects, solving the problem of poor implementation of training strategies.

[0103] Because the model uses a single-stage training mode, directly training on complex field samples, it is prone to problems such as slow convergence, insufficient feature learning, and overfitting. Therefore, a hierarchical training mechanism can be used to adapt transfer learning. A two-stage hierarchical transfer training mode, which combines general pre-training and fine-tuning on complex samples, can be adopted to adapt to the dual-branch feature extraction structure of the SSAW-Net model. This allows for the gradual learning of solid features, significantly improving the model's training efficiency and scene adaptability.

[0104] In the general feature pre-training stage, the model can be trained using a universal solid waste sample set, freezing some of the underlying network parameters. This allows the SSAW-Net model to fully learn the basic morphology and spectral common features of solid waste, strengthening the model's basic recognition capabilities and avoiding feature learning chaos in complex scenarios.

[0105] In the complex sample fine-tuning stage, all network parameters can be unfrozen, and secondary fine-tuning can be performed using sample sets of highly disturbed and complex terrains such as mountains, forests, and urban-rural fringe areas. This allows for targeted adaptation to complex and disturbed outdoor scenarios, solving the core problems of high false detection rate and weak generalization ability of traditional models in complex terrains, and upgrading the SSAW-Net model from general recognition to scene-specific recognition.

[0106] In some embodiments, the method can also expand multidimensional samples through remote sensing-specific data augmentation. Traditional image data augmentation only includes spatial dimension transformations such as flipping, cropping, and scaling, which fails to adapt to the core characteristics of remote sensing images, such as variable spectra, complex terrain, and diverse imaging environments. This results in a single dimension of sample expansion, unable to cover complex field monitoring scenarios. Therefore, three types of remote sensing-specific multidimensional data augmentation methods can be designed to simultaneously expand sample diversity from three dimensions: spectrum, scale, and terrain. This addresses the problems of scarce remote sensing solid waste samples, incomplete scene coverage, and model overfitting, fully adapting to the spectral-spatial dual-branch feature learning requirements of the SSAW-Net model.

[0107] Therefore, by adjusting the spectral band reflectance values ​​of the labeled image blocks within a preset physical range, a band-enhanced image block can be obtained. A random scaling transformation is then performed on the labeled image block to obtain a scale-transformed enhanced image block. Furthermore, random interference elements are randomly superimposed on the labeled image block to obtain a terrain simulation enhanced image block. The random scaling transformation includes at least one of random scaling and random cropping. The interference elements include at least one of natural shadows, vegetation occlusion, and surface debris coverage. An expanded sample set is then generated based on the band-enhanced image block, the scale-transformed enhanced image block, and the terrain simulation enhanced image block.

[0108] For example, in the spectral dimension, band reflectance can be randomly fine-tuned and enhanced. This involves slightly adjusting the reflectance values ​​of each spectral band within a reasonable physical range to accurately simulate the differences in satellite imaging under different atmospheric humidity, light intensity, and imaging time. This allows the SSAW-Net model to learn the essential spectral characteristics of solid waste and avoid recognition failures caused by changes in the imaging environment.

[0109] In terms of scale, scale transformation enhancement can be performed by randomly scaling and cropping image patches to simulate solid waste targets of different resolutions and sizes, adapting to the multi-scale detection characteristics of the SSAW-Net model. In terms of scene, terrain simulation enhancement can be performed. This involves randomly overlaying realistic interference elements such as natural shadows, slight vegetation occlusion, and surface debris to simulate complex field monitoring scenarios, forcing the SSAW-Net model to learn the core features of solid waste under disturbed environments, significantly improving the model's anti-interference ability and scene adaptability. Furthermore, common image enhancement methods such as Mosaic and random flipping can be used simultaneously to further enrich the sample scenes and ensure the model's training effectiveness.

[0110] The expanded sample set is then input into the remote sensing solid waste detection model to obtain the training result data output by the model. Next, a combined training loss is calculated based on the training result data; this combined training loss is a weighted sum of the classification loss, bounding box regression loss, and confidence loss. Finally, the network parameters of the remote sensing solid waste detection model are adjusted according to the combined training loss.

[0111] For example, because the native YOLO loss function uses uniform weights to constrain all targets, it is not adapted to the characteristics of remote sensing solid waste, such as a large number of small targets, easy boundary shift of large targets, and imbalance of positive and negative samples. This results in a high false negative rate for small solid waste, inaccurate fitting of large wasteyard boundaries, and a high probability of false positives. Therefore, the SSAW-Net model's structural characteristics can be combined to optimize the combined loss function with adaptive weights. By optimizing the combined loss function, an adaptive weighting mechanism for targets of different sizes can be designed to hierarchically constrain the accuracy of classification, boundary regression, and confidence discrimination, achieving accurate optimization of solid waste targets at different scales. The total loss then consists of three parts: classification loss, bounding box regression loss, and confidence loss, i.e.:

[0112] Among them, L cls This represents the classification loss. The classification loss can be the cross-entropy loss, which accurately constrains the classification prediction accuracy of various types of solid waste, such as scattered solid waste, industrial solid waste dumps, and municipal solid waste landfills, reduces the probability of classification confusion between different solid wastes in the same type of terrain, and adapts to the multi-class recognition needs of the SSAW-Net model.

[0113] L box This represents the bounding box regression loss; This represents the adaptive weighting coefficients of the bounding box regression loss. The bounding box regression loss can employ a CIoU loss that considers overlap, center distance, and aspect ratio to accurately optimize the boundary fitting effect for large, contiguous solid waste stockpiles, addressing issues such as boundary offset and incomplete contour fitting for large targets. Simultaneously, an adaptive weighting coefficient is added to the bounding box regression loss. The weight coefficient for large targets is set to 1.0, while for small, scattered solid waste targets with a pixel area of ​​less than 1024 pixels, an adaptive high-weight reinforcement regression constraint strength higher than 1.0 is configured to balance the training weights of large and small targets and alleviate the problem of small target omission in the SSAW-Net model.

[0114] L obj This represents the confidence loss. The confidence loss can be achieved using binary cross-entropy loss, which accurately distinguishes foreground solid waste targets from complex background areas, effectively suppressing false detections and false positives caused by vegetation, shadows, and bare land, and improving the accuracy of the SSAW-Net model's detection confidence.

[0115] Furthermore, an optimal model selection mechanism can be used to set training hyperparameters. By configuring dedicated training hyperparameters adapted to SSAW-Net remote sensing solid waste detection, efficient model convergence and optimal performance are ensured. The SSAW-Net model optimizer uses the AdamW optimizer, which has the advantage of weight decay and can effectively prevent overfitting. The initial learning rate of the model training is set to 0.001, combined with a cosine annealing learning rate decay strategy, to achieve rapid convergence in the early stage of training and fine-tuning in the later stage. The batch size of the model training is set to 8 or 16 depending on hardware resources, and the total number of iterations is set to 300 to 500. At the same time, a precise optimal model selection mechanism is set up. After each iteration, the model mAP50 (mean accuracy) is evaluated using a validation set, and the SSAW-Net optimal weight model with the highest accuracy is saved in real time to eliminate overfitting and underfitting models, ensuring the comprehensive detection performance of the final inference model.

[0116] In some embodiments, as a specific implementation of the intelligent identification method for solid waste from remote sensing images described in the above embodiments, some embodiments of this application also provide an intelligent identification system for solid waste from remote sensing images, such as... Figure 5As shown, the system includes: The data acquisition module is used to acquire multispectral remote sensing image data, which includes remote sensing images of multiple spectral bands. The image cropping module is used to crop the remote sensing image into multiple image blocks by a sliding window according to dynamic cropping parameters, wherein the dynamic cropping parameters are dynamically adjusted based on the image resolution and ground feature density of the remote sensing image. The solid waste detection module is used to sequentially input multiple image blocks into the remote sensing solid waste detection model to obtain the complete inference parameters output by the remote sensing solid waste detection model. The remote sensing solid waste detection model is an image detection model formed by building spatial feature branches and spectral feature branches in parallel at the front end of the basic backbone network, and embedding a dilated spatial pyramid pooling module and an efficient channel attention mechanism. The dilated spatial pyramid pooling module uses multiple sets of dilated convolutions with different dilation rates in parallel. The complete inference parameters include the pixel-level bounding box coordinates and target confidence for the solid waste target. The post-processing module is used to filter out valid bounding boxes by traversing the bounding box coordinates in the complete inference parameters, and to calculate the solid waste dump quantitative index based on the complete inference parameters corresponding to the valid bounding boxes; the valid bounding boxes are the bounding box coordinates filtered based on the post-processing threshold; the post-processing threshold is a threshold dynamically set according to different landform types according to the landform adaptive dynamic nonmaximum suppression strategy. The results output module is used to export the complete inference parameters and the quantitative indicators of the solid waste storage site in a structured manner to output a solid waste monitoring results file.

[0117] By applying the technical solutions of the above embodiments, the remote sensing image solid waste intelligent identification system described in the above embodiments can acquire multispectral remote sensing image data through a data acquisition module, and then perform sliding window cropping by an image cropping module according to dynamic cropping parameters. The solid waste detection module then sequentially inputs the multiple image blocks obtained from the cropping into the remote sensing solid waste detection model to obtain complete inference parameters. The post-processing module then filters out effective bounding boxes by traversing the bounding box coordinates in the complete inference parameters and calculates the solid waste dump quantitative index, so that the result output module can export the complete inference parameters and the solid waste dump quantitative index in a structured manner, and output a solid waste monitoring result file. The system can dynamically adjust the cropping parameters according to the image resolution and ground feature density to adapt to different solid waste targets, and improve the feature expression ability and accuracy of solid waste identification results by adding parallel spatial feature and spectral feature dual extraction branches and feature fusion in the remote sensing solid waste detection model.

[0118] It should be noted that other corresponding descriptions of the functional units involved in the intelligent identification system for solid waste from remote sensing images provided in the embodiments of this application can be found in the corresponding descriptions in the intelligent identification method for solid waste from remote sensing images provided in the above embodiments, and will not be repeated here.

[0119] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0120] The embodiments described above are merely examples of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these modifications and improvements all fall within the protection scope of this application.

Claims

1. A method for intelligent identification of solid waste from remote sensing images, characterized in that, The method includes: Acquire multispectral remote sensing image data, which includes remote sensing images of multiple spectral bands; The remote sensing image is sliced ​​into multiple image blocks using a sliding window according to dynamic cropping parameters, which are dynamically adjusted based on the image resolution and ground feature density of the remote sensing image. Multiple image blocks are sequentially input into the remote sensing solid waste detection model to obtain the complete inference parameters output by the model. The remote sensing solid waste detection model is an image detection model based on the YOLO network as its backbone network. Spatial feature branches and spectral feature branches are built in parallel at the front end of the backbone network, and a dilated spatial pyramid pooling module and an efficient channel attention mechanism are embedded in it. The dilated spatial pyramid pooling module uses multiple sets of dilated convolutions with different dilation rates in parallel. The complete inference parameters include the pixel-level bounding box coordinates and target confidence for the solid waste target. The process involves filtering valid bounding boxes by traversing the bounding box coordinates in the complete inference parameters, and calculating quantitative indicators for solid waste dumps based on the complete inference parameters corresponding to the valid bounding boxes. This includes: calculating the pixel width and pixel height of the valid bounding boxes in the complete inference parameters; obtaining the image spatial resolution of the remote sensing image; calculating the solid waste area based on the pixel width and pixel height, combined with the image spatial resolution; extracting the vertex geographic coordinates of the valid bounding boxes; and determining the actual boundary range of the solid waste dump based on the vertex geographic coordinates. The valid bounding boxes are the bounding box coordinates filtered based on a post-processing threshold. The post-processing threshold is a threshold dynamically set according to different landform types using a terrain-adaptive dynamic non-maximum suppression strategy. The complete inference parameters and the quantitative indicators of the solid waste storage site are exported in a structured manner to output a solid waste monitoring results document.

2. The method according to claim 1, characterized in that, Acquire multispectral remote sensing image data, including: Acquire raw input data, which includes multispectral remote sensing images of the area to be monitored collected by the remote sensing monitoring equipment and recorded metadata; the metadata includes at least one of raw spatial resolution, geographic coordinates, and projected coordinate system; The remote sensing images in the original input data are subjected to standardization correction processing to obtain the corrected images. The standardization correction processing includes atmospheric correction, orthorectification, and panchromatic multispectral fusion. The corrected image is subjected to remote sensing noise removal and vegetation masking to obtain the multispectral remote sensing image data.

3. The method according to claim 1, characterized in that, The remote sensing image is cropped into multiple image blocks by a sliding window according to dynamic cropping parameters, including: Define a basic clipping window, whose basic clipping parameters include window size, minimum overlap rate, and maximum overlap rate; The image resolution and ground feature density of the remote sensing image are obtained, wherein the ground feature density is the prior ground feature distribution density obtained based on prior knowledge of the remote sensing image. The basic cropping window is dynamically adjusted based on the image resolution and the density of ground features. The adjustment parameters of the basic cropping window include the image size of a single image block and the overlap rate between blocks. The number of pixels in the single-sided overlapping area is calculated based on the image size and the inter-block overlap rate to generate the dynamic cropping parameters; The entire remote sensing image is cropped using the dynamic cropping parameters to obtain multiple image blocks, and the pixel coordinates of the marker points of each image block are recorded synchronously.

4. The method according to claim 1, characterized in that, Multiple image patches are sequentially input into the remote sensing solid waste detection model to obtain the complete inference parameters output by the remote sensing solid waste detection model, including: The image block is input into the spatial feature branch to extract spatial features through the convolutional structure of the YOLO network; the spatial features include at least one of image texture features, image contour features, and image morphological features. The image block is input into the spectral feature branch to extract spectral features through a channel fusion convolution and a depthwise separable convolution cascade structure in the spectral feature branch; the spectral features include multi-band original spectral reflectance features and spectral response difference features; Learnable weight coefficients are obtained, and the spatial features and spectral features are adaptively weighted and fused based on the learnable weight coefficients to obtain fused features; The fused features are fed into the feature pyramid network and combined with the void space pyramid pooling module to output multi-scale perceptual features. According to the efficient channel attention mechanism, feature filtering is performed on the multi-scale perceptual features to obtain recalibrated features; By dividing the recalibrated features into a grid, the association between local detection features and surrounding landform features is established, and global context features are generated by combining the context information passed by the adjacency graph convolution. The detection head is used to output the complete inference parameters based on the global context features and the local detection features.

5. The method according to claim 1, characterized in that, The method further includes: Annotated image blocks are obtained by adding sample annotations to the image blocks; the sample annotations include solid waste category and target bounding rectangle boundary; A sample dataset is constructed based on the labeled image blocks. The sample dataset includes a global general solid waste sample set and a complex terrain sample set. The initial model is trained using the global general solid waste sample set to obtain a general pre-trained model, wherein some network parameters are frozen during the training process of the initial model; The general pre-trained model is fine-tuned using the complex terrain sample set to obtain the remote sensing solid waste detection model. During the fine-tuning training process, all network parameters of the general pre-trained model are unfrozen.

6. The method according to claim 5, characterized in that, The method further includes: By adjusting the spectral band reflectance values ​​corresponding to the labeled image blocks within a preset physical range, a band-enhanced image block is obtained; A random scale transformation is performed on the labeled image block to obtain a scale-enhanced image block; the random scale transformation includes at least one of random scaling and random cropping. Disturbance elements are randomly superimposed onto the labeled image blocks to obtain terrain simulation enhanced image blocks; the interference elements include at least one of natural shadows, vegetation occlusion, and surface debris coverage; An expanded sample set is generated based on the band-enhanced image block, the scale-transformed enhanced image block, and the terrain simulation enhanced image block; The expanded sample set is input into the remote sensing solid waste detection model to obtain the training result data output by the remote sensing solid waste detection model; The combined training loss is calculated based on the training results data. The combined training loss is a weighted sum of the classification loss, bounding box regression loss, and confidence loss. The network parameters of the remote sensing solid waste detection model are adjusted based on the combined training loss.

7. The method according to claim 1, characterized in that, Valid bounding boxes are selected by traversing the bounding box coordinates in the complete inference parameters, including: Extract the bounding box coordinates from the complete inference parameters; Calculate the intersection-union ratio between the two bounding boxes based on the bounding box coordinates; Extract the image slope and NDVI vegetation index of the remote sensing image, and determine the landform type of the remote sensing image based on the image slope and the NDVI vegetation index. The post-processing threshold is set according to the landform type, and the post-processing threshold includes an intersection-over-union threshold and a confidence threshold. The effective bounding boxes are obtained by filtering out bounding boxes from the complete inference parameters whose intersection-union ratio is greater than the intersection-union ratio threshold and whose target confidence is greater than the confidence threshold.

8. The method according to claim 1, characterized in that, The complete inference parameters and the quantitative indicators of the solid waste storage site are exported in a structured manner to output a solid waste monitoring results document, including: Extract valid bounding box coordinates and block offset coordinates from the complete inference parameters; The effective bounding box coordinates are calculated back to the remote sensing image according to the block offset coordinates; Different types of valid bounding boxes are labeled using different display styles on the remote sensing image, and labeled result images are generated by synchronously labeling detection result information; the detection result information includes at least one of category name, target confidence level, and area. Based on the complete inference parameters and the quantitative indicators of the solid waste storage site, standardized detection reports and target vector information are generated in batches. According to the output format set by the regulatory business requirements information, the result image, the standardized test report, and the target vector information are output as solid waste monitoring result files.

9. A remote sensing image-based intelligent identification system for solid waste, characterized in that, The system includes: The data acquisition module is used to acquire multispectral remote sensing image data, which includes remote sensing images of multiple spectral bands. The image cropping module is used to crop the remote sensing image into multiple image blocks by a sliding window according to dynamic cropping parameters, wherein the dynamic cropping parameters are dynamically adjusted based on the image resolution and ground feature density of the remote sensing image. The solid waste detection module is used to sequentially input multiple image blocks into the remote sensing solid waste detection model to obtain the complete inference parameters output by the remote sensing solid waste detection model. The remote sensing solid waste detection model is an image detection model formed by building spatial feature branches and spectral feature branches in parallel at the front end of the basic backbone network, and embedding a dilated spatial pyramid pooling module and an efficient channel attention mechanism. The dilated spatial pyramid pooling module uses multiple sets of dilated convolutions with different dilation rates in parallel. The complete inference parameters include the pixel-level bounding box coordinates and target confidence for the solid waste target. The post-processing module is used to filter out valid bounding boxes by traversing the bounding box coordinates in the complete inference parameters, and to calculate the quantitative indicators of the solid waste dump based on the complete inference parameters corresponding to the valid bounding boxes. This includes: calculating the pixel width and pixel height of the valid bounding boxes in the complete inference parameters; obtaining the image spatial resolution of the remote sensing image; calculating the solid waste area based on the pixel width and pixel height, combined with the image spatial resolution; extracting the vertex geographic coordinates of the valid bounding boxes; and determining the actual boundary range of the solid waste dump based on the vertex geographic coordinates. The valid bounding boxes are the bounding box coordinates filtered based on a post-processing threshold. The post-processing threshold is a threshold dynamically set according to different landform types using a terrain-adaptive dynamic non-maximum suppression strategy. The results output module is used to export the complete inference parameters and the quantitative indicators of the solid waste storage site in a structured manner to output a solid waste monitoring results file.

Images