Urban scale solid waste landfill site identification method based on artificial intelligence target identification algorithm

By improving the YOLOv8 model and multi-source data enhancement technology, combined with on-site verification, the problems of low efficiency and insufficient accuracy of urban-scale solid waste landfill site identification have been solved, and efficient and accurate full-domain coverage and continuous optimization have been achieved.

CN120544041APending Publication Date: 2025-08-26ZHEJIANG UNIV

Patent Information

Application Number
CN202510662969.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-05-22
Publication Date
2025-08-26

AI Technical Summary

Technical Problem

The prior art is inefficient in identifying solid waste landfill sites on the urban scale, inaccurate accuracy, limited coverage, and difficult to adapt to changes in complex backgrounds and different geographical environments.

Method used

The improved YOLOv8 model is adopted, combining multi-source data and data augmentation technology, and through multi-scale feature aggregation and transfer learning, the urban-scale solid waste landfill site identification model is trained, and the results of field survey verification are combined.

Benefits of technology

Large-scale and high-frequency monitoring of solid waste landfill sites has been achieved, identification accuracy and coverage have been improved, adapted to different environments and backgrounds, reduced the rate of misjudgment, and supported full-region coverage and continuous optimization.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120544041A_ABST
    Figure CN120544041A_ABST
Patent Text Reader

Abstract

The invention discloses an urban scale solid waste landfill site identification method based on an artificial intelligence target identification algorithm, and belongs to the technical field of urban solid waste management. Comprising the following steps: constructing a multi-source remote sensing image database of an urban scale solid waste landfill site through multi-source data; improving a YOLOv8 model to adapt to a remote sensing image recognition task, and training a solid waste landfill site recognition model for an urban scale by using the data subjected to data enhancement; acquiring a visible light remote sensing image of a target area, performing radiation correction, atmospheric correction and geometric correction, inputting the trained model for reasoning, and identifying a solid waste filling site; and sampling a model identification result, verifying the accuracy through field survey, and compiling an actual material diagram. According to the method, rapid and accurate identification of the urban scale solid waste landfill site is realized by using an artificial intelligence technology, the identification efficiency and accuracy are remarkably improved, and powerful support is provided for urban solid waste management and environmental governance.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of urban-scale solid waste landfill identification, and specifically relates to an urban-scale solid waste landfill identification method based on an artificial intelligence target recognition algorithm. Background Art

[0002] In the current context of "zero-waste city" construction, the management and disposal of solid waste has become a key issue in achieving sustainable urban development. Solid waste landfills, particularly informal ones, are both blind spots in urban waste management and a significant source of environmental pollution. Efficiently and accurately identifying these landfills and conducting dynamic risk assessments are key tasks in promoting zero-waste city development and minimizing the impact of solid waste.

[0003] Traditional solid waste landfill identification methods rely on manual surveys and bottom-up information reporting. Limited by manpower, time, and technical limitations, they often struggle to comprehensively cover all solid waste landfill sites within a city. Due to the widespread geographic distribution, uncertainty, and hidden nature of solid waste landfills, traditional statistical methods are prone to omissions and misidentification, which in turn impacts the implementation and effectiveness of relevant remediation measures. Remote sensing imagery, however, enables efficient coverage and scanning of large areas. The resulting spatial data provides a scientific basis for the precise identification of solid waste landfills. Analyzing remote sensing imagery using object recognition algorithms not only allows for rapid location pinpointing of solid waste landfills but also enables large-scale, high-frequency monitoring, significantly reducing the cost of traditional manual inspections and significantly improving inspection efficiency. This provides city managers with more comprehensive, timely, and accurate data on solid waste landfill locations, effectively addressing the gaps in existing management methods. The application of object recognition algorithms to urban-scale solid waste landfill identification primarily involves automatically extracting features of target areas from remote sensing imagery, enabling rapid and accurate identification of solid waste landfills. This method offers significant advantages over traditional manual interpretation, particularly in large-scale, frequent monitoring scenarios. Traditional solid waste landfill identification methods rely on manual analysis and interpretation of remote sensing imagery by experts. This typically requires significant time and human resources, is susceptible to human subjectivity, and struggles to efficiently and accurately process large amounts of image data. Consequently, automated methods based on target recognition algorithms have gained widespread adoption and become an effective means of addressing this problem.

[0004] Deep learning-based object recognition algorithms are a mainstream technology in the field. Convolutional neural network (CNN) algorithms, in particular, have been widely used for the automatic identification of solid waste landfills. These algorithms are trained on large amounts of annotated remote sensing imagery data to automatically learn the spatial, shape, and texture features of solid waste landfills within the images. This allows them to accurately identify and locate solid waste landfills in new, unseen imagery. Advantages of Deep Learning-Based Object Recognition Algorithms: Object recognition algorithms offer significant advantages over traditional image analysis methods. Deep learning algorithms can process large amounts of data and automatically extract useful features, avoiding the tedious and subjective nature of manual feature extraction. Therefore, deep learning-based recognition methods are both efficient and robust, adapting to the identification needs of solid waste landfills in diverse regions and types. Furthermore, these algorithms can be trained using large datasets, enabling them to not only learn target features from a global perspective but also optimize for local details, thereby improving recognition accuracy. For example, algorithms such as YOLO excel in real-time processing and rapid recognition, capable of processing hundreds of targets in seconds, making them suitable for rapid monitoring of urban-scale solid waste landfills. Multi-source data fusion improves recognition accuracy. Because the geographic characteristics and environmental context of solid waste landfills vary significantly, single-source remote sensing images may not provide sufficient discriminative information. Therefore, many studies have begun to explore the fusion of multi-source data to improve recognition accuracy. This allows object recognition algorithms to reason based on more comprehensive information, thereby improving recognition accuracy and robustness. In complex urban environments, solid waste landfills can have similar appearances to other types of landforms (such as buildings and green spaces), making them difficult to distinguish using traditional single-image data. However, integrating ground monitoring data or GIS data can effectively enhance the algorithm's background suppression capabilities, thereby improving solid waste landfill recognition accuracy. While solid waste landfill recognition methods based on object recognition algorithms offer significant advantages, they still face challenges in practical application. The quality and resolution of remote sensing images directly impact recognition performance. In particular, in low-resolution images, detailed information about solid waste landfills can be lost, resulting in inaccurate recognition results. Furthermore, the large variations in landfill size, morphology, and their relationship to the surrounding environment can lead to variations in algorithm performance across different regions. Urban solid waste landfills are often surrounded by complex backgrounds such as buildings and roads, which challenges the generalization of algorithms. To address these challenges, diverse training datasets are crucial, especially those covering landfills in diverse geographical environments and climates. This requires including sample data from a variety of backgrounds to ensure the algorithm can adapt to these complex environments. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for identifying urban-scale solid waste landfills based on an artificial intelligence target recognition algorithm to solve the technical problems of low efficiency, insufficient accuracy and limited coverage of urban-scale solid waste landfill identification in the existing technology.

[0006] To achieve the above objectives, the present invention provides the following technical solutions:

[0007] The present invention provides a method for identifying urban-scale solid waste landfills based on an artificial intelligence target recognition algorithm, comprising the following steps:

[0008] Step 1: Use multi-source data to construct a multi-source remote sensing image database of urban-scale solid waste landfill sites;

[0009] Step 2: Improve the YOLOv8 model to adapt to remote sensing image recognition tasks and use the data augmentation data to train a city-scale solid waste landfill recognition model;

[0010] Step 3: Acquire and process visible light remote sensing images of the target area and use the trained target recognition model for inference;

[0011] Step 4: Sample the model identification results and verify the results through field surveys.

[0012] Furthermore, the step 1 specifically includes:

[0013] Obtain high-confidence data through public datasets and government-issued solid waste location lists;

[0014] Select satellite image sources based on the geographic characteristics of the target area and the properties of the solid waste landfill, including spatial resolution, spectral resolution, and revisit period;

[0015] The solid waste location information is interpreted into corresponding remote sensing image data through latitude and longitude, and a dynamic database is constructed by combining multi-temporal remote sensing images.

[0016] Furthermore, the step 2 specifically includes:

[0017] Step 2.1: By embedding the LSKA module in the YOLOv8 model SPPF and adding the CCFM module at the neck of the model, the model's multi-scale feature aggregation capability is improved;

[0018] Step 2.2: Use data augmentation methods to augment the remote sensing images in the dataset. For the enhancement of remote sensing images, a data augmentation method combining Mix up, Cut out, Mosaic and random dithering is used to simulate image changes under different environmental backgrounds.

[0019] Step 2.3: Use transfer learning to train the recognition model based on the pre-trained weights to ensure the stability and convergence speed of the training process. During training, select the loss function and optimizer suitable for the remote sensing image classification task, and set the hyperparameters of the training process in a targeted manner.

[0020] Furthermore, the step 2.1 specifically includes the following steps:

[0021] Step 2.1.1: The LSKA module decomposes the two-dimensional depth convolution kernel into cascaded horizontal and vertical one-dimensional convolution kernels, processing the horizontal and vertical features of the feature map respectively. On this basis, spatial dilation convolution with different dilation rates is used to process in both horizontal and vertical directions to increase the receptive field and further extract features:

[0022]

[0023] A C =W 1×1 *Z C

[0024]

[0025] Where Z C The feature map F is obtained by using a 1×(2d-1) vertical depth convolution kernel and a (2d-1)×1 horizontal depth convolution kernel. C The characteristics of the vertical and horizontal directions are then divided into and The convolutional neural network of the deep convolutional neural network is used to compensate for the grid effect of the deep convolutional neural network. On this basis, a 1×1 convolution kernel is used to perform channel compression to obtain the attention map A. C ,Finally, calculate the attention map A C and feature map F C The final characteristic map of the Haadmann product;

[0026] Step 2.1.2: Embed the CCFM module in the YOLOv8 neck to improve the quality of feature fusion. Use two layers of cross-scale convolution to extract the features of the feature pyramid fusion. Then use the horizontal and vertical attention mechanism blocks with a size of 1 / 16 of the extracted features to compress and recalibrate the extracted features.

[0027] Step 2.1.3: Change the YOLOv8 head structure to a dynamic head structure, and use attention from three dimensions: scale perception, spatial perception, and task perception to improve feature perception capabilities:

[0028] W(F)=π C (π S (π L (F)·F)·F)·F

[0029]

[0030] π C (F)·F=max(α 1 (F)·F c +β 1 (F),α 2 (F)·F c +β 2 (F))

[0031] Where, π L (F) is scale-aware attention, which uses 1×1 convolution for feature mapping and then uses sigmoid function activation as the attention value; π S (F) is spatial perception attention, which uses variable convolution to perform sparse attention calculation, and then aggregates different layer features at the same location, so that the model always pays attention to the identifiable areas coexisting between different layers; π C (F) is task-aware attention, which uses hyperfunctions to control the activation values ​​of different channel features, summarizes different features of the same object, and improves task perception capabilities.

[0032] Furthermore, the step 2.3 specifically includes the following steps:

[0033] Step 2.3.1: Use downsampling and oversampling methods to divide the dataset into a balanced training set, a validation set, and a test set to ensure that the sample class distribution in each dataset is consistent;

[0034] Step 2.3.2: Use the pre-trained YOLOv8 weights as the initial weights for transfer learning to train the model;

[0035] Step 2.3.3: The solid waste landfill identification task is a classification task, so the loss function selected for model training can describe the cross entropy loss of the difference between prediction and label:

[0036]

[0037] Where, is the model prediction classification result, y i is the actual label, and the cross entropy describes the difference between the two distributions. Using the optimizer to minimize the loss function can make the model prediction gradually approach the actual label;

[0038] Step 2.3.4: When training the model, the optimizer selects the Adam optimizer, which dynamically adjusts the learning rate through first-order moment and second-order moment estimation:

[0039]

[0040] mt =β1·m t-1 +(1-β1)·g t

[0041]

[0042] Where θ t+1 represents the updated parameters, θ t represents the currently updated parameter, α is the learning rate, is the resulting bias-corrected first-order moment estimate of the gradient, is the bias-corrected gradient second-order moment estimate, ε is a very small constant to ensure zero stability, m t is the current momentum, g t is the current gradient, β1 is the decay factor, v t is the current second-order moment, is the square of the current gradient, β2 is the second-order moment attenuation factor. When the momentum and second-order moment deviations are corrected, as the iteration proceeds, the attenuation degree gradually approaches 1, and the correction effect on the momentum and second-order moment becomes weaker and weaker, that is, the closer the gradient is to the current step, the greater the influence on the current gradient.

[0043] Furthermore, the step 3 specifically includes:

[0044] Step 3.1: Select remote sensing images that meet the timestamp, resolution, and quality requirements based on the mission requirements;

[0045] Step 3.2: Perform radiometric, atmospheric, and geometric corrections on the image.

[0046] Step 3.3: Input the pre-processed remote sensing image data into the trained recognition model for inference to obtain the predicted results of the solid waste landfill site. Adjust the spatial resolution or map it as needed to ensure that the predicted results effectively correspond to the actual area.

[0047] Step 3.4: Post-process the results of model inference, including noise removal, boundary smoothing, and region merging.

[0048] Furthermore, the reasoning process of step 3.3 includes:

[0049] The input image is normalized to the range [0,1]. The band characteristics of the multispectral image are integrated and combined with the NDVI index to distinguish the solid waste dump from the vegetation background.

[0050] A sliding window multi-scale inference method is used to fuse prediction results of different scales.

[0051] Furthermore, after verification, an actual material map is compiled; the actual material map includes the route of field investigation and verification, the location and number of observation points, the location and number of specimen and sample collection, and the expression method is represented by the actual attribute classification code, classification code notes and map name, legend, and scale.

[0052] Based on the above technical solution, the embodiments of the present invention can produce at least the following technical effects:

[0053] (1) This invention replaces traditional manual surveys with automated AI recognition, significantly reducing labor costs and time consumption, and achieving large-scale, high-frequency monitoring of urban solid waste landfills. In combination with multi-source remote sensing data, it can quickly scan areas of thousands of square kilometers, ensuring full coverage and reducing omissions.

[0054] (2) The present invention adopts convolutional neural networks and data enhancement to enable the model to adapt to changes in different lighting, seasons and geographical environments, improve the recognition accuracy under complex backgrounds, and combine NDVI index, multi-band feature fusion and other technologies to enhance the spectral and texture feature extraction of solid waste landfills and reduce the misjudgment rate.

[0055] (3) The present invention utilizes transfer learning and multi-scale reasoning to enable the algorithm to adapt to the task of identifying solid waste landfills of different regions and sizes, thus avoiding performance degradation caused by differences in data distribution. Furthermore, through dynamic database updates and combined with multi-temporal remote sensing images, the model can be continuously optimized to meet long-term monitoring needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying any creative work.

[0057] Figure 1 This is a flow chart of the city-scale solid waste landfill identification method of the present invention;

[0058] Figure 2 This is a schematic diagram of the samples in the dataset after rotation, scaling, and mosaic data enhancement;

[0059] Figure 3 This is a schematic diagram of the regional solid waste landfill identification results of the present invention. DETAILED DESCRIPTION

[0060] The technical solutions in the embodiments of the present invention will be described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. In addition, the technical solutions between the various embodiments can be combined with each other, but they must be based on the ability of ordinary technicians in this field to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0061] The embodiment of the present invention provides a method for identifying urban scale solid waste landfill sites based on an artificial intelligence target recognition algorithm, such as Figure 1 Shown, including:

[0062] Step 1: Build a remote sensing image database of solid waste landfill sites using multi-source data including but not limited to public datasets and remote sensing interpretations;

[0063] Step 2: Improve the YOLOv8 model to adapt to remote sensing image recognition tasks and use the data augmentation data to train a city-scale solid waste landfill recognition model;

[0064] Step 3: Acquire and process visible light remote sensing images of the target area and use the trained target recognition model for inference;

[0065] Step 4: Sample the model identification results and verify the results through field surveys.

[0066] Example 1: Structural optimization is performed based on the YOLOv8 model for multi-scale and small target recognition, and an urban-scale solid waste landfill recognition model is trained using the constructed database. Appropriate hyperparameters are set during the training process to ensure the model's generalization ability to adapt to the problem of background noise interference in urban-scale solid waste landfill working scenarios, including: embedding the LSKA module, CCFM module, and dynamic head based on YOLOv8.

[0067] Step 2.1 addresses the problem of severe background noise in remote sensing images, which suppresses the characteristics of solid waste landfill sites. By embedding the LSKA module in the YOLOv8 model SPPF and adding the CCFM module to the model neck, the model's multi-scale feature aggregation capability is improved.

[0068] Step 2.1 includes:

[0069] Step 2.1.1: The LSKA module decomposes the two-dimensional depthwise convolution kernel into a cascade of horizontal and vertical one-dimensional convolution kernels, processing the horizontal and vertical features of the feature map respectively. Based on this, spatial dilated convolutions with different dilation rates are used to process the horizontal and vertical features, which can increase the receptive field and further extract features.

[0070]

[0071] Where Z C The feature map F is obtained by using a 1×(2d-1) vertical depth convolution kernel and a (2d-1)×1 horizontal depth convolution kernel. C The characteristics of the vertical and horizontal directions. Then, the size is divided into and The convolutional neural network of the deep convolutional neural network is used to compensate for the grid effect of the deep convolutional neural network. On this basis, a 1×1 convolution kernel is used to perform channel compression to obtain the attention map A. C Finally, calculate the attention map A C and feature map F C The Haadmann product of is the final feature map.

[0072] Step 2.1.2: To further enhance the model's multi-scale detection capabilities, a lightweight Cross-Scale Feature Fusion Module (CCFM) is embedded in the YOLOv8 framework to improve feature fusion quality. Two layers of cross-scale convolution are used to extract features from the feature pyramid fusion. These features are then compressed and recalibrated using horizontal and vertical attention blocks that are 1 / 16 the size of the extracted features to enhance the response to important features.

[0073] Step 2.1.3: To address the problem of insufficient feature information for remote sensing small target recognition, the YOLOv8 head structure is changed to a dynamic head structure, and attention is used from three dimensions: scale perception, spatial perception, and task perception to improve feature perception capabilities.

[0074] W(F)=π C (π S (π L (F)·F)·F)·F

[0075]

[0076]

[0077] π C (F)·F=max(α 1 (F)·F c +β 1 (F),α 2 (F)·F c +β 2(F))

[0078] Where π L (F) is scale-aware attention, which uses 1×1 convolution for feature mapping and then uses sigmoid function activation as the attention value; π S (F) is spatial perception attention, which uses variable convolution to perform sparse attention calculation, and then aggregates different layer features at the same location, so that the model always pays attention to the recognizable areas coexisting between different layers; π C (F) is task-aware attention, which uses hyperfunctions to control the activation values ​​of different channel features, summarizes different features of the same object, and improves task perception capabilities.

[0079] In step 2.2, to enhance the generalization and robustness of the recognition model and better adapt it to the task of identifying solid waste dumps in diverse backgrounds, data augmentation is performed on the remote sensing images in the dataset. This augmentation primarily uses a combination of Mix Up, Cut Out, Mosaic, and Random Jitter to simulate image variations in diverse environments, increase the diversity of the training data, and reduce the risk of overfitting.

[0080] In step 2.3, transfer learning is used to train the recognition model based on the pre-trained weights to ensure the stability and convergence speed of the training process. During training, the loss function and optimizer suitable for the remote sensing image classification task are selected, and the hyperparameters of the training process are set in a targeted manner.

[0081] Step 2.3 includes:

[0082] Step 2.3.1: Use methods such as downsampling and oversampling to divide the dataset into balanced training, validation, and test sets. Ensure that the sample category distribution in each dataset is as consistent as possible to avoid the model being biased towards a certain type of target due to uneven sample distribution, which will affect the model's generalization ability.

[0083] Step 2.3.2: Use the YOLOv8 weights pre-trained on the COCO dataset as the initial weights for transfer learning to train the model. This can avoid the gradient explosion or gradient vanishing problem caused by numerical instability in the early stages of training due to the deep number of network layers, and can also shorten the training convergence time.

[0084] Step 2.3.3: The solid waste landfill identification task is a classification task, so the loss function selected for model training can describe the cross-entropy loss of the difference between the prediction and the label.

[0085]

[0086] In the formula is the model prediction classification result, y iis the actual label, and the cross entropy describes the difference between the two distributions. Using the optimizer to minimize the loss function can make the model prediction gradually close to the actual label.

[0087] Step 2.3.4: During model training, the Adam optimizer is selected as the optimizer. The Adam optimizer dynamically adjusts the learning rate through first-order moment and second-order moment estimation to reduce the impact of the learning rate setting on the model training effect. At the same time, the average of the gradient can accelerate convergence.

[0088]

[0089] m t =β1·m t-1 +(1-β1)·g t

[0090]

[0091] Where θ t+1 represents the updated parameters, θ t represents the currently updated parameter, α is the learning rate, is the resulting bias-corrected first-order moment estimate of the gradient, is the bias-corrected gradient second-order moment estimate, and ε is a very small constant to ensure zero-division stability. t is the current momentum, g t is the current gradient, β1 is the decay factor, v t is the current second-order moment, is the square of the current gradient, and β2 is the second-order moment attenuation factor. When correcting the momentum and second-order moment deviations, as the iteration proceeds, the attenuation gradually approaches 1, and the correction effect on the momentum and second-order moment becomes weaker and weaker. That is, the gradient closer to the current step has a greater impact on the current gradient.

[0092] Step 2.3.4: The learning rate, batch size, and training epochs during model training should be determined based on computing conditions. Try to set a large batch size within the limits of your graphics memory to improve training performance. Typical batch sizes are powers of 2, such as 16, 32, or 64.

[0093] Figure 2This diagram shows a sample from the dataset after rotation, scaling, and mosaic data augmentation. This image demonstrates how data augmentation techniques are used to process remote sensing imagery in the solid waste landfill target recognition task, providing the model with diverse training data and improving its recognition and generalization capabilities. First, during the preprocessing stage, basic augmentation operations such as rotation and scaling are applied to the solid waste landfill targets in the image. These methods adjust the image's orientation and scale to generate samples from a variety of perspectives and scales. This effectively simulates the image variations caused by changes in shooting angle and target size in real-world scenarios, enhancing the model's adaptability to variations in spatial scale and orientation. Furthermore, random cropping techniques are used to extract localized information from the original remote sensing imagery, generating samples with partially missing or occluded target information. This allows the model to learn to cope with complex occlusion scenarios and blurred boundaries. Stitching techniques, on the other hand, combine images from multiple different scenes to create training samples with multiple targets or complex backgrounds, helping the model to more accurately distinguish between targets and background.

[0094] In terms of optical property enhancement, color jittering, contrast adjustment and other technologies were used to adjust the brightness, saturation and contrast of the image to simulate the differences in optical properties caused by lighting conditions, weather changes or seasonal changes in actual shooting. Combined with these data enhancement technologies, the image data in the image has been processed to successfully cover a variety of possible practical application scenarios, fully improving the diversity and richness of the data set. This processing method not only increases the number of training samples, but also effectively reduces the risk of model overfitting, making the model more accurate, stable and practical when identifying solid waste landfills. Through the above methods, the solid waste landfill targets in the image can be correctly identified by the model under various complex backgrounds and environmental conditions, providing strong support for solid waste management and environmental monitoring.

[0095] Example 2: Using the trained model, reasoning is performed on the remote sensing image of the target area to identify the solid waste landfill site in the target area. This includes:

[0096] Step 3.1 During the inference phase, select appropriate remote sensing image data based on the geographic scope of the target area and the mission requirements, ensuring that the timestamp, resolution, and quality of the image data meet the input requirements of the model;

[0097] Step 3.1 includes:

[0098] Step 3.1.1: When selecting remote sensing imagery data, prioritize the timeframe based on the target area's seasonal, weather, and sunlight conditions. Select clear, cloudless images to minimize atmospheric disturbances and ensure image quality. If the target area exhibits seasonal variations, select imagery data from multiple time points for multi-temporal analysis to improve the model's adaptability to solid waste landfills in different seasonal environments.

[0099] In step 3.1.2, the image data resolution should be selected based on the mission requirements. For solid waste landfills within complex urban environments, higher spatial resolution remote sensing images should be preferred to better distinguish the solid waste landfill from the surrounding environment.

[0100] In step 3.1.3, the quality of the imagery data is crucial. It is recommended to select images from high-quality, reliable data sources, such as commercial satellites, which typically have high spatial resolution and accurate geometric correction. When using publicly available data sources (such as Landsat and Sentinel-2), ensure that the imagery's quality indicators (such as geometric accuracy and radiometric accuracy) meet the required standards and satisfy the model input requirements.

[0101] In step 3.1.4, considering that remote sensing images from different sources may be used in the inference stage, it is preferred to use an image quality consistency calibration scheme, such as unifying images from different satellites or sensors through radiometric standardization methods to ensure data consistency and reduce the deviation caused by differences between images.

[0102] Step 3.2: Preprocess the acquired remote sensing images, including radiometric correction, atmospheric correction, geometric correction, and other possible preprocessing steps, to eliminate the impact of external factors on image quality and ensure the consistency and accuracy of the input data;

[0103] Step 3.3: Input the pre-processed remote sensing image data into the trained recognition model for inference to obtain the prediction results of the solid waste landfill site. Adjust the spatial resolution or map it as needed to ensure that the prediction results can effectively correspond to the actual area.

[0104] Step 3.3 includes:

[0105] Step 3.3.1 Before inputting the remote sensing image into the trained model, it is preferred to standardize the input data to ensure that the input data has the same scale and range as the data used in the training process.

[0106] Step 3.3.2: For multispectral or hyperspectral remote sensing imagery, it is preferred to fuse features from different bands to improve the model's recognition capabilities. Infrared bands can be used to enhance the identification of solid waste landfills, or vegetation indices such as NDVI (Normalized Difference Vegetation Index) can be used as additional input features to improve the distinction between solid waste landfills and their surroundings.

[0107] Figure 3This diagram shows the results of solid waste landfill identification in the Taizhou National Groundwater Pollution Prevention and Control Pilot Area. By leveraging the trained object recognition model and combining it with the image selection, preprocessing, and inference methods described above, the model achieved efficient identification of solid waste landfills throughout Taizhou. The model achieved an overall accuracy of 85% in practical applications, demonstrating its high reliability and adaptability for identifying solid waste landfills in complex environments.

[0108] When analyzing remote sensing imagery covering 5,700 square kilometers of Taizhou City, high-quality, clear, cloudless remote sensing imagery was selected based on the geographic scope and mission requirements. Resolution matching and preprocessing were then performed to address regional characteristics. The preprocessed remote sensing imagery was then fed into a trained model for large-scale inference, ultimately successfully identifying a total of 499 solid waste landfills within the city. The classification results showed that these included 151 general industrial solid waste sites, 41 domestic waste sites, and 307 construction waste sites.

[0109] Through the model's efficient reasoning capabilities and adaptability to multi-source data, the solid waste sites in the images were successfully distinguished from the natural background and other buildings, achieving full-area monitoring of Taizhou's solid waste landfills.

[0110] Identification results show that construction waste dumps are the primary source of solid waste landfills in Taizhou, accounting for over 60%, followed by general industrial solid waste and domestic waste dumps. This research provides important data support for the standardized management of solid waste landfills and the prevention and control of groundwater pollution in Taizhou, effectively assisting environmental regulators in carrying out targeted remediation efforts. Furthermore, this study validates the efficiency and practicality of the target recognition model for identifying solid waste sites across large regions, laying the foundation for its future application in other regions.

[0111] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for identifying urban-scale solid waste landfills based on an artificial intelligence target recognition algorithm, characterized in that: The following steps are involved: Step 1: Use multi-source data to construct a multi-source remote sensing image database of urban-scale solid waste landfill sites; Step 2: Improve the YOLOv8 model to adapt to remote sensing image recognition tasks and use the data augmentation data to train a city-scale solid waste landfill recognition model; Step 3: Acquire and process visible light remote sensing images of the target area and use the trained target recognition model for inference; Step 4: Sample the model identification results and verify the results through field surveys.

2. The urban-scale solid waste landfill identification method based on artificial intelligence target recognition algorithm according to claim 1 is characterized in that: The step 1 specifically includes: Obtain high-confidence data through public datasets and government-issued solid waste location lists; Select satellite image sources based on the geographic characteristics of the target area and the properties of the solid waste landfill, including spatial resolution, spectral resolution, and revisit period; The solid waste location information is interpreted into corresponding remote sensing image data through latitude and longitude, and a dynamic database is constructed by combining multi-temporal remote sensing images.

3. The urban-scale solid waste landfill identification method based on artificial intelligence target recognition algorithm according to claim 1 is characterized in that: The step 2 specifically includes: Step 2.1: By embedding the LSKA module in the YOLOv8 model SPPF and adding the CCFM module at the neck of the model, the model's multi-scale feature aggregation capability is improved; Step 2.2: Use data augmentation methods to augment the remote sensing images in the dataset. For the enhancement of remote sensing images, a data augmentation method combining Mix up, Cut out, Mosaic and random dithering is used to simulate image changes under different environmental backgrounds. Step 2.3: Use transfer learning to train the recognition model based on the pre-trained weights to ensure the stability and convergence speed of the training process. During training, select the loss function and optimizer suitable for the remote sensing image classification task, and set the hyperparameters of the training process in a targeted manner.

4. The urban-scale solid waste landfill identification method based on artificial intelligence target recognition algorithm according to claim 3 is characterized in that: The step 2.1 specifically includes the following steps: Step 2.1.1: The LSKA module decomposes the two-dimensional depth convolution kernel into cascaded horizontal and vertical one-dimensional convolution kernels, processing the horizontal and vertical features of the feature map respectively. On this basis, spatial dilation convolution with different dilation rates is used to process in both horizontal and vertical directions to increase the receptive field and further extract features: AND C =In 1×1 *WITH C Where Z C The feature map F is obtained by using a 1×(2d-1) vertical depth convolution kernel and a (2d-1)×1 horizontal depth convolution kernel. C The characteristics of the vertical and horizontal directions are then divided into and The convolutional neural network of the deep convolutional neural network is used to compensate for the grid effect of the deep convolutional neural network. On this basis, a 1×1 convolution kernel is used to perform channel compression to obtain the attention map A. C ,Finally, calculate the attention map A C and feature map F C The final characteristic map of the Hadamard product; Step 2.1.2: Embed the CCFM module in the YOLOv8 neck to improve the quality of feature fusion. Use two layers of cross-scale convolution to extract the features of the feature pyramid fusion. Then use the horizontal and vertical attention mechanism blocks with a size of 1 / 16 of the extracted features to compress and recalibrate the extracted features. Step 2.1.3: Change the YOLOv8 head structure to a dynamic head structure, and use attention from three dimensions: scale perception, spatial perception, and task perception to improve feature perception capabilities: W(F)=π C (p S (p L (F)·F)·F)·F p C (F)·F=max(α 1 (F)·F c +b 1 (F),a 2 (F)·F c +b 2 (F)) Where, π L (F) is scale-aware attention, which uses 1×1 convolution for feature mapping and then uses sigmoid function activation as the attention value; π S (F) is spatial perception attention, which uses variable convolution to perform sparse attention calculation, and then aggregates different layer features at the same location, so that the model always pays attention to the recognizable areas coexisting between different layers; π C (F) is task-aware attention, which uses hyperfunctions to control the activation values ​​of different channel features, summarizes different features of the same object, and improves task perception capabilities.

5. The urban-scale solid waste landfill identification method based on artificial intelligence target recognition algorithm according to claim 3 is characterized in that: The step 2.3 specifically includes the following steps: Step 2.3.1: Use downsampling and oversampling methods to divide the dataset into a balanced training set, a validation set, and a test set to ensure that the sample class distribution in each dataset is consistent; Step 2.3.2: Use the pre-trained YOLOv8 weights as the initial weights for transfer learning to train the model; Step 2.3.3: The solid waste landfill identification task is a classification task, so the loss function for model training is the cross entropy loss that describes the difference between the prediction and the label: Where, is the model prediction classification result, y i is the actual label, and the cross entropy describes the difference between the two distributions. The optimizer is used to minimize the loss function so that the model prediction gradually approaches the actual label; Step 2.3.4: When training the model, the optimizer selects the Adam optimizer, which dynamically adjusts the learning rate through first-order moment and second-order moment estimation: m t =β1·m t-1 +(1-β1)·g t Where θ t+1 represents the updated parameters, θ t represents the currently updated parameter, α is the learning rate, is the resulting bias-corrected first-order moment estimate of the gradient, is the bias-corrected gradient second-order moment estimate, ε is a very small constant to ensure zero stability, m t is the current momentum, g t is the current gradient, β1 is the decay factor, v t is the current second-order moment, is the square of the current gradient, β2 is the second-order moment attenuation factor. When the momentum and second-order moment deviations are corrected, as the iteration proceeds, the attenuation degree gradually approaches 1, and the correction effect on the momentum and second-order moment becomes weaker and weaker, that is, the closer the gradient is to the current step, the greater the influence on the current gradient.

6. The urban scale solid waste landfill identification method based on artificial intelligence target recognition algorithm according to claim 1 is characterized in that: The step 3 specifically includes: Step 3.1: Select remote sensing images that meet the timestamp, resolution, and quality requirements based on the mission requirements; Step 3.2: Perform radiometric, atmospheric, and geometric corrections on the image. Step 3.3: Input the pre-processed remote sensing image data into the trained recognition model for inference to obtain the predicted results of the solid waste landfill site. Adjust the spatial resolution or map it as needed to ensure that the predicted results effectively correspond to the actual area. Step 3.4: Post-process the results of model inference, including noise removal, boundary smoothing, and region merging.

7. The method for identifying urban-scale solid waste landfills based on an artificial intelligence target recognition algorithm according to claim 6, characterized in that: The reasoning process of step 3.3 includes: The input image is normalized to the range [0,1]. The band characteristics of the multispectral image are integrated and combined with the NDVI index to distinguish the solid waste dump from the vegetation background. A sliding window multi-scale inference method is used to fuse prediction results of different scales.

8. The method for identifying urban-scale solid waste landfills based on an artificial intelligence target recognition algorithm according to claim 1, characterized in that: After verification, an actual material map is compiled; the actual material map includes the route of field investigation and verification, the location and number of observation points, the location and number of specimen and sample collection, and the expression method is represented by the actual attribute classification code, classification code notes and map name, legend, and scale.

Citation Information

Patent Citations

  • Road defect target detection method and system based on SCDD-YOLO network

    CN118115444A

  • Denoising diffusion model-based waterlogging detection method and device under urban monitoring perspective

    CN118470659A

  • Face detection method and device based on YOLOv8 target detection model

    CN118470767A

  • Lightweight automatic driving target detection method based on YOLOV5

    CN118918557A

  • YOLOv8 improved model training method and ship detection method

    CN119131580A

Cited By

  • Solid waste storage yard remote sensing automatic identification method and system

    CN122156980A