Naked ground dust source monitoring method, device and system based on high-resolution satellite

Abstract

The invention discloses a naked ground dust source monitoring method, device and system based on a high-resolution satellite, and relates to the field of remote sensing monitoring, artificial intelligence image recognition and atmospheric environment models. The method comprises the following steps: acquiring a satellite image, training a machine learning algorithm through manual annotation in combination with transfer learning and online learning, preprocessing the image, identifying a dust source by using a multi-task learning framework and a semantic segmentation unit, and completing land typing and construction stage judgment. The device is composed of a high-resolution satellite data receiving end, a data management platform and the like, and the system comprises a satellite image receiving module, an algorithm training module and the like. The scheme can efficiently and accurately identify the flying dust source in the bare place, adapts to scene change in real time, forms closed-loop management from monitoring to early warning to law enforcement, and provides accurate data support for flying dust pollution prevention and control.

Description

Technical Field

[0001] This invention relates to the field of remote sensing monitoring and atmospheric environmental protection applications, and in particular to a method, device and system for monitoring bare ground dust sources based on high-resolution satellites. Background Technology

[0002] In the fields of remote sensing monitoring and atmospheric environmental protection, effective monitoring of dust sources from exposed ground has always been a crucial aspect of urban air quality management. With the acceleration of urbanization, the number and distribution of dust pollution sources such as construction areas and exposed ground are constantly expanding, severely impacting air quality. Traditional dust source monitoring methods mainly rely on manual on-site surveys and periodic testing, which are not only inefficient but also difficult to achieve comprehensive and real-time monitoring.

[0003] Traditional methods for monitoring dust sources from exposed land rely heavily on manual on-site surveys and periodic inspections, resulting in low efficiency. Furthermore, the algorithms lack effective techniques to accelerate convergence during training, often requiring large amounts of labeled data and struggling to update the model in real-time based on new data. This leads to poor adaptability to dynamic changes in exposed land types and construction phases. In terms of image analysis, it cannot simultaneously and efficiently handle exposed land type classification and construction phase analysis tasks, making accurate identification and boundary delineation of dust sources difficult. During dynamic monitoring, the lack of integration with atmospheric diffusion models and real-time meteorological data prevents the scientific simulation of dust diffusion trajectories. It also fails to automatically trigger timely warnings when pollutant concentrations exceed standards, thus failing to meet the requirements of remote sensing monitoring and atmospheric environmental protection applications. Therefore, this paper proposes a method, device, and system for monitoring dust sources from exposed land based on high-resolution satellite imagery. Summary of the Invention

[0004] This invention provides the following technical solution: a method for monitoring bare ground dust sources based on high-resolution satellites, comprising the following steps: S1 satellite imagery acquisition: Utilize satellite imagery to obtain information on urban bare land and construction areas; S2 algorithm training: The bare land was manually classified and the corresponding construction stage was determined. The results were combined with image data and input into the machine learning algorithm. At the same time, transfer learning technology and online learning mechanism were used to train the machine learning algorithm. S3 machine learning algorithm processing: First, the image data acquired in step S1 is preprocessed, including radiometric correction, geometric correction and noise filtering. Then, the preprocessed satellite image is analyzed and fixed dust emission sources such as construction bare land, garbage dumps, unpaved and ungreened bare land are identified through the multi-task learning framework and semantic segmentation unit inside the machine learning algorithm. S4 Land Classification and Construction Stage Determination: The identified land is categorized into types, including construction land, garbage dumps, unpaved land, ungreened land, land near water, and land under construction for roads and bridges, and the construction stage of different types of land is further determined. S5 Dynamic Monitoring and Management: Based on the trained algorithm, the monitoring area is continuously monitored periodically. The dust source diffusion model uses the dust source identification results of bare land as the input source, and combines real-time meteorological data to simulate the dust diffusion trajectory. When the predicted concentration exceeds the manually set threshold, the system automatically pushes early warning information including the scope of pollution impact to the law enforcement terminal.

[0005] This invention provides a bare land dust source monitoring device based on high-resolution satellites, employing the aforementioned bare land dust source monitoring method based on high-resolution satellites, comprising: The system includes a high-resolution satellite data receiver, a data management platform, an enforcement and supervision terminal, a data storage device, and a user interface. The high-resolution satellite data receiver is used to receive image data of the monitoring area, and the data management platform is used to store and manage the generated bare land ledger and dynamic monitoring data. The law enforcement and supervision terminal is used to receive monitoring data and assist law enforcement personnel in supervising dust sources from exposed land and tracking the rectification of problems. The data storage device is used to store the raw satellite image data, intermediate algorithm results, monitoring conclusions and early warning records in a structured manner. The user interface is used for intuitive interaction.

[0006] This invention provides a bare land dust source monitoring system based on high-resolution satellites. The bare land dust source monitoring device based on the aforementioned high-resolution satellites includes: The system includes a satellite image receiving module, a machine learning algorithm training module, a machine learning algorithm processing module, a land classification and construction stage determination module, a dynamic monitoring and management module, and an atmospheric diffusion simulation and early warning module. The satellite image receiving module is used to receive image data of the monitoring area transmitted by satellite. The machine learning algorithm training module is used to manually determine the type of bare land and the corresponding construction stage, and combine the determination results with image data, which are then input into the machine learning algorithm for training. The machine learning algorithm training module is equipped with transfer learning technology and online learning mechanism. The machine learning algorithm processing module integrates a multi-task learning framework and a semantic segmentation unit. The land classification and construction stage determination module is used to classify the identified land. The classification includes construction bare land, garbage dump, unpaved bare land, ungreened bare land, pipeline, waterfront and road and bridge construction types. The dynamic monitoring and management module is used to perform periodic and continuous monitoring of the monitoring area based on the trained algorithm. The atmospheric diffusion simulation and early warning module is equipped with an integrated dust source diffusion model.

[0007] Preferably, in step S1, images are acquired during clear weather periods when the cloud cover is less than 10%. Furthermore, based on the seasonal variation of urban dust pollution, the frequency of image acquisition is increased during the peak dust seasons in spring and autumn. According to the current status of high-resolution satellites, this can be increased to once or twice a month at this stage.

[0008] Preferably, in step S2, when manually determining the bare land type and construction stage, the texture features, geometric shapes, and surrounding environmental elements in the image are comprehensively judged according to the pre-established annotation specifications, and the annotation data must be cross-verified by at least two professionals.

[0009] Preferably, the transfer learning technique in step S2 is based on a deep learning model pre-trained on a large-scale natural scene remote sensing image dataset, and the parameters of the model's classification layer and some convolutional layers are fine-tuned to meet the specific needs of the bare land dust source monitoring task.

[0010] Preferably, the online learning mechanism in step S2 involves receiving new manually labeled data in real time during the operation of the monitoring system and iteratively updating the trained machine learning algorithm using incremental learning.

[0011] Preferably, in step S3, radiometric correction calibrates the spectral radiance of the image by establishing a sensor response model; geometric correction uses polynomial transformation combined with ground control points to eliminate deformation caused by terrain undulations and satellite attitude; and noise filtering uses median filtering and Gaussian filtering algorithms to remove random noise from the image. The multi-task learning framework within the machine learning algorithm treats bare land type identification and construction stage determination as two parallel tasks, combining a shared feature extraction layer and setting a task-specific output layer to achieve joint analysis of satellite imagery. The semantic segmentation unit adopts a deep learning network with an encoder and decoder architecture. By extracting high-level semantic features of the image in the encoder and gradually restoring the spatial resolution of the image in the decoder, pixel-level segmentation of construction bare land and garbage dump areas is achieved.

[0012] Preferably, the satellite image receiving module is equipped with a multi-source data fusion function, which is used to fuse satellite image data with other types of data, including meteorological data, topographic data, traffic data, and socio-economic data.

[0013] Preferably, the data storage device integrates a data management submodule, a data lineage tracking submodule, and an intelligent caching submodule, and the user interface is configured with an intelligent search engine, a comparison analysis tool, and an access control component.

[0014] In summary, compared with the prior art, the present invention provides a method, device, and system for monitoring bare ground dust sources based on high-resolution satellites, which has the following beneficial effects: 1. This invention acquires satellite imagery as a data source, enabling comprehensive and rapid acquisition of detailed information on dust sources from bare land. Compared to traditional manual monitoring methods, this improves the accuracy and efficiency of monitoring, achieving all-round, blind-spot-free monitoring of dust sources from bare land. This effectively avoids omissions and errors that may exist in traditional monitoring methods. Furthermore, by processing the images using machine learning algorithms, it achieves rapid and accurate identification of fixed dust emission sources such as construction bare land, garbage dumps, and unpaved and ungreened bare land. Simultaneously, by combining transfer learning technology, it utilizes pre-trained models to accelerate the convergence speed of machine learning algorithms, reducing the amount of labeled data required for training. In addition, by combining an online learning mechanism, it can receive new labeled data in real time and iteratively update the model, allowing the algorithm to continuously adapt to the dynamic changes in bare land type and construction stage, improving the timeliness and accuracy of monitoring. At the same time, by combining a multi-task learning framework with a semantic segmentation unit, the device can simultaneously handle bare land type classification and construction stage analysis tasks, achieving efficient identification and boundary delineation of dust sources in satellite imagery, providing accurate data support for subsequent classification and determination. 2. This invention enables real-time monitoring of changes in dust sources from exposed land through periodic and continuous monitoring. This system provides timely and accurate data support for pollution prevention and control, helping environmental protection departments to take timely measures to control dust pollution. When the predicted concentration exceeds air quality standards, the system can automatically push early warning information including the scope of pollution impact to law enforcement terminals, ensuring that relevant departments can respond quickly and effectively curb the spread of dust pollution. Furthermore, through the fusion analysis of dust source diffusion models and real-time meteorological data, the system achieves dynamic simulation of dust diffusion trajectories. When pollutant concentrations exceed thresholds, an early warning is automatically triggered, providing law enforcement departments with a scientific basis for the scope of pollution impact. This forms a closed-loop management system from monitoring to early warning to enforcement, thereby assisting law enforcement departments in responding quickly and shortening the pollution treatment cycle. Attached Figure Description

[0015] Figure 1 This is a flowchart of the method of the present invention.

[0016] Figure 2 This is a diagram of the device of the present invention.

[0017] Figure 3 This is a system diagram of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 This invention provides a technical solution: a method for monitoring bare ground dust sources based on high-resolution satellites, comprising the following steps: S1 satellite imagery acquisition: Information on urban bare land and construction areas will be obtained using satellite imagery. Imagery will be acquired during clear weather periods with cloud cover below 10%. Based on the seasonal variation of urban dust pollution, the frequency of image acquisition will be increased during the peak dust seasons of spring and autumn. According to the current status of high-resolution satellites, this can be increased to once or twice a month at this stage.

[0020] S2 algorithm training: The method involves manually determining the type of bare land and the corresponding construction stage, combining the determination results with image data, and inputting them into a machine learning algorithm. Simultaneously, transfer learning technology and online learning mechanisms are used to train the machine learning algorithm. When manually determining the type of bare land and the construction stage, the texture features, geometric shapes, and surrounding environmental elements in the image are comprehensively judged according to the pre-defined annotation specifications. The annotation data must be cross-verified by at least two professionals. The specific implementation of the above method is as follows: In the initial stage of algorithm training, a deep learning model pre-trained on a large-scale natural scene remote sensing image dataset is selected as the basic framework. This pre-trained model already has the ability to extract general visual features from massive amounts of natural images. For the specific task requirements of monitoring dust sources from bare land, technicians will fine-tune the parameters of this basic model. This process focuses on adjusting the weights of neurons in the model's classification layer, while also refining the filter parameters of some convolutional layers, enabling the model to more accurately identify typical dust source characteristics such as construction sites and garbage dumps. The manual annotation process strictly adheres to pre-defined annotation standards. Annotators must systematically analyze multiple elements in satellite imagery: in terms of texture features, they need to distinguish different roughness levels of exposed surfaces; construction areas often exhibit regular geometric textures, while natural bare land has irregular granular textures; geometric shape analysis requires consideration of spatial scale; large construction sites typically have rectangular or strip-shaped outlines, while small piles of materials appear as dots; surrounding environment analysis requires examining the spatial relationship between the site and surrounding roads and buildings; for example, waterfront areas require comprehensive judgment based on the distribution characteristics of water bodies. After each annotator works independently, the system randomly assigns at least two professionals to cross-validate the same set of images. When inconsistencies arise in the annotation results, a final annotation conclusion is reached through an expert consultation mechanism to ensure the quality of the training data. An online learning mechanism is implemented throughout the entire operation cycle of the monitoring system. When new types of bare land emerge or construction phases change in the monitored area, the system receives newly labeled data that has been manually verified in real time. This incremental data is input into the training system through a dynamic loading mechanism, and the model is iteratively updated using a flexible batch processing strategy. The update process employs a progressive parameter adjustment strategy, which preserves the model's ability to recognize historical features while quickly adapting to newly emerging land cover types. Through this continuous learning mechanism, the monitoring system maintains a keen awareness of the dynamic changes in bare land dust sources. The collaborative work between the multi-task learning framework and the semantic segmentation unit further enhances the algorithm's processing efficiency. In the image preprocessing stage, the radiometric correction module eliminates the influence of atmospheric scattering on the spectral signal by establishing a sensor radiometric calibration model; the geometric correction module uses a multinomial transformation algorithm combined with ground control point coordinates to achieve accurate georegistration of the image; and the noise filtering module employs adaptive median filtering technology to effectively suppress random noise while preserving feature edge information. After the preprocessed image data is input into the multi-task learning framework, the shared feature extraction layer simultaneously completes feature mapping for bare land type classification and construction stage determination, while the task-specific output layers generate pixel-level semantic segmentation results and stage attribute annotations, respectively. This joint analysis mode significantly improves the system's ability to analyze complex scenes, providing accurate basic data support for subsequent dust diffusion simulation and pollution early warning. Transfer learning technology uses a deep learning model pre-trained on a large-scale natural scene remote sensing image dataset as a basis. To meet the specific needs of the bare land dust source monitoring task, the parameters of the model's classification layer and some convolutional layers are fine-tuned. The specific implementation process of the above method is as follows: In monitoring dust sources from bare land, the implementation of transfer learning techniques begins with the selection of pre-trained models. Technicians select deep learning models pre-trained on large-scale remote sensing image datasets of natural scenes from publicly available model libraries. These models typically employ convolutional neural network architectures, such as the ResNet series or EfficientNet. Selection criteria focus on the model's performance in remote sensing image feature extraction and the suitability of its underlying architecture for the current monitoring task. After selecting the base model, the parameter fine-tuning phase begins. The implementers will freeze all parameters in the pre-trained model except for the final classification layer and specific convolutional layers. Replacing the classification layer is a crucial step, requiring the output nodes to be redesigned based on the specific category requirements of bare land dust source monitoring. For example, when the monitoring task includes six typical dust sources such as construction bare land, garbage dumps, and unpaved ground, the number of neurons in the classification layer will be set to six accordingly, and a mapping relationship from high-level semantic features to specific categories will be established. During the convolutional layer fine-tuning phase, technicians select 2-3 convolutional blocks near the end of the model to unfreeze their parameters. These convolutional layers are responsible for extracting high-level abstract features. By appropriately adjusting their filter weights, the model can capture the unique texture patterns and spatial distribution features of bare ground dust sources. The fine-tuning process employs a differentiated learning rate strategy: a higher learning rate is set for newly added classification layers to accelerate convergence, while a lower learning rate is used for unfrozen convolutional layers to preserve pre-trained features. During the training data preparation phase, an image tile library containing typical dust source scenes needs to be constructed. Each sample needs to be labeled with the type of bare land, construction stage, and spatial coordinate information, ensuring a balanced proportion of samples from each category. The training process adopts a phased optimization strategy: initially, only the classification layer parameters are trained; after the loss function stabilizes, convolutional layers are gradually introduced for fine-tuning; finally, the entire model is jointly optimized. The validation phase employs cross-validation, dividing the labeled data into training, validation, and test sets. By monitoring metrics such as accuracy and F1 score on the validation set, the selection of fine-tuning layers and the learning rate parameters are dynamically adjusted. The transfer learning process is complete when the model achieves the preset recognition accuracy on the test set. The online learning mechanism receives new manually labeled data in real time during the monitoring system's operation and iteratively updates the trained machine learning algorithm using incremental learning. Specifically, during the continuous operation of the monitoring system, the online learning mechanism receives new labeled samples from the labeling workstation in real time through a dynamic data stream interface. These samples are created by a professional labeling team based on the latest high-resolution satellite imagery, and the labeled content strictly follows a predefined classification system, including structured information such as bare land type and construction stage. After new data is received, the system first activates the data verification module to ensure the reliability of the labeled data through spatial consistency checks and temporal continuity verification. For data that passes verification, the system adopts an incremental feature extraction strategy, calculating only the depth feature vectors of newly added samples to avoid repeatedly processing historical data. In the model update phase, the online learning mechanism employs an elastic parameter update strategy. For newly emerging bare land types or construction techniques, the system dynamically expands the neuron nodes in the classification layer and uses an adaptive learning rate algorithm to initialize the parameters of the new nodes. For updates to samples of existing categories, a moving average mechanism is used to adjust the weight parameters of the corresponding category, preserving historical knowledge while reflecting data evolution trends. The incremental learning process employs mini-batch gradient descent, using only the most recently received N labeled samples in each iteration. To prevent catastrophic forgetting, the system includes a memory retention module that periodically replays representative historical samples to maintain the model's ability to recognize earlier categories. The learning rate uses a dynamic decay strategy, gradually decreasing with each update to ensure model convergence stability. The updated model needs to pass multiple verification checkpoints, including accuracy monitoring on the retained test set, recall evaluation for new types of samples, and comparative testing with historical versions of the model. When the model performance metrics reach the preset update threshold, the system automatically triggers a hot update procedure to seamlessly switch to the new model, ensuring the continuity of monitoring services. S3 machine learning algorithm processing: First, the image data acquired in step S1 is preprocessed, including radiometric correction, geometric correction and noise filtering. Then, the preprocessed satellite image is analyzed and fixed dust emission sources such as construction bare land, garbage dumps, unpaved and ungreened bare land are identified through the multi-task learning framework and semantic segmentation unit inside the machine learning algorithm. Radiometric correction calibrates the spectral radiance of the image by establishing a sensor response model. Geometric correction uses polynomial transformation combined with ground control points to eliminate deformation caused by terrain undulations and satellite attitude. Noise filtering uses median filtering and Gaussian filtering algorithms to remove random noise from the image. The multi-task learning framework within the machine learning algorithm treats bare land type identification and construction stage determination as two parallel tasks, combining a shared feature extraction layer and setting a task-specific output layer to achieve joint analysis of satellite imagery. The semantic segmentation unit adopts a deep learning network with an encoder and decoder architecture. By extracting high-level semantic features of the image in the encoder, the spatial resolution of the image is gradually restored in the decoder, achieving pixel-level segmentation of construction bare land and garbage dump areas. S4 Land Classification and Construction Stage Determination: The identified land is categorized into types, including construction land, garbage dumps, unpaved land, ungreened land, pipelines, waterfront areas, and road and bridge construction, and the construction stage of different types of land is further determined. S5 Dynamic Monitoring and Management: Based on the trained algorithm, the monitoring area is continuously monitored periodically. The dust source diffusion model uses the dust source identification results of bare land as the input source, and combines real-time meteorological data to simulate the dust diffusion trajectory. When the predicted concentration exceeds the manually set threshold, the system automatically pushes early warning information including the scope of pollution impact to the law enforcement terminal.

[0021] Please see Figure 2 This invention provides a bare land dust source monitoring device based on high-resolution satellites, employing the aforementioned bare land dust source monitoring method based on high-resolution satellites, comprising: The system includes a high-resolution satellite data receiver, a data management platform, an enforcement and supervision terminal, a data storage device, and a user interface. The high-resolution satellite data receiver is used to receive image data of the monitoring area, and the data management platform is used to store and manage the generated bare land ledger and dynamic monitoring data. The law enforcement and supervision terminal is used to receive monitoring data and assist law enforcement personnel in supervising dust sources from exposed land and tracking the rectification of problems. The data storage device is used to store the raw satellite image data, intermediate algorithm results, monitoring conclusions and early warning records in a structured manner. The user interface is used for intuitive interaction. The data storage device integrates a data management submodule, a data lineage tracking submodule, and an intelligent caching submodule. The user interface is equipped with an intelligent search engine, a comparison analysis tool, and an access control component. Please see Figure 3 This invention provides a bare land dust source monitoring system based on high-resolution satellites. The bare land dust source monitoring device based on the aforementioned high-resolution satellites includes: The system includes a satellite image receiving module, a machine learning algorithm training module, a machine learning algorithm processing module, a land classification and construction stage determination module, a dynamic monitoring and management module, and an atmospheric diffusion simulation and early warning module. The satellite image receiving module is used to receive image data of the monitoring area transmitted by satellite. The machine learning algorithm training module is used to manually determine the type of bare land and the corresponding construction stage, and combine the determination results with image data to input the machine learning algorithm for training. The machine learning algorithm training module is equipped with transfer learning technology and online learning mechanism. The machine learning algorithm processing module integrates a multi-task learning framework and a semantic segmentation unit. The land classification and construction stage determination module is used to classify the identified land, including construction bare land, garbage dumps, unpaved bare land, ungreened bare land, pipelines, waterfront areas, and road and bridge construction types. The dynamic monitoring and management module is used to conduct periodic and continuous monitoring of the monitoring area based on the trained algorithm. The atmospheric diffusion simulation and early warning module has an integrated dust source diffusion model. The satellite image receiving module has a multi-source data fusion function, which is used to fuse satellite image data with other types of data, including meteorological data, topographic data, traffic data, and socio-economic data.

[0022] This solution acquires satellite imagery as a data source, enabling comprehensive and rapid acquisition of detailed information on dust sources from bare land. Compared to traditional manual monitoring methods, it improves the accuracy and efficiency of monitoring, achieving all-round, blind-spot-free monitoring of dust sources from bare land. This effectively avoids omissions and errors that may exist in traditional monitoring methods. Furthermore, by processing the images using machine learning algorithms, it achieves rapid and accurate identification of fixed dust emission sources such as construction bare land, garbage dumps, and unpaved and ungreened bare land. Simultaneously, by combining transfer learning technology, it utilizes pre-trained models to accelerate the convergence speed of machine learning algorithms, reducing the amount of labeled data required for training. In addition, by combining an online learning mechanism, it can receive new labeled data in real time and iteratively update the model, allowing the algorithm to continuously adapt to the dynamic changes in bare land type and construction stage, improving the timeliness and accuracy of monitoring. At the same time, by combining a multi-task learning framework with a semantic segmentation unit, the device can simultaneously handle bare land type classification and construction stage analysis tasks, achieving efficient identification and boundary delineation of dust sources in satellite imagery, providing accurate data support for subsequent classification and determination.

[0023] This solution enables real-time monitoring of changes in dust sources from exposed land through periodic and continuous monitoring. This system provides timely and accurate data support for pollution prevention and control, helping environmental protection departments to take timely measures to control dust pollution. When the predicted concentration exceeds air quality standards, the system can automatically push early warning information, including the scope of pollution impact, to enforcement terminals, ensuring that relevant departments can respond quickly and effectively curb the spread of dust pollution. Furthermore, through the fusion analysis of dust source diffusion models and real-time meteorological data, the system achieves dynamic simulation of dust diffusion trajectories. When pollutant concentrations exceed thresholds, an early warning is automatically triggered, providing enforcement departments with a scientific basis for the scope of pollution impact. This forms a closed-loop management system from monitoring to early warning to enforcement, thereby assisting enforcement departments in responding quickly and shortening the pollution treatment cycle.

[0024] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0025] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for monitoring bare land dust sources based on high-resolution satellites, characterized in that, Includes the following steps: S1 satellite imagery acquisition: Utilize satellite imagery to obtain information on urban bare land and construction areas; S2 algorithm training: The bare land was manually classified and the corresponding construction stage was determined. The results were combined with image data and input into the machine learning algorithm. At the same time, transfer learning technology and online learning mechanism were used to train the machine learning algorithm. S3 machine learning algorithm processing: First, the image data acquired in step S1 is preprocessed, including radiometric correction, geometric correction and noise filtering. Then, the preprocessed satellite image is analyzed and fixed dust emission sources such as construction bare land, garbage dumps, unpaved and ungreened bare land are identified through the multi-task learning framework and semantic segmentation unit inside the machine learning algorithm. S4 Land Classification and Construction Stage Determination: The identified land is categorized into types, including construction land, garbage dumps, unpaved land, ungreened land, land near water, and land under construction for roads and bridges, and the construction stage of different types of land is further determined. S5 Dynamic Monitoring and Management: Based on the trained algorithm, the monitoring area is continuously monitored periodically. The dust source diffusion model uses the dust source identification results of bare land as the input source, and combines real-time meteorological data to simulate the dust diffusion trajectory. When the predicted concentration exceeds the manually set threshold, the system automatically pushes early warning information including the scope of pollution impact to the law enforcement terminal.

2. The method for monitoring bare land dust sources based on high-resolution satellites according to claim 1, characterized in that: In step S1, images are acquired during clear weather periods when the cloud cover is less than 10%, and the image acquisition frequency is increased during the spring and autumn dust pollution seasons, based on the seasonal variation of urban dust pollution.

3. The method for monitoring bare land dust sources based on high-resolution satellites according to claim 1, characterized in that: In step S2, when manually determining the type of bare land and the construction stage, the texture features, geometric shapes, and surrounding environmental elements in the image are comprehensively judged according to the pre-established annotation specifications, and the annotation data must be cross-verified by at least two professionals.

4. The method for monitoring bare land dust sources based on high-resolution satellites according to claim 1, characterized in that: The transfer learning technique in step S2 is based on a deep learning model pre-trained on a large-scale natural scene remote sensing image dataset. The model's classification layer and some convolutional layers are fine-tuned to meet the specific needs of the bare land dust source monitoring task.

5. The method for monitoring bare land dust sources based on high-resolution satellites according to claim 1, characterized in that: The online learning mechanism in step S2 receives new manually labeled data in real time during the operation of the monitoring system and iteratively updates the trained machine learning algorithm by combining incremental learning.

6. The method for monitoring bare land dust sources based on high-resolution satellites according to claim 1, characterized in that: In step S3, radiometric correction calibrates the spectral radiance of the image by establishing a sensor response model. Geometric correction uses polynomial transformation combined with ground control points to eliminate deformation caused by terrain undulations and satellite attitude. Noise filtering uses median filtering and Gaussian filtering algorithms to remove random noise from the image. The multi-task learning framework within the machine learning algorithm treats bare land type identification and construction stage determination as two parallel tasks, combining a shared feature extraction layer and setting a task-specific output layer to achieve joint analysis of satellite imagery. The semantic segmentation unit adopts a deep learning network with an encoder and decoder architecture. By extracting high-level semantic features of the image in the encoder, the spatial resolution of the image is gradually restored in the decoder, achieving pixel-level segmentation of construction bare land and garbage dump areas.

7. A bare land dust source monitoring device based on high-resolution satellite, employing the bare land dust source monitoring method based on high-resolution satellite as described in any one of claims 1-6, characterized in that, include: The system includes a high-resolution satellite data receiver, a data management platform, an enforcement and supervision terminal, a data storage device, and a user interface. The high-resolution satellite data receiver is used to receive image data of the monitoring area, and the data management platform is used to store and manage the generated bare land ledger and dynamic monitoring data. The law enforcement and supervision terminal is used to receive monitoring data and assist law enforcement personnel in supervising dust sources from exposed land and tracking the rectification of problems. The data storage device is used to store the raw satellite image data, intermediate algorithm results, monitoring conclusions and early warning records in a structured manner. The user interface is used for intuitive interaction.

8. A bare ground dust source monitoring device based on high-resolution satellite as described in claim 7, characterized in that: The data storage device integrates a data management submodule, a data lineage tracking submodule, and an intelligent caching submodule. The user interface is equipped with an intelligent search engine, a comparison analysis tool, and an access control component.

9. A bare land dust source monitoring system based on high-resolution satellite, comprising a bare land dust source monitoring device based on high-resolution satellite as described in any one of claims 7-8, characterized in that, include: The system includes a satellite image receiving module, a machine learning algorithm training module, a machine learning algorithm processing module, a land classification and construction stage determination module, a dynamic monitoring and management module, and an atmospheric diffusion simulation and early warning module. The satellite image receiving module is used to receive image data of the monitoring area transmitted by satellite. The machine learning algorithm training module is used to manually determine the type of bare land and the corresponding construction stage, and combine the determination results with image data, which are then input into the machine learning algorithm for training. The machine learning algorithm training module is equipped with transfer learning technology and online learning mechanism. The machine learning algorithm processing module integrates a multi-task learning framework and a semantic segmentation unit. The land classification and construction stage determination module is used to classify the identified land. The classification includes construction bare land, garbage dump, unpaved bare land, ungreened bare land, water-adjacent land, and road and bridge construction types. The dynamic monitoring and management module is used to perform periodic and continuous monitoring of the monitoring area based on the trained algorithm. The atmospheric diffusion simulation and early warning module is equipped with an integrated dust source diffusion model.

10. A bare ground dust source monitoring system based on high-resolution satellites according to claim 9, characterized in that: The satellite image receiving module is equipped with a multi-source data fusion function, which is used to fuse satellite image data with other types of data, including meteorological data, topographic data, traffic data, and socio-economic data.

Images