A PM based on Sunflower satellite data and machine learning 10 Concentration Inversion Methods and Systems

Abstract

A PM concentration inversion method and system based on sun-synchronous satellite data and machine learning 10 The application relates to the fields of atmospheric environment remote sensing monitoring and artificial intelligence technology, and solves the problem that the prior art lacks real-time and spatial continuous quantitative inversion technology for the PM concentration of a target area. 10 The application comprises the following steps: S1, constructing a PM concentration dataset and preprocessing the same; S2, selecting satellite full-channel data, calculating a plurality of derived remote sensing features based on extracted original observation data, screening the plurality of derived remote sensing features by adopting a feature screening method, and obtaining a final feature set; S3, training an XGBoost model by adopting the final feature set of step S2; and S4, inputting the PM concentration dataset preprocessed in step S1 into the XGBoost model trained in step S3, and completing continuous real-time monitoring of the PM concentration of the target area. 10 10 10 The application has the advantages that the PM concentration inversion method and system based on sun-synchronous satellite data and machine learning can realize real-time and spatial continuous quantitative inversion of the PM concentration of a target area, and can provide a basis for the prevention and control of air pollution.​​

Description

Technical Field

[0001] This invention relates to the fields of atmospheric environment remote sensing monitoring and artificial intelligence technology, specifically to PM2.5 based on Himawari satellite data and machine learning. 10 Concentration inversion method and system. Background Technology

[0002] PM 10 Particulate matter (PM2.5) with a diameter ≤10μm is an important indicator for evaluating air quality. It can be used to characterize dust concentration and classify dust levels, which is of great significance for real-time monitoring of dusty weather. Ground monitoring stations can obtain high-precision PM2.5 concentration data. 10 Concentration data are available, but their spatial coverage is limited and their distribution is uneven, making it difficult to reflect PM levels over a large area. 10 Spatiotemporal distribution characteristics. Satellite remote sensing can provide continuous and large-scale atmospheric observation information, and has gradually become a research hotspot. For example, Hu Xiuqing et al. (2007) used channel brightness temperature difference characteristics to identify dust areas and used infrared channels to construct the Infrared Difference Dust Index (IDDI) to semi-quantitatively describe dust concentration; Xue Zhengzheng et al. (2022) used FY-2G multi-channel remote sensing data for dust identification. Such methods perform well in dust distribution identification and pollution level determination, but mainly focus on qualitative detection and are difficult to achieve ground-based PM2.5 detection. 10 Quantitative inversion of concentration. Furthermore, some studies indirectly characterize particulate matter concentration by inverting aerosol optical thickness (AOD). For example, She Lu (2018) conducted AOD inversion based on the Himawari satellite and used it for dust storm intensity assessment. However, AOD is an integral part of the atmospheric column and is not directly related to near-surface PM2.5 concentration. 10 The concentration of PM2.5 is influenced by multiple factors, including boundary layer height, relative humidity, and particle size distribution, making it still difficult to achieve PM2.5 concentration control. 10 Accurate quantitative inversion.

[0003] With the increase in data scale and computing power, statistical regression and machine learning methods have been gradually introduced into PM (Production Management). 10 Concentration estimation research. Tan et al. (2022) used multiple linear regression and various machine learning models to estimate PM2.5 concentration at station scales in the Tibetan Plateau region. 10 Concentration prediction; T r nk et al. (2023) used models such as XGBoost and combined them with ground-level emission inventories to construct PM2.5 concentrations. 10 The prediction model achieved high accuracy; Zhang Xiang (2025) used the XGBoos method to invert visibility during dusty weather. These studies demonstrate that machine learning methods have good performance in predicting particulate matter-related elements, but their research focuses primarily on concentration trends or predictions, lacking specific applications for PM2.5 prediction. 10 Its real-time, spatially continuous inversion capability.

[0004] Therefore, by combining multi-source remote sensing data with machine learning methods, PM2.5 levels in the study area can be determined. 10 Real-time, spatially continuous quantitative inversion of concentration is of great significance for improving the monitoring capabilities of dust pollution. Summary of the Invention

[0005] This invention solves the problem that existing technologies lack targeted PM in specific target areas. 10 The problem of real-time, spatially continuous quantitative inversion of concentration.

[0006] The present invention describes a PM based on Sunflower satellite data and machine learning. 10 The concentration inversion method includes the following steps: Step S1, construct PM 10 Concentration datasets were obtained and preprocessed. Step S2: Select satellite full-channel data, calculate multiple derived remote sensing features based on the extracted raw observations, and use a feature filtering method to filter the multiple derived remote sensing features to obtain the final feature set; Step S3: Train the XGBoost model using the final feature set from step S2; Step S4, process the PM preprocessed in step S1. 10 The concentration dataset is input into the XGBoost model trained in step S3 to complete the PM2.5 concentration analysis of the target region. 10 Continuous real-time monitoring of concentration.

[0007] Furthermore, in one embodiment of the present invention, in step S1, the construction of PM 10 The concentration dataset includes satellite remote sensing data and site monitoring data; Satellite remote sensing data is automatically connected to the JAXA P-Tree remote sensing data service platform via FTP protocol, and the satellite full disk NetCDF data for the target time period is downloaded in batches. The data is automatically converted to UTC time range according to the Beijing time folder, and the target file is located and downloaded to the local folder according to the remote directory structure. Site monitoring data is collected hourly for PM. 10 The observation data, along with the station name, longitude, and latitude information, is serialized by date and output as a NetCDF format file.

[0008] Furthermore, in one embodiment of the present invention, the preprocessing in step S1 specifically includes: Samples at or above the severe pollution level were selected, and outlier samples with excessively large spatial matching distances were removed. PM2.5 levels were then analyzed. 10 The concentration was subjected to a logarithmic transformation.

[0009] Furthermore, in one embodiment of the present invention, in step S2, the step of using a feature filtering method to filter multiple derived remote sensing features to obtain a final feature set specifically involves: Feature selection methods include MI mutual information analysis, SHAP value analysis, XGBoost feature importance analysis and ablation experiments. The importance ranking of multiple derived remote sensing features is obtained using feature selection methods. The screening results of the feature screening method are normalized, a comprehensive feature consistency score index is constructed, and the comprehensive feature contribution is obtained by weighted fusion of the screening results of the feature screening method. At the same time, a ranking dispersion penalty mechanism is introduced to reduce the comprehensive score of features with large differences in ranking among feature selection methods; By combining correlation analysis to remove redundant features with high correlation, the final feature set is obtained.

[0010] Furthermore, in one embodiment of the present invention, the weights corresponding to the MI mutual information analysis, SHAP value analysis, XGBoost feature importance analysis, and ablation experiment are 0.25, 0.35, 0.30, and 0.10, respectively.

[0011] The present invention describes a PM based on Sunflower satellite data and machine learning. 10 The concentration inversion system is a PM2.5 system based on the above-described method, using Himawari satellite data and machine learning. 10 The concentration inversion method includes the following modules: Module S1, build PM 10 Concentration datasets were obtained and preprocessed. Module S2 selects full-channel satellite data, calculates multiple derived remote sensing features based on the extracted raw observations, and uses a feature filtering method to filter the multiple derived remote sensing features to obtain the final feature set; Module S3 uses the final feature set from Module S2 to train the XGBoost model; Module S4 processes the PM preprocessed by module S1. 10 The concentration dataset is input into the XGBoost model trained by module S3 to complete the PM2.5 concentration analysis of the target region. 10 Continuous real-time monitoring of concentration.

[0012] The electronic device of the present invention includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus. Memory, used to store computer programs; A processor, when executing a program stored in memory, implements any of the methods described above.

[0013] The present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any of the methods described above.

[0014] This invention solves the problem that existing technologies lack targeted PM in specific target areas. 10 The challenges of real-time, spatially continuous quantitative inversion of concentration. Specific beneficial effects include: The present invention describes a PM based on Sunflower satellite data and machine learning. 10 To address the technical limitations of existing technologies, this concentration inversion method constructs a high-quality dust sample dataset based on multi-source data and performs preprocessing. It then extracts key influencing factors through various feature engineering methods and introduces an XGBoost model to establish PM concentration retrieval. 10 A concentration inversion model, combined with multiple indicators to evaluate the inversion effect, achieves PM2.5 concentration inversion. 10 A method for real-time, continuous, and quantitative inversion of concentration. Attached Figure Description

[0015] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 It is the PM based on Himawari satellite data and machine learning described in Implementation Method 1. 10 Concentration inversion flowchart; Figure 2 This is the SHAP value analysis diagram of the feature quantity described in Implementation Method 1; Figure 3 It is the feature quantity correlation analysis diagram described in Implementation Method 1; Figure 4 It is the XGBoost feature importance graph described in Implementation Method 1; Figure 5 This is a diagram showing the experimental results of the characteristic quantity ablation described in Implementation Method 1; Figure 6 This is a single-moment dust detection image as described in the embodiment; Figure 7 The PM described in the embodiment 10 Inversion value plot; Figure 8 The dust event process described in the embodiment involves dust detection (lines 1 and 3) and PM2.5. 10 Comparison chart of inversion values ​​(rows 2 and 4). Detailed Implementation

[0016] Various embodiments of the present invention will now be clearly and completely described with reference to the accompanying drawings. The embodiments described with reference to the drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0017] Implementation Method 1: Sandstorm weather is characterized by its sudden onset, rapid transport speed, and wide impact range. PM2.5 10 Concentrations often change rapidly within a short period of time.

[0018] Chinese patent CN106931583A discloses an "Indoor Ambient Air Purification System and Method". By selecting different filters and sensors, the system can purify existing and future pollutants simultaneously, and adds the detection and control of indoor oxygen concentration. However, it is a fixed-point quantitative measurement in a small indoor area and cannot monitor the concentration in a large area.

[0019] Chinese patent CN120255025A discloses "a method for joint monitoring of sandstorm weather by space and ground". Based on the automatic monitoring network of ground ambient air quality, it calculates the abnormal coefficient of particulate matter concentration change at each monitoring station. It mainly relies on the stations to identify sandstorms and classify them into semi-quantitative levels (sandstorm intensity index), but it cannot achieve quantitative inversion.

[0020] Therefore, existing monitoring methods relying on ground-based monitoring stations or sensors are limited by station distribution and data update cycles, and can only provide observation results from discrete locations, making it difficult to continuously reflect PM levels over large areas. 10 Spatial distribution and evolution of PM2.5 concentration. When dust storms develop rapidly, a time lag can easily occur between monitoring results and the actual pollution state, resulting in a lack of timely and reliable data support for air quality warnings, pollution process analysis, and environmental emergency responses, thus affecting timeliness. Therefore, achieving PM2.5 concentration monitoring is crucial. 10 Real-time continuous monitoring of PM2.5 concentration is of great significance. Meanwhile, PM2.5 concentration... 10 Accurate quantification of PM2.5 concentration is a crucial basis for pollution level assessment, risk warning, and pollution control decisions. Only by achieving PM2.5 concentration... 10 Only through large-scale, real-time quantitative inversion of concentrations can we obtain continuous spatial concentration field information that traditional site monitoring cannot provide, thereby solving technical problems such as insufficient spatial coverage, monitoring lag, and difficulty in conducting regional-scale quantitative pollution assessments in existing technologies.

[0021] To solve the above technical problems, such as Figure 1 As shown, this embodiment proposes a PM based on Himawari satellite data and machine learning. 10 The concentration inversion method, based on multi-source data, constructs a high-quality dust sample dataset and preprocesses it. It then extracts key influencing factors by combining various feature engineering methods and introduces an XGBoost model to establish PM concentration retrieval. 10A concentration inversion model was developed, and the inversion effect was evaluated using multiple indicators, ultimately achieving full-time PM2.5 concentration analysis for dust storms in China. 10 Real-time concentration monitoring specifically includes the following steps: Step S1, Dataset Construction and Preprocessing The data used includes satellite remote sensing data and site monitoring data. Satellite remote sensing data comes from the Himawari satellite. It is automatically connected to the JAXA P-Tree remote sensing data service platform via FTP protocol to batch download the full-disk NetCDF data (NC_H0X_YYYYMMDD_HHMM_R21_FLDK.02401_02401.nc format) for the target time period from the Himawari satellite. The data is automatically converted to UTC time range based on the Beijing time folder, and the target files are located and downloaded to the local folder according to the remote directory structure. Site monitoring data comes from the national urban air quality real-time release platform, collecting hourly PM2.5 data. 10 The observation data, along with information such as station name, longitude, and latitude, is serialized by date and output as a NetCDF format file for subsequent spatiotemporal matching.

[0022] Based on annual and monthly meteorological analysis literature, the occurrence periods and intensity levels of typical dust storm events were determined, with events reaching dust storm levels or above since 2016 selected as the research subjects. Satellite imagery data and ground station data were extracted according to the start and end times of dust storm events, and a sample dataset was constructed through spatiotemporal matching. Considering only PM2.5 concentrations in typical dust storm events in the Chinese region... 10 For concentration retrieval, the selected region is 0°–60° North latitude and 80°–140° East longitude. Regarding the data range selection, in accordance with national Air Quality Index (AQI) technical regulations, samples at the severe pollution level and above were selected, retaining PM2.5 concentrations. 10 Concentration between 350 and 4000 The samples were processed, and outlier samples with excessively large spatial matching distances were removed. Finally, the PM... 10 The concentration is logarithmically transformed to reduce the adverse effects of its long-tailed distribution on the stability and accuracy of model training.

[0023] Step S2, Feature Engineering Screening Select all satellite channels, extract raw observations, and calculate multiple derived remote sensing features based on these, including but not limited to the following: 6 visible / near-infrared albedo channels (albedo_01 to albedo_06); 10 thermal infrared brightness temperature channels (tbb_07 to tbb_16); solar-satellite geometric parameters: solar zenith angle (SOZ), solar azimuth angle (SOA), satellite zenith angle (SAZ), and satellite azimuth angle (SAA); brightness temperature difference features: BT 11 BT 12 BT 13 BT 11 etc.; Normalized difference characteristics: Standardized brightness temperature difference ND13_11; Vegetation and surface indices: Normalized Difference Vegetation Index (NDVI), Soil-Adjusted Vegetation Index (SAVI), Normalized Difference Humidity Index (NDMI), Normalized Difference Building Index (NDBI), Building Index (BU); Dust-specific indices: Dust Reflectance Index (DRI), Mid-Infrared Dust Index (MIDI); Geometric angular derivatives: Relative Azimuth Angle (REL_AZI), Cosine of Solar Zenith Angle (COS_SOZ); Static surface characteristics: Land cover type (LU_INDEX), Landmass Mask (LANDMASK), Altitude (HGT_M); Temporal characteristics: Year (year), Day of the Year (DOY).

[0024] Existing studies often employ a single method for feature engineering screening, resulting in relatively limited evaluation dimensions. To address this technical issue, this implementation method utilizes multiple feature screening methods, including MI mutual information analysis, SHAP value analysis, XGBoost feature importance analysis, and ablation experiments. However, since MI mutual information analysis, SHAP value analysis, XGBoost feature importance analysis, and ablation experiments evaluate feature contributions from different perspectives, the same feature may exhibit inconsistent importance rankings under different evaluation methods, thus affecting the stability and reliability of the feature screening results.

[0025] To address the aforementioned technical problems, this embodiment further improves feature selection by proposing a multi-dimensional feature consistency evaluation method, specifically as follows: First, the importance ranking of each feature was obtained by using MI mutual information analysis, SHAP value analysis, XGBoost feature importance analysis, and ablation experiments. The mutual information (MI) method can measure the non-linear dependency between features and the target variable. .

[0026] like Figure 2 As shown, SHAP value analysis can decompose the prediction results into the marginal contributions of each feature, predicting the contribution of each feature to the model: .

[0027] like Figure 4 As shown, XGBoost feature importance analysis, based on the gradient boosting tree model, measures the importance of a feature by how it is used in the tree structure, and automatically completes feature selection during training.

[0028] like Figure 5 As shown, the ablation experiment method removes individual features one by one and retrains the model to evaluate the actual contribution of the feature to the model's predictive performance.

[0029] The evaluation results were then normalized to construct a comprehensive feature consistency scoring index. The comprehensive feature contribution was obtained by weighted fusion of the evaluation results of each method. The weights of MI mutual information analysis, SHAP value analysis, XGBoost feature importance analysis and ablation experiment were 0.25, 0.35, 0.30 and 0.10, respectively.

[0030] Simultaneously, a ranking dispersion penalty mechanism is introduced to reduce the overall score of features with significant differences in ranking among different evaluation methods, thereby improving the consistency and stability of feature selection results. The ranking dispersion penalty mechanism calculates the standard deviation based on multiple ranking results from the aforementioned selection methods and removes features with excessively large standard deviations.

[0031] Finally, correlation analysis was used to remove redundant features with high correlation, resulting in the final feature set.

[0032] like Figure 3 As shown, correlation analysis measures the degree of linear correlation between a feature and the target variable. .

[0033] The following 13 features remain after filtering and will be used for subsequent machine learning training: Geographic information categories: HGT_M, longitude; latitude; Satellite data types: BT08, BT12, BTD14_11, BTD13_11; Geometric parameter classes: SAA-SOA; SOA; SAZ; COS(SOZ); Time-related information: year, DOY.

[0034] Therefore, the above screening method effectively reduces feature redundancy and improves the stability, interpretability and generalization ability of the feature screening results while ensuring the accuracy of model prediction.

[0035] Step S3, XGBoost model training; The machine learning training part uses the XGBoost model, an efficient and scalable nonlinear ensemble model widely used in regression and classification problems. Through a gradient boosting framework and regularization terms, the model effectively controls overfitting while maintaining its expressive power, making it suitable for prediction tasks with complex nonlinear relationships. Furthermore, this model has been previously applied to PM (Productivity Management). 2.5 inversion and PM 10The objective function for the prediction is as follows: ; .

[0036] To prevent outliers from hindering model learning, consistent outlier filtering was first implemented. A lightweight XGBoost base model was trained using 5-fold KFold cross-validation to obtain Out-of-Fold (OOF) predictions. The relative residuals of each sample were calculated, and samples were binned according to matching distance. Outliers within each bin were pruned using a 95th percentile threshold, resulting in a removal rate of approximately 5%, effectively eliminating training noise caused by matching errors or site anomalies. 10 Concentration range (350-4000) The sample was divided into 7 intervals, with weights increasing linearly from 1.0 to 2.0. For high-concentration dust samples (PM2.5), the weights were adjusted accordingly. 10 >500 (This can be used to assign higher training weights, thereby improving the model's ability to fit in high concentration ranges.)

[0037] As shown in Table 1, for the selection of optimal hyperparameters, a large parameter range is first set, and a computationally efficient random search is used to narrow the hyperparameter range. Then, BayesSearchCV (Bayesian hyperparameter optimization) is used to perform automated search and optimization on key hyperparameters such as learning rate, maximum tree depth, subsample ratio, and column sampling ratio. The optimal model, feature list, and transformation parameters are persistently saved in joblib format for use by the inference module. The final hyperparameters are as follows: Table 1. Optimal hyperparameters of the model (after Bayesian optimization)

[0038] Step S4: Model output and inversion effect evaluation; By inputting the satellite data received in real time by the ground station into the inversion model, the PM2.5 concentration in the target area can be obtained in real time. 10 Concentration inversion results, achieving PM2.5 concentration in the target area. 10 Continuous real-time monitoring of concentration. The validation set results are compared with the true values ​​to evaluate the inversion effect. Mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²) are used. 2 As an indicator for evaluating predictive effectiveness, the calculation formula is as follows: ; ; ; in, For the true value, For predicted values, The average of the true values. This represents the number of samples.

[0039] A 10-fold cross-validation method grouped by site (to prevent spatial leakage) is employed. Training is performed in the transformed space, and MAE, RMSE, and R² are evaluated after inverse transformation at the original scale. (Note: The last part, "PM," appears to be an error and doesn't translate directly. It likely refers to a specific method or approach.) 10 The segmentation error for each concentration range was calculated (by statistically analyzing the MAE and systematic bias for each range), serving as a quantitative basis for adjusting the sample weighting scheme. The final model MAE was 93.46, and R0 was... 2 The value is 0.873, the RMSE is 175.4, and the segmented MAE is shown in Table 2.

[0040] Table 2 PM 10 Piecewise MAE of inversion values

[0041] Therefore, existing methods are insufficient to achieve PM. 10 Addressing the challenge of achieving continuous spatial and real-time quantitative concentration retrieval, this implementation method combines satellite data and on-site measured data to construct an inversion model, enabling the retrieval of PM2.5 concentrations over large areas. 10 Real-time quantitative monitoring of concentration.

[0042] To better illustrate the PM based on Himawari satellite data and machine learning described in this embodiment... 10 The technical advantages of the concentration inversion method are illustrated in detail through the following experimental data: like Figure 6 and Figure 7 As shown, through comparative analysis of single-time inversion results and dust monitoring areas, it was found that PM2.5 concentrations... 10 High-value areas and dust distribution areas show a high degree of spatial consistency; such as Figure 8 As shown in the comparison results of dust storm processes, the method also demonstrates good overall performance in dust storm process monitoring.

[0043] In summary, this implementation method achieves PM based on machine learning methods. 10 Quantitative inversion of concentrations has, for the first time, constructed a full-time PM2.5 concentration model for dust storm processes in China based on Himawari satellite data. 10 Real-time concentration monitoring method. This implementation method does not rely on auxiliary data such as difficult-to-obtain or complex pollutant emission inventories; it can achieve PM2.5 concentration monitoring of the target area solely based on satellite observation data. 10Quantitative inversion of concentration breaks through the "qualitative" limitations of existing methods and is applicable to large-scale environmental monitoring. The feature engineering stage innovatively integrates multiple methods such as statistical analysis, information theory, model-driven interpretation, and performance ablation to construct a multi-dimensional comprehensive evaluation system, improving model prediction performance while ensuring the conciseness and interpretability of the feature set. Simultaneously, combined with real-time satellite data received from ground stations, PM2.5 concentration can be determined. 10 Online quantitative inversion and real-time monitoring of concentrations were performed. The results showed that the model underestimated concentrations in the high concentration range, but had good inversion performance in the medium and low concentration range. The overall inversion error was controlled within 10%, demonstrating good performance in related studies.

[0044] Implementation Method Two: This implementation method describes a PM based on Himawari satellite data and machine learning. 10 The concentration inversion system is based on the PM2.5 concentration analysis method described in Implementation Method 1, which uses Himawari satellite data and machine learning. 10 The concentration inversion method includes the following modules: Module S1, build PM 10 Concentration datasets were obtained and preprocessed. Module S2 selects full-channel satellite data, calculates multiple derived remote sensing features based on the extracted raw observations, and uses a feature filtering method to filter the multiple derived remote sensing features to obtain the final feature set; Module S3 uses the final feature set from Module S2 to train the XGBoost model; Module S4 processes the PM preprocessed by module S1. 10 The concentration dataset is input into the XGBoost model trained by module S3 to complete the PM2.5 concentration analysis of the target region. 10 Continuous real-time monitoring of concentration.

[0045] Implementation Method 3: An electronic device according to this implementation method includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus. Memory, used to store computer programs; The processor, when executing a program stored in memory, implements the method described in Embodiment 1.

[0046] Implementation Method 4: A computer-readable storage medium according to this implementation method, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, it implements the method described in Implementation Method 1.

[0047] The above describes the PM based on Himawari satellite data and machine learning proposed in this invention. 10The concentration inversion method and system are described in detail. Specific examples are used in this paper to illustrate the principle and implementation of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation and application scope based on the idea of ​​the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A PM based on Himawari satellite data and machine learning 10 The concentration inversion method is characterized by, Includes the following steps: Step S1, Build PM 10 Concentration datasets were obtained and preprocessed. Step S2: Select satellite full-channel data, calculate multiple derived remote sensing features based on the extracted raw observations, and use a feature filtering method to filter the multiple derived remote sensing features to obtain the final feature set; Step S3: Train the XGBoost model using the final feature set from step S2; Step S4, process the PM preprocessed in step S1. 10 The concentration dataset is input into the XGBoost model trained in step S3 to complete the analysis of PM in the target region. 10 Continuous real-time monitoring of concentration.

2. A PM based on Himawari satellite data and machine learning as described in claim 1 10 The concentration inversion method is characterized by, In step S1, the construction of PM 10 The concentration dataset includes satellite remote sensing data and site monitoring data; Satellite remote sensing data is automatically connected to the JAXA P-Tree remote sensing data service platform via FTP protocol, and the satellite full disk NetCDF data for the target time period is downloaded in batches. The data is automatically converted to UTC time range according to the Beijing time folder, and the target file is located and downloaded to the local folder according to the remote directory structure. Site monitoring data is collected hourly for PM. 10 The observation data, along with the station name, longitude, and latitude information, is serialized by date and output as a NetCDF format file.

3. A PM based on Sunflower satellite data and machine learning as described in claim 1 10 The concentration inversion method is characterized by, In step S1, the preprocessing specifically includes: Samples at or above the severe pollution level were selected, and outlier samples with excessively large spatial matching distances were removed. PM2.5 levels were then analyzed. 10 The concentration was subjected to a logarithmic transformation.

4. A PM based on Himawari satellite data and machine learning as described in claim 1 10 The concentration inversion method is characterized by, In step S2, the use of a feature filtering method to filter multiple derived remote sensing features to obtain a final feature set specifically involves: Feature selection methods include MI mutual information analysis, SHAP value analysis, XGBoost feature importance analysis and ablation experiments. The importance ranking of multiple derived remote sensing features is obtained using feature selection methods. The screening results of the feature screening method are normalized, a comprehensive feature consistency scoring index is constructed, and the comprehensive feature contribution is obtained by weighted fusion of the screening results of the feature screening method. At the same time, a ranking dispersion penalty mechanism is introduced to reduce the comprehensive score of features with large differences in ranking among feature selection methods; By combining correlation analysis to remove redundant features with high correlation, the final feature set is obtained.

5. A PM based on Himawari satellite data and machine learning as described in claim 4 10 The concentration inversion method is characterized by, The weights for the MI mutual information analysis, SHAP value analysis, XGBoost feature importance analysis, and ablation experiment are 0.25, 0.35, 0.30, and 0.10, respectively.

6. A PM based on Himawari satellite data and machine learning 10 The concentration inversion system is based on the PM2.5 concentration analysis method described in claim 1, which uses Himawari satellite data and machine learning. 10 The concentration inversion method is characterized by, Includes the following modules: Module S1, build PM 10 Concentration datasets were obtained and preprocessed. Module S2 selects full-channel satellite data, calculates multiple derived remote sensing features based on the extracted raw observations, and uses a feature filtering method to filter the multiple derived remote sensing features to obtain the final feature set; Module S3 uses the final feature set from Module S2 to train the XGBoost model; Module S4 processes the PM preprocessed by module S1. 10 The concentration dataset is input into the XGBoost model trained by module S3 to complete the PM2.5 concentration analysis of the target region. 10 Continuous real-time monitoring of concentration.

7. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, when executing a program stored in memory, implements the method described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method described in any one of claims 1-5.

Images