Near-surface particulate matter pollution monitoring method based on multi-task deep learning

Abstract

The invention discloses a near-surface particulate matter pollution monitoring method based on multi-task deep learning, belongs to the technical field of particulate matter pollution monitoring, is used for particulate matter pollution monitoring, and comprises sample processing and construction and model construction and training. The sample processing and construction comprises the steps of obtaining ground observation data, obtaining PM concentration, performing space-time matching in combination with a multi-task sample data set, and performing preprocessing on a space-time matching result in combination with stationary meteorological satellite observation data and auxiliary data; the model construction and training comprises the steps of constructing a multi-task deep learning model and training the multi-task deep learning model, the multi-task deep learning model is based on a full-connection neural network, and an output layer of the multi-task deep learning model is divided into two branches which are respectively used for two PM inversion; a double-task loss function is introduced in the training process of the multi-task deep learning model.

Description

Technical Field

[0001] This invention discloses a near-surface particulate matter pollution monitoring method based on multi-task deep learning, belonging to the field of particulate matter pollution monitoring technology. Background Technology

[0002] In particulate matter pollution monitoring, most existing technologies typically employ diverse modeling methods and machine learning, especially deep learning, which has made significant progress in surface PM estimation, particularly excelling in handling high-dimensional, nonlinear, multi-source data. However, machine learning or deep learning methods still face an unavoidable limitation in improving accuracy: the cumulative effect of AOD inversion errors propagating to the final PM estimation. Furthermore, most existing technologies typically focus only on single pollutants and have not yet explored how to directly estimate PM based on satellite observations. 2.5 With PM 10 Joint inversion. Joint inversion holds significant potential within the deep learning framework, not only reducing uncertainties between different models but also improving the efficiency of satellite data processing (e.g., obtaining multiple results from a single training iteration). This strategy, known as multi-task learning, has yielded remarkable results in computer science, resulting in a series of classic studies. However, no research has yet applied multi-task learning methods to the simultaneous inversion of atmospheric pollutants, particularly surface PM2.5. 2.5 and PM 10 The joint estimate. Summary of the Invention

[0003] The purpose of this invention is to provide a near-surface particulate matter pollution monitoring method based on multi-task deep learning, so as to solve the problem in the prior art that near-surface particulate matter pollution monitoring does not combine multi-task learning, resulting in the inability to perform inversion of multiple particulate matter types.

[0004] A near-surface particulate matter pollution monitoring method based on multi-task deep learning, including sample processing and construction and model construction and training; Sample processing and construction include acquiring ground observation data, and obtaining... and The concentration was determined, and spatiotemporal matching was performed using a multi-task sample dataset. The spatiotemporal matching results were then preprocessed using geostationary meteorological satellite observation data and auxiliary data. Model building and training include building a multi-task deep learning model and training the multi-task deep learning model. The multi-task deep learning model is based on a fully connected neural network, and its output layer is divided into two parts, each used for... and The inversion branch introduces a dual-task loss function during the training of multi-task deep learning models. Make full use of and The interaction between them.

[0005] The auxiliary data includes angle data, meteorological data, geospatial data, and economic data.

[0006] The preprocessing includes secondary projection, resampling, and cloud masking.

[0007] include: ; In the formula, yes The weight, yes loss function, yes The weight, yes The loss function.

[0008] and The weights are automatically determined using a dynamic weighted averaging method.

[0009] A fully connected neural network in a multi-task deep learning model includes multiple hidden layers.

[0010] Multi-task deep learning model The calculation formula is: ; In the formula, It is the output of a multi-task deep learning model. It is the top atmospheric reflectivity of the first six bands. It's the angle. It is the near-surface air pressure. It is the wind field component at a height of 10 meters. It is the boundary layer height. It is a digital elevation model. It is the normalized vegetation index. It is the annual land coverage. There are two datasets. It is a spatiotemporal characteristic.

[0011] The multi-task deep learning model contains six hidden layers, each with 1024 neurons. During training, it iterates 1800 times, with a learning rate of 0.001 and a batch size of 1024.

[0012] Pearson correlation coefficient, root mean square error, normalized root mean square error, and mean absolute error are used to quantify multi-task deep learning models and measure the fitting effect and error level.

[0013] Compared with the prior art, the present invention has the following beneficial effects: the present invention reduces the inconsistency inherent in building a model alone and improves the inversion accuracy, providing a valuable reference for future quantitative inversion of aerosols and air pollutants based on deep learning, and showing that by sharing input features, a single trained model can achieve accurate inversion of multiple atmospheric parameters. Attached Figure Description

[0014] Figure 1 For multi-task models under the sample benchmark Estimation results.

[0015] Figure 2 Multi-task model based on site baseline Estimation results.

[0016] Figure 3 For a multi-task model based on time Estimation results.

[0017] Figure 4 The task model is based on the sample benchmark. Estimation results.

[0018] Figure 5 For the site-based order task model Estimation results.

[0019] Figure 6 The time-based order task model Estimation results.

[0020] Figure 7 For multi-task models based on sample benchmarks Estimation results.

[0021] Figure 8 Multi-task model based on site baseline Estimation results.

[0022] Figure 9 For a multi-task model based on time Estimation results.

[0023] Figure 10 The task model is based on the sample benchmark. Estimation results.

[0024] Figure 11 For the site-based order task model Estimation results.

[0025] Figure 12 The time-based order task model Estimation results. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention are described clearly and completely below. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0027] A near-surface particulate matter pollution monitoring method based on multi-task deep learning, including sample processing and construction and model construction and training; Sample processing and construction include acquiring ground observation data, and obtaining... and The concentration was determined, and spatiotemporal matching was performed using a multi-task sample dataset. The spatiotemporal matching results were then preprocessed using geostationary meteorological satellite observation data and auxiliary data. Model building and training include building a multi-task deep learning model and training the multi-task deep learning model. The multi-task deep learning model is based on a fully connected neural network, and its output layer is divided into two parts, each used for... and The inversion branch introduces a dual-task loss function during the training of multi-task deep learning models. Make full use of and The interaction between them.

[0028] The auxiliary data includes angle data, meteorological data, geospatial data, and economic data.

[0029] The preprocessing includes secondary projection, resampling, and cloud masking.

[0030] include: ; In the formula, yes The weight, yes loss function, yes The weight, yes The loss function.

[0031] and The weights are automatically determined using a dynamic weighted averaging method.

[0032] A fully connected neural network in a multi-task deep learning model includes multiple hidden layers.

[0033] Multi-task deep learning model The calculation formula is: ; In the formula, It is the output of a multi-task deep learning model. It is the top atmospheric reflectivity of the first six bands. It's the angle. It is the near-surface air pressure. It is the wind field component at a height of 10 meters. It is the boundary layer height. It is a digital elevation model. It is the normalized vegetation index. It is the annual land coverage. There are two datasets. It is a spatiotemporal characteristic.

[0034] The multi-task deep learning model contains six hidden layers, each with 1024 neurons. During training, it iterates 1800 times, with a learning rate of 0.001 and a batch size of 1024.

[0035] Pearson correlation coefficient, root mean square error, normalized root mean square error, and mean absolute error are used to quantify multi-task deep learning models and measure the fitting effect and error level.

[0036] This invention enables hourly surface PM2.5 analysis based on FY-4A / AGRI imagery. 2.5 With PM 10 Joint estimation is performed using advanced multi-task deep learning methods. The model simultaneously optimizes two objectives during training and imposes constraints on them during inference, thereby improving sample utilization efficiency and enhancing model consistency and robustness. Multi-task learning (MTL) is a type of machine learning strategy that trains multiple related tasks in parallel through shared representations and inductive transfer mechanisms. Unlike well-known single-task learning (STL) methods, where each input corresponds to only a single output, MTL effectively considers the learning process of multiple related tasks. Although PM... 2.5 and PM 10Multi-task learning (MTL) can improve the overall performance of tasks by sharing similar data sources and distributions, but their learning objectives differ, thus they cannot be considered as a single task. Through joint modeling, MTL can improve the overall performance of each task. MTL can be implemented through hard parameter sharing or soft parameter sharing. Hard parameter sharing uses a common feature extraction layer in the early stages of the neural network, while setting independent output layers for each task in the later stages. This reduces model complexity, mitigates overfitting risk, and enables information sharing between tasks, while enhancing the robustness of the learning process. In contrast, soft parameter sharing ensures that each task maintains independent model parameters, while promoting similarity between tasks through constraints such as L2 distance. Multi-task learning can facilitate knowledge transfer between related tasks, allowing learning gains from one task to be transferred to other tasks, making it particularly suitable for learning performance management (PM). 2.5 and PM 10 The joint inversion task. In this invention, a hard parameter sharing strategy is adopted, that is, the backbone network is shared among tasks, while the output layer of each task learns independently.

[0037] Despite the use of advanced deep learning models, surface PM 10 With PM 2.5 Satellite inversion accuracy has been significantly improved, but previous studies typically constructed independent models for each pollutant, neglecting PM2.5. 10 With PM 2.5 The physical connection between them. Therefore, on some satellite pixels (especially under high pollution conditions and with large errors), the PM2.5 output of a single model may be different. 10 Concentration lower than PM 2.5 Concentration. This contradicts PM2.5 concentration. 10 The concentration should always be greater than or equal to PM2.5. 2.5 Physical constraints on concentration. This inconsistency may stem from random factors during training, including parameter initialization and batch sampling, both of which can lead to prediction fluctuations. Furthermore, single-task models have limited ability to handle rare or complex patterns, potentially resulting in greater discrepancies under extreme weather or pollution events.

[0038] Considering PM 10 With PM 2.5 Based on the correlation between different tasks and their similar impact on multispectral observations, this invention constructs a multi-task deep learning model. Due to its simple structure, significant performance, and good scalability, this invention uses a fully connected neural network (FCNN) as the basic model, with the output layer divided into two branches, one for PM and the other for P. 10 and PM 2.5 The inversion. Considering PM 2.5With PM 10 To address differences in concentration levels, this invention assigns different loss weights to different tasks during training to improve the model's convergence speed and prediction accuracy. However, accurately determining PM... 2.5 With PM 10 Optimal weighting remains a challenge. Therefore, a Dynamic Weight Averaging (DWA) method is used to automatically determine the weights. This method assigns higher weights to tasks with slower loss reduction, while setting a minimum weight of 0.4 to avoid any task having an excessively low weight.

[0039] This invention uses the top-level atmospheric reflectance of the first six bands (0.45–2.35 μm, TOAR1–6) of FY-4A / AGRI as key input features. These bands are highly sensitive to atmospheric aerosol information, especially in the short-wavelength band (such as the blue light band), thus effectively capturing changes in aerosol optical properties. By combining corresponding observational geometric parameters (solar zenith angle SZA, satellite zenith angle VZA, solar azimuth angle SAA, and satellite azimuth angle VAA), the differences in radiative transfer paths under different surface and atmospheric conditions are considered to improve inversion accuracy. Key atmospheric variables include near-surface pressure (SP), wind components at 10-meter height (U10 and V10), and boundary layer height (BLH). These variables are important indicators for assessing pollutant diffusion, migration, and accumulation. Surface-related data included MODIS 1 km monthly Normalized Differential Vegetation Index (NDVI) products, 500 m annual land cover (LUC) products, and a 90 m digital elevation model (DEM) provided by Spaceborne Radar Topographic Survey (SRTM). These data characterize surface features and land cover changes, influencing particulate matter deposition and resuspension processes, and help distinguish the surface contribution of aerosols in satellite TOA signals. Socioeconomic variables, primarily traffic and population, reflect anthropogenic emissions and their impact on particulate matter concentrations in urban areas. Ultimately, this invention selected 22 features, including temporal information and geographic coordinates (latitude and longitude), for model construction. Detailed information for each dataset is listed in Table 2. All auxiliary data were unified to a 0.04° × 0.04° grid using bilinear interpolation and cropped to the coverage experimental area to align with satellite observations. (Ground PM) 10 With PM 2.5The observed data were matched temporally and spatially with FY-4A / AGRI imagery and auxiliary variables, resulting in a total of 843,260 samples. To prevent overfitting and ensure optimal parameters, an early stopping strategy was employed during training. To improve computational efficiency, this invention uses five-fold cross-validation, dividing the sample data into five mutually exclusive subsets. Four subsets are used for training each time, with the remaining subset used for validation. This process is repeated five times, and the average performance of the five-fold cross-validation is used as the model evaluation metric. Specifically, three data partitioning strategies are employed: random partitioning by sample, hourly partitioning by time, and site-based partitioning. Sample-based validation is used to evaluate the overall model performance, while time-based and site-based validation are used to evaluate the model's spatiotemporal predictive capabilities. Model performance is quantified using the Pearson correlation coefficient (R), root mean square error (RMSE), normalized root mean square error (NRMSE), and mean absolute error (MAE) to measure the fit and error level.

[0040] This invention applies five-fold cross-validation at the sample level to evaluate the performance of multi-task models on surface PM2.5. 2.5 With PM 10 Estimated predictive performance. Model for PM 2.5 The concentration prediction performance was good, showing strong correlation (R=0.88) and low error (RMSE=14.57μg / m³). 3 MAE = 9.57 μg / m 3 ),like Figure 1 As shown. In contrast, the traditional single-task model has lower accuracy (R=0.86) and larger error (RMSE=15.41μg / m). 3 MAE = 10.12 μg / m 3 ),like Figure 2 As shown. For PM 10 The multi-task model also outperformed the single-task model (R=0.87, RMSE=26.25μg / m). 3 MAE = 16.72 μg / m 3 The maximum increase can reach 6%, such as Figure 7 and Figure 10 As shown. This improvement may stem from the introduction of PM during the training and inference processes in the multi-task framework. 2.5 With PM 10 The mutual constraints between tasks improve prediction accuracy. Overall, the multi-task model performs robustly within the experimental range, and PM... 2.5 With PM10 The R-values ​​were above 0.7 in over 85% and 80% of the regions, respectively. Regions with lower correlation (R < 0.5) were mainly distributed in scattered areas of the southwest, central, and northwest, which may be related to the sparse monitoring stations and complex environmental conditions. Nevertheless, the multi-task model showed good correlation for PM2.5 in over 90% of the regions, including the aforementioned areas. 2.5 and PM 10 The normalized root mean square error (NRMSE) remains at a low level (<0.1).

[0041] This invention employs time- and site-based validation to evaluate the model's spatiotemporal generalization ability. For site validation, the multi-task model in PM... 2.5 In predictions, it outperformed the single-task model, with its correlation coefficient R increasing from 0.82 to 0.85. Figure 2 and Figure 5 The prediction error was also significantly reduced, with RMSE decreasing from 17.62 μg / m². 3 Reduced to 16.01 μg / m 3 MAE from 11.50 μg / m 3 Decreased to 10.39 μg / m 3 This represents a reduction of approximately 9%. These results indicate that the model effectively mitigates the overfitting problem caused by regional heterogeneity, partly due to the sharing of spatial feature representations. (Compared to PM) 2.5 In comparison, PM 10 The predictive performance improvement was more significant, with the correlation coefficient R increasing from 0.79 to 0.83, such as... Figure 8 and Figure 11 As shown, the improvement was 5%. Meanwhile, RMSE decreased from 33.74 μg / m 3 Decreased to 29.77 μg / m 3 MAE increased from 21.89 μg / m 3 Reduced to 18.94 μg / m 3 These figures were reduced by 12% and 13% respectively. This greater improvement may be attributed to PM2.5. 10More sensitive to diverse emission sources (such as dust storms and industrial activities), leading to greater spatial variability, the model benefits more from the multi-task framework. Overall, the multi-task model exhibits robust spatial generalization performance. Over 70% of the stations have correlation coefficients (R) higher than 0.7, and most stations have NRMSE lower than 0.1, indicating that the model maintains stable prediction accuracy within the experimental range. However, at some stations, the prediction accuracy is relatively low (R < 0.4), with an NRMSE of approximately 0.4. These results may reflect the limitations of complex terrain and its associated local meteorological anomalies on the model's generalization ability. Furthermore, large areas of snow cover and high surface reflectivity (such as glaciers and bare rocks) may introduce noise into the top-level atmospheric reflectivity signal, increasing the difficulty of atmospheric information extraction and thus amplifying the prediction error in these areas.

[0042] In temporal cross-validation, the model also demonstrated strong temporal generalization ability. For PM... 2.5 The correlation coefficient R of the multi-task model reached 0.83. Figure 3 As shown, it is higher than the 0.77 of the single-task model, such as Figure 6 As shown, the improvement was 8%. Simultaneously, RMSE decreased from 19.58 μg / m². 3 Reduced to 16.59 μg / m 3 MAE from 12.88 μg / m 3 Reduced to 10.50 μg / m 3 These figures decreased by 15% and 18% respectively. In PM 10 A similar pattern was observed in the estimation. The correlation coefficient R of the multi-task model reached 0.83, significantly higher than the 0.75 of the single-task model, representing an improvement of 11%. Furthermore, the RMSE and MAE of the multi-task model were 29.82 μg / m³. 3 and 18.44 μg / m 3 It is significantly better than the traditional model's 36.94 μg / m 3 and 24.24 μg / m 3 The errors were reduced by 19% and 24% respectively, such as Figure 9 and Figure 12 As shown. Overall, regarding PM 2.5 and PM 10 The multi-task model achieved an R-value greater than 0.7 at over 81% of the monitoring sites. NRMSE remained stable at most sites (generally between 0.1 and 0.2), with slightly higher values ​​only observed at a few sites in the central Sichuan Basin and southeastern Tibet.

[0043] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A near-surface particulate matter pollution monitoring method based on multi-task deep learning, characterized in that, This includes sample processing and construction, and model construction and training. Sample processing and construction include acquiring ground observation data, and obtaining... and The concentration was determined, and spatiotemporal matching was performed using a multi-task sample dataset. The spatiotemporal matching results were then preprocessed using geostationary meteorological satellite observation data and auxiliary data. Model building and training include building a multi-task deep learning model and training the multi-task deep learning model. The multi-task deep learning model is based on a fully connected neural network, and its output layer is divided into two parts, each used for... and The inversion branch introduces a dual-task loss function during the training of multi-task deep learning models. Make full use of and The interaction between them.

2. The near-surface particulate matter pollution monitoring method based on multi-task deep learning according to claim 1, characterized in that, The auxiliary data includes angle data, meteorological data, geospatial data, and economic data.

3. The near-surface particulate matter pollution monitoring method based on multi-task deep learning according to claim 2, characterized in that, The preprocessing includes secondary projection, resampling, and cloud masking.

4. The near-surface particulate matter pollution monitoring method based on multi-task deep learning according to claim 3, characterized in that, include: ； In the formula, yes The weight, yes loss function, yes The weight, yes The loss function.

5. The near-surface particulate matter pollution monitoring method based on multi-task deep learning according to claim 4, characterized in that, and The weights are automatically determined using a dynamic weighted averaging method.

6. The near-surface particulate matter pollution monitoring method based on multi-task deep learning according to claim 5, characterized in that, A fully connected neural network in a multi-task deep learning model includes multiple hidden layers.

7. The near-surface particulate matter pollution monitoring method based on multi-task deep learning according to claim 6, characterized in that, Multi-task deep learning model The calculation formula is: ； In the formula, It is the output of a multi-task deep learning model. It is the top atmospheric reflectivity of the first six bands. It's the angle. It is the near-surface air pressure. It is the wind field component at a height of 10 meters. It is the boundary layer height. It is a digital elevation model. It is the normalized vegetation index. It is the annual land coverage. There are two datasets. It is a spatiotemporal characteristic.

8. The near-surface particulate matter pollution monitoring method based on multi-task deep learning according to claim 7, characterized in that, The multi-task deep learning model contains six hidden layers, each with 1024 neurons. During training, it iterates 1800 times, with a learning rate of 0.001 and a batch size of 1024.

9. The near-surface particulate matter pollution monitoring method based on multi-task deep learning according to claim 8, characterized in that, Pearson correlation coefficient, root mean square error, normalized root mean square error, and mean absolute error are used to quantify multi-task deep learning models and measure the fitting effect and error level.

Images