关键词: PM_2.5估算, 物理约束, 卫星遥感, 时空建模, 深度学习

Abstract:

[Objective] Fine particulate matter (PM_2.5) pollution has become one of the most important environmental challenges worldwide due to its significant impacts on human health, ecosystems, and regional climate systems. Long-term exposure to elevated PM_2.5 concentrations has been associated with respiratory and cardiovascular diseases, premature mortality, and increased risks of hospitalization. Atmospheric PM_2.5 can also affect visibility, cloud properties, and regional climate processes. However, obtaining spatially continuous and temporally consistent PM_2.5 observations remains challenging because ground-based monitoring networks are sparse and unevenly distributed, especially in regions with complex terrain and heterogeneous emission sources. Satellite remote sensing provides an effective approach for large-scale PM_2.5 estimation, and aerosol optical depth (AOD) has been widely used as an important proxy for ground-level PM_2.5 concentrations. Nevertheless, existing AOD-based inversion methods still have several limitations. Traditional statistical and machine learning models generally focus on either spatial feature extraction or temporal variation analysis, while lacking the ability to simultaneously characterize complex spatiotemporal relationships. Moreover, most deep learning approaches are purely data-driven and do not explicitly incorporate atmospheric physical mechanisms, which may reduce their generalization capability and lead to physically inconsistent predictions across different temporal or spatial conditions. Therefore, this study develops a novel physics-constrained U-Net-LSTM deep learning framework for high-resolution spatiotemporal reconstruction of ground-level PM_2.5 concentrations. [Methods] The proposed framework integrates a U-Net encoder-decoder architecture with a long short-term memory (LSTM) network to simultaneously capture spatial patterns and temporal evolution of PM_2.5 pollution. The U-Net architecture is employed to extract multiscale spatial features and preserve detailed spatial information through skip connections, while the LSTM network is used to model temporal dependencies and dynamic variations in air pollution processes. To improve feature representation and model adaptability, convolutional block attention modules (CBAMs) are embedded in the encoder and decoder pathways, enabling the model to automatically learn channel and spatial attention weights. This mechanism allows the network to focus on critical environmental variables and important geographic regions while reducing the influence of irrelevant information. Furthermore, a physics-based residual constraint derived from the atmospheric advection-diffusion equation is incorporated into the training loss function. This constraint introduces atmospheric mechanisms, including horizontal wind-driven transport, turbulent diffusion, and pollutant removal processes, into the deep learning framework. By combining data-driven feature learning with physical knowledge, the model is encouraged to generate PM_2.5 predictions that are not only statistically accurate but also consistent with atmospheric dynamics, thereby improving prediction reliability and generalization capability. The proposed model was trained and evaluated using observations from 97 air quality monitoring stations across the Sichuan Basin from 2015 to 2020. The Sichuan Basin was selected because of its complex topography, severe pollution conditions, and frequent haze events. The input dataset included multisource environmental variables at a spatial resolution of 1 km, including satellite AOD products, ERA5 meteorological parameters (wind speed, temperature, relative humidity, surface pressure, and boundary layer height), a digital elevation model (DEM), the normalized difference vegetation index (NDVI), land-use classification, and geographic coordinates. A total of 85,205 spatiotemporal samples were generated. Three validation strategies, namely sample-based, time-based, and site-based validation, were used to evaluate model accuracy and generalization capability. Model performance was assessed using the coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE). [Results] The proposed physics-constrained U-Net-LSTM model achieved superior performance compared with benchmark models, including CNN, LSTM, and ConvLSTM, under all validation scenarios. In sample-based validation, the model achieved an R² of 0.874, an RMSE of 8.12 μg·m^-3, and an MAE of 8.01 μg·m^-3. In time-based and site-based validations, the model maintained strong predictive capability, achieving R² values of 0.86 and 0.83, respectively, with prediction errors reduced by approximately 5%‒25% compared with baseline models. These results demonstrate the strong temporal transferability and spatial generalization capability of the proposed framework. Ablation experiments further confirmed the effectiveness of each model component. The baseline U-Net-LSTM model without CBAM or physical constraints achieved an R² of 0.79, with an RMSE of 16.62 μg·m^-3 and an MAE of 14.63 μg·m^-3. After introducing CBAM, model performance improved to R² = 0.82, demonstrating the contribution of the attention mechanism to the extraction of important environmental features. With the further incorporation of the physics-based residual constraint, the model achieved R² = 0.87, RMSE = 11.21 μg·m^-3, and MAE = 8.83 μg·m^-3, with more than 80% of monitoring stations achieving R² values greater than 0.7. These findings confirm that physical guidance significantly enhances model accuracy and robustness. Permutation importance analysis indicated that AOD was the most influential predictor, contributing for 40.63% of the total permutation importance, followed by surface pressure, air temperature, DEM, and relative humidity. These results highlight the combined effects of aerosol loading, meteorological conditions, and topographic characteristics on PM_2.5 accumulation and dispersion in the Sichuan Basin. The reconstructed PM_2.5 maps successfully captured regional pollution patterns and long-term temporal trends. The annual mean PM_2.5 concentration decreased from 38.95 μg·m^-3 in 2015 to 25.91 μg·m^-3 in 2020, indicating the effectiveness of regional air pollution control measures. Spatially, a persistent pattern of high concentrations in the center and low concentrations at the periphery was observed, with the Chengdu Plain and surrounding urban agglomerations consistently exhibiting higher PM_2.5 levels. Seasonal analysis showed that winter had the highest PM_2.5 concentration (44.71 μg·m^-3), mainly because of unfavorable atmospheric dispersion conditions, temperature inversion, and increased heating emissions, whereas summer had the lowest concentration (15.95 μg·m^-3) because of enhanced atmospheric ventilation and precipitation scavenging. Additionally, the proposed framework successfully reproduced a severe haze episode in December 2020, accurately capturing pollutant accumulation, rapid enhancement, and dissipation following the intrusion of a cold front. The results further revealed the influence of regional transport processes and identified major urban clusters as persistent pollution hotspots. [Conclusion] This study proposes a physics-constrained deep learning framework that effectively integrates spatiotemporal feature learning, attention mechanisms, and atmospheric physical constraints for high-resolution PM_2.5 reconstruction. The proposed physics-constrained U-Net-LSTM model achieves improved accuracy, generalization capability, and interpretability compared with conventional data-driven approaches. By bridging the gap between deep learning prediction and physical understanding of atmospheric processes, this framework provides an effective solution for generating continuous, high-resolution PM_2.5 distributions in regions with complex terrain and offers valuable support for air quality assessment, pollution control strategies, and environmental management.

Key words: PM_2.5 estimation, physics-constrained, satellite remote sensing, spatiotemporal modeling, deep learning