A Spatial-Temporal Interpretable Deep Learning Model for improving interpretability and predictive accuracy of satellite-based PM2.5

Highlights: • A new Spatial-Temporal Interpretable Deep Learning Model (SIDLM) was developed in this study. • In contrast to traditional deep learning models, the SIDLM is "wide" and "deep.". • SIDLM was applied to predict PM_2.5 and was validated at national and urban scales. • SILDM exhibited higher predictive accuracy than traditional machine learning methods. • SIDLM can automatically extract spatiotemporal characteristics from data. Being able to monitor PM_2.5 across a range of scales is incredibly important for our ability to understand and counteract air pollution. Remote monitoring PM_2.5 using satellite-based data would be incredibly advantageous to this effort, but current machine learning methods lack necessary interpretability and predictive accuracy. This study details the development of a new Spatial-Temporal Interpretable Deep Learning Model (SIDLM) to improve the interpretability and predictive accuracy of satellite-based PM_2.5 measurements. In contrast to traditional deep learning models, the SIDLM is both "wide" and "deep." We comprehensively evaluated the proposed model in China using different input data (top-of-atmosphere (TOA) measurements-based and aerosol optical depth (AOD)-based, with or without meteorological data) and different spatial resolutions (10 km, 3 km, and 250 m). TOA-based SIDLM PM_2.5 achieved the best predictive accuracy in China, with root-mean-square errors (RMSE) of 15.30 and 15.96 μg/m³, and R² values of 0.70 and 0.66 for PM_2.5 predictions at 10 km and 3 km spatial resolutions, respectively. Additionally, we tested the SIDLM in PM_2.5 retrievals at a 250 m spatial resolution over Beijing, China (RMSE = 16.01 μg/m³, R² = 0.62). Furthermore, SIDLM demonstrated higher accuracy than five machine learning inversion methods, and also outperformed them regarding feature extraction and the interpretability of its inversion results. In particular, modeling results indicated the strong influence of the Tongzhou district on the principle PM_2.5 in the Beijing urban area. SIDLM-extracted temporal characteristics revealed that summer months (June–August) might have contributed less to PM_2.5 concentrations, indicating the limited accumulation of PM_2.5 in these months. Our study shows that SIDLM could become an important tool for other earth observation data in deep learning-based predictions and spatiotemporal analysis.