Tracking hourly PM2.5 using geostationary satellite sensor images and multiscale spatiotemporal deep learning

Spatially continuous fine particulate matter (PM2.5) mapping with hourly updated is essential for monitoring environmental pollution and promoting public health. The intensive observation of geostationary satellite enables accurate estimation of PM2.5 at a fine-scale. However, current estimation models are still limited by their weak transferability and hard to provide a robust hourly PM2.5 estimation. In this research, we aim to estimate the daytime PM2.5 concentrations at fine spatial and temporal resolution (1 km and hourly) in mainland China using an improved deep learning algorithm and the AOD products from geostationary satellite Himwari-8. An Adaptive Spatio-Temporal Multiscale Neural Network (ASTMNN) which contains three sub-networks and an adaptive weight was proposed to capture the spatiotemporal heterogeneity of hourly PM2.5. The three subnetworks of ASTMNN are spatial adjacency module (SaM), temporal adjacency module (TaM) and global module (GM), which used to incorporate the information from spatial neighborhood, temporal neighborhood, and global spatiotemporal range, respectively. And the weight function combines the outputs from the three subnetworks, where the weights were adaptively trained from the model optimization. The proposed model outperformed most current hourly PM2.5 estimation models with the sample-based, time-based, and site-based cross-validation (CV) R2 of 0.94, 0.89 and 0.83, respectively. Besides, we used our PM2.5 product to track extreme dust events. Our findings provide valuable implications for tracking continuous variation in particulate pollution using geostationary satellites.