Deriving PM2.5 from satellite observations with spatiotemporally weighted tree-based algorithms: enhancing modeling accuracy and interpretability

Abstract Tree-based machine learning algorithms, such as random forest, have emerged as effective tools for estimating fine particulate matter (PM 2.5 ) from satellite observations. However, they typically have unchanged model structures and configurations over time and space, and thus may not fully capture the spatiotemporal variations in the relationship between PM 2.5 and predictors, resulting in limited accuracy. Here, we propose geographically and temporally weighted tree-based models (GTW-Tree) for remote sensing of surface PM 2.5 . Unlike traditional tree-based models, GTW-Tree models vary by time and space to simulate the variability in PM 2.5 estimation, and they can output variable importance for every location for the deeper understanding of PM 2.5 determinants. Experiments in China demonstrate that GTW-Tree models significantly outperform the conventional tree-based models with predictive error reduced by >21%. The GTW-Tree-derived time-location-specific variable importance reveals spatiotemporally varying impacts of predictors on PM 2.5 . Aerosol optical depth (AOD) contributes largely to PM 2.5 estimation, particularly in central China. The proposed models are valuable for spatiotemporal modeling and interpretation of PM 2.5 and other various fields of environmental remote sensing.