A spatiotemporal Downscaling Framework based on machine learning for hourly 1 km PM2.5 mapping in China

PM2.5 pollution is a global environmental problem, and its hourly exposure characteristics are closely related to short-term health risks. Traditional estimation methods are mainly based on satellite AOD, which are limited by AOD’s daily timescale and cloud/snow interference, resulting in difficulties in meeting the short-term prediction needs of PM2.5 pollution and in achieving high spatial resolution. This study proposes a spatiotemporal downscaling framework based on the Light-GBM (ST-Light-GBM) algorithm that integrates multi-source data. It innovatively integrated the daily 1 km PM2.5 data derived from AOD and other auxiliary predictors as a primary predictor for the hourly modeling instead of using the satellite AOD directly. Based on this, coupling with meteorological high-temporal-resolution data, this study successfully constructed a 1 km, hourly PM2.5 concentration prediction model. Testing on China in 2019, cross-validation results showed that the model was significantly superior to traditional methods in three dimensions (the random 10-fold cross-validation (10CV) R² reached 0.94, the spatial 10CV R² was 0.85, and the temporal 10CV R² was 0.92). The modeling process results indicated that incorporating the daily average variation in PM2.5 is important in capturing the hourly fluctuation characteristics, with a 0.84 correlation coefficient with hourly measurements and ranking top in variable importance analysis. The framework developed in this study realizes the importance of daily pm2.5 in the dynamic downscaling modeling of hourly concentration, providing a theoretical paradigm for building a "daily constraint-hourly response" PM2.5 prediction model, and produces gap-free pm2.5 data with both high spatial and temporal resolution for supporting refined pollution prevention and control and health risk assessment.