Observation-driven meta-learning framework for near-real-time mapping of hourly multi-pollutant air quality across China

High-resolution, temporally continuous information on multiple air pollutants is essential for understanding atmospheric processes, tracking emission-control outcomes, and assessing population exposure. Here we present an observation-driven framework for hourly, 1-km, near-real-time reconstruction of PM 2.5 , PM 10 , SO 2 , NO 2 , CO and O 3 and generate the Hourly Air Multi-species Pollutants product for China (HAMP). This framework combines neural-process-based meta-learning, long-range dependency modeling and multi-objective optimization to fuse ground observations with satellite retrievals, meteorology, emissions and geographic covariates. In a temporally independent 2025 evaluation, HAMP reproduced in-situ observations with \({R}^{2}\) values of 0.58–0.95, and reduced RMSE by 59.2–83.5% relative to CAMS analysis data; blind validation at 148 newly established stations confirmed its spatiotemporal transferability. Removing non-target pollutant observations increased RMSE by 17.8–44.9%, showing that cross-species information provides useful constraints. The resulting 2015–2025 dataset reveals strong sub-daily variability and rapid cross-regional transport that daily averages obscure. A modified hourly AQHI-based comparative exposure analysis indicates that the population-weighted metric decreased by 16.03–23.80%, while O 3 became an increasingly important exposure driver.