SCAGAT: A scene-aware ensemble graph attention network for global PM2.5 pollution mapping via land–atmosphere interactions

Abstract

The sparse and uneven distribution of ground-based air quality monitoring stations poses significant challenges for large scale PM_2.5 pollution mapping. Spatially heterogenous land–atmosphere interactions often lead to large uncertainties in satellite-based PM_2.5 estimations from global modeling strategies. To enhance global PM_2.5 mapping accuracy, particularly in poorly monitored regions, we propose a novel ensemble learning framework called the SCene-Aware ensemble Graph ATtention network (SCAGAT), which integrates locally trained PM_2.5 prediction models across regions using a graph attention network and transfer learning concept. Unlike popular global modeling strategy, SCAGAT first constructs thousands of site-specific PM_2.5 estimation models at individual monitoring station using the random forest (RF) method. For each target grid, raw PM_2.5 estimates are predicted by the 32 site-specific RF models with the most similar geographic scene attributes, characterized by nine variables relevant to haze pollution levels, land cover, and climate characteristic. A graph attention network then aggregates these initial estimates to produce an optimal PM_2.5 prediction through ensemble learning. By taking advantage of the strength of SCAGAT, global daily gap-free PM_2.5 concentrations over land from 2000 to 2021 were finally mapped based on a long-term gap-filled aerosol optical depth dataset. Cross-validation shows that SCAGAT achieves high global PM_2.5 modeling accuracy, with a correlation coefficient of 0.909 and a root-mean-squared error of 9.87 μg m⁻³. Intercomparison results demonstrate SCAGAT’s superiority over other widely used global modeling methods, reducing PM_2.5 modeling bias by 44.2 %, 12.7 %, 32.4 %, 44.4 %, and 48.3 % in China, the USA, Europe, India, and a global product, respectively. Overall, SCAGAT provides a robust solution for large-scale air quality mapping and effectively resolves data imbalance related low accuracy in poorly monitored areas by accounting for geographic scene similarity. Furthermore, this method can be readily adapted to other data-driven Earth observing applications facing similar challenges.