Nationwide spatiotemporal modeling of air pollutants in Japan using generalized random forest models

• Growing demand for efficient and accurate estimation of diverse air pollutants • We developed generalized random forest models for multiple air pollutants. • The models of PM2.5, NO2, and Ox achieved high performance at high resolution • Models using only IDW-interpolated values demonstrated comparable performance • We demonstrated the potential of using IDW-interpolated values as input variables Recent advancements in mixture modeling of air pollutants and their health impacts have underscored the importance of estimating a wide range of air pollutants with high spatiotemporal resolution. This study aimed to develop high-performance and computationally efficient models for estimating various air pollutants across Japan. We constructed separate generalized random forest (GRF) models to predict daily mean ambient concentrations of particulate matter ≤2.5μm in diameter (PM 2.5 ), nitrogen dioxide (NO 2 ), sulfur dioxide (SO 2 ), carbon monoxide (CO), non-methane hydrocarbons (NMHCs), total hydrocarbons (THCs), and photochemical oxidants (Ox) at a spatial resolution of 0.74 km² from 2010 to 2022. Input variables included inverse-distance-weighted (IDW) interpolated pollutant concentrations, satellite data, emission inventories, land use, traffic patterns, meteorological data, and population density. Additionally, we developed simplified GRF models using only IDW-interpolated pollutant concentrations. Model performance was evaluated using the leave-one-location-out cross-validated statistics. Approximately 2,000 monitoring stations were distributed across Japan. For PM 2.5 , NO 2 , and Ox, both the full and reduced models demonstrated strong predictive performance, characterized by low median absolute errors and root mean squared errors, as well as high R 2 values (0.91, 0.71, and 0.89, respectively) at the daily level. Notably, reduced models relying solely on IDW interpolation achieved predictive accuracy comparable to that of the full models, suggesting that in regions with dense monitoring networks, cross-pollutant information may be useful to maintain strong model performance without incorporating additional environmental covariates. GRF models for other pollutants showed lower performance. In conclusion, we developed high-performing and computationally efficient GRF models for PM 2.5 , NO 2 , and Ox.