Explainable PM2.5 forecasting for decision support: a Bayesian-optimized extra trees approach for South Asia

This study utilizes a predictive hyperparameter-tuned extra-tree regression (ETR) to estimate PM2.5 levels in 15 major South Asian cities from 2012 to 2024. We incorporated meteorological, land cover, and socioeconomic data to provide a more comprehensive understanding of how pollution varies across spatial and temporal dimensions. Our findings indicate that the mean PM2.5 levels are incredibly high, exceeding 70 µgm−3, and that the pollution levels are highest (125 µgm−3) during the winter. The enhanced ETR model outperforms the baseline decision tree model, with an R 2 of 0.819 for the generic model and city-specific R 2 values (0.882 in Patna), with the lowest observed RMSE of 2.552 (Colombo). An AI expandability analysis using SHAP values shows that surface pressure and vegetation indices are critical predictors. This research shows that South Asia needs rapid air quality regulations and that ensemble models may be able to anticipate pollution, with significance for groundwork for future environmental monitoring efforts.