A GeoML-XAI framework for identifying high PM_2.5 areas and source attribution: Application to an agricultural...

Air pollution remains a major environmental health concern, with fine particulate matter (PM_2.5) posing significant risks to human health. Although emission sources in urban and industrial areas have been widely investigated, spatial assessment and source attribution in agricultural landscapes remain limited. This study developed a GeoML-XAI framework integrating IoT microsensor networks, machine learning, and explainable artificial intelligence to identify agricultural areas with elevated PM_2.5 concentrations and their associated emission sources. PM_2.5 observations from IoT microsensors were first calibrated against fixed monitoring stations. The selected optimal model, Extreme Gradient Boosting Regression (XGBR), was then constructed incorporating meteorological factors, co-pollutants, and land-use predictors. The model demonstrated high predictive accuracy (R² = 0.93; RMSE = 1.94 μg/m³) and revealed elevated PM_2.5 levels in western and southwestern coastal agricultural regions dominated by dry farmland and rice fields. Explainable artificial intelligence analysis indicated that co-pollutants, meteorological conditions, and agricultural land use, particularly dry farmland, were the primary contributors to PM_2.5 variability. Twenty high-priority emission-control sites were consistently identified in Jiali District, Shanhua District, Liuying District, and Xinshi District. These findings provide a data-driven framework for agricultural emission source attribution and support targeted air-quality management in rural and peri-urban environments.