Artificial intelligence–driven air quality mapping and forecasting for extreme PM2.5 events

Abstract

Extreme air pollution events have emerged as a major environmental hazard in rapidly urbanizing regions, driven by climate variability, emission intensification, and complex atmospheric interactions. Fine particulate matter (PM2.5) is of particular concern due to its severe impacts on human health, ecosystems, and climate feedback. This study presents an Artificial Intelligence (AI)-based Air Quality Mapping and Forecasting Program (AQMFP) to characterize and predict extreme PM2.5 events over the Indore urban airshed in central India. A multi-source dataset integrating reanalysis-based PM2.5 reconstructions (1980–2023) with meteorological observations was developed for long-term analysis and forecasting. Temporal feature engineering incorporated lagged variables, while SHAP-based explainable AI identified sustained thermal stagnation and ventilation efficiency as dominant drivers. An ensemble of machine learning and deep learning models was evaluated, achieving strong predictive performance (R2 = 0.80–0.92) after bias correction. Extreme event analysis revealed a consistent intensification across indices, with regulatory exceedances (>40 μg/m3) increasing at 2.06 days per year and persistent multi-day events rising significantly, including during the post-NCAP period. Winter showed the most pronounced deterioration, with Sen's slopes three to six times higher than other seasons. Extreme value modeling indicated shortened return intervals for hazardous concentrations, with 20–50-year events reaching 170-220 μg m-3. The proposed AQMFP framework provides a scalable and policy-relevant tool for forecasting and mitigating extreme air pollution under changing climatic conditions. © 2026 Elsevier Ltd.