Global PM2.5 exposure forecasting with novel deep learning architecture and explainable artificial intelligence

Abstract

Particulate matter (PM) of fine size (≤2.5 μm) remains one of the most significant global environmental risk factors for early mortality and morbidity, and more than 90% of the global population currently lives in areas exceeding the World Health Organization 2021 guideline value of 5 μg/m3. This study introduces a temporally constrained transformer-based forecasting model to anticipate annual population-weighted PM2.5 exposure across 204 countries and territories between 1990 and 2020, aimed at supporting evidence-based air quality and climate policy development. The framework is based on a filtered dataset from the State of Global Air, comprising 6,323 country–year observations with harmonized exposure estimates and uncertainty bounds, allowing the model to capture long-range temporal variations and enduring heterogeneity among countries in exposure trends. A time-aware expanding-window cross-validation approach was strictly implemented to prevent information leakage and ensure realistic predictive conditions. Five-fold temporal validation demonstrates strong performance across geographical locations, with mean squared error ranging from 0.00043 to 0.00115, root mean squared error from 0.0207 to 0.0339 μg/m3, and mean absolute error from 0.0094 to 0.0193 μg/m3, with Nash–Sutcliffe efficiencies exceeding 0.95 on average. Continental-scale evaluation shows similar high accuracy in Europe and Oceania (root mean squared error <0.01 μg/m3; R2 > 0.98), while systematically higher errors are observed in Asia and Africa, which bear a higher pollution burden. The attention-weight inspection offers clear decompositions of temporal trends and country-specific patterns that drive predictions. The proposed framework is, therefore, a methodological and practical addition to transformer-based environmental forecasting and policy-relevant global health-risk assessment.

Topics & Concepts