Assessing air quality in Tehran: Explainable artificial intelligence for megacities

Urban air pollution poses significant public health challenges in megacities like Tehran, where complex emission sources and topographical constraints amplify exposure risks. This study introduces a transparent, interpretable machine learning framework for district-level air quality forecasting, leveraging a decade of high-resolution hourly monitoring data (April 2015-April 2025) across six socio-spatially stratified districts. Using tree-based ensembles (LightGBM and CatBoost), the framework forecasts six criteria pollutants simultaneously, with PM₂.₅-the dominant AQI driver across 83.3% of hours citywide-achieving strong predictive accuracy (MAE: 3.12-8.95 µg/m³; R²: 0.659-0.909) for 1-h-ahead predictions aligned with operational needs for real-time exposure warnings. SHAP analysis reveals district-specific predictive dependencies for next-hour forecasting that reflect Tehran's emission geography: CO/NO₂ dominance in traffic corridors during winter inversions and O₃-driven photochemical activity in receptor basins during summer-without conflating statistical patterns with causal source attribution. EPA NowCast-compliant AQI conversion enables spatially resolved health risk estimation with classification accuracy exceeding 88.9% across districts. The framework transforms existing monitoring infrastructure into actionable forecasting intelligence, offering a globally transferable model for megacities seeking evidence-informed environmental governance. Its modular architecture supports seamless integration of meteorological covariates and remote sensing inputs, promising enhanced accuracy and refined spatial resolution when deployed with richer observational infrastructure across diverse global urban contexts-while maintaining robust performance under current data constraints typical of operational monitoring networks worldwide.