Adaptive expert-guided deep imbalanced regression for global PM2.5 forecasting with temporal convolutional networks and GEOS-FP inputs

Accurate forecasting of fine particulate matter (PM2.5) remains a global challenge due to spatial gaps, data imbalance, and limited representation of extreme events. This study presents an enhanced Deep Imbalanced Regression (DIR) framework that integrates NASA’s GEOS-FP forecasts with global ground-based PM2.5 observations using a Temporal Convolutional Network (TCN) and a Residual Mixture-of-Experts (ResMoE) architecture. The model was trained on 378,000 samples (2021–2025) from U.S. Embassy AirNow sites and OpenAQ sensors, increasing geographic diversity. To address the imbalance, Label Distribution Smoothing (LDS) and weighted loss were applied, while ResMoE adaptively routed samples to specialized experts across meteorological-aerosol regimes. This configuration achieved strong performance (R² = 0.88, MSE = 23.4, MAE = 2.86 µg/m³) and generalized well across polluted and clean regions, including unseen sites. During the May 2025 Minnesota wildfire, the model captured both temporal evolution and peak magnitude missed by the TCN baseline, demonstrating improved responsiveness to extreme events. Uncertainty quantification and sensitivity analysis confirm model consistency. Beyond forecasting, the framework enables spatiotemporally consistent PM2.5 reconstruction for exposure assessment and policy analysis in data-scarce regions. This study provides a scalable and interpretable pathway for next-generation global air-quality forecasting.