Reliability-Aware PM2.5 Mapping in Africa via Satellite–Reanalysis Fusion and Sparse Monitors

Africa’s sparse ground-based monitoring limits exposure assessment for fine particulate matter (PM2.5) and delays evidence-based air-quality interventions. We present a pan-African PM2.5 mapping pipeline that fuses public ground observations with satellite and reanalysis covariates and produces reliability-aware uncertainty for decision support. Using 2,068,901 quality-controlled PM2.5 records from 404 monitoring locations across 29 African countries (2016–2025), we integrate aerosol optical thickness, satellite NO2, planetary boundary layer height, meteorology, and population density. We evaluate generalization with leakage-resistant 5-fold location-grouped spatial cross-validation that holds out entire monitoring locations. Under this protocol, LightGBM achieves RMSE 30.83 +/- 5.07 ug/m3 and R^2 0.134 +/- 0.023 and provides stronger balance for AQI-style category prediction, while XGBoost yields slightly better regression accuracy. We quantify uncertainty with split-conformal prediction targeting 90% marginal coverage and demonstrate substantial regional heterogeneity, including severe degradation in East Africa consistent with covariate shift. We operationalize these findings with deterministic reliability flags, an uncertainty-and-population-based monitor prioritization score, and out-of-fold SHAP analyses to communicate when and why predictions should not be trusted.