Satellite-reanalysis PM2.5 fusion system trained on 2 million+ records from 404 African monitoring locations reveals severe geographic generalization limits and covariate shift in air quality prediction.

A LightGBM model combined satellite data with reanalysis data and conformal prediction, trained on 2,068,901 records from 404 monitoring locations across 29 African countries (OpenAQ, 2017-2022), to map PM2.5 air pollution while quantifying prediction uncertainty.

Under spatial cross-validation (which tests geographic generalization rather than random splits), the model achieved RMSE = 30.83 +/- 5.07 ug/m3 and R2 = 0.134 +/- 0.023—substantially lower than random-split benchmarks (>0.90)—because satellite and weather data availability varies geographically; East Africa showed severe degradation (actual coverage 65.3% vs. target 90%), tied to medium-strength covariate shift in humidity (KS = 0.2237) and atmospheric boundary layer height (KS = 0.2558).

The system operationalizes these limits through regional reliability flags (High/Medium/Low/Unreliable) and a monitor prioritization score to direct infrastructure expansion toward unmonitored populations with highest air pollution burden, supporting Africa's green industrial transition and UN Sustainable Development Goals 3.9, 7.1.2, 9, 11.6.2, and 13.