Conformal PM2.5 Mapping Under Spatial Covariate Shift: Satellite-Reanalysis Fusion for Africa's Green Industrial Transition

Honest uncertainty quantification is critical for real-world deployment: The system uses conformal prediction to flag when predictions are unreliable rather than presenting false confidence. East Africa showed only 65.3% coverage versus the intended 90%, which the authors transparently reported and flagged as "Unreliable"—preventing decision-makers from trusting inaccurate air quality estimates in regions where monitoring is most needed.

Geographic generalization is harder than standard benchmarks suggest: The model's R² of 0.134 under realistic spatial cross-validation (where training and test locations are geographically separated) is far lower than the >0.90 achieved with random data splits. This demonstrates that responsible AI evaluation must test how models perform in genuinely new geographic areas, not just randomly shuffled data.

Covariate shift—when environmental conditions differ between regions—undermines model reliability: The system identified that humidity and atmospheric boundary layer height differ significantly between training and test regions (Kolmogorov-Smirnov test values of 0.22-0.26), explaining why predictions fail in East Africa. Acknowledging these shifts prevents overconfident deployment in dissimilar environments.

Transparent limitations enable better resource allocation: Rather than deploying an unreliable system uniformly, the authors created regional reliability flags and a monitor prioritization score to direct infrastructure investment toward high-pollution areas with the worst data coverage—directly supporting public health and environmental justice in undermonitored African regions.