AI Model Maps PM2.5 Across Africa for Green Industrial Policy
A study available on arXiv (2604.22787) introduces a machine learning framework that integrates satellite observations and reanalysis to assess PM2.5 air pollution levels throughout Africa, aiding the continent's shift towards greener industries. The model utilized 2,068,901 data entries from 404 monitoring sites across 29 African nations (OpenAQ, 2017–2022) and employed LightGBM with spatial cross-validation and conformal prediction techniques. It recorded RMSE = 30.83 ± 5.07 µg/m³, MAE = 14.54 ± 1.66 µg/m³, R² = 0.134 ± 0.023, and macro F1 = 0.336 ± 0.018 during 5-fold location-grouped spatial cross-validation. The low R² indicates challenges in geographic generalization rather than model inadequacy, as random-split benchmarks surpass 0.90. Notably, split conformal prediction aimed at 90% marginal coverage showed significant deterioration in East Africa (actual PICP = 65.3% vs. nominal 90%), aligning with a medium-strength covariate shift in humidity. This system effectively quantifies predictions and outlines their geographic applicability limits, essential for monitoring infrastructure.
Key facts
- arXiv paper 2604.22787 presents a satellite-reanalysis PM2.5 fusion system for Africa.
- Trained on 2,068,901 records from 404 monitoring locations in 29 African countries.
- Data source: OpenAQ, covering 2017–2022.
- Uses LightGBM with leakage-resistant spatial cross-validation and conformal prediction.
- 5-fold location-grouped spatial cross-validation results: RMSE = 30.83 ± 5.07 µg/m³, MAE = 14.54 ± 1.66 µg/m³, R² = 0.134 ± 0.023, macro F1 = 0.336 ± 0.018.
- Random-split benchmarks exceed R² > 0.90, indicating geographic generalization difficulty.
- Conformal prediction targeting 90% marginal coverage shows East Africa degradation (PICP = 65.3% vs. nominal 90%).
- Covariate shift in humidity is identified as a factor in East Africa.
Entities
Institutions
- arXiv
- OpenAQ
Locations
- Africa
- East Africa
- 29 African countries