High Spatiotemporal Resolution Modeling of PM2.5 in West Africa Using Satellite Data and Machine Learning

Exposure to ambient fine particulate matter (PM_2.5) is a leading environmental risk factor for premature death. In Africa, surface PM_2.5 data is sparse, hindering pollution mitigation plans and human health improvement. NASA satellite observations provide near-complete spatial coverage, but their columnar nature is imperfect representations of surface pollution. To estimate PM_2.5 concentrations at high spatiotemporal resolution (1 km², daily) across West Africa over the past two decades, we trained, tested, and fine-tuned a machine learning (XGBoost) model with the following data: PM_2.5 from reference-grade and calibrated low-cost monitors, aerosol optical depth from MODIS MAIAC satellite retrievals, five meteorological features from ERA5, and seven tropospheric trace gas column or aerosol property features from TROPOMI/OMI satellite retrievals. Preliminary results show that the model performs reasonably well (r² = 0.73, mean absolute error = 11 µg m^-3) in predicting daily PM_2.5 compared to observations in six cities across West Africa. Likewise, we will develop a machine learning model focused on Ghana for more localized epidemiological and environmental justice studies. These novel PM_2.5 datasets will enable us to identify and explain long-term trends, cycles (seasonal, weekly, and diurnal), and significant sources of PM_2.5 in both urban and rural areas. As air quality monitoring networks expand across Africa, the spatial predictions are expected to improve.