Machine Learning-Based Estimation of Surface PM2.5 in Nigeria Using Integration of Satellite, Ground Observations and Reanalysis Data
TL;DR
It is demonstrated that multi-source machine learning data fusion substantially improves surface PM2.5 estimation over Nigeria with scalable framework and identifies relative humidity as the dominant predictor, followed by MERRA-2 PM2.5.
Abstract
Accurate estimation of surface PM2.5 across Nigeria remains challenging because satellite AOD and global reanalysis products alone cannot adequately represent near-surface particulate concentrations owing to complex aerosol-meteorology interactions, regional transport, and limited ground observations for calibration. This study addresses these limitations by developing a multi-source data fusion framework based on the Random Forest (RF) algorithm that integrates low-cost sensor measurements, satellite observations, atmospheric composition, meteorological variables, and reanalysis data to improve PM2.5 estimation. Ground-based observations from Purple Air and Clarity sensors were combined with satellite-derived aerosol optical depth, atmospheric trace gases, meteorological variables, and MERRA-2 reanalysis products. The RF model was trained and evaluated using a spatial cross-validation (leave location out) framework and assessed using RMSE, MAE, coefficient of determination (R²), Index of Agreement (IOA), and correlation coefficient. The RF model consistently outperformed MERRA-2 across all monitoring stations, yielding substantially lower prediction errors (RMSE: 14-35 µg m⁻³ versus 30-126 µg m⁻³; MAE: 9-31 µg m⁻³ versus 18-89 µg m⁻³) and stronger agreement with observations (IOA up to 0.68). Whereas MERRA-2 produced large negative R² values at several locations, the RF model achieved improved predictive performance, including a positive R² of 0.29 in Lagos and higher correlation coefficients across most stations. Feature importance analysis identified relative humidity as the dominant predictor, followed by MERRA-2 PM2.5, O3, NO2, and AOD, highlighting the combined influence of aerosol hygroscopic growth, atmospheric chemistry, and regional transport. These findings demonstrate that multi-source machine learning data fusion substantially improves surface PM2.5 estimation over Nigeria with scalable framework...