Model Estimation of Ground-based PM2.5 Concentration Over Nigeria and Its Assessment Using MERRA-2 Reanalysis

International audience Accurate estimation of fine particulate matter (PM2.5) remains a major challenge in data-sparse regions such as West Africa because of limited ground-based monitoring networks and highly variable atmospheric conditions. This study evaluated statistical and machine-learning models for predicting ground-level PM2.5 concentrations across seven monitoring stations representing diverse ecological zones in Nigeria. Predictor variables included satellite-derived aerosol optical depth (AOD), meteorological parameters, gaseous pollutants, and temporal features. Ordinary Least Squares (OLS), Random Forest (RF), Extreme Gradient Boosting (XGBoost), Long Short-Term Memory (LSTM), and stacked ensemble models were developed and evaluated using a consistent time-based training and testing strategy. Model performance was assessed using the coefficient of determination (R2), root mean square error (RMSE), mean absolute error (MAE), correlation coefficient (r), bias, normalized mean bias (NMB), and the Seasonal Stability Index (SSI). The results revealed substantial spatial and seasonal variability in model performance across Nigeria. Random Forest consistently achieved the highest predictive accuracy, producing the lowest regional mean RMSE (10.56 µg m⁻3) and ranking as the best-performing model at four of the seven monitoring stations. OLS demonstrated competitive performance in several locations, indicating that linear relationships remained important under certain environmental conditions, whereas XGBoost and LSTM generally exhibited lower predictive performance. In contrast, the MERRA-2 reanalysis dataset showed considerably larger prediction errors than the developed models. Seasonal analysis further demonstrated that model performance varied across ecological zones, with greater instability observed in the Sahel and more consistent predictions in the Guinea Coast. Overall, the findings demonstrated that ensemble tree-based machine learning models provided robust and reliable PM2.5 predictions in Nigeria and outperformed conventional statistical models and coarse-resolution reanalysis products. The proposed framework provides a practical approach for improving air quality assessment, exposure estimation, and evidence-based air pollution management in Nigeria and other data-sparse regions.