Spatial estimation of particulate matter (PM2.5) in selected African cities using machine learning-based models for improved air quality assessment

Nine (9) out of ten (10) people around the world breathe air that does not meet WHO’s recommended air quality standards. This study aims to create an accurate and scalable machine learning model using satellite-derived Aerosol Optical Depth (AOD) data, temporal and seasonal derived features to estimate Particulate Matter (PM2.5) levels, enhance air quality monitoring and inform interventions for public health and environmental sustainability. The geographical locations considered in this study include Lagos (Nigeria), Bujumbura (Burundi), Nairobi (Kenya), and Kampala (Uganda). Predictive Regression models used in this study are XGBoost, LightGBM, Ridge, Polynomial, and Feedforward Neural Network (FNN). XGBoost emerged as the best-performing model, which achieves an RMSE of 12.01 µg/m³ and an R² of 0.76. Spatial analysis using Local Indicators of Spatial Association (LISA) and Global Moran’s I statistic revealed varying degrees of spatial clustering of PM2.5 concentrations across the cities. Lagos, which exhibits the strongest positive spatial autocorrelation with Moran’s I statistic of 0.686 and Nairobi the weakest with Moran’s I statistic of 0.046. This study shows the effectiveness of combining satellite AOD data with temporal and seasonal variables in enhancing the predictive accuracy of PM2.5 estimation models. It provides critical insights for air quality management and highlights the importance of spatially informed models. This is important to identify localised pollution hotspots for more effective environmental health interventions.