Modelling spatiotemporal concentrations of PM2.5 over Nigerian cities using machine learning algorithms and open-source data

Increasing attention has been drawn to the concentrations of particulate matter in cities because of the consequent health burden and environmental impacts. Due to its importance, particulate matter has been integrated into the UN SDG 11 as a target for monitoring and fostering sustainable communities. However, the paucity and irregularity of data pose a challenge to achieving the SDGs. This study aims to investigate the concentrations of particulate matter (PM2.5) in Nigerian cities and compare the predictions using machine learning models and open-source data, including satellite-derived data. The influence of meteorological factors, population growth, and human activities on PM2.5 emissions in eleven locations in Nigeria was investigated. The algorithms are linear regression (LR), K nearest neighbor (KNN), decision tree regression (DTR), support vector regression (SVR), artificial neural networks (ANN) and CatBoost (CBT). Hyperparameter optimization of the models was carried out by an exhaustive search of possible values and fivefold cross-validation. The results showed that the SVR, CBT, and ANN mostly predict PM2.5 to be more correlated to the actual targets than the KNN, DTR and LR during the training and test phases. The CatBoost is the best predictor with a root mean square error (RMSE) of 9.88 whereas decision tree regression had the highest error with an RMSE of 15.75. Precipitation made the highest contribution to the CatBoost prediction model, followed by temperature and nighttime light. The findings can be used in the management of particulate matter concentrations in the context of ground data paucity.