Machine Learning-Based Prediction and SHAP Interpretation of NO2 Concentrations in Kaduna Metropolis, Nigeria

Abstract

Nitrogen dioxide (NO₂) pollution remains a major environmental concern in rapidly urbanizing cities, particularly where ground-based air quality monitoring is limited. This study modelled and interpreted the spatial distribution of NO₂ concentrations across Kaduna Metropolis, Nigeria, using a Random Forest (RF) machine learning approach. Predictor variables, including proximity to industrial sites, proximity to infrastructure, proximity to roads, NDWI, NDVI, land surface temperature, NDBAI, and NDBI, were processed and partitioned into 70% training and 30% independent testing subsets. The RF model showed strong predictive performance, with R² values of 0.964 for training, 0.773 for spatial cross-validation, and 0.817 for independent testing, indicating good model robustness and generalization. SHAP analysis showed that proximity to industrial sites, infrastructure, and roads were the main drivers of NO₂ concentrations in Kaduna Metropolis, highlighting the strong influence of anthropogenic activities, while environmental variables such as NDWI, NDVI, LST, NDBAI, and NDBI had comparatively smaller effects. SHAP dependence plots showed that areas closer to industrial sites and roads had higher NO₂ contributions, while NDVI generally reduced predicted NO₂. LST showed a positive influence. The study offers useful evidence for air quality monitoring, industrial emission control, transport planning, and sustainable urban environmental management in Kaduna Metropolis.