A close look at using national ground stations for the statistical modeling of NO₂

Abstract. Air pollution causes a manifold of negative health and societal problems. It is therefore essential to model and predict air pollution over space. An increasing number of statistical models of air pollution have been developed using geospatial variables associated with air pollution emission and dispersion processes. However, the increasing number of air pollution models does not always equate to an increase in prediction accuracy and uncertainty reduction. An important aspect that is often disregarded is the spatial heterogeneity. In this study, we aim to evaluate and compare various spatial and non-spatial statistical and machine learning methods, with attention given to different spatial groups. Spatial groups are identified by the predictor variables. We found that prediction accuracy differs substantially in different spatial groups. Predictions in places close to roads with high populations show poor prediction accuracy, while prediction accuracy increases in low population density areas for both local and global models. Prediction accuracy is further increased in places that are far from roads for global models. This division into spatial groups also shows that global non-linear methods are capable of higher prediction accuracy than global linear methods. The spatial prediction patterns show that non-linear methods generally predict more smoothly than linear methods. Additionally, clusters of predicted air pollution differ within and between cities. Lastly, applying the same methods to the local dataset yields poor metrics, especially for the non-linear methods.