A close look at using national ground stations for the statistical modeling of NO ₂

Abstract. Air pollution leads to various health and societal issues. Modeling and predicting air pollution over space have important implications in health studies, urban planning, and policy-making. Many statistical models have been developed to understand the relationships between geospatial data and air pollution sources. An important aspect often neglected is spatial heterogeneity; however, the relationships between geographically distributed variables and air pollutants commonly vary over space. This study aims to evaluate and compare various spatial and non-spatial statistical modeling (including machine learning) methods within different spatial groups. The spatial groups are defined by traffic- and population-related variables. Models are classified into local and global models. Local models use air pollution measurements from the Amsterdam area. Global models use ground station observations in Germany and in the Netherlands. We found that prediction accuracy differs substantially in different spatial groups. Predictions for places near roads with high populations show poor prediction accuracy, while prediction accuracy increases in low-population-density areas for both local and global models. The prediction accuracy is further increased in places far from roads for global models. Modeling of air pollution in different spatial groups shows that nonlinear methods can have higher prediction accuracy than linear methods. The spatial prediction patterns of global models show that nonlinear methods generally are less sensitive to extreme values compared to linear methods. Additionally, clusters of predicted air pollution differ between models within cities despite similar prediction accuracy. Also, the influence of predictors on NO2 concentrations varies across different cities. Using the local dataset of our study and explicitly accounting for spatial autocorrelation in the universal and ordinary kriging models does not improve accuracy; however, analyzing prediction performance across spatial groups provides valuable insights. Comparing local and global prediction patterns reveals that local models capture regional clusters of high air pollution, which are not detected by global models. These findings highlight the fact that solely relying on overall prediction accuracy can be insufficient and potentially misleading, underscoring the importance of considering spatial variability and model performance within different spatial groups.