A Comparison of Spatial and Nonspatial Methods in Statistical Modeling of NO2: Prediction Accuracy,Uncertainty Quantification,and Model Interpretation

${\text{NO}}_2$ is a traffic-related air pollutant. Ground ${\text{NO}}_2$ monitoring stations measure ${\text{NO}}_2$ concentrations at certain locations and statistical predictive methods have been developed to predict ${\text{NO}}_2$ as a continuous surface. Among them, ensemble tree-based methods have shown to be powerful in capturing nonlinear relationships between ${\text{NO}}_2$ measurements and geospatial predictors but it is unclear if the spatial structure of ${\text{NO}}_2$ is also captured in the response-covariates relationships. We dive into the comparison between spatial and nonspatial data models accounting for prediction accuracy, model interpretation and uncertainty quantification. Moreover, we implement two new spatial and a nonspatial methods that have not been applied to air pollution mapping. We implemented our study using national ground station measurements of ${\text{NO}}_2$ in Germany and the Netherlands of 2017. Our results indicate heterogeneous levels of importance of modeling the spatial process in different areas. The prediction intervals predicted with ensemble tree-based methods are more satisfactory than the geostatistical methods. The two new methods implemented each obtained better prediction accuracy compared to the original ensemble tree-based and stacking methods. The probabilistic distribution of the spatial random field estimated by the geostatistical methods could provide useful information for analyzing emission sources and the spatial process of observations.