Hierarchical Multi-Resolution Graphs for Hyperlocal NO2 Mapping: An evaluation using Graph Neural Networks on Amsterdam Road Networks

Abstract

Air pollution varies sharply at fine spatial scales, challenging traditional modeling approaches. This thesis evaluates whether Graph Neural Networks (GNNs) can improve prediction of long-term NO2 concentrations across Amsterdam’s road network using mobile monitoring data. Road segments were represented as nodes with 97 features and connected by spatial adjacency. Baseline GCN and GAT models were compared to a mixed-effects model, and a custom Hierarchical Multi-Resolution Graph (HMRG) was developed to enlarge receptive fields via coarse supernodes. The HMRG GAT model achieved the best internal validation (R2 = 0.762), while external validation showed similar RMSE across models, with GNNs better capturing variability in observed concentrations. Predic- tion maps confirmed that GNNs balance smoothing and hotspot preservation more effectively than mixed-effects baselines. Although challenges remain in stability, extrapolation, and interpretability, the findings highlight GNNs as a promising tool for fine-scale air quality mapping.