Exploring urban environmental semantics for air quality prediction using explainable multi-view spatiotemporal graph neural networks

Accurate air quality forecasting is essential for urban management and public health, yet it remains challenging due to the complexity of spatiotemporal air pollution dynamics and the interplay of static and dynamic urban factors. Traditional models often neglect the influence of static urban drivers, such as built environment features, which are critical to understanding pollution patterns. Addressing this limitation, we propose a multi-view, multi-modal Spatio-Temporal Graph Neural Network (STGNN) framework that integrates dynamic pollution data with static urban environmental semantics to improve predictive performance and interpretability. By embedding static features, such as Points of Interest (POIs), into the graph structure and leveraging a self-attention mechanism, our model captures complex spatial dependencies and temporal dynamics. Furthermore, an integrated Explainer module enhances transparency by revealing the spatial and feature-level influences driving air quality predictions. Experimental results demonstrate that our approach not only achieves superior predictive accuracy compared to benchmark models but also provides actionable insights into the relationships between urban features and air quality. This study highlights the importance of integrating multi-modal data and interpretability in advancing air quality prediction, offering valuable implications for urban planning and pollution mitigation strategies.