How does spatial structure affect psychological restoration? A method based on graph neural networks and street view imagery

Abstract

The Attention Restoration Theory (ART) proposed four essential indicators (being away, extent, fascinating, and compatibility) for understanding urban and natural restoration quality. However, previous studies have overlooked the impact of spatial structure (the visual relationships between scene entities) and neighboring environments on restoration quality as they mostly relied on isolated questionnaires or images. This study introduces a spatial-dependent graph neural networks (GNNs) approach to address this gap and explore the relationship between spatial structure and restoration quality at a city scale. Two types of graphs were constructed: street-level graphs using sequential street view images (SVIs) to capture visual relationships between entities and represent spatial structure, and city-level graphs modeling the topological relationships of roads to capture the spatial features of neighboring entities, integrating perceptual, spatial, and socioeconomic features to measure restoration quality. The results demonstrated that spatial-dependent GNNs outperform traditional models, achieving an accuracy (Acc) of 0.742 and an F1 score of 0.740, indicating their exceptional ability to capture features of adjacent spaces. Ablation experiments further revealed the substantial positive impact of spatial structure features on the predictive performance for restoration quality. Moreover, the study highlighted the greater significance of naturally relevant entities (e.g., trees) compared to artificial entities (e.g., buildings) in relation to high restoration quality. This study clarifies the association between spatial structure and restoration quality, providing a new perspective to improve urban well-being in the future.