Deep learning assessment of street spatial quality in old residential communities of Wuchang, Wuhan, China.

Perceptions of old residential communities—urban neighborhoods typically built before 2000 with compact layouts and limited facilities—reflect residents’ intuitive experiences of their surrounding physical environment and spatial atmosphere. Traditional assessment methods have mainly emphasized residents’ behaviors and facility use; however, with ongoing community renewal and rising living standards, more perception-oriented approaches are required. This study employs street view images (SVIs)—panoramic photographs captured from eye-level perspectives—together with deep learning and semantic segmentation to quantitatively evaluate street spatial quality and propose optimization strategies. A high-resolution SVI dataset was established for Wuchang District, Wuhan, and 19 street elements were automatically extracted. Six subjective perception dimensions—boring, beautiful, depressing, lively, safe, and wealthy—were predicted using the Place Pulse 2.0 dataset, a large-scale crowdsourced visual perception database developed by the Massachusetts Institute of Technology (MIT). To link subjective and objective indicators, Local Moran’s I spatial autocorrelation and Geographic Information System (GIS) mapping were used to identify perception clusters, while Shapley Additive Explanations (SHAP) analysis quantified the influence of visual features on perception outcomes. Results demonstrate strong correlations between street elements and perceptual dimensions, with clear spatial variations across neighborhoods. This research contributes a reproducible, interpretable, and culturally adaptive framework for evaluating street spatial quality. By integrating global perception data with localized validation, the study offers practical guidance for improving both the functionality and aesthetic experience of streets in aging urban communities, advancing inclusive and sustainable neighborhood renewal.