AI-Empowered Research on Healthy Streets: An Iterative Path of Streetscape Perception, Evaluation, and Optimization

Abstract

This study presents a closed-loop AI framework for assessing and enhancing “healthy streets,” applied to downtown Haining City, China. First, semantic segmentation of street-view images yields eight quantitative indicators (e.g., greenery, enclosure, openness). Second, an Elo-based pairwise survey combined with Random Forest and CNN models reveals how these indicators predict perceived street healthiness. Third, we use a Low-Rank Adaptation model (Lora) in Stable Diffusion generator to optimise low-scoring scenes by adding health-supportive elements. Results indicate that (1) the eight indicators capture key environmental qualities; (2) Elo-scoring aligns with machine-learning feature importances—especially greenery, human-scale enclosure, and visual diversity; and (3) the generative model improves predicted health scores by enhancing greenery and openness. This iterative approach offers planners a rapid tool for healthy-street design.