Assessing children’s perceptions of school-route streets using child-height street-view imagery and explainable machine learning

Abstract School-route streets shape children’s experiences, yet street-view assessments may miss children’s eye-height environments and conflate direct ratings with model-inferred perceptions. Our framework combined reconstructed 1.00 m imagery, children’s ratings, and explainable machine learning at nine primary schools in Harbin, China. Reconstruction was compared with real 1.00 m photographs at 45 sites. Seventy-two children rated 150 scenes for safety, accessibility, comfort, and interest; associations with 15 environmental indicators were analysed separately from proxy-model results. A frozen SigLIP–SVR model generated perception-proxy scores at 1,908 sites, analysed with spatially blocked XGBoost and out-of-fold SHAP across three urban spatial types. Reconstruction performance was mixed; only sky proportion showed a supported improvement under a validity mask. Green View Index had the highest predictive importance (20.26%), while indicator rankings and responses varied by dimension and spatial type. Among the 60 direct-rating tests, 25 remained significant after global BH-FDR correction. Permutation-based hotspot analysis identified six persistent Tier I hotspots and 22 non-persistent Tier III cells, with six Tier I hotspots in the New Planned Large-Block Area. Findings support context-sensitive street assessment, but turning points and hotspots are model-derived screening signals, not validated thresholds or confirmed deficiencies.