The urban greening paradox: Measuring day-night perceived safety divergences from paired street view images using generative AI

Source abstract

Whether nighttime is the better half of life varies across cities depending on how safe residents perceive the neighborhood environment. Although street view images (SVIs) are instrumental in auditing perceived safety, city-wide nighttime SVIs do not exist. Consequently, to what extent day-night safety perceptions diverge remains unclear. Although emerging studies use day-to-night (D2N) translations for nighttime SVIs, the model convergence and transferability are unknown. Using paired day-night SVIs from U.S. and Chinese cities, we confirmed the converging size (2,000 pairs) and transferability of mainstream GenAI models. Medium-density SVIs facilitated cross-density transferability, and US-trained models outperformed Chinese-trained ones, indicating a possible influence from American urban planning ideology. Differences in day-night safety perception revealed a 'vegetation safety paradox': green spaces enhance daytime perceived safety but reduce it at night. Daytime perceived safety relies on visibility and openness, whereas nighttime perceived safety shifts dramatically to human activity and brightness. Building and tree views universally influence nighttime safety perceptions, but sky view roles differ significantly between US and Chinese cities. Mapping day-night safety perceptions in Boston, we underscored the significant spatial heterogeneity, especially with many neighborhoods showing a marked decline after dark. This study highlights the overlooked divergence in day-night safety perceptions, extending urban scene studies to nighttime with a scalable framework integrating GenAI and GeoAI to support inclusive nocturnal urban planning.