Generative AI and causal-spatial modeling for understanding nighttime pedestrian risk in urban systems

Urban environments exhibit intricate spatial and visual structures that complicate efforts to model and interpret human-environment interactions. This complexity challenges our ability to capture how perceptual and spatial factors jointly shape movement and exposure patterns in cities, particularly under low-light conditions. However, most spatial analytics frameworks lack high-resolution nighttime visual data and rely on context-insensitive models that overlook spatiotemporal heterogeneity. This study developed an integrated computational framework that combines Generative AI-based visual synthesis with causal-spatial machine learning to characterize urban visual environments and their influence on pedestrian crash risk. To this end, a citywide dataset of AI-synthesized nighttime street view imagery was established using a CycleGAN model fine-tuned on local Hong Kong data, enabling a realistic representation of street-level illumination and visibility. The derived visual metrics were then incorporated into a geographically weighted causal random forest integrated with double machine learning to estimate localized causal effects while addressing confounding and nonlinearity. The results reveal that visual overload from dense built structures increases pedestrian crash risk, whereas clear spatial boundaries and balanced lighting contrast reduce hazards in low-light conditions. Nighttime visibility of walking paths shows stronger protective effects than overall brightness, suggesting that excessive illumination may impair rather than enhance safety. The study highlights the value of combining generative visual analytics with causal-spatial inference for large-scale, perception-aware analysis of urban environments. The proposed framework advances computational urban analytics and provides transferable tools for data-driven, context-sensitive street design and pedestrian safety planning. • AI-synthesized nighttime imagery for citywide pedestrian risk modeling. • Causal machine learning integrated with geographically weighted modeling. • Visual overload from dense infrastructure linked to higher daytime risk. • Path visibility outperforms brightness in modeling nighttime perception. • Localized causal effects reveal spatially uneven pedestrian vulnerability patterns.