Street-scale heat exposure mapping for urban cyclists: Integrating vision transformer-GAN with aggregated Strava cycling activity data

Street-scale variation in the thermal environment is difficult to represent using spatially coarse meteorological data, limiting the assessment of cycling environments during extreme heat. This study develops a street-scale heat-exposure mapping framework and applies it alongside aggregated cycling activity records in Melbourne. The framework combines street-view imagery and environmental metadata using a Vision Transformer and a conditional generative adversarial network to predict calibrated operative-temperature-equivalent fields. These fields serve as environmental heat-exposure proxies. Direct assessment of thermal comfort and physiological heat stress requires additional subjective and physiological observations. The model achieved a mean absolute error of 3.42°C against calibrated reference label fields on the independent test set. Analysis of 35,416 street-view images identified six landscape types, with higher predicted exposure in dense built-up environments and lower exposure in green-dominated environments. Multi-kernel support vector regression was used to examine the predictive importance of streetscape indicators, while aggregated Strava records from 4 to 13 March 2024 were used to compare cycling activity during the heatwave and adjacent comparison days. The principal contribution is a framework for characterising fine-scale environmental heat-exposure patterns and supporting their investigation alongside cycling activity. Its application informs the screening of potentially heat-vulnerable cycling environments, while the short observation window and aggregate data limit inference about individual adaptation and causal effects.