Mapping high-resolution spatio-temporal patterns of pedestrian thermal comfort at different scales using street view imagery and deep learning

Abstract

Accurate evaluating pedestrian-level thermal comfort is important for improving resident well-being and mitigating urban heat island (UHI). This study proposes a framework for calculating high-resolution pedestrian-level thermal comfort at both neighborhood and urban scales. We identified the geolocations and three-dimensional (3D) structure of street trees using triangulation and deep learning models of mask region-based convolutional neural network (Mask R-CNN) and Monodepth2 from street view imagery. By incorporating meteorological forcing data, land use classification, building information and digital elevation model (DEM), the mean radiant temperature (MRT) and universal thermal climate index (UTCI) were obtained. To explore the feasibility of the framework, we calculated the spatio-temporal UTCI patterns in Shenzhen, China at 2 m and 50 m resolutions at 9:00, 12:00, 15:00 and 17:00 during heatwave and non-heatwave periods, along with local climate zone (LCZ) classification. Results indicated that the proposed framework not only accurately identified the high-resolution pedestrian-level thermal comfort in neighborhoods, but also rapidly captured the spatio-temporal distribution of pedestrian thermal comfort across the city. We also found that the thermal comfort in high-rise areas (LCZs 1 and 4) was better than in open low-rise areas (LCZ 6). These findings contribute to guiding sustainable and resilient urban development.