Using machine learning and mobile transects to assess microscale variability in wet bulb globe temperature

Wet Bulb Globe Temperature (WBGT) quantifies heat stress by integrating the effects of air temperature, humidity, wind speed, and solar radiation. Although WBGT is typically forecast at coarse (e.g., 2.5 km x 2.5 km) spatial resolutions, each subvariable can vary markedly within a single forecast grid cell. This sub-grid heterogeneity is driven by fine-scale variations in land characteristics and the influence of the built environment. We conducted one of the longest citizen-science campaigns to date, utilizing walking, bicycling, and vehicle mobile transects to gather heat stress data and assess fine-scale variations in WBGT across days with a wide variety of weather patterns across the 2024–2025 heat seasons in Carrboro, Chapel Hill, and South Durham, North Carolina. Random-forest regression models were trained on mobile transect data and land-surface characteristics to generate spatially continuous fields of WBGT-related variables, which were then combined using the Liljegren method to calculate daily WBGT (R² = 0.798). Machine-learning analysis was used both to model WBGT sub-variables (R² = 0.7–0.99) and to identify the land characteristics most strongly influencing local heat stress. Results indicate that there can be extreme heterogeneity in heat stress across development gradients within a study area. The most extreme WBGT difference observed across a single study area on a given day was 8.5 °C and spanned all of the heat vulnerability risk levels, indicating that land characteristics play a large role in the heat stress of a microclimate. This study provided one of the first high-resolution, season-long analyses of WBGT using multi-modal mobile transects and provided insights into how urban form and the built environment shape the magnitude of heat stress.