Machine learning to characterize and explain the fine-scale temporal variability of UHI

While urban heat islands have mostly been studied as climatological phenomena, the temporal variability of their signal hints at significant meso-to-synoptic scale dynamics over periods ranging from hours to days. In addition, health and other impacts of urban overheating tend to be concentrated during regional heatwaves and in specific neighborhoods, and mitigation of these impacts thus also needs to focus on their hot spots and spells, rather than climatic averages. In this talk, we analyze air temperature data from Phoenix, focusing on the UHI signal at hourly and daily scales using 20 years of data from the NOAA ASOS stations. The probability density function of the UHI signal shows significant variability around the mean, with peaks around 12ºC while the mean is only around 2ºC. Spectral and time series analyses show that only about 30% of this time variability is directly linked to the diurnal cycle, with 70 % occurring at scales of hours or multiple days. Attribution analysis using gradient boosting identifies atmospheric transmissivity (clouds, pollution) as the primary driver of variability at these finer scales. The findings, while strictly only applicable to Phoenix, underline the importance of this fine-scale variability and the gap in our understanding of its scales and main drivers.