Constructing air temperature and relative humidity-based hourly thermal comfort dataset for a high-density city using machine learning

Abstract

Global warming causes new challenges for urban citizens and metropolitan governments in adapting to the changing thermal environment. However, fine-scale spatiotemporal mapping of urban thermal environments has been inadequate. Therefore, this study takes a typical high-density city, Hong Kong, as an example and utilises a machine learning algorithm, the random forest (RF), to carry out 100 m resolution hourly thermal environment mapping, including air temperature (Ta), relative humidity (RH) and the net effective temperature (NET), for the summer season (May to September) of 2008–2018, considering meteorological drivers, topography and local-climate-zone-based landscape drivers. The validation results show that the developed dataset achieves satisfactory accuracy. The mean values of R2, root mean square error (RMSE) and mean absolute error (MAE) for Ta achieve 0.8723, 1.1160 °C and 0.8227 °C, respectively, while those for RH reach 0.7970, 5.3816% and 3.8641%. In addition, the thermal comfort index, NET, reveals that people in built-up areas feel hotter than measured by Ta during the night due to the urban heat island effect. We believe this newly developed thermal comfort dataset can provide novel, reliable and fine-grained data support for urban climate research areas such as urban heat islands, heat exposure, heat-related health risk assessment, and urban energy consumption estimation.