Hourly air temperature mapping in Guangdong province utilizing machine learning

Air temperature is a crucial variable in urban climate and relevant to many studies, such as urban heat islands, heat waves, climate change, energy consumption, and health-related heat exposure risk studies. Previous studies used land surface temperature (LST) and inversion methods to obtain air temperature maps with spatial detail or used weather station observations and spatial interpolation to obtain air temperature maps with high temporal resolution. However, fine spatial detail and high temporal resolution have not been resolved simultaneously. Moreover, there are differences in LST and air temperature definitions, which cannot be equated. Therefore, in this study, we carried out hourly air temperature mapping at 1-km resolution over a multi-year summer period for Guangdong Province, China, employing machine learning algorithms as well as meteorological and landscape data. The meteorological data were hourly observations from 86 weather stations in Guangdong containing variables such as air temperature, relative humidity, precipitation, barometric pressure, and wind speed. The landscape data were mainly from the landscape indices calculated based on local climate zone (LCZ) maps, mapped via Google Earth Pro and Google Earth Engine. Then, we employed the random forest (RF) algorithm for the hourly air temperature mapping. The validation results showed that the hourly air temperature maps achieved good accuracy from 2008 to 2019 with a mean R² value of 0.8001. The importance assessment of the driving factors showed that meteorological factors, especially relative humidity, make the most outstanding contribution to air temperature mapping. Simultaneously, the landscape factors also played a non-negligible role. Further analysis revealed that the maps steadily maintained high accuracy at nighttime (20:00-7:00), which is the most critical period for studying urban heat islands. In addition, the air temperature patterns showed a correlation with the landscape. Air temperatures in contiguous mountainous areas with dense trees were significantly lower than those in the plains. Moreover, there is a correlation between nighttime air temperature changes and urban morphology, and urban-rural differences exist simultaneously. Air temperatures tend to fall more slowly in the core of metropolitan areas than in the urban fringe. Overall, this study employed machine learning to reliably improve the temporal resolution of air temperature mapping with more spatial detail. Furthermore, it reveals spatially explicit air temperature patterns in and around cities at different times in a day during the summer. In addition, it provides a new valuable and advantageous dataset for relevant applications.