A machine learning algorithm for converting land surface temperature to air temperature and testing in the determination of the urban heat island effect over the city of Rome

Air temperature (Ta) plays a crucial role in numerous applications, including studies on physical stress conditions and understanding phenomena such as urban heat island (UHI). Ta measurements, acquired from in situ sensors often distributed unevenly, are limited in describing the spatial temperature field pattern. On the other hand, land surface temperature measurements (LST) obtained from geostationary satellites provide a more detailed spatial overview, but represent a different variable. In this work, a method based on machine learning algorithms is presented for converting LST detected from geostationary satellites MSG, into air temperature. To perform the conversion, a gradient boosting algorithm, which is part of the tree-structured family of machine learning algorithms, was implemented. The method is applied to LST and Ta data available for the city of Rome (Italy) during the summers of 2019 and 2020. The Ta data are sourced from 17 weather stations, predominantly consisting of amateur stations whose quality has been verified. Using predictive variables such as instantaneous LST and with delays ranging from 1 to 4 hours, along with other parameters like altitude, imperviousness, land cover, tree cover, grassland, NDVI, and temporal parameters such as time of day, Ta was estimated, designated as the target variable, at points where no in situ measurement sensors are available. The Ta predicted by the model exhibits an average error of 1.2°C during the daytime and 0.8°C at night. This model output has improved the accuracy and spatial resolution of temperature pattern analysis across the city of Rome, compared to analyses based solely on in situ measurements. Furthermore, the spatiotemporal pattern of the UHI, which can now be measured at high resolution, aligns well with the expected pattern.