A Machine earning Algorithm to Convert Geostationary Satellite LST to Air Temperature Using In Situ Measurements: A Case Study in Rome and High-Resolution Spatio-Temporal UHI Analysis

Accurate air temperature (T_a) measurements are essential for analyzing phenomena such as the urban heat island (UHI) effect, which can lead to critical conditions in cities during summer. However, in-situ sensors offer limited spatial coverage due to their uneven distribution. In contrast, satellite-derived land surface temperature (LST) provides detailed spatial information but does not directly correspond to T_a. This study introduces a machine learning approach to derive T_a from LST using data from geostationary satellites, with Rome, Italy, as a case study. A gradient boosting algorithm, trained on T_a observations from 15 weather stations, was applied. The model incorporated variables such as instantaneous and lagged LST (1–4 hours) alongside other factors to estimate T_a in areas lacking direct measurements. The predicted T_a achieved an average error of 0.9°C, with a spatial resolution of 3 km and an hourly temporal resolution. This dataset allowed for a more detailed investigation of UHI intensity and dynamics during summer, significantly improving both spatial and temporal resolution compared to previous studies based solely on in-situ observations. The findings also indicate a slightly higher nocturnal UHI intensity than previously reported, likely due to the inclusion of rural areas with minimal impervious surfaces, made possible by the comprehensive T_a mapping now available across the study domain.