Applicability Assessment of Passive Microwave LST Downscaling over Semi–Homogeneous Desert Underlying Surface Based on Machine Learning

The spatial and temporal resolution of remote sensing products in land surface temperature (LST) studies can be improved using the downscaling method. This is a crucial area of research as it provides basic data for the study of climate change. However, there have been few studies evaluating the applicability of downscaling methods using underlying surfaces of varying complexities. In this study, we focused on the semi–homogeneous underlying surface of Gurbantunggut Desert and evaluated the applicability of five classical, passive microwave, downscaling methods based on the machine learning of Catboost, using 365 days of AMSR–2 and MODIS data in 2019, which can be scanned once during the day and night. Our results showed four main points: (1) The correlation coefficients between feature vectors and the LST of the semi–homogeneous underlying surface were clearly different from those of the surrounding oases. The correlation coefficient of the semi–homogeneous underlying surface was high, and that of the surrounding oases was low. (2) At the same frequency, the correlation coefficient between vertically polarized BT and LST was greater than that between horizontally polarized BT and LST. Considering the semi–heterogeneous underlying surface, 23.8 GHz and 36.5 GHz may be more suitable for passive microwave LST retrieval than 89 GHz according to physical mechanisms. (3) The fine–scale LST downscaling accuracy achieved with all BT channels of AMSR–2 was higher than that achieved with the other four classical models. The day and night RMSE values verified with MYD11A1 data were 2.82 K and 1.38 K, respectively. (4) The correlation coefficients between downscaled LST and the soil temperature of the top layer of the site were the highest, with daytime–nighttime R2 values of 0.978 and 0.970, and RMSE values of 3.42 and 4.99 K, respectively. The all–channel–based LST downscaling method is very effective and can provide a theoretical foundation for the acquisition of all–weather, multi–layer soil temperature.