Improved hourly all-sky land surface temperature estimation: Incorporating the temporal variability of cloud-radiation interactions

Abstract

Land surface temperature (LST) is an indispensable factor for comprehending of surface equilibrium state on the Earth. In particular, satellites can continuously provide LST data and support the large-scale monitoring of LST with a high temporal resolution; however, satellite data may be easily contaminated by clouds. Previous satellitebased all-sky LST reconstruction approaches have inherent limitations, such as low temporal resolution and insufficient consideration of cloud effects. Therefore, this study aims to propose a novel methodology for all-sky 2-km hourly LST reconstruction from GEO-KOMPSAT-2A (GK2A) using machine learning and timely weighted accumulated radiation to reflect the temporal variation of cloud effects. The light gradient boosting machine approach used the European Center for Medium-Range Weather Forecasts Reanalysis-Land variables (i.e., LST, 2 m air temperature, evaporation, and wind), GK2A products (i.e., short and longwave radiation, and binary cloud cover), and auxiliary variables including geographic variables as independent variables. The GK2A LST and in situ measurements were used as dependent variables. The proposed model showed robust spatial agreement with GK2A LST under clear-sky conditions when conducting five-fold spatial cross-validation, with coefficient of determination (R2) values of 0.97-0.99. In the leave one station-out cross-validation using 36 in situ data under all-sky conditions, the proposed model showed high performance with R2 values of 0.86-0.97, root mean square error values of 1.42-2.60 degrees C, and bias values of - 0.49-0.23 degrees C. In a comparison of the proposed model with two scenarios and previous research investigating the effect of accumulated radiation, we demonstrated that the use of accumulated radiation was effective in reconstructing cloudy-sky LST, particularly during the daytime, as evident from the variable analysis conducted through Shapley additive explanations. Using the proposed model, we successfully reconstructed a spatiotemporally seamless LST, which can serve as a fundamental dataset for hourly heat-related research, such as hourly heat flow estimation and urban heat island analysis.