Balancing prediction accuracy and generalization ability: A hybrid framework for modelling the annual dynamics of satellite-derived land surface temperatures

Annual temperature cycle (ATC) models enable the multi-timescale analysis of land surface temperature (LST) dynamics and are therefore valuable for various applications. However, the currently available ATC models focus either on prediction accuracy or on generalization ability and a flexible ATC modelling framework for different numbers of thermal observations is lacking. Here, we propose a hybrid ATC model (ATCH) that considers both prediction accuracy and generalization ability; our approach combines multiple harmonics with a linear function of LST-related factors, including surface air temperature (SAT), NDVI, albedo, soil moisture, and relative humidity. Based on the proposed ATCH, various parameter-reduction approaches (PRAs) are designed to provide model derivatives which can be adapted to different scenarios. Using Terra/MODIS daily LST products as evaluation data, the ATCH is compared with the original sinusoidal ATC model (termed the ATCO) and its variants, and with two frequently-used gap-filling methods (Regression Kriging Interpolation (RKI) and the Remotely Sensed DAily land Surface Temperature reconstruction (RSDAST)), under clear-sky conditions. In addition, under overcast conditions, the LSTs generated by ATCH are directly compared with in-situ LST measurements. The comparisons demonstrate that the ATCH increases the prediction accuracy and the overall RMSE is reduced by 1.8 and 0.7 K when compared with the ATCO during daytime and nighttime, respectively. Moreover, the ATCH shows better generalization ability than the RKI and behaves better than the RSDAST when the LST gap size is spatially large and/or temporally long. By employing LST-related controls (e.g., the SAT and relative humidity) under overcast conditions, the ATCH can better predict the LSTs under clouds than approaches that only adopt clear-sky information as model inputs. Further attribution analysis implies that incorporating a sinusoidal function (ASF), the SAT, NDVI, and other LST-related factors, provides respective contributions of around 16%, 40%, 15%, and 30% to the improved accuracy. Our analysis is potentially useful for designing PRAs for various practical needs, by reducing the smallest contribution factor each time. We conclude that the ATCH is valuable for further improving the quality of LST products and can potentially enhance the time series analysis of land surfaces and other applications.