Abstract. Land Surface Temperature (LST) is one of the most important parameters for monitoring surface energy and water balance, playing a crucial role in analyzing environmental changes at local and global scales. Despite significant advancements in remote sensing technologies and the availability of time-series data, limitations in sensor design and the inherent trade-off between spatial and temporal resolution still pose challenges for generating accurate LST time series. To overcome these challenges, various Spatiotemporal Fusion (STF) algorithms have been proposed. In this study, ensemble learning algorithms, including Extreme Gradient Boosting (XGBoost), Gradient Boosting Machine (GBM), Random Forest (RF), and Deep Forest, were employed as the core of an STF framework to simulate daily LST data from Landsat 8/9—an approach not widely explored in spatiotemporal fusion until now. Additionally, the classical Enhanced Spatial and Temporal Adaptive Reflectance Fusion Model (ESTARFM) was included for comparison. The results indicated that XGBoost achieved the highest accuracy in predicting LST, with Root Mean Square Error (RMSE) values of 2.76, 1.68, and 1.54 ˚K for 2016, 2022, and 2023, respectively, outperforming ESTARFM and other ensemble learning methods. Conversely, GBM showed the weakest performance among the ensemble models, with RMSE values of 2.83, 1.74, and 1.69 ˚K. The ESTARFM algorithm produced RMSE values of 2.78, 1.69, and 2.64 ˚K, reflecting a noticeably different performance compared to ensemble-based approaches. Overall, this research highlights the potential and advantages of ensemble learning techniques, particularly XGBoost, in spatiotemporal fusion for enhancing LST prediction accuracy.
Haghshenas et al. (Fri,) studied this question.