Spatiotemporal Fusion of Landsat and MODIS Land Surface Temperature Data Based on Ensemble Methods

Land Surface Temperature (LST) is a key parameter for monitoring the Earth’s surface energy balance and water cycle, playing an important role in assessing environmental changes from local to global scales. Despite advances in remote sensing and the availability of time-series data, limitations in sensor design and the trade-off between spatial and temporal resolution remain major challenges for generating accurate LST time series. To address these issues, various spatiotemporal fusion (STF) methods have been developed. In this study, four ensemble learning algorithms, namely Extreme Gradient Boosting (XGBoost), Gradient Boosting Machine (GBM), Random Forest (RF), and Deep Forest, were used as the core of a spatiotemporal fusion framework to simulate daily LST from Landsat 8 and Landsat 9 imagery. The Enhanced Spatial and Temporal Adaptive Reflectance Fusion Model (ESTARFM) was also employed as a benchmark for comparison. The results showed that XGBoost achieved the best performance, with Root Mean Square Error (RMSE) values of 2.76 K, 1.60 K, 1.68 K, and 1.54 K for 2016, 2021, 2022, and 2023, respectively. Deep Forest showed the highest errors among the ensemble models, while GBM and Random Forest performed at an intermediate level. In addition, ESTARFM produced higher RMSE values ranging from 1.69 K to 2.91 K, indicating lower accuracy compared to the machine learning approaches. Overall, the results demonstrate that XGBoost provides the most accurate and robust performance for LST spatiotemporal fusion, outperforming both other ensemble methods and the ESTARFM model.