講演情報
[U06-P03]Ensemble Learning for Land Surface Temperature Downscaling Using Multi-Sensor Remote Sensing Data
*Nguyen-Thanh Son1、Chi-Farn Chen1、Cheng-Ru Chen1、Shu-Ling Chen2、Shih-Hsiang Chen2 (1.Center for Space and Remote Sensing Research, National Central University, Taiwan、2.Department of Finance and Cooperative Management, National Taipei University, Taiwan)
キーワード：
Land surface temperature (LST)、Ensemble learning-based downscaling、Multi-source remote sensing integration、Taiwan
Accurate downscaling of land surface temperature (LST) is essential for regional climate analysis, environmental monitoring, and applications such as precision agriculture and water resource management. Satellite-derived LST products often have low spatial resolution, limiting their utility in heterogeneous landscapes. In this study, an ensemble learning-based approach using the random forest algorithm was applied to downscale monthly Landsat land surface temperature (LST) over Taiwan for the year 2024. Sentinel-2 spectral indices and DEM-derived topographic variables were incorporated as predictor features to enhance the spatial resolution. The integration of spectral and terrain information enables the model to effectively capture fine-scale temperature variability associated with vegetation dynamics, land cover heterogeneity, and elevation effects. The downscaled LST products were validated against MODIS LST data, exhibiting strong seasonal fluctuations in model performance. Winter months showed lower errors (RMSE = 2–3 °C) and stronger correlations (r = 0.75–0.79), while transitional and summer periods exhibited higher errors (RMSE = 5–7 °C), and stronger biases, reflecting increased thermal heterogeneity and the influence of vegetation and soil moisture during warmer periods. The cross-validation results also indicated a systematic underestimation of downscaled LST, with an MBE of -3.3 °C, highlighting the importance of multi-metric evaluation for robust assessment. Spatial analysis further demonstrated that ensemble model effectively captures fine-scale temperature patterns associated with variations in land cover, topography, and human activity, providing detailed insights into local thermal dynamics. Overall, these findings indicate that the ensemble learning-based approach offers a reliable and stable approach for downscaling LST across different seasons, producing high-resolution temperature datasets suitable for regional climate and environmental applications. The study emphasizes the value of integrating multi-source remote sensing data and considering sensor-specific characteristics to enhance the accuracy, interpretability, and practical utility of downscaled LST products. Such high-resolution temperature information can support informed decision-making in agriculture, water management, and regional climate adaptation strategies.