Downscaling Land Surface Temperature by Using Random Forest Regression Algorithm

This study proposes a land surface temperature (LST) downscaling method to downscale the LST of Moderate Resolution Imaging Spectroradiometer (MODIS) from 990m to 90m by using random forest (RF) regression algorithm. The LST product of Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER), with 90m resolution, serves as the validation reference at the finer scale. The proposed method is based on the relationship between LST and a variety of surface parameters including band reflectance, spectral indices, land cover types and terrain factors. The proposed downscaling method is evaluated in Segovia, Spain, the Pefiarora mountain region. Comparison between downscaled LST and referenced LST proves that the proposed method shows a great accuracy in downscaling LST. Furthermore, another downscaling method, an algorithm for sharpening thermal imagery (TsHARP), is also implied to get finer resolution LST to make a more complicated comparison with the proposed method. The results are evaluated by root mean squared error (RMSE) and bias, which demonstrate that the accuracy and robustness of RF downscaling method compared with TsHARP.