Fusion of All-weather Land Surface Temperature from AMSR-E and MODIS Data Using Random Forest Regression

On the basis of preceding study of microwave (MW) land surface temperature (LST) downscaling, this letter proposed an all-weather LST fusion method based on random forest (RF) and evaluated it using MODIS and AMSR-E LSTs in four areas of China that represent different landscapes. The results show that RF method can effectively avoid the problem of over-smoothing patterns derived by the widely used Bayesian maximum entropy (BME) method and obtained LSTs more consistent with reality. Taking MODIS LST in the Yunnan-Guizhou Plateau (YGP) region and the border of Shanxi and Henan Provinces (BSH) region as reference, the accuracy of RF method improved up to 13% and 11% compared with those of BME method under different cloud proportions. Taking field observations in the Heihe River Basin (HRB) and the Naqu area as references, the accuracy of RF-derived LST under cloudy conditions is basically consistent with that of MODIS LST in clear sky, differing by only 0.004 K to 0.067 K. Due to the introduction of environmental variables, the performance of RF method is more stable than the BME method under different cloud proportions. In summary, RF is promising for fusing MW and thermal infrared (TIR) LSTs.