Downscaling Land Surface Temperatures Using a Random Forest Regression Model With Multitype Predictor Variables

In this paper, a random forest regression model with multitype predictor variables (MTVRF) was utilized with four kinds of input variables, including surface reflectance, spectral indices, terrain factors, and land cover types, to establish the nonlinear relationship between land surface temperature (LSTs) and other land surface parameters. The main objective of this paper is to analyze the superiority of MTVRF model in multivariable regression wherever on the simple or complex underlying surface and further to demonstrate the robustness of random forest (RF) regression downscaling model trained in one study area while being applied to another area. The spatial resolution of the Moderate Resolution Imaging Spectroradiometer LST product was downscaled by MTVRF from 990 to 90 m. A comparison with two other downscaling methods, such as the basic RF model and the thermal sharpening algorithm, was also made. By computing the mean error, the determination coefficient (
$R^{2}$
), and the root mean square error (RMSE) between the downscaled and referenced LSTs, the MTVRF model achieved a satisfied performance. Further satisfactory results were also obtained for the MTVRF to downscale LSTs for different land covers and evaluate the training model in various regions. The RMSE of the MTVRF model trained on study area B and evaluated on study area A was 3.13k, while the RMSE trained on study area A and evaluated on study area B was 2.11k; this shows the MTVRF model trained in a specific region is thought to be robust enough to downscale LSTs under other various surface conditions.