Global comparison of diverse scaling factors and regression models for downscaling Landsat-8 thermal data

Statistical downscaling of land surface temperature (SDLST) algorithms with diverse scaling factors and regression models have been used to produce high spatial resolution LSTs based on Landsat-8 LST. However, the optimal choice of scaling factors and regression models and their associated combinations over various land surfaces, especially from a global perspective, remain unclear and even controversial. To cope with this issue, we compare 35 SDLST algorithms derived from a combination of seven scaling factors and five frequently used regression models over 32 geographical regions worldwide. The seven scaling factors, at varying degrees, make use of the LST-related information embedded within the visible and near-infrared and short-wave infrared bands of Landsat-8 data. The five regression models involved are multiple linear regression, partial least squares regression, artificial neural networks, support vector regression, and random forest (RF). Our main findings are: (1) The performance of the scaling factors and regression models are highly dependent on each other. Nevertheless, for most scaling factors, especially for high-dimension scaling factors with numerous LST-related variables, the downscaling algorithms that use RF as the regression model achieve the highest accuracy. (2) RFT21, a newly proposed SDLST algorithm based on the comparison results and further optimization, has high global operability and sufficiently high accuracy. RFT21 requires only Landsat-8 data as the inputs, and achieves the highest accuracy in comparison with the thermal sharpening (TsHARP) and high-resolution urban thermal sharpener (HUTS) algorithms, with the mean root-mean-square error (RMSE) reduced by more than 15%. These findings will facilitate the generation of high spatial resolution LSTs worldwide and associated applications.