Downscaling MODIS land surface temperature over a heterogeneous area: An investigation of machine learning techniques, feature selection, and impacts of mixed pixels

Remotely sensed Land Surface Temperature (LST) is of paramount importance in numerous environmental applications. Although, coarse spatial resolution sensors provide frequent LST measurements, their applicability is rather limited for many applications. Downscaling methods are therefore applied to improve the spatial resolution of LST products. A number of Machine Learning (ML) methods have already been used in the LST downscaling studies. Nevertheless, the literature lacks a suitable inter-comparison of different ML methods, as well as the impact of the feature selection process on downscaling results. This study aims at comparing downscaled 1ST from 1 km daily MODIS 1ST product (MOD11A1) to 240 m using Random Forest Regression (RFR), Support Vector Regression (SVR), Extreme Learning Machine (ELM) and Temperature Sharpening (TsHARP) approaches with 11 predictor variables at a heterogeneous area in different seasons. In addition, by implementing a feature selection with the Support Vector Machine-Recursive Feature Elimination (SVM-RFE) method, the most important variables were selected and used as inputs of the models. The results were evaluated against the 1ST derived from Landsat-8 thermal imageries using a split window method, showing that all the ML methods perform well in 1ST downscaling (with an average RMSE = 2.5 and MAE = 1.74) with marginal differences, outperforming the TsHARP method (RMSE similar to 3.02, MAE similar to 2.19). Among all the methods, ELM required the least computational effort, and when it was combined with SVM-RFE, general efficiency of the downscaling procedure was increased substantially.