A Machine Learning Method to Correct the Terrain Effect on Land Surface Temperature in Mountainous Areas

In mountainous areas, land surface temperature (LST) shows significant terrain effect, which can be directly reflected by the spatial distribution associated with the change of topographic factors (elevation, slope, and aspect). By the way, the terrain effect diminishes the impacts from the differences in surface water and heat fluxes, and influences their comparison or estimation over complex terrain. In this study, a practical way to reduce the terrain effect is proposed based on the random forest method with datasets from MODIS products, which is used to build a LST prediction model instead of the previous model developed based on some numerical model or empirical method. The results indicates that the constructed LST model shows a good performance in predicting LST with the R2 of 0.93 and the RMSE lower than 2.0 K for four selected days. Corrected LST maps are compared with the original LST map, which presents a preliminary correction results with an obvious correction on pixels with significant terrain effect.