An Improved Method for Soil Moisture Monitoring With Ensemble Learning Methods Over the Tibetan Plateau

Abstract

Soil moisture (SM) is a key parameter of the hydrological process, which affects exchanges of water and heat at the land/atmosphere interface. The “trapezoid” (or “triangle”) method has been widely applied to SM monitoring based on the pixel distribution within the thermal and optical remote sensing observations. However, the trapezoid method is a linear empirical model highly related to the retrieval accuracy of the surface temperature. In the article, the moderate-resolution imaging spectroradiometer (MODIS) data were applied to retrieve SM through an improved method over the Tibetan Plateau. The improved method is integrated with the “trapezoid” model and multiple learning techniques, Random Forest (RF) and Extreme Gradient Boosting (XGBoost). Meanwhile, RF and XGBoost were both trained with SM target data (the scale of SM and soil temperature) derived from the Tibetan Plateau observations, and the input variables were derived from MODIS observations. Compared with the SM measured, the results showed the root mean square error, the mean absolute error, and the correlation coefficient of the ensemble retrievals were 0.046–0.081 m3m−3, 0.030–0.065 m3m−3, and 0.60–0.87, respectively, which is better than that of the separate model. The ideas to implement the combination of traditional inversion algorithms and machine learning methods are helpful for researches in remote sensing fields.