Multi-model comprehensive inversion of surface soil moisture based on model averaging method

Soil moisture has been one of the most important factors for crop and pasture growth, depending mainly on global climate change and the water cycle. It is very necessary to accurately monitor the soil moisture for the sustainable management of agriculture and livestock. In this study, a combined inversion of Multivariable Linear Regression (MLR) and Random Forest (RF) Soil moisture model was established to compare with the Inverse Distance Weighting (IDW) spatial interpolation of soil moisture using the Landsat 8 OLI multispectral imagery, surface biophysical properties, and topographic parameters. Field measurement of soil water content was also used to construct an empirical model before that. Two indicators were selected to evaluate the models using the Granger-Ramanathan (GR), including the coefficient of determination (R2) and Root Mean Square Error (RMSE). The inversion of the spatial distribution of soil moisture was then determined using single and combined models. There was a significant correlation between the different biophysical/topographic parameters and the surface soil moisture. In the surface biophysical characteristics, the correlation between the surface temperature, vegetation index, building index, water body index, albedo, brightness, greenness and surface soil moisture were significant. The correlation between the elevation/slope of topographic parameters and the soil surface moisture were extremely significant. There were different sensitivities of biophysical/topographic parameters to the surface soil moisture. Among them, there was the most significant correlation between the surface temperature and surface soil moisture. The interpolation model presented a higher accuracy in April and lower in August. The MLR and RF models presented higher accuracy in August and lower in April. The reason was that the vegetation growth was received a significant effect on the surface temperature in the sand area in August, where the model as the main variable was more sensitive to the prediction of surface soil moisture. The R2 of the RF model was above 0.8 in August and April, which was better than the MLR and IDW model, respectively, whereas the RMSE was the smallest. The RMSE of the RF model was roughly comparable to that of the MLR model, and differed significantly from that of the IDW model, which was lower than those of the IDW model, respectively. The highest inversion accuracy was achieved by the combined model, in which the R2 was above 0.85 in August and April, with the small increase compared with the optimal single RF model. The surface temperature and elevation were the dominant factors to determine the soil moisture content in different months. The overall soil moisture content was low, containing less than 20% (more than 95% of the total area) in most areas. The spatial distribution of soil moisture was closely related to the topographic landscape type, with the higher soil moisture content on the high and steep slopes in the north and southeast areas, whereas, the lower soil moisture content in the flat areas in the north-central region.