In order to construct the modeling indicators of soil moisture content in the Mu Us sandy land using multi-source data, totally 17 variables, including microwave backscattering coefficient, surface temperature, silk hat transform factor, band reflectance, drought index and topographic factor were used as modeling factors. PLSR, extreme learning machine (ELM) and random forest (RF) were used to construct soil water content inversion models, verify and compare the models, and map soil water distribution in the study area. The results showed that the determination coefficient of temperature vegetation drought index was 0.64, followed by land surface temperature (0.6),σVV(0.38), vegetation index (0.38), band 7 reflectance (0.35),σVH(0.32), band 6 reflectance (0.3) and Albedo (0.26). Compared with the model constructed with unscreened variables, the accuracy of the model constructed with best subset selection (BSS) variables was improved. PLSR had the best performance in collinearity, and ELM regression model was the most stable. RF model had the highest accuracy, with a determination coefficient of 0.74, root mean square error of 8.85% and mean absolute error of 7.86% in April. In August, the determination coefficient was 0.75, the root mean square error was 8.86%, and the mean absolute error was 7.41%. There was no significant difference in soil water distribution trend between different methods. The highest soil water content occurred in the north and southeast of the study area, and the lower soil water content occurred in the flat area in the central and northern part of the study area. Using spectral index, environmental factor and topographic data, the multi-factor and multi-index comprehensive model can accurately retrieve the surface soil moisture in the Mu Us sandy land, which had reference value for the study of land desertification and ecological environment control in this area.