Eliminating the soil background in multispectral images with unmanned aerial vehicles (UAV) to improve the inversion accuracy of soil water content (SWC) in crop root zone is an effective method. The winter wheat (in the jointing stage) under different water treatments was used as the research object. Firstly, the UAV-borne multispectral cameras was used to obtain the high-resolution multispectral images at five moments (09:00, 11:00, 13:00, 15:00 and 17:00). Secondly, the improved vegetation index threshold method was used to determine the classification threshold to divide vegetation pixels and soil pixels quickly, and the soil background was eliminated with the classification threshold. According to the threshold changes of the vegetation index threshold method, the effect of soil background on the canopy reflectance was studied. Finally, the inversion models of SWC with vegetation indices were established before and after eliminating the soil background. The research results showed that the improved vegetation index threshold method could eliminate the soil background in multispectral images effectively, and the elimination accuracy of vegetation index RDVI was the highest (the overall accuracy was above 91.32%); the effect of soil background on the canopy reflectance in the near-infrared band was the biggest, followed by it in the red edge band and the effect in the visible light band was the lowest; there was a linear relationship between the vegetation index and SWC before and after eliminating the soil background, and the inversion accuracy of SWC in winter wheat root zone was improved significantly after eliminating the soil background. The performance of NGRDI at the depth of 10~20cm was the best with R2 and RMSE of calibration dataset of 0.739 and 2.0%, and these of validation dataset were 0.787 and 2.1%, respectively.