Vegetation moisture content is an important characterization of the sensitivity of farmland ecosystem. The spectral reflectance of two vegetation covers, alfalfa and corn were extracted, based on the UAV multispectral image data, and on the basis of which the red-edge band was introduced to calculate the improved spectral indices in order to increase the efficiency and accuracy of the inversion of vegetation water content by near-earth remote sensing. A back-propagation neural network (BPNN) was created after the five spectral bands and 25 indices were filtered by using the variable importance in projection (VIP), gray relational analysis (GRA), and Pearson’s correlation analysis. To find the optimum inversion model for vegetation water content under various crop covers, back-propagation neural network, partial least squares regression (PLSR), support vector regression (SVR), and random forest (RF) were used. The findings indicated that, among the three screening algorithms, the accuracy of the models following GRA and VIP was significantly higher than that of Pearson’s correlation analysis, and the inversion results were less volatile. Among the four machine learning algorithms, the SVR algorithm had a stronger nonlinear problem resolution ability and model robustness than BPNN, PLSR, and RF algorithms. In the nonlinear problem, the SVR algorithm outperformed the BPNN, PLSR, and RF algorithms in terms of analytical ability and model robustness. The validation set coefficient of determination R2 reached above 0.77 and its results can offer more accurate feedback on vegetation water content. The GRA-SVR based inversion model for vegetation water content had the highest accuracy in the two sample sites. The GRA-SVR validation set R2 of alfalfa cover reached 0.889, RMSE of 0.798%, and MAE of 0.533%;the inversion result validation set R2 of corn cover was 0.848, RMSE of 0.668%, and MAE of 0.542%. The research results can provide a theoretical basis for rapid and accurate inversion of vegetation water content.