Soil moisture content is an important indicator that reflects the coupled surface water-heat-solute transport in arid regions. The visible and near-infrared spectroscopy has been widely used for soil moisture content prediction owing to its rapid response. The soil moisture content and corresponding spectral data are obtained in the laboratory; then, the calibration datasets (n=77) are selected using Monte Carlo cross-validation algorithm. The competitive adaptive reweighted sampling algorithm is used to optimize spectral variables. Three machine learning algorithms, namely back propagation neural network, random forest regression, and extreme learning machine are used to construct predicting models. The results reveal that competitive adaptive reweighted sampling algorithm can effectively filter and eliminate massive irrelevant variables. Herein, a total of 20 feature bands are divided from all spectral bands, where the band of R1848 is the most prominent (the maximum correlation coefficient is 0.531). The performance of models based on machine learning algorithms is superior to those based on partial least squares regression, with the optimal prediction of the coefficient of determination (R2), root mean square error of prediction (RMSE), residual predictive deviation (RPD), and ratio of performance to interquartile range (RPIQ). Compared with the predictive effects of all the models, the extreme learning machine-based predicting model is the most effective (R2=0.918, RMSE=0.015, RPD=3.123, and RPIQ=3.325). Compared with common linear models, the machine learning algorithms can effectively improve the precision and stability of the quantitative estimation of soil moisture content. The results provide scientific guidance and baseline data for the accurate monitoring of soil moisture content and precision agriculture in arid regions.