The soil salinity is essential for the morphological development, growth process and final yield of crops in the irrigation area. With present methods, satellite remote sensing though was noninvasive, dynamic, rapid and macroscopic, estimated soil salinity of soil covered by vegetation have less significant effect, yet. In order to improve the estimation effect, soil salinity at different depths at Hetao Irrigation Area was collected. GF-1 image simultaneous was downloaded as the data source. Best subset selection was used to analyze the sensitivity of different bands and spectral indices to soil salinity at different depths. RMSE, R2, AIC and SIC were used to determine the optimal combination mode of the sensitive independent variables number at different depths. Based on these, artificial neural network (ANN), support vector machine (SVM) and quantile regression (QR) were used to construct soil salinity inversion model at such depths as: 0~20cm, 20~40cm, 0~40cm, 40~60cm and 0~60cm before and after best subset selection. The determination coefficient for calibration set before best subset selection (R2c0), determination coefficient for calibration set after best subset selection (R2c1), determination coefficient for verification set before best subset selection (R2v0), determination coefficient for verification set after best subset selection (R2v1), root mean square error for calibration set before best subset selection (RMSEc0), root mean square error for calibration set after best subset selection (RMSEc1), root mean square error for verification set before best subset selection (RMSEv0) and root mean square error for verification set after best subset selection (RMSEv1) were used to evaluate the effects of the models. The results showed that B4, BI, SI1 and SI3 were sensitive variable combinations of soil salinity at depths of 0~20cm and 0~40cm. B4, BI and NDVI were sensitive variable combinations of soil salinity at depths of 20~40cm, 40~60cm and 0~60cm. QR inversion model showed its good performance because of its strong robustness. With R2c1 and R2v1 were both above 04, and RMSEc1 and RMSEv1 were both under 04%; followed by SVM, and ANN was the worst. Compared with other depths, the QR inversion model performed best at depths of 20~40cm, with R2c1 of 0611, R2v1 of 0.671, RMSEc1 of 0.177%, and RMSEv1 of 0.160%. The combination of best subset selection and QR method in the modeling analysis of soil salinity provided a new approach to optimize the satellite multispectral model and quickly measure the soil salinity. The research result provided a reference for the widescale soil salinity monitoring of soil covered by vegetation.