Abstract:With the aim to monitor the salinization of soil salt content timely and effectively, taking Shahaoqu District of Hetao Irrigation Area as study area, the Sentinel-1 image as a data source, synchronous acquisition different depths of soil salinity data, by combining the two groups of radar backscatter coefficient to build a variety of indices, by using gray correlation degree (GCD) index to exclude the index with strong collinearity, and partial least squares regression (PLSR), quantile regression (QR) and support vector machine regression (SVM) were used to construct soil salinity inversion models at different depths of 0~10cm and 10~20cm. The results showed that among the three regression methods the accuracy of SVM regression model was the highest, the model modeling set determination coefficient R2c and the validation set determination coefficient R2p were all above 04, the modeling set root mean square error RMSEc and the validation set root mean square error RMSEp were all less than 03%, QR regression model was the next, PISR regression model was the worst. At each inversion depth, the inversion accuracy of 0~10cm was higher than that of 10~20cm, among which the SVM inversion model was better than other models at 0~10cm depth, R2c and R2p were 0.568 and 0.686, respectively, and RMSEc and RMSEp were 0.201% and 0.151%, respectively. The results could provide a reference for monitoring soil salinization in bare soil stage by radar remote sensing.