A total of 191 soil samples were collected from Wensu, Hetian and Baicheng counties for analysis of original reflectances, which were then subjected to continuum removal and first derivative of continuum removal for analysis of hyperspectral absorption characteristics and sensitive bands of salts. On such a basis, a number of hyperspectral inversion models for quantification of soil salt contents (S_t) were established. Results show that there was no obvious relationship between soil reflectance and S_t in 400~2 400 nm bands, and an apparent negative relationship between the data of continuum removal and S_t in 640~700 nm bands and a positive one in 710~780 nm bands. Prediction accuracies of the inversion models were improved significantly after the original reflectances underwent continuum removal. The inversion models established on spectral indices and spectral absorption features were inferior in stability and prediction accuracy to the PLSR models based on spectral reflectance in 400~2 400 nm bands or sensitive bands after continuum removed (CR) and first derivative of continuum removal (FDCR). Among all the models, the inversion models between S_t and FDCR were higher in determination coefficient and lower in root mean square error than the inversion models between S_t and CR. But the former was lower than the latter in prediction accuracy. Among all the inversion models, only the ones based on full spectrum and 640~700 nm continuum removal were higher than 2.5 in RPD, reaching 2.62 and 2.52, respectively and moreover, little difference was found between the two in all the other evaluation indices. The PLSR inversion model based on the continuum-removal data of the 640~700 nm bands is the most effective one in predicting St of paddy soils in Southern Xinjiang Uygur Autonomous Region of China.