Abstract: Salinization in frozen soil has posed a serious threat to the emergence and growth of crops in the next growing period. It is of great importance for the accurate detection of the content and composition of soil salt during the freeze-thaw period. Fortunately, the spectral data can be used to monitor the soil salinity during the crop growing period, particularly for the soil under an unfrozen state. However, the monitoring models under the unfrozen state cannot suitable for the frozen soil, due to the variation in the soil reflectance during freezing. In this study, an inversion model was established for the soil water-soluble salt ions in the frozen state using hyperspectral technology. A systematic analysis was made to compare the accuracy of the model in the frozen and unfrozen states. The soil samples were first collected with different salinity gradients from the Jiefangzha Irrigation Area of Hetao Irrigation District in the Inner Mongolia of China. The contents of major water-soluble salt ions (i.e., HCO3-, Cl-, CO32-, SO42-, K+, Na+, Ca2+, and Mg2+) were then measured in the unfrozen soil. The hyperspectral reflectance of soil samples was also measured by the ASD FieldSpec 3 instrument. Secondly, the soil samples were then frozen at -15°C for 12h. After that, the above-mentioned hyperspectral reflectance and ion contents were measured once again after freezing. The raw spectral data was also processed using standard normal variable (SNV) for the subsequent model construction, in order to make the hyperspectral curves smoother. Thirdly, two-thirds of the soil samples were used for the modeling (n = 78), while one-third for the validation (n = 39). The concentration gradient method was utilized to ensure the statistical characteristics of the modeling and the validation sets resembled that of the whole sample set. At the same time, the sensitive spectral intervals of each water-soluble salt ion were selected using variable importance in projection (VIP). The hyperspectral inversion model was formulated for the major water-soluble salt ions content with the sensitive spectral bands using partial least squares regression (PLSR), support vector regression (SVR), and extreme learning machine (ELM). Finally, the performances of these models were evaluated by the determination coefficient of calibration sets (RC2), determination coefficient of prediction sets (RP2), root mean square error (SRMSE), and residual predictive deviation (SRPD). The results showed that the VIP hyperspectral monitoring model managed to invert the most content of the water-soluble salt ions in the frozen soil, but the inversion accuracy of different ions varied greatly. Among them, the prediction accuracies of Cl-and K+ were extremely high with an RPD of above 2.5. There was a reasonably good prediction accuracy of SO42-, Ca2+, and Na+ (2.0