Abstract: The high water content of iron ore will reduce its machinability, which is not conducive to the smooth progress of mineral processing, sintering, smelting and tailings treatment. Therefore, it is very important to control the water content of iron ore reasonably for improving mining production efficiency, reducing energy consumption and reducing waste of raw materials. However, due to the complexity of iron ore composition and properties, traditional detection techniques (such as loss on drying method and resistance method) have shortcomings in sensitivity and accuracy. Based on this, this paper selects three kinds of Anshan-type magnetite from a certain area in Tangshan, Hebei Province, and tests its hyperspectral data under different water contents (0%-40.0%), and applies S-G smoothing filtering (S-G), multivariate scattering correction (MSC), standard normal transformation (SNV), second derivative (SD), reciprocal logarithm (LR) and envelope removal. In order to further improve the prediction ability of the model, the competitive adaptive reweighting method (CARS) is used to optimize the characteristic band, and a prediction model is established by combining random forest regression (RFR), least square support vector regression (LSSVR) and particle swarm optimization least square support vector regression (PSO-LSSVR). The prediction effects of different magnetite water content models are compared, and finally the best model is selected to improve the accuracy of water content detection in mineral processing and smelting. The results show that: (1) when the water content of Anshan-type magnetite samples with different particle sizes changes, the change trend of their spectral curves is generally consistent, and the reflectivity is negatively correlated with the water content, and it shows obvious absorption characteristics around 990nm, 1440nm and 1920nm; The Pearson correlation coefficient (r) of spectral data pretreated by MSC and SNV can reach -0.950(412nm) and -0.964 (421nm) respectively. (2) Among the three models, PSO-LSSVR model is the most stable, and SNV-CARS-LSSVR model with granularity of 0.3-0.5mm and MSC-CARS-PSO-LSSVR model with granularity of 0.5-2mm are preferred. The prediction set determination coefficient (R²) of the models is 0.778 and 0.789, and the root mean square error of prediction set (RMSEP) is 5.45% and 5.41%. Compared with previous studies, this paper constructs a more stable water content prediction model of Anshan magnetite by combining data preprocessing, CARS feature screening and nonlinear regression algorithm, which provides higher precision support for water content detection in mining production.