In order to build a robust model for predicting water content in fresh pork based on hyperspectral imaging technology, the influence of sample set partition, spectral preprocessing and wavelengh selection on model prediction result was discussed. The pork sample set was parted by concentration gradient(CG) firstly, and then the reflectance spectra was pre-processed with multiple scattering correlation(MSC), first derivative and autoscale successively. The partial least square regression (PLSR) model was built, which had the best prediction abilities. The water content values were predicted with the cross-validation and prediction correlation coefficients (Rc and Rp) of 0.814 and 0.804, with the root mean square error (RMSECV and RMSEP) values of 0.726% and 0.686%, respectively. The feature wavelengths were identified by using competitive adaptive reweighted algorithm. The proposed PLSR model was again built by using the feature wavelengths, which had remarkable prediction abilities with Rc and Rp of 0.926 and 0.924, with RMSECV and RMSEP of 0.467% and 0.438%, respectively.