Seeds are the basis of the agricultural industry. The viability of seedsis a very important index of seed quality, which is closely related to resistance to biotic and abiotic stress, germination percentage, plant performance, and which decreases with increasing storage period.Increased understanding of wheat seed viability would be beneficial to the wheat industry by ensuring a higher yield for farmers and reducing crop variability. Seed companies would also benefit from enhanced viability by being able to ensure a higher quality product. As the viability of seeds was gradually brought to the public attention, the rapid detection of seed viability without destroying has been a research hot spot. This study aimed at investigating the possibility of using visible and near-infrared (VIS/NIR) hyperspectral imaging (HSI) technique to discriminate viable and nonviable wheat seeds. Firstly, 190 wheat seeds treated by high temperature and high humidity aging (128 germination samples and 62 non-germination samples) were prepared as experimental materials. The visible and near-infrared hyperspectral imaging acquisition system （400~1 000 nm） was constructed to acquire the hyperspectral images of the wheat seeds. After HSI spectra collection of the wheat seeds, a germination test was implemented to check for seed viability. We recorded a seed as germinated (yes=1) if the plumule and radicle were both over 2 mm long, and non-germinated (no=2) if not. The average reflectance data of the region of interest were extracted for spectral characteristics analysis. Secondly, different pre-processing algorithms including the first derivative (FD), orthogonal signal correction (OSC), multiplicative scatter correction (MSC), mean centering (MC) were conducted to build partial least squares discriminant analysis (PLS-DA) model of the viability of wheat seeds. Lastly, three variable selection methods including the uninformative variables elimination (UVE), competitive adaptive reweighted sampling (CARS) and successive projections algorithm (SPA) were used to screen the characteristic wavelengths related to seed viability. PLS-DA models were established by these characteristic wavelengths. The results showed that, the classification accuracies of different pre-processing algorithms were diverse. Among them, the MC method was the best pre-processing algorithm, from which the overall classification accuracy were 82.5% and 83.0%, and the viability classification accuracy were 94.8% and 90.6%, in calibration and prediction sets, respectively. Among the single variable selection methods, UVE method was superior to other two variable selection methods while maintaining an excellent performance of the model for overall classification accuracy (84.6%, 83.0%) and viability classification accuracy (86.5%, 78.1%) in the calibration and prediction sets. This model could promote the germination percentage of the seed lot from 67.4% to 96.2%. Comparing all variable selection methods comprehensively, the UVE-CARS-SPA method selected only 8 variables (473, 492, 811, 829, 875, 880, 947 and 969 nm) from the all 688 spectral variables. The PLS-DA model built by using UVE-CARS-SPA method exhibited the optimal performance with overall accuracy of 86.7% and 85.1% for calibration and prediction, respectively, and accuracy for viable seed was 93.8% and 84.4%. After screening by this model, the germination percentage of the seed lot enhanced from 67.4% to 93.1%. The results indicated that appropriate variable selection could improve the performance of a model, simplify the classification models, and increase the classification accuracy of viable and nonviable wheat seeds. In the future, combining the visible and near-infrared hyperspectral imaging technique with MC-UVE-CARS-SPA-PLS-DA can be used as a feasible and reliable method for the determination of seed viability during the storage. The result can provide the theoretical reference for rapid detection of seed viability during grain storage using spectral information.