Maize subnormal kernels are generated from haploid induction and have lower weight and viability than normal kernels. The germination percentage of subnormal kernels is below 40%. In the present work, a new approach to discriminating the viability of the subnormal maize kernels based on near infrared spectroscopy was developed and the feather selection method based on Kolmogorov-Smirnov (KS) test was applied into near infrared spectra analysis for the first time. The partial least squares model was established and validated with 600 spectral samples from 200 maize subnormal kernels, half of which have the ability to geminate within 3 days through the standard germination test and the other half cannot geminate within 7 days. The spectra were collected by a FT-NIR spectrometer in the diffuse reflectance mode. Ten models established via different preprocessing and feather selection methods were compared. Each model with 1 134 different parameter sets were evaluated through Monte Carlo cross validation. The optimum model was obtained by using 482 wavelengths in the range of 4 027 to 5 500 cm-1 and 6 858 to 9 088 cm-1 via the combination of the smoothing, vector normalization, KS feature selection, and low-signal-to-noise-ratio wavelength elimination. The highest correct discrimination rates for the seeds with germination ability and without germination ability arrive at 92.20% and 84.86%, respectively.