The effective sample partition in the near-infrared spectroscopy model is studied.When classical Kennard Stone (KS) algorithm uses the distance metric to describe the difference between high-dimensional spectral data, the effect is unsatisfactory or even meaningless.To solve the problem, and considering the shortcomings of current similarity measurement methods, we constructed a new similarity measure function.The spectral features and property features were combined to calculate the difference between the samples.An improved KS algorithm was thus proposed to find the best expression of sample difference. The improved algorithm was analyzed from the aspects of effectiveness and the impact on the near-infrared spectroscopy model by comparing with other improved methods, and the rationality and superiority of the proposed algorithm were verified.