The side effects in spectral multivariate modeling caused by the uneven distribution of sample numbers in the region of the calibration set and validation set were analyzed, and the “average” phenomenon that samples with small property values are predicted with larger values, and those with large property values are predicted with less values in spectral multivariate calibration is showed in this paper. Considering the distribution feature of spectral space and property space simultaneously, a new method of optimal sample selection named Rank-KS is proposed. Rank-KS aims at improving the uniformity of calibration set and validation set. Y-space was divided into some regions uniformly, samples of calibration set and validation set were extracted by Kennard-Stone(KS) and Random-Select(RS) algorithm respectively in every region, so the calibration set was distributed evenly and had a strong presentation. The proposed method were applied to the prediction of dimethylcarbonate (DMC) content in gasoline with infrared spectra and dimethylsulfoxide in its aqueous solution with near infrared spectra. The “average” phenomenon showed in the prediction of multiple linear regression (MLR) model of dimethylsulfoxide was weakened effectively by Rank-KS. For comparison, the MLR models and PLS1 models of MDC and dimethylsulfoxide were constructed by using RS, KS, Rank-Select, sample set partitioning based on joint X- and Y-blocks (SPXY) and proposed Rank-KS algorithms to select the calibration set, respectively. Application results verified that the best prediction was achieved by using Rank-KS. Especially, for the distribution of sample set with more in the middle and less on the boundaries, or none in the local, prediction of the model constructed by calibration set selected using Rank-KS can be improved obviously.