The work aims to study a more effective training sample selection method to improve the reconstruction accuracy of spectral reflectance. The sample was proposed to be regarded as a vector, and the similarity between the test sample and the training sample was determined according to the angle between them. Then, the included angle was used as the similarity weight of the training sample. The experiment used the Munsell color card as a sample set, and the training samples were selected by Mohammadi method and the proposed method. With color difference and spectral root mean square error as the evaluation indexes, the two sample selection methods were compared and verified from two aspects: reconstruction accuracy and the effectiveness of sample selection. Through the Matlab software simulation experiment, the average color difference of the proposed method could be reduced to 0.7945, the maximum color difference was 2.1569, the mean spectral root mean square error was reduced to 0.011 42, and the maximum spectral root mean square error was 0.0218. The sample selection based on vector angle is simple and accurate, which can meet the requirements of high precision color reproduction and provide a reference for how to select samples quickly and accurately and improve the accuracy of color reproduction.