The work aims to study the selection of training samples for spectral reflectance reconstruction to improve the reconstruction accuracy of spectral reflectance. The similarity between the test sample and the training sample was determined according to their Euclidean distance and the dimensional components of the sample vector were ''normalized'' to the average and the variance to be equal, so that the dimensions respectively met the standard normal distribution, and the reciprocal of the variance was given to the training samples as the weight. In the experiment, the Munsell color card was used as the total training sample set, the samples selected by Mohammadi method, Cao et al method and the method presented hereinwere used as the final training samples, and Color Rendition Chart 24 was used as the test samples. Spectral reflectance reconstruction was performed on the training samples respectively selected by pseudo-inverse method. Through the Matlab software simulation experiment, the average color difference of the proposed method was 0.7918 , the maximum color difference was 1.7148 , the average root-mean-square error was 0.0060, and the maximum spectral root-mean-square error was 0.0127. The selection of training samples based on weighted Euclidean distance can effectively improve the reconstruction accuracy of the spectrum and achieve better color reproduction.