Developing an efficient training sample selection method is one of the goals of spectral reflectance reconstruction. In spectral reflectance reconstruction, the training set selection method and sample capacity are strongly related to the reconstruction accuracy. The accuracy of the spectral reconstruction is affected by the unstable clustering results due to the randomness of the starting value selection and the outliers. K-means clustering has low computing complexity and great computational efficiency. Based on this, this paper proposes an improved K-means clustering training sample selection method. Firstly, the geometric center of the training sample set is used as the initial value of the clustering center; secondly, the probability density function of the spatial distribution of the samples is constructed based on the Gaussian function, and the Euclidean (Euclidean) distance is used as the measure of other clustering centers; finally, the similarity between the spectral reflectance samples in the training sample set is measured based on the intra-cluster squared difference, and the sample with the closest distance to the center in each clustering subset is used as the training samples are used to verify the effectiveness of the method. The spectral reconstruction was performed by principal component analysis. The experimental results show that the proposed method has been improved significantly compared with the traditional method, the average root-mean-square error of the reconstructed spectra is less than 4%, and the CIEDE2000 color difference is less than 3.756 7. The improved training sample selection method of K-mean clustering proposed in this paper can improve the spectral reconstruction accuracy to some extent and meet the requirements of reproducing the reproduced images.