In a multispectral color imaging system, the spectral reflectance of the object being imaged always needs to be accurately reconstructed by employing the training samples on specific color charts. Considering that the workload is heavy when all those color samples are used in practical applications, it is important to select only a limited number of the most representative samples. This is possible as the color charts are usually designed to cover the range of commonly imaged colors, and the color samples are redundant for spectral image reconstruction. We propose an eigenvector-based method and a virtual-imaging-based method for representative color selection by minimizing the total reflectance root-mean-squares errors. The effectiveness of the proposed methods is confirmed by experimental results when compared with existing techniques.