This experiment aimed to study the influence of the training sample selection method and the training sample number on the accuracy of spectral reconstruction. A real six-channel spectral imaging workflow was set up and calibrated where the R-matrix method was employed for spectral reconstruction. The database that consisted of 1687 mineral pigment samples whose spectral reflectance was known was used for spectral reconstruction. The spectral reconstruction accuracy and the computational efficiency of the training sample selection method introduced and proposed in this study were compared with the current values based on the established six-channel spectral imaging systems and the prepared sample set. The experimental results showed that the segment description method of maximum gamut boundary could be well applied to selection of training samples in spectral imaging workflow, which can greatly improve the efficiency of the training sample selection and keep the spectral reconstruction accuracy at the same time. The experimental results also illustrated that the proposed selection method of training samples based on gamut maximization was practical for reproduction-oriented spectral imaging workflow. The results had a certain referencing value for the selection of training samples in spectral imaging process.