Objective The spectral image was chosen as a test sample, to discuss the effects of different number and distribution of training samples on the reconstruction of spectral image. Methods The spectral reflectance of ColorChecker SG, Color Rendition Chart and Munsell color cards were chosen as training samples and were analyzed by principal component analysis (PCA). And spectral image reconstruction was carried out using the principal component extracted by the analysis to color cards. Results Experimental results showed that the spectral image reconstructed using the seven principal components of the color card of ColorChecker Color Rendition Chart (24) had the highest precision, and the color difference was smaller than those of the other two color cards, with a largest color difference of less than 3. Conclusion The reconstruction precision of the spectral image did not increase with the increasing training sample number and distribution range under the same reconstruction condition. The reconstruction precision of red and purple was relatively low for the three kinds of training samples.