The work aims to solve the problem of redundancy and heaviness caused by the large number of factors in current spectral reconstruction, so as to prove that the clustering algorithm can be well applied in spectral selection sample analysis, and achieve higher reconstruction chromaticity accuracy and physical accuracy. PCA method was used to carry out simulation experiments. Firstly, the number of principal components was explored and the clustering number was determined. Then, the clustering method and three common sample selection methods were compared. Finally, the effects of types of light source on the reconstruction results were analyzed and compared. When the number of principal components determined by experiment was 6 and the clustering number was 20, the reconstruction effect using KFCM algorithm under light source A was the best. At this time, the mean color difference was 0.35ΔE00, the mean RMSE was 0.0078, and the mean GFC was 99.94%. Clustering algorithm can be well applied to the selection of training samples in spectral imaging process, and help improve the computing speed and accuracy of spectral reconstruction.