There are many training sample selection methods based on multi-channel imaging system. The selection of training sample is closely related to the accuracy of spectral reconstruction. Aiming at the problem that the number of training samples is subjective, this paper proposes a semi-supervised neighbor propagation clustering algorithm based on entropy weight method. Firstly, the nearest neighbor propagation algorithm determines the clustering center according to the reference degree. It does not need to specify the number of clusters and has high computational efficiency. Secondly, the close clustering centers are combined to output the fixed cluster number and determine the new clustering centers. Finally, the independent weight of each spectral reflectance in each class was determined, and the spectral reflectance sample with the largest weight was selected to participate in spectral reconstruction. The experimental results show that the proposed training sample selection method based on weighted distance is involved in the reconstruction of spectral reflectance. Compared with other methods, the spectral accuracy is increased by 34% and the chromaticity accuracy is increased by 16%, achieving good color reproduction effect.