To solve the problem brought by high dimensionality of multi-spectral images during color reproduction, a nonlinear dimensionality reduction method for multi-spectral images was presented. Firstly, according to the characteristics of human visual system, the CIE standard observer color matching functions were weighted to the source spectral reflectance and a Principal Component Analysis (PCA) method was used to the weighted spectrum to effectively improve the colorimetric precision and color difference stability of dimensionality reduction. Then, for the spectral reflectance loss caused by weighting color matching functions, a PCA method was imposed on the lost spectrum to compensate the lost spectral accuracy caused by the improvement of colorimetric precision to effectively improve the spectral precision of dimensionality reduction. Finally the principal components obtained from the first two steps were combined to form the low-dimensional spectral data. Experiments show that the proposed method can offer the average spectral precision in 0.013 9, average colorimetric precision in 0.705 8, and the color difference stability in 1.950 6,which is increased by 14% and 15%,47% and 68%,as well 84% and 82% as comparied those of the PCA cand LabPQR methods. The method outperforms the existing methods in the colorimetric accuracy, spectral accuracy and color difference stability under different illuminants.