The current major multispectral image dimensionality reduction methods (principal component analysis, LabPQR, WSPCAplus) cannot meet the need of the multispectral image high-fidelity reproduction. This paper researched a multispectral image dimensionality deduction method for high-fidelity reproduction. Based on the facts that binary wavelet decomposition of the signal matches the human vision characteristics and the nonnegative principal component analysis method can better ensure the spectral accuracy of the dimension reduction image, a composite dimensionality reduction method based on the discrete binary wavelet change and the nonnegative principal component analysis was put forward. Based on the spectral accuracy, chroma precision and chromatic aberration stability of changing light source, the standard color difference of CIELAB, the spectral fidelity and the image average gradient were proposed to evaluate the dimensionality reduction efficacy. After the multispectral image test, the composite dimensionality reduction method based on the discrete wavelet transform and nonnegative principal component analysis could better ensure the spectral accuracy, chroma precision and image definition when compared with the other three methods. This method could better realize the multispectral image high-fidelity production, besides providing a new theoretical explanation for the color vision cognitive process.