The work aims to study a more effective spectral reconstruction method to improve the reconstruction precision. A spectral reconstruction method based on broad learning system was proposed. With Munsell matt color set of 1 269 color blocks and Agfa IT8.2 of 289 color blocks as experimental samples, the proposed method was verified by the simulation system of commercial color digital camera. The root mean square error of spectrum, goodness of fit coefficient of spectrum and two color difference formulas were used as the evaluation indexes to compare the proposed algorithm with the existing spectral reconstruction methods. The experimental results showed that the root mean square error of the method was 0.4%, the average spectral goodness of fit coefficient was 99.9%, and the average color difference was 0.147 and 0.112. The spectral precision and chromaticity precision were significantly better than those of other two methods. The spectral reconstruction algorithm based on broad learning system can effectively improve the precision of spectral reconstruction and meet the requirements of higher precision spectral color representation and reproduction.