The spectral information of the scene is affected by different illumination conditions, hence the spectral reflectance reconstruction of multispectral images taken under scenes with uncontrollable illumination requires illumination spectrum estimation. Therefore, a general method based on a single multispectral image is proposed to accurately predict the illumination spectrum of the scene. First, by analyzing the response features of each pixel, the chroma weight map is designed and calculated to find the pixels that contain more illumination information. Then, the component analysis of the weighted image is carried out to extract the illuminant response features in the channel domain. Finally, benefiting from the innovative introduction of the dictionary learning method trained by illumination spectrum library, the relative spectral power distribution of the scene illuminant can be estimated. The average angular errors of the illumination spectrum estimation obtained by the proposed method on simulated data and real data are 0.29 and 3.42, respectively. Compared with the existing counterparts, the proposed method shows better accuracy and robustness.