Bronze surface patina is prone to mixed stacking and it is not easy to examine the distribution of bronze patina, making it difficult to identify harmful patina. At present, conservators mainly rely on human observation to identify patina, which relies heavily on expert experience and is not easy to record. With the development of imaging technology and deep learning, some intelligent methods have begun to be applied to bronze patina identification. However, the bronze training samples are also mainly labelled and extracted by visual observation, and the quality of the selected samples is uncertain. Therefore, this paper proposes a bronze patina identification method based on hyperspectral image unmixing. The method decomposes the overlapping spectral features based on hyperspectral image unmixing to make the extracted spectral features more representative, and combines the spectral library to not only identify specific types of patina, but also select training samples based on the abundance map to provide accurate training samples for the classification task, and improve the efficiency and accuracy of bronze surface patina labelling. The experiments show that the method has good results in the intelligent annotation of bronze corrosion categories.