Hyperspectral remote sensing is a cutting-edge technology in remote sensing, which has the characteristics of multi-band and high spectral resolution, so it is increasingly widely used in rock identification and classification. In the current study of hyperspectral rock classification, many rocks are easily confused because of their similar mineral composition, and the classification accuracy is not always high. In the study of high spectral lithology in a wide range of field conditions, there is a lot of interference from the external environment, such as ground cover, pixel mixing and so on, so the spectral characteristics of rocks need to be further studied. The rocks with similar spectra are reclassified. In this study, from the point of view of the laboratory hyperspectral remote sensing system, the HySpex hyperspectral images of 81 common magmatic and metamorphic rock samples were taken as the research data images, and the images were preprocessed such as reflectance correction. Combined with the spectra of rock samples measured by ASD to verify the extraction of corresponding sample spectral curves in the images, the spectral information representing each rock sample was extracted, and the spectral similarity was classified. Finally, the preliminary classification system of 9 large and 28 small categories based on 81 rock samples is obtained. The initial classification system has similar composition properties and spectral characteristics of rock samples in large classes. The spectral characteristics of small classes are more similar than those of large classes. In order to verify the effect of preliminary classification experience on computer lithology classification, the follow-up study is based on the initial classification system of rock samples, and the minimum noise separation technique is used to extract the feature information of hyperspectral images. Finally, the computer classification algorithm model uses the maximum likelihood method and random forest classification, and the training samples set each rock as a single rock book and each subclass in the initial classification system as a sample. Complete the hyperspectral image classification of common magmatic and metamorphic rocks. The experimental results show that the accuracy of maximum likelihood method and random forest classification based on traditional model is 83.21% and 83.63%, while the accuracy of maximum likelihood classification and random forest classification based on initial classification can be improved to 85.46% and 89.39%. Random forest classifier is superior to the traditional maximum likelihood method, while the rock primary classification system has some advantages compared with simple original rock classification. It can be used as a reference for future rock classification work.