The spectrum of rock is a comprehensive reflection of the physical chemistry properties, composition and structure of the rock. Rock spectral data have been applied to the study of rock classification. But unlike the mineral spectrum, the Rock spectrum has no standard database and is influenced by many disturbing factors, for example, mineral composition, structure, chemical composition, weathering strength, the error of measuring instrument, etc. The traditional rock spectrum classification model firstly preprocesses the rock spectrum to eliminate the interference. Then, some spectral features are analyzed by different methods to achieve the classification goal. However, the loss of spectral data features makes the classification of low accuracy and cumbersome operation process; efficiency is not high. Therefore, it is of great significance to establish a simple, fast and accurate automatic classification model of the rock spectrum. Machine learning can learn all the data obtained; there is no omission, greatly improving the classification accuracy. And is the direct operation to the original data, does not need the pretreatment, simplifies the process. Therefore, Xingcheng city of Liaoning Province, China was chosen as the study area, and several typical rock samples were collected. Based on the measured spectral data from the ASD Portable Spectrometer, 608 pieces of data were obtained. According to the spectral characteristics of rocks, the study is divided into three types. Firstly, the decision tree and the upgrade model of the decision tree are used as a the random forest, But when the data noise is large, random forest is easy to get into overfitting. Therefore, the knearest neighbor model, which is not sensitive to outlier is used. But KNN needs to consider every sample when the data is large, the computation will be very large, inefficient. So use Support vector machine to improve classification accuracy. The experimental results show that the order of accuracy of the four classification models is: SVM>KNN>Random Forest>Decision Tree. In order to further improve the automatic classification accuracy of rock spectral features. By fusing several different models. That is to vote on the classification results of different models, choose the most votes as the final classification results. Since hard voting can reduce the occurrence of over-fitting to a certain extent, it is more suitable for classification models. In this paper, we use a hard voting method to fuse three machine learning models: RF, KNN and SVM. The final classification accuracy can reach 99.17%. To sum up, it is feasible, accurate and efficient to classify rock spectral features automatically based on the fusion learning model.