Rock spectrum is the comprehensive embodiment of rocks physical and chemical properties, composition and structure. Now, it has been widely used in rock classification research. Due to the difficulty of collecting the data on the rock spectrum, it often needs to be collected manually, which not only causes great labor cost but also leads to the limited data on the rock spectrum collected. When the rock spectral classification model is trained with a limited number of samples, the dimensional disaster phenomenon will generally occur. That is, the accuracy of classification will decrease with the rise of the feature dimension, and the rock spectral data coincides with this feature, with a high dimensional number of features. Therefore, to achieve good classification results, a large number of training samples are needed to be used in the training of traditional rock spectral classification models, usually more times than the feature dimension. If the number of samples is small, we must reduce the features to obtain the ideal classification accuracy. Therefore, when the number of samples is small, obtaining a more accurate classification effect on rock spectral data has become a hot research topic. This paper collects the spectral data of typical rocks in Xingcheng, Liaoning Province. Based on the Python programming language, the Siamese Network classification model is constructed with few training samples, and the Triplet Loss is used as the loss function to realize the 3-way-1-shot classification model, and the prediction accuracy of 97.8% is achieved in the verification set. At the same time, four traditional machine learning methods, which include Decision Tree, Random Forest, Support Vector Machine and K-Nearest Neighbor, were used to establish the classification model under the same training samples and compared with them. By drawing the learning curve, it is verified that these four traditional machine learning methods do not have good classification functions in the case of small samples. Since converting the original spectral data into image data will not affect the classification effect of the Siamese Network classification model, the rock spectral classification problem can be transformed into the problem of image classification. Then the image classification methods and means can be used. The experimental results show that the Siamese Network classification model in the case of fewer rock spectral samples can still achieve excellent classification effect, which effectively makes up for the shortcomings of the traditional machine learning model in the case of small samples. Because the data input is paired, it can effectively reduce the overfitting problem caused by too few training samples.