Laser-induced breakdown spectroscopy has the characteristics of small-invasive, in situ and rapid analysis. It has wide application prospects in the field of sample identification and component analysis. In order to explore the feasibility of the technology in the automatic identification of natural geological samples, a method of identifying and sorting LIBS spectral of natural geological samples by self-organizing feature map neural network combined with correlation is proposed in this paper. In order to reduce the interference of unrelated data such as background noise in the whole spectrum and the computational complexity, the feature spectral line is extracted on the basis of elemental to achieve the dimensionality reduction of high dimensional spectral data. The network training model is established by using the feature spectrum data as input, then the weight vectors which have the feature of input samples are obtained. Finally the geological sample classification is achieved by the correlation analysis between the weight vectors and the samples to be tested. The classification results of the 16 kinds of natural geological samples prove that the feature spectrum is superior to full spectrum and PCA dimension reduction, especially in the aspects of descending dimension and extracting the main features of data. The algorithm proposed in this paper has a better classification effect on the feature spectrum data of 16 samples than SVM and SOM neural network algorithm. Moreover, the validity of the proposed method is initially verified in this paper.