Remote sensing lithology mapping is of great significance to basic geological research and mineral exploration. In this study, ASTER data was used to compare and analyze a variety of machine learning algorithms, combining spectral features with spatial features, and a high-precision automatic lithology classification method was proposed. In this method, watershed segmentation algorithm is introduced into the extreme learning machine (ELM) model to extract the spatial features of images, and the maximum voting method is combined with the spectral feature-based classification results of standard ELM. At the same time, principal component analysis and regularization methods are used to reduce the feature space dimension, simplify the hidden layer structure, and improve the generalization ability of the model. The experimental results show that ELM has the advantage of no need to adjust parameters, and the classification accuracy and speed are better than SVM, maximum likelihood method and Mahalanobis distance method. The watershed segmentation algorithm was introduced to extract spatial features, which effectively reduced the wrong zoning and improved the lithology identification accuracy. The classification accuracy of the watershed + principal component regularization limit learning machine model can reach 92.3%, and the time is short, which has good popularization value for the automatic classification of lithology in the exposed area.