Abstract: Hyperspectral image classification methods make best use of spectral information, ignoring objects’ spatial information. This paper proposes a novel classification algorithm based on spectral-spatial information through composite kernels of least squares support vector machine (LS-SVM). The algorithm adopts principal component analysis (PCA) for feature extraction. The selected PCs are used to extract targets’ spatial information by mathematical morphology approach. Different combination strategies are adopted to combine the spatial features with spectral features. Then composite kernels are constructed. By introducing spatial information into classification, the accuracy is improved. Instead of using the standard SVM, the proposed algorithm adopts LS-SVM which changes the quadratic programming problem into linear equations. The algorithm makes the advantages of LS-SVM’s training speed and classification efficiency. The superiors of the algorithm are verified by the hyperspectral experiment. Compared with the algorithms only using one kind of features in classification, the proposed shows some advantages. So it can be applied to large scale hyperspectral image classification.