In recent years, several supervised learning methods have been introduced in hyperspectral image (HSI) classification. However, these methods use only spectral information without taking into account the spatial features and manifold structures of HSIs. To overcome this problem, a new classification method was proposed for HSI classification, combining spatial-spectral features and manifold reconstruction. Based on the spatial consistency of ground objects distribution in HSIs, the proposed algorithm used a small number of labeled samples and large number of unlabeled spatial neighbor samples to perform semisupervised learning, and utilized the reconstruction error of test samples in each submanifold to represent the similarities for discriminant classification. Experimental results obtained from the Indian Pines and University of Pavia data set reveal that the proposed method exhibits a higher classification accuracy compared to other classification algorithms under various training conditions, the highest overall accuracy achieved in the two cases being 95.67% and 91.92%, respectively. The proposed method integrates spatial-spectral information to represent the submanifold structure of different land objects, exhibits superior discrimination performance, especially for a small number of training samples, and effectively improves the performance of HSI classification.