Abstract: Hyperspectral imaging can provide rich spectral information about land covers, allowing detailed analysis of the materials on the earth. To address the high dimensionality, lack of sufficient labeled samples, and computationally intensive problems involved in hyperspectral image classification, this article presents a simple and efficient method to realize high accuracy classification with a limited training set. Taking advantage of the rich spectral features and the spatially homogeneous property of land covers' distribution, the proposed method firstly employs feature extraction and spatial filtering to reduce the dimension and the noise of the original hyperspectral data, and then uses the proposed neighborhood collaborative support vector machine (NC-SVM) to classify each pixel. The NC-SVM further exploits the spatially homogeneous property of land covers' distribution by combining the discriminant information of neighboring pixels to make reliable class judgement for the central one. Neighborhood collaboration can efficiently reduce the probability of misclassification when a small training set is available. Experiments show that the proposed algorithm can achieve a higher classification accuracy without significant increment in computational cost than several state-of-the-art methods. It can realize fast and high accuracy classification of hyperspectral images with limited training samples.