Multi-linear Local and Global Preserving Embedding and its Application in Hyperspectral Remote Sensing Image Classification

Traditional manifold learning methods assume that hyperspectral data may reside on one single manifold,but data from different classes may reside on different manifolds of possible different intrinsic dimensions.In order to explore multiple low-dimensional manifolds in hyperspectral images,a multi-manifold learning algorithm based on local and global preserving embedding(LLGPE) is proposed.First,the manifolds of different classes are learned by LLGPE for each class separately,and the data are projected onto low-dimensional spaces.Then,the optimal dimensionality of each class is founded by genetic algorithm(GA) from the viewpoint of classification.At last,classification is performed under a minimum reconstruction error based classifier.The experimental results on the HYDICE hyperspectral data show the effectiveness of the proposed algorithm,when 2,4 and 6 samples of each class are randomly selected for training and 90 samples of each class for testing,the overall accuracy of the proposed algorithm is improved by 3.5%,6.9% and 7.2% respectively,as compared with other methods.