The Research Of Manifold Learning For Hyperspectral Image Classification

The hyperspectral remote sensing images have detailed space geometry information and abundant spectral information, the characteristics of large amount of data, high dimensions and strong redundancy. To classify hyperspectral remote sensing image directly is easy to fall into "dimension disaster". Therefore, It becomes a key problem that how to extract useful information of discriminant characteristics features to improve the classification accuracy of hyperspectral sensing images.Based on the nonlinear property of hyperspectral remote sensing data and mainly from the perspective of manifold learning, this paper makes an in-depth research in feature extraction and classification of hyperspectral remote sensing data. Related research work is as follows:① To sum up and introduce the traditional linear feature extraction algorithms, the local manifold learning algorithms, commonly used classification algorithms and evaluation criteria of classification results;② Semi-Supervised Laplace Discriminant Embedding algorithm for hyperspectral sensing image classification is researched and proposed in this paper. As the traditional global linear feature extraction algorithms can’t find the local manifold structure of the data set, and with ignoring class information of labeled samples, the purely local manifold learning algorithm destroys the separability of data set easily. Thus, this paper presents a Semi-Supervised Laplace Discriminant Embedding(SSLDE) algorithm based on protecting the separability of sample set and manifold structure characteristics. The proposed algorithm makes use of the class information of labeled samples to maintain the separability of sample set, and discovers the local manifold structure in sample set by constructing Laplace matrix of labeled and unlabeled samples, which can achieve semi-supervised manifold discriminant. The experimental results on KSC and Urban database show that the algorithm has higher classification accuracy and can effectively extract the information of discriminant characteristics;③ Semi-Supervised Bundle Manifold Learning algorithm for hyperspectral sensing image classification is researched and proposed in this paper. Based on theoretical ideas of bundle manifold learning, and multi-class characteristics of hyperspectral remote sensing data, this paper presents a Semi-Supervised Bundle Manifold Learning(SSBML) algorithm. The algorithm makes use of labeled samples and unlabeled samples to construct two neighborhood graphs which are used to keep "whole" structure(the relationship between the various sub-manifolds) of bundle manifold in the data set and intrinsic structure characteristics within each sub-manifold, which can achieve semi-supervised bundle manifold learning. The experimental results on KSC and Pavia U hyperspectral database show that the algorithm can efficiently discover subtle character of the bundle manifold structure in hyperspectral remote sensing database, and enhance the hyperspectral remote sensing image classification accuracy.In summary, based on the nonlinear characteristics of hyperspectral remote sensing data, this paper proposed two new feature extraction algorithms, and verify the effect of feature extraction by doing the classification experiments on hyperspectral remote sensing data.