Semi-supervised Classification For Large-scale Hyperspectral Data

The rapid development of hyperspectral remote sensing bring probalility for precise classification of surface features. For hyperspectral data, unsupervised classification method can not obtain good classification results, while the classifier trained by supervised classification method has bad generalization ability due to the considerable cost of labeling samples. Therefore, semi-supervised method which can utilize both labeled and unlabeled samples has become a hot spot. And the graph-based has attracted widespread concern because that it can effectively reflect the intrinsic character and manifold structure of data. However, high spectral feature dimensions and high data scale lead to computing and storage pressure for traditional graph-based algorithm. In this paper, several semi-supervised classification methods for large-scale hyperspectral data based on anchor graph regularizer are proposed to solve this problem:(1) A hyperspectral image classification algorithm based on fuzzy anchor graph regularizer is proposed. The number of selected anchor is much less than the total number of samples, which can effectively reduce the computing and storage cost of traditional graph-based semi-supervised method. Firstly, the cluster centers and membership matrix are obtained by using space neighbor based kernel weighted fuzzy C-means clustering. Then cross similarity matrix is calculated and used to constructe fuzzy anchor graph regularizer. By solving the optimization problem of objective function, the labels of anchors are acquired and used to calculate the labels of unlabeled samples according to the linear relationship. The experimental results show that the proposed algorithm has better classification accuracy compared with other algorithms.(2) A hyperspectral image classification algorithm based on local fuzzy anchor graph regularizer and relaxed clustering is proposed. Based on the algorithm of the previous chapter, the cross similarity matrix is used to cluster hyperspectral data into several clusters which are divided into more super-pixels by spatial distance. Then, each super-pixel and its neighborhood super-pixels form a set of super-pixels, and samples in this set are used to calculate the similarity matrix which will be used to construct a local fuzzy anchor graph regularizer. The relaxed clustering assumption is introduceded into our algorithm to solve the misclassification caused by mixed pixels. It's need to calculate the labels of anchors and probability vector iteratively when solve optimization problem of objective function. Finally, we solve optimization promblem of objective function on each super-pixel set to get anchors' labels and unlabeled samples' labels. The experimental results show that the proposed algorithm has high accuracy and local consistency.(3) The distributed implementation of hyperspectral classification algorithm based on fuzzy anchor graph regularizer is proposed. The data are allocated to subsystems by random sampling and regular sampling. Then, select the anchors by using space neighbor based kernel weighted fuzzy C-means clustering. Different distributed algorithms are achieved according to different sampling strategies. The experimental results show that the distributed approach is available and can be implemented.