Graph embedded discriminant analysis for the extraction of features in hyperspectral images

摘要

In remote-sensed hyperspectral imagery, class discrimination has been a major concern in the process of reducing the dimensionality of hyperspectral images. Local discriminant analysis (LDA) is a widely accepted dimensionality reduction technique in hyperspectral image processing. LDA discriminates between classes of interest in order to extract features from the hyperspectral image (HSI). However, the drawbacks of its application to HSI is the presence of few libelled samples and its inability to extract an equivalent number of features for the classes in the image, i.e., it can only extract ( c - 1) features provided there are c classes in the image. This paper proposes a new graphical manifold dimension reduction (DR) algorithm for HSI. The proposed method has two objectives: to maximise class separability using unlabeled samples and preserve the manifold structure of the image. The unlabeled samples are clustered and the labels from the clusters are used in our semi-supervised feature extraction approach. Classification is then performed using support vector machine and neural networks. The analysis of the result obtained shows that proposed algorithm can preserve both spatial and spectral property of HSI while reducing the dimension. Moreover, it performs better in comparison with some related state-of-the-art dimensionality reduction methods.