Subspace Detection Approaches for Hyperspectral Image Classification

Abstract

Hyperspectral data provides rich information and is very useful for a range of applications from ground-cover types identification to target detection. With many benefits they also present some challenges including high storage cost, intensive computational load and difficulties in machine assisted interpretation, namely, in classification. The limited number of training samples may cause a significant loss in classification accuracy. This thesis investigates effective and feasible approaches to reduce the dimensionality of the hyperspectral images while keeping the intrinsic structure of the input data intact. The first study is concerned with finding a subspace which consists of the most informative features for reliable hyperspectral image classification. In this study, a hybrid approach which combines both feature extraction and feature selection is proposed. Principal Component Analysis (PCA) is applied first to generate new features from the complete set of the original spectral bands. Feature selection is then performed effectively using a normalized mutual information measure with two constraints to maximize the general relevance and minimize redundancy to the target class identification in the selected subspace. Improvement of the existing nonlinear feature extraction method is undertaken in the second study. In this study, the input features are decorrelated at the first step by applying nonlinear kernel principal component analysis. The spatial properties of the input features are then incorporated to select a subset of features which better reveal object structures and provide good separation among the classes of interest. The third contribution of this study is the evaluation of a number of recent approaches for kernel selection and an improved and computationally efficient approach is proposed. The alignment between the target kernel matrix and input kernel matrix is used to select the kernel parameter(s) for each candidate kernel function. Cross-validation is used at the final stage to search for the best kernel function using the selected kernel parameter(s) for each function. Experiments were carried out on both real and synthetic data. The results show that the proposed approaches provide an improved classification performance.