Studies Of Some Key Techniques In Hyperspectral Classification

The emergence and the rapid development of the hyperspectral remote sensing technology, which can provide higher spectral resolution data about the Earth's surface, is very helpful to the achievement of the quantitative remote sensing. Hyperspectral imagery generally consists of dozens or hundreds of narrow, also contiguous, spectral bands, which accounts for the computational problem and the phenomenon where the response of bands tends to be highly correlated. Consequently, advanced techniques are needed to exploit the extensive information contained in hyperspectral data. My studies aim to enhance the performance of land cover classification, which is one of the most prominent applications of hyperspectral remote sensing. Reletive studies demonstrate that classification accuracy depends on the number of training samples, class separability, dimensionality, and the performance of classifier. Based on thoroughly analyzing the affecting principles of the four factors, an application-oriented hyperspectral classification scheme, which involves several efficient models, is proposed. The main works are as follows:1. The IEM algorithm is developed based on the investigation of the EM algorithm. It is common that only limited training samples are available, which will substantially affect the parameters estimation. The EM algorithm, which takes the advantage of class information contained in unlabeled samples, has been studied and applied to pattern recognition analysis. In general, more accurate estimates can be obtained. Unfortunately, one of the EM algorithm's weaknesses is that it is very sensitive to noise. When unlabelled samples contain noise, the performance of the EM algorithm will be seriously deteriorated. As for the hyperspectral data, the presence of noise is unavoidable. Consequently, identifying the noise is a key task when the EM algorithm is utilized in hyperspectral classification. The IEM algorithm applies a distance threshold test to eliminate the large number of noise. Furthermore, the IEM algorithm assigns appropriate weights to unlabelled samples. As a result, the influence of training samples in the estimation of parameter is increased and the phenomenon where one of the classes dominates the other classes is alleviated. Also, the IEM algorithm adopts the lowpass filtering technique, which will improve the class separability. Experiments show that the IEM algorithm is highly reliable in obtaining accurate class distribution estimates.2. The Tabu search algorithm and its application in hyperspectral dimensionality reduction are studied. Dimensionality reduction is often an essential part of solving a classification task. It is desirable that informative features are retained and redundant features are eliminated. However the current dimensionality reduction techniques do have their limitations, such as the problem of computation, the complexity of the methods, even the loss of the original spectral information. In this chapter, a novel dimensionality reduction technique, i.e. Tabu search...