A Study On Band Selection For Hyperspectral Image Of Vegetable

Rapid advances in sensor technology have made it recently possible to collect hyperspectral remote sensing data of hundreds of contiguous spectral bands which spans typically from visible light to near infrared energy range. Such high dimensionality, on one hand, provides us with more potential discrimination power between classes with subtle differences. On the other hand, it is more difficult to get better classification performance, due to the fact that the available number of training samples is unable to catch up with the increase of dimensionality, which causes poor estimation accuracy for classifiers' parameters. Maybe, we need to strive to make two aspects of tasks. The one is that using dimensionality reduction techniques to select or extract bands subset from original data, which can improve classifier's performance. The other is that to develop special methods of image processing which could employ the potentials and advantages of hyperspectral remote sensing data. So, firstly we study systematically the band selection theory, which is one of dimensionality reduction methods, including a fast algorithm, evaluation method, and a comparative study in this dissertation. And then, based on higher dimensionality with better separability, we make heuristic study as to easy-mixed class problem, and achieve some satisfied results by experiments.The concrete contents of this dissertation are the following. To begin with, we recite some band selection methods developed by other researchers, comprising content-based methods and separability-based methods (Chapter 2). Then, we analysis their characteristics and furthermore propose a fast algorithm and develop an evaluation method for band selection algorithms (Chapter 3). And then, a comparative study is implemented by design three experiments with utilizing the fast algorithm and evaluation method (Chapter 4). Lastly, aim at the easy-mixed class problem, we propose class confusion index (CCI) to decide the confusion degree between classes. The index is simpler and more precise than the confusion matrix. It can provide significant guidance for pre-processing step before classification. As an important discriminant criterion of separability, several class separability measures, such as divergence, Mahalanobis distance, Chernoff Distance, and Bhattacharyya Distance, etc., are deduced...