Study On Features And Classifications Of Hyperspectral Image

In the context of feature extraction and classification algorithms for hyperspectral data, this paper studies some important problems of spectral feature curve shape and classification algorithms for applications. The most important innovations are as follows: Thanks to the high dimensionality of hyperspectral data, it is now possible to analyze the shape of feature curve. Moreover, the shape of feature curve may result in great practical applications. Such as using Lab sample data to learn the hyperspectral data recently from satellites,the two kinds of data have different magnitudes, but remain the similar curve shape. However, there are only a few curve shape algorithms, where the SAM is one of the most useful tools. We enumerate the mathematical characteristics of SAM. Based on the differences and similarity among samples in every spectral library class, we derive an algorithm called similar band selection (SBS) that selects similar band subset, which will be used for next step of classification.The Gaussian maximum likelihood classification (GMLC) algorithm is one of the most useful methods for hyperspectral classification. However, the classification speed and high dimensionality of hyperspectral data hamper its farther practical applications. We analyze the relationship between the GMLC error and Bhattacharyya distance, and enumerate the addition property of Bhattacharyya distance under the condition of uncorrelated features. Based on such analysis, a new feature selection algorithm is derived. It effectively decreases the dimensionality of the feature set, which saves the computation time, and gets rid of the Hughes phenomenon,where the classification accuracy decreases as feature number increases.Hyperspectral classification often involves a large number of classes, for the high dimensionality of hyperspectral increases the potential of discriminant similar classes. Thus the definition of class becomes very important. Most of the classification algorithms often classify all the classes at one step, which may increases the difficulty of decision boundary. A novel pairwise decision tree (PDT) framework is proposed, where no partitions and clustering are needed. The original C-class problem is divided into a set of two-class problems, where all the two classes are information classes. For one classification, the proposed algorithm needs far fewer binary tests, which saves the classification time, and needs fewer layers, which decreases the potential of accumulated errors.Based on such innovations, three papers have been published and accepted, where the paper of the third part has been accepted for publication by the SCI source journal of International Journal of Remote Sensing.