Hyperspectral Image Classification Method By Combining Spatial Information

As a new type of remote sensing, hyperspectral imaging plays an important rolein the fields of civilian and military applications. Hyperspectral image classificationis an important research topic for the hyperspectral data processing, wherehigh-precision classification algorithm is the foundation for a variety of applications.Traditional approaches usually concentrate on using the spectral characteristics fromthe hyperspectral imagery, and the spatial information provided by the hyperspectraldata cubes is barely considered. Constrained by the fundamental idea of classifiersdesigning，traditional classification methods can hardly improve the results of theobject classification and identification for hyperspectral remote sensing images．Thenew research shows that exploring a priori knowledge derived from thehyperspectral spatial information will help to improve the classification accuracy.This thesis focuses on the hyperspectral image classification methods by combiningspatial information, and the main works include:Firstly, this thesis analyzes the hyperspectral image classification principle,class separability, and other factors affecting the classification accuracy. In addition,the traditional classification methods based on spectral information and a typicalclassification method based on spatial information are surveyed. And theeffectiveness of the methods is verified based on a real hyperspectral dataset.Secondly, traditional approaches usually concentrates on using the spectralcharacteristics from the hyperspectral imagery, and the spatial information providedby the hyperspectral data cubes is barely considered. This paper introduces a methodto improve the hyperspectral image classification accuracy by adding on new spatialinformation. First, support vector machines are used to classify the hyperspectraldata and an initial result is obtained. Second, parts of mis-classified pixels areremoved by using a region-connection template. Finally, the initial classificationresult is improved by combining the spatial neighborhood information, which isdrawn from our new constructed potential functions. Simulations are carried outbased on two hyperspectral datasets, and the results show that the proposed methodcan increase the classification accuracy, while the false alarm rate is reduced.Finally, on the basis of previous studies, this paper improves a hyperspectral image classification method based on edge information, which is typical spatialinformation. The first step is hyperspectral image classification based on spectralinformation and SVM, then a method for multi-bands edge extraction is proposed,Finally, the classification results of SVM are improved by combining the extractededge information and the internal expansion method.