Fine Classification Of Hyperspectral Imagery Based On Kernel Methods

Hyperspectral imagery is widely used in landcover classification and targetdetection area, due to its outstanding property of high spectral resolution. Althoughthe hyperspectral imaging spectrometer is able to acquire a large amount of finespectral information, the spatial resolution is always low at the same time. Eventhough the spatial resolution of hyperspectral image has been improved greatlythanks to the fast development of remote sensing technology, the low spatialresolution and corresponding phenomenon of mixing pixels are still main problemsof hyperspectral image processing. Besides, the high dimensionality ofhyperspectral data has brought great difficulty in fine classification, and thetraditional statistical and machine learning methods always show poor performance.In order to solve the above problems, this thesis mainly studies the techniques forfine classification of hyperspectral images through an analysis of its spectral andspatial properties and the state-of-art kernel methods.Since the spatial resolution is relatively low for hyperspectral images, a pixelcould contain more than one class of landcovers. While the traditional hardclassification methods are pixel-wise, which means a pixel could only be classifiedinto one class, and thus the classification results are not accurate. In order todiscriminate and quantitate different materials in a pixel, this thesis explores a softclassification method called spectral unmixing technique. This thesis first analyzesand compares several typical unsupervised endmember extraction techniques, andthen proposes a supervised method based on probabilistic support vector machine.By experimental comparison, the latter supervised method is more accurate inextracting endmembers. Finally, the result spectra of endmembers are used as inputsfor unmixing, and thus get the rich information of landcover classes.Kernel methods are famous for their efficiency and robustness in processingnon-linear machine learning problems in the high dimensional feature space, andthus widely applied in hyperspectral image classification and detection. The paperstarts to discuss the expressions and properties of kernel methods theoretically, andthen focuses on several typical efficient kernel methods, which provides technicalbase for fine classification.Although kernel methods have many advantages in hyperspectral imageclassification, these methods as well as other traditional classification methods onlyemploy the spectral information of hyperspectral images and neglect the spatialinformation. The classification results are thus not accurate. This thesis considers combining spectral features with spatial features extracted by gray levelcooccurrence matrix and applying kernel classifiers with some fusion strategies forspatial-spectral fine classification. For fusion problems, the thesis proposes adecision fusion algorithm based on the results of multiple classifiers and a featurelevel fusion strategy using the concept of composite kernels. By using both spatialand spectral features, we can get richer information about the landcovers, and thusachieve the purpose of fine classification for hyperspectral images.