Spatial Contextual Information Aware Hyper-spectral Image Classification Methods.pdf

1、Hyper-spectral image (HSI) classification can be defined as identification of objects in a scene, is an important task in many application domain such as:Mineralogy, Agriculture, Military, etc.HSI generally contains high

2、 dimensional enormous amount of data due to hundreds continuous narrow bands, which causes the Hughes phenomenon.Furthermore, the limited labeled samples and spatial variability of spectral signatures are the key challen

3、ges and problems for HSI classification methods.In the last decade, a lot of new developments for HSI classifications have been investigated and proposed.Among these methods discriminative approaches in machine learning

4、fields, in particular SVMs and SMLR have shown success in HSI classification.Recently, the integration of spatial context in the classification of HSI is known to be very effective in improving classification accuracy.

5、r>　　This thesis discussed the challenges and problems for the hyper-spectral image classification methods.The focus of this thesis is the inclusion of the spatial contextual information into the spectral-only classificat

6、ion algorithm.Benefited from the discriminative approaches in machine learning field, i present the spatial aware HSI classification techniques.The main contributions of the thesis are as follows.
　　a.First, i combine

7、 the Support Vector Machines (SVM) and Markov Random Fields (MRF) in an integrated framework for spectral-spatial hyperspectral image classification.Experiments demonstrate that the spatial context can be used to further

8、 improve the performance of the spectral only SVM method.I also present the experimental results for the SVM and composite kernel framework for comparison analysis in the context of spatial aware classification.
　　b.S

9、econd, i investigate the classification methods using SMLR based classifier with spatial contextual aware information.We analyzed and reported the results for the SMLR-MLL: the MLR based segmentation approach with spatia

10、l multilevel logistic (MLL) prior in the hyper-spectral framework.I also incorporated the spatial regions with spectral information in a HSI classifier with composite kernels, denoted as SMLR-CK.Comparative results have

11、shown that the composite kernel aware spatial regions outperform the MLL prior.
　　c.Finally, following low computational complexity in term of time and load, i proposed a simple yet powerful filtering based post-smoot

12、hing approach for spectral-spatial hyper-spectral image classification by coupling the regularized ELNR for spectral only classifier.I mainly contributed to present a two phase hyper-spectral image classification algorit

13、hm combining Elastic Net Regression (ELNR) and spectral-spatial bilateral filtering.Initially, the proposed method computes the pixel-based classification to evaluate the quality of the selected bands identified by Elast

14、ic Net regularized multinomial logistic regression.To incorporate the spatial information,the optimization probabilities problem are obtained by performing bilateral filtering on the initial probability maps, with the pr