Research On Image Restoration And Classification Methods For Hyperspectral Remote Sensing Imagery

Hyperspectral images(HSI)can provide rich spatial and spectral information for accurate land-cover analysis,thus,they have been widely used in different applications,e.g.,envi-ronmental monitoring and management,geological exploration,precision agriculture,military detection.In recent years,rapid advances of hyperspectral imaging technique offer new oppor-tunities to the development of earth observation system.However,some new problems have been raised up at the same time,which mainly include the following three aspects.(i)Affected by the factors of spectral imaging device and imaging environment,it is always an unavoidable problem to introduce noise in hyperspectral images.The existence of noise not only degrades image quality and affects visual interpretation,but also affects subsequent image analysis and applications,(ii)Different classes of land-covers may have similar spectral characteristic while the materials of the same class may have various spectral characteristics.This phenomenon can be commonly observed in HSIs mainly due to spectral mixture,which makes accurate scene in-terpretation difficult if only spectral information is used.(iii)The“Hughes phenomenon”arises due to the high dimensionality of HSIs,leading to the discrimination capability of classifiers to be negatively affected,especially in the case of small training samples.This thesis firstly summarized and analyzed the current hyperspectral processing works.Then,we focused on the aforementioned main problems in hyperspetcral image acquirement and scene classification,and proposed novel restoration and classification methods of high-performance.The proposed methods can take full advantage of the spectral-spatial character-istics in HSI,and provide superior image restoration and classification performance.The main contributions of this thesis are as below:(1)Aiming at solving the noise contamination problem in degraded HSI,a novel spectral-spatial adaptive sparse representation based restoration method is proposed here.This approach first generates spectral-correlated and spatial-similar pixel-clusters via spectral-adaptive band grouping and spatial-adaptive homogeneous region searching.Then,all pixels within each pixel-cluster are jointly represented and reconstructed by joint sparse representation model,leading to the estimation of final denoised results.The sparsity of spectral signals,spatial-similarity,and spectral-correlation can be simultaneously taken into account by this way.As a result,noise can be effectively removed from observed HSI while the important spectral-spatial information can be preserved well.Compared with both the band-by-band restoration algorithms and other classical spectral-spatial restoration methods,the proposed method can always achieve a significant improvement in terms of both visual impression and quantitative measurements.(2)Taking object based land-cover classification with spectral-spatial information as an orientation,this thesis developed a novel set-to-set distance based spectral-spatial classification(SD-SSC)method to category the scene in a superpixel-by-superpixel manner.On the one hand,traditional pixel-wise classifiers are easily affected by the spectral mixture phenomenon,leading to inaccurate classification results.On the other hand,it is easily observed that local spatial region always contains multiple pixels of the same class.Based on the above analysis,it is proposed to use superpixel based segmentation technique to extract test sets of local similar pixels first,meanwhile each training set is constructed based on each class of training samples.Then,both training and test sets are represented by affine hull model.Finally,each test set is assigned a common label according to the minimum distance between affine hulls of training and test sets.By the proposed method,the shape-adaptive local spatial information as well as joint spectral information can contribute to eliminate "salty" classification results.In addition,the superpixel-by-superpixel classification strategy has a higher calculation efficiency than the pixel-by-pixel manner.Compared with conventional spectral-spatial classifiers,the proposed set-to-set distance based classification methods can effectively improve the overall classification accuracies.What's more,the obvious superior in classifying cropland scene by the proposed SD-SSC method reflects the practical value in accurate agricultural mapping.(3)Considering that single classifier has a limited classification performance,this thesis attempts to boost classification performance via adaptively combining multiple classifiers.In specific,a novel classification method is proposed by the fusion of pixel-wise support vector ma-chine classifier and superpixel-wise joint sparse representation classifier.There are three main steps.First,the support vector machine is used to obtain pixel-wise class probabilities.Then,the superpixel segmentation technique combined with joint sparse representation classifier is used to acquire superpixel-wise class probabilities.Finally,the two levels of class probabilities are adaptively combined in a maximum-a-posteriori(MAP)estimation model,and the classifi-cation map is obtained by solving the maximum optimization problem.The proposed method can take full advantage of both classifiers,to simultaneously remove "salty" phenomenon as well as improve the discrimination capability of pixels around edges.Comparing with both pixel-wise and superpixel-wise classifiers,the overall classification accuracies by the proposed method is higher.The comparison results with different spectral-spatial classifiers also demon-strate the effectiveness and superiority of the proposed multi-classifier fusion based method in both cropland and city scenes.(4)Focused on the problem of increasing classification performance when training sam-ples are limited,this thesis deeply analyzes the complementary information provided by subpixel-level,pixel-level and superpixel-level features,and designs two novel fusion schemes to in-tegrate the three levels of features to further boost classification performance.The spectral unmixing techniques offer the possibility to extract spectral mixture information at a subpixel level.Besides,it has been demonstrated that the discrimination between different materials will be improved by integrating the geometry and structure information,which can be derived from the variance between neighbouring pixels.By incorporating the spatial context,the superpixel based spectral-spatial similarity information can be used to smooth classification results in ho-mogeneous regions.Therefore,adaptive combination of subpixel,pixel and superpixel based complementary information with the introduced feature-level and decision-level fusion schemes can contribute to higher discrimination between different classes.For the feature-level fusion scheme,the multiple-feature induced composite kernel is incorporated with a support vector machine(SVM)classifier for label assignment.For the decision fusion scheme,class probabil-ities based on three different features are estimated by the probabilistic SVM classifier firstly.Then,the class probabilities are adaptively fused to form a probabilistic decision rule for clas-sification.Experimental results demonstrate the effectiveness and superiority of the proposed multi-feature based classification methods,especially when the number of training samples is quite limited.