Research On Superpixel Feature Extraction And Classification Methods For Hyperspectral Image

Hyperspectral image(HSI)can provide rich spectral and spatial information,which enables accurate identification of ground objects.Therefore,HSI has been widely used in different applications,e.g.,environmental monitoring,geological exploration,urban planning,precision agriculture.However,there are still some key problems in the HSI classification to resolve,which consists primarily of as:(i)Shapes of ground objects in HSI are varied.(ii)The spatial-spectral structures of HSI are very complex.The spectral mixing phenomenon,that different classes of ground objects may have similar spectrum and the ground objects of the same class may have various spectrum,can be observed in HSI.(iii)Due to the high dimensionality of HSI,the curse of dimensionality is arisen.A superpixel in an image is an image patch with local characteristics consistency,which has adaptively spatial shape and size.The introduction of superpixel into HSI classification can effectively deal with the complex and changeable ground objects.At the same time,the superpixel ability of maintaining spatial information contributes to alleviate the trouble in HSI classification caused by the distortion problem of pixel spectral curve.In addition,the homogeneity of superpixel implicitly expands the number of label samples,which makes it possible to overcome the problem of dimensionality curse.However,for the various shapes and sizes of superpixels,most of classical spatial-spectral feature extraction methods can not be directly applied on superpixels.In addition,It is a nontrivial work to determine the optimal superpixel size.Moreover,in the image obtained by superpixel segmentation,some superpixels may consist of pixels from different classes.The thesis focuses on the aforementioned problems in HSI classification.By utilizing the superpixel segmentation technique and fully exploring the spatial-spectral structure information of HSI,feature extraction and classification methods based on superpixels have been proposed.The effectiveness and superiority of the proposed methods have been verified in multiple real scenarios.The main works of this thesis are as follows.(1)To handle the problem that different objects may have similar spectral characteristic and the same objects may have different spectral characteristics in HSI,a novel superpixel-based local subspace representation method is proposed for HSI classification.This method first uses the superpixel segmentation and filtering processing to generate a 3D-superpixel image.Then,each superpixel block in the image is represented by the corresponding low-dimensionality subspace.Finally,in a framework of projection embedding,a kernel extreme learning machine is adopted to obtain the classification results.Subspace representation can well accommodate the complex variations of pixels within one superpixel and capture the common spatial-spectral characteristics of them.To learn the feature representation,the proposed method takes advantage of the merit of superpixel in characterizing shape-adaptive structure and the powerful capability of subspace representation in engraving the common characteristics among pixels of the same class.By this way,the deteriorate of classification performances caused by the spectral mixture problem is alleviated and the salt/pepper noise is effectively eliminated in the classification results.Compared with various spatial-spectral feature extraction or classification methods,the proposed method can effectively improve the overall classification accuracies.What’s more,the superior in classifying datasets containing a lot of large-size homogeneous regions is obvious.(2)Focusing on the curse of dimensionality caused by the high-dimensional HSI,this thesis proposes a HSI feature extraction method,which fuses the superpixel-based local and nonlocal discriminant analyses.To explore the local/nonlocal correlation information in HSI and project the HSI from the high-dimensional space into a low-dimensional feature space,this method makes full use of the spatial-spectral shape-adaptive advantage of superpixel and the ability of discriminant analysis in enhancing class-separability.In specific,discriminant analysis is first conducted from superpixel-level local and nonlocal viewpoints,respectively.On the one hand,spectral-similarity information within each superpixel and diversity information among adjacent superpixels are explored in the superpixel-level local discriminant analysis(SLDA).On the other hand,relationship information among different superpixels is excavated in the superpixel-level nonlocal discriminant analysis(SNDA).After that,the SLDA and the SNDA are effectively fused to learn a projection matrix and the class-discriminative ability of data is reinforced by the corresponding projection transformation.In experiments,various spatial-spectral feature extraction or classification methods are used for comparison.Experiments demonstrate that the proposed method can more fully explore spatial-spectral information for HSI classification and effectively boosts the capability in separating different classes of data sets.(3)Aiming at the irregular shapes of ground objects and linear/nonlinear spectral structure,a novel superpixel-based statistical feature and kernel sparse representation method for HSI classification is proposed.The proposed method first generates superpixel-based sample blocks by utilizing the shape-adaptive superpixel segmentation on the HSI.Then,the statistical feature(i.e.the Brownian descriptor)for each sample block is calculated,which can explore the linear/nonlinear correlation between different bands of the sample block.Finally,the obtained Brownian descriptor is input a Log-Euclidean kernel sparse representation classifier to yield the classification result.Experiments are carried out to compare the proposed method with the other spatial-spectral feature extraction or classification methods.Experiments show that the proposed method can achieve higher classification accuracy in both city scenes with rich texture and the croplands containing large homogeneous regions.(4)Aiming at the complex and diverse land-covers in HSI,a multiscale superpixel-based joint sparse representation method for the HSI classification is proposed.Specifically,to gain the superpixels on different scales in HSI,a modified segmentation strategy of multiscale superpixels is firstly applied on the HSI.Then,the joint sparse representation classification is used to classify the generated HSIs marked by multiscale superpixels.Finally,a decision-level fusion approach is adopted to obtain the final classification result.The proposed method fully integrates the spatial-spectral structure-adaptive information of land-covers on different scales.Moreover,it better solves the problem to decide the superpixel number on different scale for different HSI datasets and alleviates the degradation of classification accuracy caused by some superpixels containing pixels from different classes.Experiments compare the proposed method with the other spatial-spectral classification methods.Experimental results show that the proposed method significantly improves the classification performances.In addition,compared with the classification methods based on single scale superpixel,the proposed method achieves higher classification accuracies on both homogeneous regions and edge regions in HSI.Finally,several representative case studies are used to verify the effectiveness of multiple feature extraction and classification methods in this dissertation,including: 1)The proposed feature extraction techniques are utilized for mapping crops in Longkou area,Hubei province.2)The proposed feature extraction methods are used to identify different land-covers in Botswana wetland and make detection map of the water in this area.3)The proposed methods are employed to classify and plot a spatial distribution map of the land-covers around the Houston university.4)The proposed approaches are utilized to identify different types of vegetation around the Kennedy space center and plot the spatial distribution map of the vegetation in this area.