Research On Sparse Representation Based Spectral-spatial Compression And Classification Methods For Hyperspectral Image

Hyperspectral imaging is a new remote sensing monitoring technique for earth observation.Hyperspectral image(HSI)can provide not only rich spatial information(e.g.,size,shape,texture,and position),but also continuous spectral curves which reflect the material and component information.The combination of spatial and spectral information can improve the accuracy and reliability of remote sensing quantitative analysis.Recently,HSI processing technique has become research hot-spot and advanced area in remote sensing domain.It has attracted much attention in various applications and provides promising prospects.Along with the development of hyperspectral imaging and aerospace technique,the hyperspectral data number has increased dramatically and spectral/spatial resolution has improved greatly.As a result,the massive hyperspectral data of high-dimension brings great burden in sampling,storage,and transmission.Therefore,it is necessarily required to develop effective HSI compression methods.Another important research in remote sensing domain is accurate HSI classification methods,which have played important role in subsequent image analysis and scene understanding.However,there are a few problems which make it difficult to accurately discriminate various classes of land-covers,e.g.,curse of dimensionality,complex spectral-spatial structure.To solve these problems,it is required to develop novel and effective spectral-spatial classification methods.Sparse representation(SR),as an effective image representation model,which imitates the sparse coding mechanism of mammalian primary visual system and can represent high-dimensional and complex signals with only a few non-zero coefficients.Due to the capability of SR in extracting the intrinsic characteristics and representing data in a more effective manner,the SR presents promising potential in solving problems of HSI compression and classification.This thesis firstly summarizes and analyzes the current HSI compression and classification methods.Then,we focus on the aforementioned main problems in HSI compression and classification,and finally propose effective SR-based spectral-spatial compression and classification methods.The details are shown as below.First,aiming at removing spectral and spatial redundancy of massive high-dimensional HSIs,this thesis proposes a new sparse representation based spectral-spatial compression method.First,the superpixel technique is used to cluster similar pixels within each local region.Then,a joint sparse representation model is constructed for each cluster,to joint represent superpixel with a dictionary and sparse coefficients.Finally,the quantization and entropy coding is utilized to produce bitstream.On one hand,due to the high spectral-spatial correlation of pixels within each superpixel,the estimated sparse coefficients share the same sparse pattern,i.e.,the same non-zero positions.By this way,the complexity of sparse data can be reduced,leading to the reduction of the bit number used for encoding sparse coefficients.As a result,the compression performance can be effectively improved.On the other hand,as compared to pixel-wise SR method,the superpixel-wise manner can contribute to better spectral-spatial structure preservation capability.In experimental section,the proposed method is tested on several real HSIs and compared to state-of-the-art HSI compression methods.Experimental results show the effectiveness of the propose method on HSI compression.Second,to solve the problem of effectively reconstructing the original HSI from downsampling data,this thesis proposes a new contextual information based compressed sensing method for HSIs.Compressed sensing(CS)theory is constructed based on SR,which samples lower-dimensional data and recover original image from sampled data.One main advantage of CS over traditional compression method is the effective combination of sampling and compression,which can reduce the burden in data sampling,transmission and storage.Besides,the contextual information in spatial domain and sparsity of spectral are used to construct measurement matrixes and spectral-spatial dictionary learning model.The dictionary is well learned from sampled data,which contributes to sparser representation.A constraint item derived from spatial structures is introduced in image reconstruction model,which help to correct isolate badly-reconstructed spectral and smooth the local homogeneous regions.By the effective dictionary learning model and the use of contextual information,we can better recover original HSI with low sampling rate.In experimental section,we compare the proposed method with several state-of-the-art HSI compressive imaging methods.The quantitative and visual comparison results demonstrate the superiority of the proposed method on the reconstruction performance of HSIs.Third,in order to improve HSI classification performance via the joint use of spectralspatial information,this thesis proposes a new sparse representation based spectral-spatial classification method.On one hand,a shape-adaptive algorithm is developed to make full use of spatial-spectral similarity information.The highly-similar pixels within each shape-adaptive area are jointly represented,contributing to more accurate sparse representation coefficients,which will help to correct isolate “noisy” classification results.On the other hand,effective dimension reduction process is introduced,to overcome the negative effect of curse of dimensionality.By this way,the classification accuracy and efficiency can be further improved.In experiments,the proposed method is compared with some classical or state-of-the-art HSI classification methods.The experimental results show that the proposed method outperforms those classification methods in terms of classification accuracy.Finally,with the purpose of learning a dictionary to better represent complex spectralspatial structures for HSI classification,this thesis proposes a new spectral-spatial information based dictionary learning method.On one hand,a patch based online dictionary learning algorithm is used to learn compact and powerful dictionary,which can better represent complex spectral-spatial structures and improve efficiency of dictionary learning.On the other hand,a constraint used to preserve spectral-spatial structures is applied to dictionary learning model.The introduced constraint can push similar pixels in neighborhood to share similar sparse coefficients with the well learnt dictionary.By this way,the yielding sparse coefficients are structured and have high intra-class correlation,which can help to improve classification performance.In experiments,several representative spectral-spatial classification methods and some state-of-the-art dictionary learning based classification methods are compared.Experimental results show that the proposed method is comparable to other classification methods in terms of classification accuracy and running time,especially for dictionary learning based methods.