Research On Target Detection And Compression For Hyperspectral Imagery

Hyperspectral remote sensing,which combines the imaging spectral processing techonology and target detection technology,can simultaneously capture hundreds of narrow and contiguous spectral bands from a wide range of the electromagnetic spectrum.And the"data cube" of hyperspectral imagery(HSI)contains one dimension spectral information and two dimensions spatial information for the ground material.With the development of spectrometer,HSI can provide more features in high-resolution level.As a results,it has great capability in deriving comprehensive details about the spectral and spatial information of the ground material.In HSI,the pixels consist of different objects have special spectal responses.This is the foundation theory of target detection in HSI.Besides,the data volume in the 3-D hypercube increases dramatically with the trend for an increase in spatial/spectral resolution,radiometric precision and a wider spectral range,resulting in challenges for data transmission,storage,and processing.To reduce the volume of data,effective coding and compression become a natural choice in this context.Based on the basic hyperspectral remote sensing imagery processing theories and relevant subjects,this study mainly focuses on the target detection and compression of hyperspectral remote sensing imagery.For target detection in HSI,the methods based on sparse representation are proposed.The study mainly focus on two parts of sparse representation.First,the spatial information is considered and explored for sparse presentation models.And then,the dictiontary used in sparse presentation calculation is designed to improve targets results.Regarding the sparse presentation using spatial information support,as target detector of conventional sparse representation is applied to each pixel in the test region,independently,without considering the correlation between its neighbouring pixels.How to incorporate the spatial correlation information into sparse representation for improved target detection forms our proposed work in this research.So several sparese representation method for target detection are proposed by considering the spatial information in 4-neighbourhood,adaptive spatial region and non-connected spatial region.Comprehensive experiments using both visual inspection and quantitative evaluation are carried out.And the results have indicated that the proposed methods help to generate improved results in terms of efficacy and efficiency for HSI target detection.Regarding the dictionary used in sparse presentation,As a HSI pixel in basic sparse presentation is represented as a sparse vector whose entries correspond to the weights of atoms from the given over-complete dictionary.And most of atoms are useless as their corresponding sparse representation weights are zero.So adaptive sub-dictionary which contain much less atoms is exploited to improve detection efficacy and efficiency.The searching strategy,k-nearest neighbours(k-NN)algorithm,is utilized to select the atoms in the sub-dictionary with the criteria of spectral similarity.Comprehensive experiments are carried out on three different datasets using both visual inspection and quantitative evaluation.The results from these datasets have indicated that the proposed approach help to generate improved results in terms of efficacy and efficiency.For data compression in HSI,this part of research also divides into two sections.First,an object-based compression framework is proposed to consideres the target distribution in HSI.Secondly,an improved vector quantization method for hyperspectral image compression is explored.Regarding the object-based compression framework,as in conventional compression approaches,all the pixels are treated equally.As a result,information loss to the content of interest,particularly the targets of interest,becomes unavoidable.To achieve this,an object-based coding and compression framework is proposed,where targets of interests are detected as objects and separately coded with the remaining background compressed for transmission.As a result,the objects can still be successfully detected from the decoded images.In the proposed framework,an improved sparse representation with adaptive spatial support is proposed for effective target detection.In addition,transform based techniques are embedded for coding and compression of HSI,where 2D/3D DCT are used for efficiency and efficacy.In comparison to non-object based compression using the same approaches,our proposed method yields much improved results in blind detection of objects.Regarding vector quantization method for hyperspectral imagery compression,a multivariate vector quantization(MVQ)approach is proposed for the compression of HSI,where the pixel spectra is considered as a linear combination of two codewords from the codebook,and the indexed maps and their corresponding coef cients are separately coded and compressed.A strategy is proposed for effective codebook design,using the fuzzy C-mean(FCM)to determine the optimal number of clusters of data and selected codewords for the codebook.Comprehensive experiments on several real datasets are used for performance assessment,including quantitative evaluations to measure the degree of data reduction and the distortion of reconstructed images.Our results have indicated that the proposed MVQ approach outperforms conventional VQ and several typical algorithms for effective compression of HSI,where the image quality measured using mean squared error(MSE)has been signicantly improved even under the same level of compressed bitrate.