TRANSFORM CODING TECHNIQUES FOR HYPERSPECTRAL DATA COMPRESSION

Abstract

ftc cui-rcnt research work aims at the study of various transform based methods and bit lanc encoders used for efficient transform coding of hyperspcctral data to get better results in tenhls of compression. Various imaging spectrometer like AVIRIS, 1IRS generate numerous band images with various spatial resolutions, having wavelengths lying in a particular region of - electromagnetic spectrum. Due to the advent of smart sensors and extremely fast processors with improved acquisition capabilities and radiometric resolution (spectral as well as spatial), the amount of data (,enerated is extremely large; making it inevitable to design the systems incorporating efficient image compression algorithms. Three dimensional nature of hyperspectral data with huge amount of correlation in spatial and spectral domain makes transform coding methods more efficient; pi'o'ides good results with low complexity and concentrates power in a few coefficients resulting in better low bit rate perfonnance which is desirable. As hyperspectral data carries a lot of information in spectral domain, hence different transform in spectral and spatial domain can be applied so as to yield better results. For this hybrid 3D transforms are used by first applying one complete transform in one dimension and then taking 2D scivare transform in other two dimensions. A popular approach is to employ the combination of 1 D spectral decorrelator such as KI.T, PCA, DC'T, D\VT l)WP'l' etc.: followed by 21) spatial decorrelator such as 2D-I)CT,2D-DWT, 21)-DWPT. etc.: followed by rate allocator and entropy coder such as SPIl IT, SPECK, EI3COT. SPBH'I' etc. In this research work, various combinations of spectral and spatial dccorrelator transforms are investigated to compare their energy compaction and comparative results are shown in terms of PSNR versus the iiumbcr of coefficients zeroed. Also, the performance of various 2D entropy encoders based on 2D-Set Partitioning methods like 2D-SPII IT. 2D-'l'BE. 2D-SPECK and 2DSP13I IT for each band image is investigated. In signal orocessing. KLT is an optimal transform due to its signal adaptive nature and thus can be used to decorrelate data spectrally. 1 lowever. because of covariance matrices estimation for each spectral vector, solution of eigen vector problems. and computing matrix-vector products: in past it was resisted to be used for real time signal processing applications. Thus, by applying transforms like DW'l' or DWPT or DCT along spectral domain, time complexity reduces tremendously, but these results in poor transform perfomlance. Nowadays due to the advent of smart and superfast processors. KLT practical application has rediscovered to exploit its superior dccorreiation capabilities. Since hyperspectral data call be seen in terms of' stack of different band images taken together for analysis. simple 2D image coding techniques can be used to analyze each uncon'elatcd band image. In waveict compression system, the entire image is transformed as a single object which provides thorough spatial correlation exploitation. Wavelet based methods provide hierarchical subband system, where subbands are logarithmically spaced in frequency and hence represent octave band representation. in real wavelet transform, there is a parent child relationship between the coellicients of the octave frequency bands. Bit plalle encoders exploit such parent child relationship and transmit the transformed coefficients oil bit j)lane basis. Various methods have been investigated in the state of art for hyperspectral image compression and in this research work 2D-SPIIIT. 21)-TBE, 21)-SPECK and 2D-SPBH'F are thoroughly studied. 2D-SPII-IT and 2D-TBE exploit the magnitude correlation existing across the subbands in the form of' SOT's: llowevcr they fail to exploit spatial correlation. Similarly, 21)-SPECK exploits energy clustering properties of transform in frequency domain and processes the image in the lorill of blocks: however fails to exploit parent child relationship between the frequency bands. By combilling tile features of 213- - SPIlIF and 2D-l'l3E to exploit parent child relationships and features of 2D-SPECK to exploit energy clustering, 21)-SPBI IT provides best results in terms of compression especially at low bit rates. Performance evaluation is carried out On I 2-bit radiance Scene 0 of hawaii of A\'lRIS data cube. i\VlRIS scene has 224 bands with spatial resolution 614x5 12 but for analysis pwposc it is cropped to 51 2x5 I 2x20. As for lilter selection in DW] and DWP1. (9. 7) lllortilOg011al waveiet tiller pair is used and it is constructed from liftillg scheme as it provides excellent lossy compression performance. It is found that 11)-KIT (along spatial domain) and 2D-DWl' (along spatial domain) provides good results in terms 01' energy compaction as comparedi to otller 31) hybrid transformation. It is also seen that 2D-SPBFIT perlonlls better at all bit rates as conlparcd a to otller 2D-set partitioning methods irrespective of the 3D trailsiormation used.