The International Arab Journal of Information Technology, Vol. 13, No. 4, July 2016 435

Hyperspectral Based on DWT and TD with ALS Method

Kala Rajan 1 and Vasuki Murugesan 2 1Department of Electronics and Communication Engineering, Sri Subramanya College of Engineering and Technology, India 2Department of Electronics and Communication Engineering, Velammal College of Engineering and Technology, India

Abstract : Compression of Hyper Spectral Image (HSI) is an important issue in remote sensing applications due to its huge data size. An efficient technique for HSI compression is proposed based on Discrete Transform (DWT) and Tucker Decomposition (TD) with Adaptive Least Squares (ALS) method. This technique exploits both the spatial and spectral information in the images. ALS method is used to compute the TD which is applied on the DWT coefficients of HSI spectral bands. DWT is used to segment the HSIs into various subimages, while TD is used to conserve the energy of the subimages. Run Length Encoding (RLE) performs quantization of the component matrices and encoding of core tensors. The experiments are conducted with HSI compression based on DWT, TD with ALS method and HSI compression methods based on lossless JPEG (JPEGLS), JPEG2000, Set Partitioning Embedded Block (SPECK), Object Based (OB)SPEC and 3DSPECK and the results of our work are found to be good in terms of Compression Ratio (CR) and Peak SignaltoNoise Ratio (PSNR).

Keywords : ALS, CR, DWT, HSI, PSNR, RLE, TD.

Received June 17, 2013; accepted September 26, 2013; published online August 22, 2015

1. Introduction are preferred in hyper spectral imaging because of the huge quantity of data and data loss must be Many military and civil applications involve the negligible. detection of an object or activity such as a military Many traditional compression algorithms for HSIs vehicle or vehicle tracks. Satellites must transmit a have only considered the spectral value in a feature great amount of information to control on earth for its space whose dimension were spectral bands. Basic later processing. If the satellite carries out a compression methods for HSIs includes transform preprocessing step in order to extract only the relevant coding based algorithm, Vector Quantization (VQ) information the communication process will be highly based algorithm, Differential Pulse Code Modulation optimized. (DPCM) algorithm and Adaptive DPCM (ADPCM). Hyper Spectral Images (HSIs) are used in several Some compression algorithms may increase the applications such as soil analysis, forest monitoring, computational complexity and also lead to distortion. river flow analysis, environmental studies and other DPCM compression techniques involve expensive geographical analysis. HSI sensors are advanced image decoder and multiple sampled signals rather digital color cameras with spectral resolution at a than one. specific illumination wavelength. These sensors The advantage of Discrete measure the radiation reflected by each pixel in a large (DWT) is the temporal resolution in both frequency no of visible or invisible frequency (or wavelength) and time. The decomposition into sub-bands is highly bands. HSIs are considered as 3D data in compression flexible in terms of resolution scalability. methods known as third-order tensor, composing of play a significant role in many image processing two spatial dimensions and a spectral dimension. HSI applications. Tucker Decomposition (TD) obtains a applications are oriented toward classifying or higher Compression Ratio (CR). This decomposition grouping similar pixels, many instances of which are technique permits the allocation of any values for each typical of each class of pixel since, interpretation of the dimension of the core tensor. TD is one of the most scene is based on the clustering of the majority of popular tensor decomposition methods for the pixels. compression of HSIs. The compression of HSIs can be effected by The proposed HSI compression algorithm is based detecting the spatial redundancies and spectral on DWT and TD with Adaptive Least Squares (ALS) redundancies. The two main divisions of HSI method. Applying 2DWT to each spectral band using compression methods are and biorthogonal wavelet will take care of the first stage . Lossless compression methods compression by applying along the rows of the image

436 The International Arab Journal of Information Technology, Vol. 13, No. 4, July 2016

first and then the results are decomposed along the the multi-linear predictions with the response columns. TD is applied to the four wavelet sub-images variables. Shoham and Malah [21] used a to enhance the CR . The ALS method is used to Concatenative Text-To-Speech (CTTS) compressor for compute the TD which is applied on the DWT a larger acoustic database. This compressor employing coefficients of HSI spectral bands. Run Length 3D Shape-Adaptive Discrete Cosine Transform (3D- Encoding (RLE) is used for the quantization of the SADCT) can also be extended to hyper spectral component matrices and encoding of the elements of imaging. Here, the coefficients of the 3D quantization the core tensors. matrix are coded using run-length encoding scheme. The experiments were conducted for an HSI of the Hendrix et al. [11] proposed an enclosing algorithm little Colorado River with the HSI compression based for abundance determination and end-member on DWT, TD with ALS method and existing HSI identification in HSIs. The noise in the HSI is compression methods based on Lossless JPEG (JPEG- minimized by a least squares method. Acevedo and LS), JPEG2000, Set Partitioning Embedded Block Ruedin [2] proposed a lossless HSI compression (SPECK), Object Based (OB)-SPECK and 3D-SPECK technique employing dynamic Look-Up Tables and our work is found to be good in terms of CR and (LUTs). A least squares estimator is used for the Peak Signal-to-Noise Ratio (PSNR). determination of scaling factor for the neighboring The remaining part of the paper is organized as pixels in an HSI. In [1] a least-squares estimator was follows: Section 2 involves the works related to HSI used to fasten the involved differential Huffman compression techniques. Section 3 involves the encoding. generation of the three-dimensional HSI. Section 4 The existing compression techniques employ involves the detailed description of the DWT-TD encoding techniques like, JPEG-LS [25], Joint (ALS)-RLE based HSI compression technique. Section Photographic Experts Group (JPEG2000) [5], SPECK 5 involves the performance analysis and comparison of [22], 3D-SPECK [23], Lattice Vector Quantizer the proposed HSI compression method and existing (LVQ)-SPECK [8], Discrete Wavelet Packet (DWP)- compression techniques based on JPEG-LS, SPECK [8], Shape-Adaptive Reversible Integer JPEG2000, SPECK, OB-SPECK and 3D-SPECK. The Lapped Transform (SA-RLT) [16] and OB-SPECK paper is concluded in section 6. [16]. Zhou et al . [27] designed uni-chip VLSI architecture for wireless image sensing. It involves 2. Related Works Color Filter Array (CFA) preprocessing and JPEG-LS compression method. Song et al . [24] proposed a This section deals with the works related to recent HSI differential prediction and JPEG-LS based lossless HSI compression techniques. Hou et al. [13] proposed an compression technique. Yin-Tsung et al . [26] proposed HSI lossy-to-lossless compression using 3D Embedded a lossless HSI compression method based on ZeroBlock Coding (3D-EZBC) algorithm. This hardware/software code sign. The median predictor algorithm involves 3D transform based on spatial used in this model is defined in JPEG-LS format for integer 2DWT and spectral integer 1D Karhunen- intraband predication. Loève Transform (1D-KLT). Context-based Adaptive Lilian and Leila [17] review the (AAC) and bit-plane zero-block systems involved in satellite imagery. Lossless JPEG is coding are used for entropy coding. Lopez et al . [18] basically a low complexity lossless compression for reviews the various reconfigurable hyper spectral images. It consists of a modeling phase and an imaging techniques in onboard systems. DWT is used encoding phase. The encoding includes run length, to transform the HSI from the spatial domain to differential coding, and Huffman code. Sriraam and another domain and for both spatial and spectral decor Shyamsunder [23] proposed a 3D medical image relation. DWT-based compression techniques involve compression involving multiple 3D wavelet coders. JPEG2000, Set Partitioning in Hierarchical Trees Four variants of daubechies and Cohen-Daubechies- (SPIHT) and Embedded Zero-tree Wavelet (EZW). Feauveau (CDF) wavelets are used in the first stage García-Vílchez et al . [9] the impact of lossy HSI and encoders such as 3-D SPECK, 3-D SPIHT and 3-D compression is studied using supervised Support BISK (binary set-partitioning using k-d trees) used in Vector Machine (SVM) classification and spectral the second stage for the compression. Chang et al . [5] unmixing. Here, DWT is used to decor relate the input proposed a group and region based parallel HSI HSI in the spatial domain. compression technique. This method employs a band Mažgut et al . [19] proposed a decomposition clustering and signal subspace projection. A method of binary tensors. A Tucker model denoted as combination of Principal Component Analysis (PCA) tensor-to-tensor projections are used for the and JPEG2000 are used to enhance the CR . decomposition of real tensors. The tucker model gives Wavelet coders such as SPECK and BISK are also lesser reconstruction errors than the Parallel Factor used volumetric coding based EEG image compression analysis (PARAFAC) model. The tucker model is used [22]. They possess progressive resolution and quality as a link function and multi-linear predictor to correlate during the compression process. Dutra et al . [8]

Hyperspectral Image Compression Based on DWT and TD with ALS Method 437 proposed a successive approximation wavelet coding The 2DWT of the function h(x, y) with size for HSI compression. It involves two modules, first a coordinates S1 and S2 is given in terms of coefficients LVQ codebook for processing of multiple samples and as: the next one is DWP-SPECK for DWP decomposition. S −1 S −1 1 2 1 DWP-SPECK involves 1D-DWT and simultaneous Wss(, ) = ∑ ∑ hxy (,)(,)ϕ xy ϕ 1 2 s1, s 2 (1) encoding of the huge number of spectral bands. Jiao et y =0 x =0 S1 S 2 al . [16] proposed a lossless Region of Interest (ROI) S −1 S −1 coding for two-dimensional remote sensing images m 1 2 1 Wqss(,,)= ∑ ∑ hxy (,)Ψ m (,) xy ψ 1 2 q, s , s (2) based on SA-RLT. This model can be a 3D model for S S y =0 x =0 1 2 HSIs by performing the following steps: Compute the 1 2 SA-RLT in spatial domain, décor relate the spectral Here, m= { H, V, D}, φ is the scaling function and ψ is redundancy using integer DWT, and encode the the wavelet function. The coefficients Wφ(s1,s2) give an transformed coefficients using 3D OB-SPECK or 3D approximation of h(x, y). The coefficients Wψ(q, s1, s2) OB-SPIHT. Rawat and Meher [20] introduced a hybrid provide the H, V and D information. Conventionally, image compression scheme (i.e., DCT and fractal q s1=s2=2 is selected so that q=0, 1, …, Q–1. The image compression). The color image was compressed m functions and are expressed in ϕs, s (x, y ) ψq, s , s (x, y ) using DCT. Fractal image compression was employed 1 2 1 2 for preventing the compression on the same blocks of terms of wavelet filter coefficients cφ and cψ the image. respectively. The wavelet filtering is performed by biorthogonal 3. Generation of 3D HSI wavelet. The inverse 2DWT is calculated as: 1 The source of an HSI is an imaging spectrometer hxy(,)= Wss (,)ϕ (,) xy + ∑∑ ϕ 1 2s1 , s 2 which collects the images simultaneously in several S S s s 1 2 2 1 (3) spectral bands that can reach an approximately 1 Q −1 ∑ ∑Wssm(,)ψ m (,) xy contiguous spectral sample. Higher amount of spectral ψ 12 q , s1 , s 2 S S m ={}H,V,D q =0 bands increases the processing time and complexity. 1 2 HSIs can be characterized as 3D data represented as: S1 × S 2 × S 3 S1 × S 2 × S 3 DR , where R is an S1×S2×S3 3.2. Computation of TD dimensional real vector space and S1×S2×S3 is the size S× S × S of the HSI. The third-order tensor AR 1 2 3 is decomposed by TD HSIs contain two types of simultaneous correlation into an unknown core tensor MR Q1× Q 2 × Q 3 multiplied by namely, spectral correlation within the spectral bands a set of component matrices A, B and C. The TD and spatial correlation within the images. Generally, process is shown in Figure 2. spectral correlation is higher than spatial correlation. 3.1. 2DWT DWT decomposes the signals into lower resolution with finer details. DWT consists of consecutive Low- Pass Filter (LPF) and High-Pass Filter (HPF). At each decomposition level, the HPF generates main information given as: Horizontal (H), Vertical (V) and Figure 2. Third-order tensor decomposition by TD. Diagonal (D) information. The LPF affiliated with the A= M.A.B.C+ E= Aˆ + E (4) scaling function generates the finer details represented Tensor Aˆ is an estimation of tensor A and is as Approximate (A) information. 2DWT is applied to dependent on the dimensions of the core tensor, i.e., each band of HSIs. An image band comprises of S1 (Q1, Q2, Q3). E gives the estimation error during the rows and S2 columns. On application of 2DWT four sub-band images namely A, H, V and D, each decomposition process. ALS algorithm is used to compute the TDs and containing S1/2 rows and S2/2 columns are obtained. Images of sub-band A possess the highest energy CANDECOMP (CANonical ECOMPosition)/ among the coefficients of the remaining sub-band PARAFAC (PARAFAC analysis) decompositions. images, as shown in Figure 1. This algorithm is adaptive in nature which in turn computes the TDs easily for dynamic tensors and suitable for higher order tensors. The algorithm is provided for third order tensors and is stated in its general form as:

1. Given an order-r tensor Aˆ , declare a set of projection matrices U(1) , U(2) , …, U(r) . Figure 1. Decomposition of HSI. 2. Let ii = 1.

438 The International Arab Journal of Information Technology, Vol. 13, No. 4, July 2016

3. The equation for U(ii ) is solved keeping the other with ALS method is given in Figure 3. matrices constant.

The matricized form of A on mode i is represented as U(ii ). It is equal to the product of two terms, where the first term is an n-ary Khatri-Rao product and the second term is a moore-penrose pseudo inverse. The aforementioned three steps are repeated for all ii from 1 to rr until convergence is attained. While the above algorithm improves the convergence in all iterations, it is inefficient as it requires huge amount of required iterations for convergence. The PARFAC components are generally estimated by the minimization of the quadratic cost function. 2 Figure 3. Working of proposed HSI compression based on DWT R f(, A B , C ) = M ∑ arobrocr (5) and TD with ALS method. r = 1 The HSI is simultaneously split into several spectral The minimization of Equation 5 involves two bands which are obtained from the imaging problems: spectrometer. The HSI compression involves four • The decomposition of core tensor M can be stages. The first stage involves the application of computed when Equation 5 becomes zero. biorthogonal wavelet based 2DWT to each band of HSIs to obtain four sub-band images ( A, H, V, D) for • The best rank R approximation to the core tensor M each spectral band. The tensor modeling process can be calculated when the minimum of Equation 5 structures a tensor for each sub-band image ( A, H, V, is distinct from zero. D). Equation 5 is the generally minimized by the ALS In the second stage, the four tensors are subjected to algorithm in which the components are updated per TD algorithm constructed upon the ALS method. The mode. The components of the PARAFAC size of the core tensor is selected manually for each decomposition are 2×2×2 tensors. The tensors may be tensor. Tensor A contains the lowest frequency 2×2 matrices or 2×3 matrices, depending on the rank R components with the majority of the wavelet of the tensor. The rank of the tensor is 1, 2 or 3. The coefficient energy. The values of ( Q1A, Q2A, Q3A) are component matrices are defined as: higher than those of other tensors. The values of (Q1D, Q2D, Q3D) are higher than the other remaining tensor Aaa=(, , … , a ) (6) 1 2 R values ( Q1H, Q2H, Q3H) and ( Q1V, Q2V, Q3V). CR is defined as the ratio of the number of bits in B=( bb1, 2 , … , b R ) (7) the original HSI to the number of transmitted bits. The

Ccc=(1, 2 , … , c R ) (8) number of transmitted bits is denoted as P.

With vectors ai, bi, ci and i= 1, 2, …, R as columns. CR=S1 S 2 S 3 /P (10) The quadratic cost function is rewritten as: S S P=1 ( QQQQ +++ )( +2 QQQQ +++ ) 2 111ADV1H 2222 ADVH f(, A B , C=M) [, A B, C ] (9) 2 2 +SQ( +++ Q Q Q ).. + QQQ ALS algorithm is used to solve Equation 9 as shown in 33A 3 D 3 V 3 H 123 AAA (11) +QQQ.. + QQQ .. + QQQ .. the following steps: 123DDD 1V23 VV 1 HHH 2 3 • Initially, the ALS algorithm determines A under The four core tensors for A, V, D, H and the component constant B and C. matrices for each core tensor are transmitted in the • The component matrix B is updated under constant third stage. In the fourth stage, RLE performs quantization of the component matrices A, B, C and A and C. Q× Q × Q encoding of core tensors MR 1 2 3 . The quantization • Once component matrix B is updated, component process involves the compression of the data values in matrix C is updated under constant A and B. the component matrices to a single quantum value. The • The updating process is repeated till a convergence components in the frequency domain divide by a criterion is satisfied. constant and then rounded to the nearest integer. Each core tensor is scanned row-wise for 4. Proposed Compression Method determination of repetitive pixels and they are grouped. These groups contain the pixel value and the frequency TD decreases the spectral and spatial correlations of repetition. When a pixel value occurs only once it is simultaneously to enhance the CR . The flow of the not replaced with its frequency value as it would cause proposed HSI compression based on DWT and TD an overhead in compression. Consider the following

Hyperspectral Image Compression Based on DWT and TD with ALS Method 439

3 example in a core tensor: (O(QASA )+ O(S1S2S3)), whereas the computational Input Stream: 55 55 55 98 98 98 98 45 12 12 12 12 complexity of HSI compression using DWT and TD Encoded Stream: 355 498 45 412 alone resulted in a higher total computational complexity of ( O(4 Q S 3)+ O(9 S S S /7)), because it The fifth stage involves the decoding of the transmitted A A 1 2 3 involves 4 core tensors and a (9/7) biorthogonal compression data. The final sixth stage involves the wavelet [15]. inverse 2DWT for obtaining the reconstructed HSIs The computational complexities of existing HSI Dˆ . techniques like 3D-EZBC [13], Kronecker sensing [7], Iterative Spectral Mixture Analysis (ISMA) [14] and 5. Performance Analysis DWT-TD [15] are analyzed and compared in Table 1. The experiments were conducted for an HSI of the 3D-EZBC and DWT-TD employed AVIRIS dataset little Colorado River with the proposed HSI images of low altitude and cuprite respectively [10]. compression method and existing HSI compression Kronecker sensing used Smoothed Particle methods based on lossless JPEG [25], JPEG2000 [5], Hydrodynamics Code (SPHC) data cube [7]. ISMA SPECK [22], 3D-SPECK [23] and OB-SPECK [16]. method applied the splib06 dataset mentioned in [4]. It The performance of the compression techniques is is observed that the DWT-TD (ALS)-RLE analyzed and compared in terms of computational compression method performs satisfactorily in terms of complexity, CR and PSNR. computational complexity compared to the other existing methods.

5.1. Experimental Results Table 1. Performance analysis-computational complexity. Image/Data Method/Application Computational Complexity The HSI compression steps of the little Colorado River 2 Low Altitude 3D-EZBC O(SA ) + O(2( S – 2)) + O(S – 2) 3 using the HSI compression method based on DWT-TD Cuprite DWT-TD O(4 QASA ) + O(9 S1S2S3/7) SPHC data cube Kronecker sensing O(SA log SA) (ALS)-RLE is highlighted in Figure 4. The input HSI 3 splib06 ISMA O(( S1S2S3) ) 3 is loaded and is split into seven spectral bands. The Little Colorado River Our work-Avg. O(QASA ) + O(S1S2S3) DWT for band 7 of the input HSI with Level 1 decomposition is shown in Figure 5. 5.3. Compression Ratio

The CR of the existing JPEG-LS compression [25] and the DWT-TD ALS-RLE compression is analyzed and compared for four HSI datasets in Figure 6. CR for HSI-1 in our work is significantly higher than that of JPEG-LS compression. Then, CR decreases gradually a) Input HIS. b) Spectral band 7 of split input HIS. in our work compared to JPEG-LS compression, for datasets HSI-2, HSI-3 and HSI-4. It is observed that on an average basis CR for the DWT-TD ALS-RLE compression is slightly higher than that of JPEG-LS compression. The small rise in CR for the proposed compression over JPEG-LS compression is because c) Spectral band 7 of transformed HIS. d) Reconstructed HIS. they are only lossless compression techniques Figure 4. Compression steps for a HSI of little colorado river. concentrating mainly on conservation of the huge data.

Compression Ratio

Figure 5. DWT for band 7 of HSI with level 1 decomposition. HSI-Datset Figure 6. CRs of JPEG-LS compression and DWT-TD ALS-RLE 5.2. Computational Complexity compression. The DWT of the proposed HSI compression technique CR obtained by other researchers using techniques involves biorthogonal wavelet and a unity like wireless image sensors [27], differential prediction decomposition level. This possesses a computational [24] and lossless interceding [5] are analyzed and complexity of the order O(S1S2S3). The maximum compared with DWT-TD (ALS)-RLE compression in computational complexity of TD is of the order Table 2. These methods employed compression 3 O(QASA ), where SA is the average number of pixels of methods like least squares, JPEG-LS and JPEG2000. the A tensor and QA is the average of the size They employed test images like Airplane [3], polar iris dimensions of the A core tensor. images [12], etc., the test image files were also, taken The total computational complexity of the (DWT- in BIL (Band Interleaved by Line) format and BSQ TD (ALS)-RLE) compressor is estimated as (band Sequential) format. The comparative analysis

440 The International Arab Journal of Information Technology, Vol. 13, No. 4, July 2016

showed that the CR of DWT-TD (ALS)-RLE Table 3. These methods employed compression compression is better compared to other techniques. methods like JPEG2000 [16], SPECK [24], SA-RLT Table 2. Performance analysis-CR. [16] and SA-DWT [16]. They employed test images like motor movement imagery and Shelter Island-BG Image/Data Method/Application Main Focus CR Airplane Wireless Image Sensor JPEG-LS 2.99 (Background) and ROI. The comparative analysis Avg. -BIL, BSQ Differential Prediction-Spatial JPEG-LS 2.51 showed that the PSNR of DWT-TD ALS-RLE Differential Prediction- Avg. -BIL, BSQ JPEG-LS 2.66 Spectral compression is better compared to other techniques. CASIA V1 Polar Iris JPEG2000 2.42 Party Scene Lossless Intra Coding JPEG-LS 1.79 Table 3. Compression of PSNR. Little Colorado River Our Work-Avg. DWT-TD (ALS)- RLE 3.38 Method/Application Main Focus Image/Data PSNR (dB) Motor movement- Avg. step Volumetric Coding SPECK 34.26 size 5.4. PSNR JPEG2000- ROI JPEG2000 Shelter Island- ROI 32.49 OB-SPECK SA-DWT Shelter Island- BG 27.12 OB-SPECK SA-RLT Shelter Island- ROI 34.39 PSNR (dB) is the ratio of the maximum possible power OB-SPECK SA-RLT Shelter Island- BG 27.72 of input HSI to the power of error introduced by the Our Work-Avg. DWT-TD (ALS)- RLE Little Colorado River 36.67 compression. The quality of the reconstructed HSI is higher when the PSNR is higher. The quality metric 6. Conclusions analysis for the DWT-TD ALS-RLE compression in terms of PSNR for the seven bands of input HSI is The proposed HSI compression based on DWT, TD given in Figure 7. It is observed that the PSNR value with ALS method reduces the size of the 3D tensors varies between 31.1dB and 32.1dB. The maximum calculated from the four wavelet sub-images of the PSNR value is obtained in the band 6 of the input HSI spectral bands of HSIs. The experiments were and minimum PSNR value is obtained in the band 2 of conducted for an HSI of the little Colorado River with the input HSI. the HSI compression based on DWT, TD with ALS method and existing HSI compression methods based on JPEG-LS, JPEG2000, SPECK, OB-SPECK and

3D-SPECK and our work was found to be good in terms of CR and PSNR value. The future work involves

PSNR(db) the development of RLE and simplification of the tensor calculations to decrease the memory

Band Index consumption, computational load and processing time Figure 7. Quality metric analysis for the DWT-TD ALS-RLE in for the HSI compression. terms of PSNR. References The PSNR values of the existing 3D-SPECK compression [23] and DWT-TD ALS-RLE [1] Abrardo A., Barni M., Magli E., and Nencini F., compression is analyzed and compared for five HSI “Error-Resilient and Low-Complexity Onboard datasets in Figure 8. PSNR for HSI-1 in the case of Lossless Compression of Hyperspectral Images proposed compression is significantly higher than that by Means of Distributed Source Coding,” IEEE of 3D-SPECK compression. Then, CR increases Transactions on Geoscience and Remote Sensing , gradually in the case of the 3D-SPECK compared to vol. 48, no. 4, pp. 1892-1904, 2010. DWT-TD ALS-RLE compression, for datasets HSI-2, [2] Acevedo D. and Ruedin A., “Lossless HSI-3, HSI-4 and HSI-5. It is observed that on an Compression of Hyperspectral Images: Look-Up average basis PSNR for the DWT-TD ALS-RLE Tables with Varying Degrees of Confidence,” in compression is higher than that of JPEG-LS Proceedings of IEEE International Conference compression by approximately 5.96dB. on Acoustics Speech and Signal Processing , Dallas, pp. 1314-1317, 2010.

[3] Anonymous, The USC-SIPI Image Database- Volume 3: Miscellaneous., available at: http://sipi.usc.edu/database/database.php?volume PSNR(db) =misc, last visited 2013.

[4] Anonymous2008, USGS Digital Spectral Library HSI-Datset 06., available at: Figure 8. PSNR values for DWT-TD ALS-RLE compression and http://speclab.cr.usgs.gov/spectral.lib06/ds231/, 3D-SPECK compression. last visited 2013. PSNR obtained by other researchers using [5] Cai Q., Song L., Li Guichun., and Ling N., techniques like volumetric coding [22], JPEG2000- “Lossy and Lossless Intra Coding Performance ROI [16] and OB-SPECK [16] are analyzed and Evaluation: HEVC, H.264/AVC, JPEG 2000 and compared with DWT-TD ALS-RLE compression in JPEG LS,” in Proceedings of Signal and Information Processing Association Annual

Hyperspectral Image Compression Based on DWT and TD with ALS Method 441

Summit and Conference , Hollywood, pp. 1-9, Coding of Remote Sensing Two-Dimensional 2012. Images,” IEEE Geoscience and Remote Sensing [6] Chang L., Chang y., Tang Z., and Huang B., Letters , vol. 8, no. 2, pp. 326-330, 2011. “Group and Region based Parallel Compression [17] Lilian N. and Leila F., “Performance evaluation Method using Signal Subspace Projection and of Data Compression Systems Applied to Band Clustering for Hyperspectral Imagery,” Satellite Imagery,” Journal of Electrical and IEEE Journal of Selected Topics in Applied Computer Engineering , vol. 2012, no. 18, pp. 1- Earth Observations and Remote Sensing , vol. 4, 15, 2012. no. 3, pp. 565-578, 2012. [18] Lopez S., Vadimirova T., Gonzalez C., Resano [7] Duarte M. and Baraniuk R., “Kronecker J., Mozos D., and Plaza A., “The Promise of Compressive Sensing,” IEEE Transactions on Reconfigurable Computing for Hyperspectral Image Processing , vol. 21, no. 2, pp. 494- Imaging Onboard Systems: A Review and 5042012. Trends,” Proceedings of the IEEE , vol. 101, no. [8] Dutra A., Pearlman W., and Silva E., “Successive 3, pp. 698-722, 2013. Approximation Wavelet Coding of AVIRIS [19] Mazgut J., Tino P., Boden M., and Yan H., Hyperspectral Images,” IEEE Journal of Selected “Dimensionality Reduction and Topographic Topics in Signal Processing , vol. 5, no. 3, pp. Mapping of Binary Tensors,” Pattern Analysis 370-385, 2011. and Applications , vol. 17, no. 3, pp. 1-19, 2013. [9] García-Vílchez F, Muñoz-Marí J., Zortea M., [20] Rawat M. and Sukadev S., “A Hybrid Image Blanes I., González-Ruiz V., Camps-Valls G., Compression Scheme using DCT and Fractal Plaza A., and Serra-Sagristà J., “On the Impact of Image Compression,” the International Arab Lossy Compression on Hyperspectral Image Journal of Information Technology , vol. 10, no. Classification and Unmixing,” Geoscience and 6, PP. 553-562, 2013. Remote Sensing Letters , vol. 8, no. 2, pp. 253- [21] Shoham T., Malah D., and Shechtman S., 257, 2011. “Quality Preserving Compression of a [10] Gowey K. and S Al., “AVIRIS Data-Ordering Concatenative Text-To-Speech Acoustic Free AVIRIS Standard Data Products,” available Database,” IEEE Transactions on Audio, Speech , at: http://aviris.jpl.nasa.gov/data/free_data.html, and Language Processing , vol. 20, no. 3, pp. last visited 2013. 1056-1068, 2012. [11] Hendrix E., García I., Plaza J, Martin G., and [22] Srinivasan K., Dauwels J., and Reddy M., Plaza A., “A New Minimum-Volume Enclosing “Multichannel EEG Compression: Wavelet- Algorithm for Endmember Identification and Based Image and Volumetric Coding Approach,” Abundance Estimation in Hyperspectral Data,” IEEE Journal of Biomedical and Health IEEE Transactions on Geoscience and Remote Informatics , vol. 17, no. 1, pp. 113-120, 2013. Sensing , vol. 50, no. 7, pp. 2744-2757, 2012. [23] Sriraam N. and Shyamsunder R., “3-D Medical [12] Horvath K., Stögner H., Uhl A., and Weinhandel Image Compression using 3-D Wavelet Coders,” G., “Lossless Compression of Polar Iris Image Digital Signal Processing , vol. 21, no. 1, pp. 100- Data,” in Proceedings of the 5 th Iberian 109, 2011. Conference , Spain, pp. 329-337, 2011. [24] Song H., Song Z., Deng G., Ma Y., Ma P., [13] Hou Y. and Li Y., “Hyperspectral Image Lossy- “Differential Prediction-Based Lossless to-Lossless Compression using 3D SPEZBC Compression with Very Low-Complexity for Algorithm Based on KLT and Wavelet Hyperspectral Data,” in Proceedings of the 13 th Transform,” in Proceedings of the 2 nd Sino International Conference on Communication foreigninterchange Workshop , China, pp. 713- Technology , Jinan, pp. 323-327, 2011. 721, 2012. [25] Wahl S., Haitham A., Tantawy H, Wang Z., [14] Iordache M., Bioucas-Dias J., and Plaza A., Werner P., and Simon S., “Exploitation of “Sparse Unmixing of Hyperspectral Data,” IEEE Context Classification for Parallel Pixel Coding Transactions on Geoscience and Remote Sensing , in JPEG-LS,” in Proceedings of the 18 th vol. 49, no. 6, pp. 2014-2039, 2011. International Conference on Image Processing , [15] Karami A., Yazdi M., and Mercier G., Brussels, pp. 2001-2004, 2011. “Compression of Hyperspectral Images using [26] Yin-Tsung Y., Cheng-Chen H., and Ruei-Ting Discerete Wavelet Transform and Tucker L., “Lossless Hyperspectral Image Compression Decomposition,” IEEE Journal of Selected System-based on HW/SW Codesign,” Embedded Topics in Applied Earth Observations and Systems Letters , vol. 3, no. 1, pp. 20-23, 2011. Remote Sensing , vol. 5, no. 2, pp. 444-450, 2012. [27] Zhou R., Liu L., Yin S., Luo A., Chen X., and [16] Jiao L., Wang L., Wu J., ; Bai J., Wang S., and Wei S., “A VLSI Design of Sensor Node for Hou B., “Shape-Adaptive Reversible Integer Wireless Image Sensor Network,” in Proceedings Lapped Transform for Lossy-to-Lossless ROI

442 The International Arab Journal of Information Technology, Vol. 13, No. 4, July 2016

of IEEE International Symposium on Circuits and Systems , Paris, pp. 149-152, 2010.

Kala Rajan is Associate Professor in Sri Subramanya College of engineering and technology, Palani, Tamilnadu, India. She holds M.E degree in Applied Electronics in Anna University, Trichy and she is a fellow membership of MISTE. She has also, completed her master degree in business administration in the year 2002. She has more than 11 years of experience including extensive project management experience in planning and leading a range of Electronic projects during these years, she shouldered a number of teachings, and societal based assignments. She also leads and teaches modules at both BE and ME levels in electronics and communication

Vasuki Murugesan completed her PhD degree in Color Image Processing from Anna University, Chennai. Presently, she is working as Professor and Head in Electronics and Communication Engineering Department at Velammal College of Engineering and Technology, Madurai, Tamilnadu, India. She has 26 years of teaching experience. Her research interests are in computer vision, wavelet based image analysis, bio medical image processing and pattern recognition schemes. She has published 30 papers in International Journals and presented 200 papers in International and National conferences. She is a Life Member of Indian Society of Technical Education (ISTE) and Fellow member of Institution of Electronics and Telecommunication Engineers (IETE) and Member in IEEE.