This PDF file contains the front matter associated with SPIE Proceedings Volume 7455, including the Title Page, Copyright information, Table of Contents, Introduction (if any), and the Conference Committee listing.

This paper investigates a new concept, called Hyperpsectral Information Compression (HIC) as opposed to Hyperspectral Data Compression (HDC) commonly used in the literature. A key feature that differentiates the HIC from the HDC is its focus on compression of information rather than data size. Such information compression is completely determined by a specific application in data exploitation. Accordingly, the HIC can be also referred to as exploitationbased compression. In order to substantiate the utility of the HIC, experiments are conducted to demonstrate that the HIC is indeed more effective than traditional HDC.

For optimum compression performances of multispectral and hyperspectral images, algorithms must exploit both spectral and spatial correlation of the data. To do this, different approaches are possible: a spectral decorrelation preprocessing stage followed by application of an image compressor to the decorrelated bands or, an integrated solution dealing with the three dimensions simultaneously. For example, Part II of JPEG2000 standard introduces a multi-component transform capability applied prior Part I spatial wavelet decomposition and coding. This article proposes to use the CCSDS Image Data Compression Recommendation together with a spectral transform to perform a multicomponent compression. Depending on the number of spectral bands, an efficient spectral transform, such as the DCT, is applied first and the CCSDS algorithm encodes each decorrelated bands. We compare the performances of such a scheme with JPEG2000 and also with a comparable scheme with a very simple decorrelation stage. Thanks to a bit plane coding of blocks of wavelet coefficients, the CCSDS encoder is a good tool to control the quality or the rate of these transformed bands. We present performances for different types of sensors: multispectral and hyperspectral. This work is part of the CNES contribution to the new CCSDS Multispectral and Hyperspectral Data Compression Working Group.

Hyperspectral image has weak spatial correlation and strong spectral correlation. As to exploit spectrum redundancy sufficiently, it must be pre-processed. In this paper, a new algorithm for lossless compression of hyperspectral images based on adaptive band regrouping is proposed. Firstly, the affinity propagation clustering algorithm (AP) is chosen for band regrouping according to interband correlation. Then a linear prediction algorithm based on context prediction is applied to the hyperspectral images in different groups. Finally, the experimental results show that the proposed algorithm achieves performance gains of 1.12bpp over the conventional algorithm.

It has passed more than a decade since the Consultative Committee for Space Data Systems (CCSDS) made its recommendation for lossless data compression. The CCSDS standard is commonly used for scientific missions because it is a general-purpose lossless compression technique with a low computational cost which results in acceptable compression ratios. At the core of this compression algorithm it is the Rice coding method. Its performance rapidly degrades in the presence of noise and outliers, as the Rice coder is conceived for noiseless data following geometric distributions. To overcome this problem we present here a new coder, the so-called Prediction Error Coder (PEC), as well as its fully adaptive version (FAPEC) which we show is a reliable alternative to the CCSDS standard. We show that PEC and FAPEC achieve large compression ratios even when high levels of noise are present in the data. This is done testing our compressors with synthetic and real data, and comparing the compression ratios and processor requirements with those obtained using the CCSDS standard.

In this paper, we propose a bit allocation method for 2D compression of hyperspectral images to enhance classification performance. First, we select a number of classes from original hyperspectral images. It is noted that the classes can be automatically selected by applying an unsupervised segmentation method. Then, we apply a feature extraction method and determine discriminately dominant feature vectors. By examining the feature vectors, we determine the discriminant usefulness of each spectral band. Finally, based on the discriminant usefulness of the spectral bands, we determine bit allocation of each spectral band. Experimental results show that it is possible to enhance the discriminant information at the expense of PSNR. Depending on applications, one can either minimize the mean squared error or choose to preserve the classification capability of the hyperspectral images.

This paper develops to a new concept, called Progressive Dimensionality Reduction (PDR) which can perform data dimensionality progressive in terms of information preservation. Two procedures can be designed to perform PDR in a forward or backward manner, referred to forward PDR (FPDR) or backward PDR (BPDR) respectively where FPDR starts with a minimum number of spectral-transformed dimensions and increases the spectral-transformed dimension progressively as opposed to BPDR begins with a maximum number of spectral-transformed dimensions and decreases the spectral-transformed dimension progressively. Both procedures are terminated when a stopping rule is satisfied. In order to carry out DR in a progressive manner, DR must be prioritized in accordance with significance of information so that the information after DR can be either increased progressively by FPDR or decreased progressively by BPDR. To accomplish this task, Projection Pursuit (PP)-based DR techniques are further developed where the Projection Index (PI) designed to find a direction of interestingness is used to prioritize directions of Projection Index Components (PICs) so that the DR can be performed by retaining PICs with high priorities via FPDR or BPDR. In the context of PDR, two well-known component analysis techniques, Principal Components Analysis (PCA) and Independent Component Analysis (ICA) can be considered as its special cases when they are used for DR.

Hyperspectral image compression has received considerable interest in recent years. However, an important issue that has not been investigated in the past is the impact of lossy compression on spectral mixture analysis applications, which characterize mixed pixels in terms of a suitable combination of spectrally pure spectral substances (called endmembers) weighted by their estimated fractional abundances. In this paper, we specifically investigate the impact of JPEG2000 compression of hyperspectral images on the quality of the endmembers extracted by algorithms that incorporate both the spectral and the spatial information (useful for incorporating contextual information in the spectral endmember search). The two considered algorithms are the automatic morphological endmember extraction (AMEE) and the spatial spectral endmember extraction (SSEE) techniques. Experimental results are conducted using a well-known data set collected by AVIRIS over the Cuprite mining district in Nevada and with detailed ground-truth information available from U. S. Geological Survey. Our experiments reveal some interesting findings that may be useful to specialists applying spatial-spectral endmember extraction algorithms to compressed hyperspectral imagery.

The amount of data generated by hyper- and ultraspectral imagers is so large that considerable savings in data storage and transmission bandwidth can be achieved using data compression. Due to the large amount of data, the data compression time is of importance. Increasing programmability of commodity Graphics Processing Units (GPUs) allows their usage as General Purpose computation on Graphical Processing Units (GPGPU). GPUs offer potential for considerable increase in computation speed in applications that are data parallel. Data parallel computation on image data executes the same program on many image pixels on parallel. We have implemented a spectral image data compression method called Linear Prediction with Constant Coefficients (LP-CC) using Nvidia's CUDA parallel computing architecture. CUDA is a parallel programming architecture that is designed for data-parallel computation. CUDA hides the GPU hardware from the developers. Moreover, CUDA does not require the programmers to explicitly manage threads. This simplifies the programming model. Our GPU implementation is experimentally compared to the native CPU implementation. Our speed-up factor was over 30 compared to a single threaded CPU version.

Wavelet image compression provides high compression ratios without producing noticeable visible artifacts. However, wavelet image compression can be expensive in terms of memory usage and computational complexity. This paper describes a low-cost algorithm suitable for implementation in Field-Programmable Gate Arrays (FPGAs).

In this paper, we propose a lifting architecture based on a basic lifting unit, whose structure performs lifting operations in a repetitive way. By analyzing computational processes in lifting in detail, the reusable Basic Lifting Element (BLE) is presented. The BLE structure is designed and optimized from the viewpoint of hardware implementation. The proposed lifting processor can be executed by arranging BLEs repeatedly. Experimental results show that the proposed architecture can transform any size of tiles with 9/7 filter and 5/3 filter for lossy and lossless compression, respectively. The lifting processor is designed in Verilog HDL and synthesized into Xilinx FPGA, which can run up to 130MHz.

In this paper, we develop a computationally efficient approach for lossy compression of remotely sensed hyperspectral images which has been specifically tuned to preserve the relevant information required in spectral mixture analysis (SMA) applications. The proposed method is based on two steps: 1) endmember extraction, and 2) linear spectral unmixing. Two endmember extraction algorithms: the pixel purity index (PPI) and the automatic morphological endmember extraction (AMEE), and a fully constrained linear spectral unmixing (FCLSU) algorithm have been considered in this work to devise the proposed lossy compression strategy. The proposed methodology has been implemented in graphics processing units (GPUs) of NVidia^TM type. Our experiments demonstrate that it can achieve very high compression ratios when applied to standard hyperspectral data sets, and can also retain the relevant information required for spectral unmixing in a computationally efficient way, achieving speedups in the order of 26 on a NVidia^TM GeForce 8800 GTX graphic card when compared to an optimized implementation of the same code in a dual-core CPU.

Principal component analysis (PCA) is the most efficient spectral decorrelation approach for hyperspectral imagery. In conjunction with JPEG2000 for optimal bit allocation and spatial coding, the resulting PCA+JPEG2000 can yield superior rate-distortion performance and the following data analysis performance. However, the involved overhead bits consumed by the large transformation matrix may affect the performance at low bitrates, particularly when the image spatial size is relatively small compared to the spectral dimension. In this paper, we propose to apply the segmented principal component analysis (SPCA) to mitigate this effect. The resulting SPCA+JPEG200 may improve the compression performance even when PCA+JPEG2000 is applicable.

Hyperspectral images used in remote sensing can reach hundreds of megabytes due to the large number of components and high spatial and bit-depth resolution. When these images have to be transmitted, interactive transmission is necessary to deliver only those portions of the image that the client has requested. In such a scenario, compression is a useful tool to reduce the required amount of network bandwidth. JPEG2000 is a powerful image and video coding standard that, among other features, provides scalability by spatial location, component, quality, and resolution. This is used by the JPEG2000 Interactive Protocol (JPIP) to enable the interactive transmission of imagery. One of the most important aspects of a JPIP server is the rate-control algorithm used to select the portions of the compressed code-stream that will be delivered to client. The purpose of this research is to introduce new rate-control methods for the JPIP server aimed to achieve optimal performance. We focus our attention in the adaptation of two rate-control methods developed for the JPEG2000 coder and decoder. Experimental results suggest that both methods significantly improve coding performance without penalizing computational load.

A novel theory of information acquisition-"compressive sampling" has been applied in this paper, and goes against the common wisdom in data acquisition of Shannon theorem. CS theory asserts that one can recover certain signals and images perfectly from far fewer samples or measurements than traditional methods use. This paper presents an improvement on genetic algorithm instead of match pursuit algorithm in consideration of the enormous computational complexity on sparse decomposition. Then the whole image is divided into small blocks which can be processed by sparse decomposition, and an end to decomposition is determined by PSNR threshold adaptively. At last, the experiment results show that good performance on image reconstruction with less computational complexity has been achieved.

By using fixed-length codewords and not using an adaptive dictionary, the Tunstall coding stands out as an errorresilient lossless compression technique. Our previous work has shown that use of the non-exhaustive parse tree can improve the compression ratio of the Tunstall coding without compromising its error resilience. As change of codeword assignment will only affect the error propagation, not the compression ratio, we experimented in this work the various codeword assignment methods to minimize the error propagation of our revised Tunstall coding. The result shows that the breadth first assignment of codewords to the Tunstall parse tree gives the lowest percent error rate. Also the linear prediction followed by non-exhaustive Tunstall parse tree, and the alignment marker (LP+Tunstall non-exhaustive) has the best performance in terms of compression ratio and percent error rate when compared to other LP+Bitcut+Tunstall methods, JPEG-2000 and CCSDS Rice coding schemes.

Endmember extraction has recently received considerable attention in hyperspectral data exploitation since they represent crucial and vital information for hyperspectral data analysis. So far, no work has been reported on how to implement endmember extraction algorithms in real-time. An endmember is defined as an idealized signature and may or may not exist as a data sample or an image pixel. The interest of endmember extraction arises in the use of hundreds of contiguous spectral channels that allows a hyperspectral imaging sensor to uncover many subtle substances in diagnostic bands. However, finding such substances also presents a great challenge to hyperspectral data analysts and becomes imperative when it comes to satellite communication if a hyperspectral imaging sensor is operated in space platform where bandwidths used by satellite links may be very limited and downloading all the data may not be realistic in many practical applications. In order to address this need many endmember extraction algorithms have been developed and designed in the past, but no work has been reported on how to implement endmember extraction algorithms in real-time. This paper investigates this issue in designing algorithms for real time processing of endmember extraction and developed several endmember extraction algorithms derived from the widely used N-finder algorithm (NFINDR) that can be implemented in real time.

The extraction of roads from high spatial resolution remote sensing images remains a problem though lots of efforts have been made in this area. High spatial resolution remote sensing images represent the surface of the earth in detail. As spatial resolution increases, spectral variability within the road cover units becomes complex and traditional remote sensing image processing methods on pixel basis are no longer suitable. This paper studies automatic road extraction from remote sensing images based on methods of Pulse-Coupled Neural Network and mathematical morphology. PCNN is a useful biologically inspired algorithm, and has the properties of linking field and dynamic threshold which make similar neurons generate pulses simultaneously. PCNN has the ability of a neuron to capture neighboring neurons which are in similar states and the independency of the pulses within unattached neuron regions. The method of mathematical morphology has the prime principle which is using a certain structure element to measure and extract the corresponding form in an image. In this paper, the simplified PCNN is applied as the image segmentation algorithm, and morphological transformation is used to purify the roads' information and to extract the road centerlines. Experimental results show that this method is efficient in road extraction from remote sensing images.

In this paper, we develop several parallel techniques for hyperspectral image processing that have been specifically designed to be run on massively parallel systems. The techniques developed cover the three relevant areas of hyperspectral image processing: 1) spectral mixture analysis, a popular approach to characterize mixed pixels in hyperspectral data addressed in this work via efficient implementation of a morphological algorithm for automatic identification of pure spectral signatures or endmembers from the input data; 2) supervised classification of hyperspectral data using multi-layer perceptron neural networks with back-propagation learning; and 3) automatic target detection in the hyperspectral data using orthogonal subspace projection concepts. The scalability of the proposed parallel techniques is investigated using Barcelona Supercomputing Center's MareNostrum facility, one of the most powerful supercomputers in Europe.

According to the data characteristics of remote sensing stereo image pairs, a novel compression algorithm based on the combination of feature-based image matching (FBM), area-based image matching (ABM), and region-based disparity estimation is proposed. First, the Scale Invariant Feature Transform (SIFT) and the Sobel operator are carried out for texture classification. Second, an improved ABM is used in the area with flat terrain (flat area), while the disparity estimation, a combination of quadtree decomposition and FBM, is used in the area with alpine terrain (alpine area). Furthermore, the radiation compensation is applied in every area. Finally, the disparities, the residual image, and the reference image are compressed by JPEG2000 together. The new algorithm provides a reasonable prediction in different areas according to characteristics of image textures, which improves the precision of the sensed image. The experimental results show that the PSNR of the proposed algorithm can obtain up to about 3dB's gain compared with the traditional algorithm at low or medium bitrates, and the subjective quality is obviously enhanced.

In order to effectively store and transmit MODIS multispectral data, a lossless compression method based on mix coding and integer wavelet transform (IWT) is proposed in this paper. Firstly, the algorithm computes the correlation coefficients between spectrums in MODIS data. Using proper coefficient threshold, the original bands will be divided two groups: one group use spectral prediction method and then compress residual error, while the other group data is directly compressed by some standard compressor. For the spectral prediction group, we can find the current band that has greatest correlation with the previous band by the judgments of correlation coefficient, thus the optimal spectral prediction sequence is obtained by band reordering. The prediction band data can be computed with the previous band data and optimal linear predictor, so the spectral redundancy can be eliminated by using spectral prediction. In order to reduce residual differences in further, the block optimal linear predictor is designed in this paper. Next, except for the first band of the spectral prediction sequence, the residual errors of other bands are encoded by IWT and SPIHT. The direct compression bands and the first band of spectral prediction sequence are compressed by JPEG2000. Finally, the coefficients of block optimal linear predictor and other side information are encoded by adaptive arithmetic coding. The experimental results show that the proposed method is efficient and practical for MODIS data.

Adaptive Group of Pictures (GOP) is helpful for increasing the efficiency of video encoding by taking account of characteristics of video content. This paper proposes a method for adaptive GOP structure selection for video encoding based on motion coherence, which extracts key frames according to motion acceleration, and assigns coding type for each key and non-key frame correspondingly. Motion deviation is then used instead of motion magnitude in the selection of the number of B frames. Experimental results show that the proposed method for adaptive GOP structure selection achieves performance gain of 0.2-1dB over the fixed GOP, and has the advantage of better transmission resilience. Moreover, this method can be used in real-time video coding due to its low complexity.

Our research focuses on reducing complexity of hyperspectral image codecs based on transform and/or subband coding, so they can be on-board a satellite. It is well-known that the Karhunen-Loève Transform (KLT) can be sub-optimal in transform coding for non Gaussian data. However, it is generally recommended as the best calculable linear coding transform in practice. Now, the concept and the computation of optimal coding transforms (OCT), under low restrictive hypotheses at high bit-rates, were carried out and adapted to a compression scheme compatible with both the JPEG2000 Part2 standard and the CCSDS recommendations for on-board satellite image compression, leading to the concept and computation of Optimal Spectral Transforms (OST). These linear transforms are optimal for reducing spectral redundancies of multi- or hyper-spectral images, when the spatial redundancies are reduced with a fixed 2-D Discrete Wavelet Transform (DWT). The problem of OST is their heavy computational cost. In this paper we present the performances in coding of a quasi optimal spectral transform, called exogenous OrthOST, obtained by learning an orthogonal OST on a sample of superspectral images from the spectrometer MERIS. The performances are presented in terms of bit-rate versus distortion for four various distortions and compared to the ones of the KLT. We observe good performances of the exogenous OrthOST, as it was the case on Hyperion hyper-spectral images in previous works.

Satellite images and aerial images have different radiation attributes. Conclusions on aerial image compression effects may not be applicable to satellite images. In this paper we study the effects of lossy compression on the digital terrain model (DTM) derived from a selected stereo pair of SPOT-5 satellite images. The satellite images are compressed by the Kakadu JPEG2000 compression software at the compression ratios of 2, 4, 6, and 8. The Imageinfo Pixelgrid V2.0 software is used as a DTM generator. The DTM results derived from original images are compared with that derived from compressed images in terms of the mean error and the standard deviation. This paper's experiment indicates that when the compression ratio rises to 4, both the DTM generation and the stereo observation by human will be affected dramatically.

This paper addresses assessment of different processing techniques for hyperspectral images target detection. An ad-hoc quality assessment approach is adopted to compare different noise reduction techniques of hyperspectral images for target detection applications. Two different noise reduction techniques are applied to a datacube collected over a well-studied area with human made targets. The quality of these noise reduced datacubes in preserving the identity of the targets of interest is compared with that of the original datacube. This is achieved by applying different measures on the datacubes. First, the Virtual Dimensionality is used and the results for both of the noise reduction methods are compared with those of the original datacube for several false-alarm probabilities. Then Maximum Noise Fraction is applied to the datacubes and its capability in finding a transform in which the information of the datacube is represented in a smaller number of bands is assessed. Finally using set measures and knowing the location of the targets, different classes are defined and the intraclass and interclass distances for each datacube is measured.

Automatic target detection in hyperspectral images is a task that has attracted a lot of attention recently. In the last few years, several algoritms have been developed for this purpose, including the well-known RX algorithm for anomaly detection, or the automatic target detection and classification algorithm (ATDCA), which uses an orthogonal subspace projection (OSP) approach to extract a set of spectrally distinct targets automatically from the input hyperspectral data. Depending on the complexity and dimensionality of the analyzed image scene, the target/anomaly detection process may be computationally very expensive, a fact that limits the possibility of utilizing this process in time-critical applications. In this paper, we develop computationally efficient parallel versions of both the RX and ATDCA algorithms for near real-time exploitation of these algorithms. In the case of ATGP, we use several distance metrics in addition to the OSP approach. The parallel versions are quantitatively compared in terms of target detection accuracy, using hyperspectral data collected by NASA's Airborne Visible Infra-Red Imaging Spectrometer (AVIRIS) over the World Trade Center in New York, five days after the terrorist attack of September 11th, 2001, and also in terms of parallel performance, using a massively Beowulf cluster available at NASA's Goddard Space Flight Center in Maryland.

In this paper, we investigate automated display methods for hyperspectral images with unsupervised segmentation. First, we apply an unsupervised segmentation method, which will produce a number of unlabeled classes. Then, we choose the classes whose sizes are larger than a threshold value. Then, we apply a feature extraction method to the chosen classes and find dominant discriminant features, which are used to display the hyperspectral images. We also exploit the use of the principal component analysis for the display of hyperspectral images. Experimental images show that the color images produced by the proposed methods show interesting characteristics compared to the conventional pseudo-color image.

Cloud is one of common noises in MODIS remote sensing image. Because of cloud interference, much important information covered with cloud can't be obtained. In this paper, an effective method is proposed to detect and remove thin clouds with single MODIS image. The proposed method involves two processing-thin cloud detection and thin cloud removal. As for thin cloud detection, through analyzing the cloud spectral characters in MODIS thirty-six bands, we can draw the conclusion that the spectral reflections of ground and cloud are different in various MODIS band. Hence, the cloud and ground area can be separately identified based on MODIS multispectral analysis. Then, the region label algorithm is used to label thin clouds from many candidate objects. After cloud detection processing, thin cloud removal method is used to process each cloud region. Comparing with traditional methods, the proposed method can realize thin cloud detection and removal with single remote sensing image. Additionally, the cloud removal processing mainly aims to the cloud label region rather than the whole image, so it can improve the processing efficiency. Experiment results show the method can effectively remove thin cloud from MODIS image.

In the most application situation, signal or image always is corrupted by additive noise. As a result there are mass methods to remove the additive noise while few approaches can work well for the multiplicative noise. The paper presents an improved MAP-based filter for multiplicative noise by adaptive window denoising technique. A Gamma noise models is discussed and a preprocessing technique to differential the matured and un-matured pixel is applied to get accurate estimate for Equivalent Number of Looks. Also the adaptive local window growth and 3 different denoise strategies are applied to smooth noise while keep its subtle information according to its local statistics feature. The simulation results show that the performance is better than existing filter. Several image experiments demonstrate its theoretical performance.

Signal space diversity (SSD) has been viewed as a very efficient method to increase the diversity order and the reliability of wireless transmission over fading channels. The nonbinary LDPC code can provide powerful error protection over fading channels. The combination of non-binary LDPC codes and SSD is an effective technique to fading and strong noise for fading channel. This paper discusses a signal space diversity with a precoder by adding a simple encoder before applying the rotation matrix that can be viewed as a SSD with rate below 1. We propose three design criteria and two implementation method. And we further consider a nonbinary LDPC-coded signal space diversity with precoding and their rate re-allocation between channel coding and SSD. It is shown that our scheme outperforms the conventional coded- SSD over the fading channel when we get a better rate allocation that depends on the specific channel coding.

In this paper, we propose an embedded satellite image compression method using the weighted zeroblock coding and optimal sorting. Unlike the conventional quad-tree coding methods such as Set Partition Embedded block (SPECK) and Embedded ZeroBlock Coder (EZBC), in the proposed weighted zeroblock coding (WZBC), 1) we use the unfixed scanning order to code a significant block-set with fewer bits and achieve variable-length quad-tree coding; 2) we exploit the significance degree of sub-blocks which is predicted by a novel context-based weighted strategy to optimize the scanning order of the variable-length quad-tree coding and obtain the weighted zeroblock coding; 3) the rate-distortion performance of WZBC is also optimized by using the above mentioned weight. In the binary mode, the new method does not employ the arithmetic coding and has a fairly low complexity. Experimental results show that the proposed WZBC in the binary coding mode can provide an excellent coding performance compared with SPECK and Set Partitioning In Hierarchical Trees (SPIHT) with arithmetic coding, and can even closely approach JPEG2000. When the arithmetic coding is extensively used, the proposed method can obtain more obvious gain.

In this paper, we propose a method based on both 3D-wavelet transform and low-density parity-check codes to realize the compression of hyperspectral images on the framework of DSC (Distributed Source Coding). The new approach which combines DSC and 3D-wavelet transform technique makes it possible to realize low encoding complexity at the encoder and achieve efficient performance of hyperspectral image compression. The experimental results for hyperspectral image coding show that the new method can performs better than 3D-SPIHT and can outperform than 2D-SPIHT and JPEG2000.