This PDF file contains the front matter associated with SPIE Proceedings Volume 7334, including the Title Page, Copyright information, Tabe of Contents, Introduction (if any), and the Conference Committee listing.

A large number of hyperspectral detection algorithms have been developed and used over the last two decades. Some algorithms are based on highly sophisticated mathematical models and methods; others are derived using intuition and simple geometrical concepts. The purpose of this paper is threefold. First, we discuss the key issues involved in the design and evaluation of detection algorithms for hyperspectral imaging data. Second, we present a critical review of existing detection algorithms for practical hyperspectral imaging applications. Finally, we argue that the "apparent" superiority of sophisticated algorithms with simulated data or in laboratory conditions, does not necessarily translate to superiority in real-world applications.

The hyperspectral matched filter (MF) is a popular tool in remote sensing problems for locating objects that extend over several pixels. However, it is ideally suited only for the detection of sub-pixel targets with known mean signature in radiance space. Here we develop an alternative affine matched filter (AMF) that is more appropriate for detecting extended targets, and which accommodates practical uncertainties in target signature knowledge. In particular, AMF is ideal when used in conjunction with another new method we develop, called Virtual Relative Calibration (VRC), for generating a radiance space representation of a laboratory reflectance signature. VRC is related to the QUick Atmospheric Correction(QuAC) method but is much simpler. We also devise and test an extension of AMF that is meant to reduce false alarms caused by dark pixels. This Joint AMF (JAMF) expands the standard statistical model of hyperspectral backgrounds to accommodate variable illumination levels, analogously to how AMF treats target level uncertainty.

An anomaly detector for hyperspectral imaging based on partialling out the effect of the clutter subspace is devised. The partialling maximizes the squared correlation between each spectral component and a linear predictor, with no restrictions on the form of the probability distribution. The detection step is defined by thresholding a Mahalanobis measure of the prediction error. The method is compared to conventional anomaly detectors using VNIR hyperspectral imagery.

Hyperspectral anomaly detector Subspace RX (SSRX), a version of the RX algorithm, has been extensively investigated and its real time implementation performs reasonably well with hyperspectral imagery. However, the problem of false alarms arising at the boundaries between different backgrounds yet needs to be addressed. We consider two paths to the solution: virtual imagery rotation to change the scanning direction and introduction of latency into the data processing. The former approach is based on analysis of the detection result's dependence on the orientation of background boundaries relative to the scanning direction. The latter approach enables incorporation of various background statistics before detection reaches the boundary. Performance of improved SSRX is investigated using collected and simulated hyperspectral imagery.

Projection Pursuit (PP) is a component transform technique which seeks a component whose projection vector points to a direction of interestingness in data space which can be specified by a Projection Index (PI). Two most popular component analysis-based techniques, Principal Components Analysis (PCA), Independent Component Analysis (ICA) can be considered as special cases with their PIs specified by data variance and statistical independency respectively. Despite the fact that various component analysis-based techniques have been used for Dimensionality Reduction (DR) the components are generally generated by a specific technique. Even in the case of PP, the same PI has been used to generate project components. This paper explores the utility of PP in DR where various projection indexes are used for DR in context of PP. It further lays out a general setting for PP-based DR and develops algorithms to perform one dimension reduction at a time by using different PIs. In order to substantiate our findings, experiments are conducted to demonstrate advantages of the PP with mixed PIs-based DR over traditional PCA-based, ICA-based and PP-based DR techniques.

Hyperspectral band selection extracts several bands of importance in some sense by taking advantage of high spectral correlation. Driven by detection or classification accuracy, one would expect that using a subset of original bands the accuracy is unchanged or tolerably degraded while computational burden is significantly relaxed. When the desired object information is known, i.e., supervised band selection, this task can be achieved by finding the bands that contain the most information about these objects. When the desired object information is unknown, i.e., unsupervised band selection, the objective is to select the most distinctive and informative bands. It is expected that these bands can provide an overall satisfactory detection and classification performance. We propose an unsupervised band selection algorithm based on band similarity measurement, which can yield a better result in terms of information conservation and class separability than other widely used techniques. We also extend this algorithm to the case when the desired object information is known. The experimental result shows the effectiveness of this new algorithm.

This paper reports a comparative study of current lossless compression algorithms for data from a representative selection of satellite based earth science multispectral imagers. The study includes the performance of compression algorithms on Advanced Very High Resolution Radiometer(AVHRR), SEVIRI, the Moderate Resolution Imaging Spectroradiometer(MODIS) imager, as well as a subset of MODIS bands as a proxy for the upcoming GOES-R series. SEVIRI aboard the ESA/EUMETSAT operated Meteosat Second Generation (MSG) satellites is a geostationary imager. The AVHRR aboard the NOAA Polar Orbiting Environmental Satellites and MODIS aboard the NASA Terra and Aqua satellites have polar orbits. Thus this study will present representatives from both polar and geostationary orbiting imagers. The imagers we include have sensors for both reflected and emissive radiance. We also note that the older satellites have coarser quantizations and present our conclusions on the impact on compression ratios. Faced with a enormous growing large volume of data on a new emerging current generation images from faster scanning, finer spatial resolution, and greater spectral resolution, this study provides a comparison of current compression algorithms as a baseline for future work. With growing satellite Earth science multispectral imager volume data, it becomes increasingly important to evaluate which compression algorithms are most appropriate for data management in transmission and archiving. This comparative compression study uses a wide range standard implementations of the leading lossless compression algorithms. Examples include image compression algorithms such as PNG and JPEG2000, and widely-used file compression formats such as BZIP2 and 7z. This study includes a comparison with the Consultative Committee for Space Data Systems (CCSDS) recommended Szip software which uses the extended-Rice lossless compression algorithm as well as the most recent recommended compression standard which relies on a wavelet transform followed by an entropy coder. To establish statistical significance of our analysis, we have developed a system to acquire and manage a large number of imager granules: currently over 1000 MODIS granules, over 2400 AVHRR granules, and over 220 SEVIRI granules.

Hyperspectral technology is currently being used by the military to detect regions of interest where potential targets may be located. Weather variability, however, may affect the ability for an algorithm to discriminate possible targets from background clutter. Nonetheless, different background characterization approaches may facilitate the ability for an algorithm to discriminate potential targets over a variety of weather conditions. In a previous paper, we introduced a new autonomous target size invariant background characterization process, the Autonomous Background Characterization (ABC) or also known as the Parallel Random Sampling (PRS) method, features a random sampling stage, a parallel process to mitigate the inclusion by chance of target samples into clutter background classes during random sampling; and a fusion of results at the end. In this paper, we will demonstrate how different background characterization approaches are able to improve performance of algorithms over a variety of challenging weather conditions. By using the Mahalanobis distance as the standard algorithm for this study, we compare the performance of different characterization methods such as: the global information, 2 stage global information, and our proposed method, ABC, using data that was collected under a variety of adverse weather conditions. For this study, we used ARDEC's Hyperspectral VNIR Adverse Weather data collection comprised of heavy, light, and transitional fog, light and heavy rain, and low light conditions.

Tracking of vehicles, in both radiance and reflectance hyperspectral imagery, was analyzed using traditional reflectance domain and forward predicting physics based target detection algorithms. The investigation consisted of locating vehicles in one image followed by locating the same (relocated) vehicles in another image that was collected at a different time. Traditional approaches to hyperspectral target detection involve the application of detection algorithms to atmospherically compensated imagery. Rather than compensate the imagery, a more recent approach uses physical models to generate radiance signature spaces. The signature space is actually a representation of what the target might look like to the sensor as the reflectance propagates through the atmosphere. The model takes into account atmospherics, illumination conditions and target reflectivity. Different collection times have varying atmospheric conditions leading to changes in the sensor-reaching radiance spectrum. A forward predicting physical model, applied in the radiance domain, handles this variation. The data used in this study is a new ground truthed hyperspectral data set, collect with the airborne HyMap sensor, and which is now freely available to the community through the web for evaluation, testing, and algorithm development. Vehicles were relocated across various flight lines. This paper compares vehicle detection results, across these multiple images collected within a short time period, using both the physical model and the traditional atmospheric compensation approach.

We use a bank of three-dimensional Gabor filters to represent the spectral/spatial properties of hyperspectral data. The orientation and scale selective properties of the filters allow the development of new algorithms that are invariant to rotation and scale. Since a large set of three-dimensional filters can be defined, we develop methods for reducing the number of features that are used to represent a region. The data reduction process is defined to optimize the features for classification. We demonstrate the efficacy of the approach using a large set of AVIRIS hyperspectral data.

An application of including spatial information into a principle components analysis (PCA) image classification through an eigenvector spatial filter is given. The spatial filter is created as a linear combination of the chosen eigenvectors based on the spatial information in the image. These eigenvectors are extracted from the defined connectivity of the image pixels and are different from those eigenvectors used in conventional PCA. The surface connectivity of the image is based on the binary matrix of image pixel neighbors (i.e. if i and j are neighbor pixels in an image and i does not equal j than matrix entry c_ij = 1 other c_ij = 0). The proposed methodology is applied to an image example (i.e. Flightline C1) with 12 spectral bands from an airborne sensor as a selected case study dataset. An explanation of how a spatial filter should be incorporated into PCA classification is given and two possibilities for gaining efficiency in the algorithm, through distributed computing and sampling is described. The advantages and drawbacks of both approaches, in terms of computing time, and amount of variance accounted for in the image is discussed.

Several studies have shown that the use of nonlocal operators can significantly remove noise and improve the quality of natural images. These operators are built on similarities between small local neighborhoods that are not necessarily spatially close, which plays a very important role in preserving the image structure, and are closely related to the kernel methods used in manifold learning and nonlinear dimension reduction. This serves as our motivation for exploring the use of nonlocal, linear, and nonlinear diffusion processes on high dimensional imagery (e.g. hyperspectral) that do not require the computation of eigenfunctions. We utilize the same iterative scheme to perform a semi-supervised multi-class classification and segmentation, only by changing the initial conditions. Furthermore, we compare the denoising performance of these algorithms with other PDE-based methods like anisotropic diffusion and compare classification accuracies for different materials on real Hyperspectral Image (HSI) cubes.

Hyperspectral data represents a mixture of several component spectra from many classifiable sources. The knowledge of the contributions of the underlying sources to the recorded spectra is valuable in many remote sensing applications. Traditional Hyperspectral classification and segmentation algorithms have used Markov random field (MRF) based estimation in recent investigations. Although, this method reflects plausible local, spatial correlation in the true scene, it is limited to using supervised learning schemes for parameter estimation. Besides, the expectation-maximization (EM) for the hidden MRF is considerably more difficult to realize due to the absence of a closed form formulation. In this paper, we propose a hidden Markov model (HMM) based probability density function (PDF) classifier for reduced dimensional feature space. Our approach uses an unsupervised learning scheme for maximum-likelihood (ML) parameter estimation that combines both model selection and estimation in a single algorithm. The proposed method accurately models and synthesizes the approximate observations of the true data in a reduced dimensional feature space.

Ann Arbor Sensor Systems is developing and experimenting with a Long Wave Multispectral Diffractive Optic Imaging Spectrometer that operates in the 7 to 14 micron band. The instrument currently under development is for fixed mount applications. The focal plane is an uncooled 320 x 240 30 micron pitch amorphous silicon microbolometer array. With a spectral bandwidth of 220nm the unit is capable of resolving moderate to broad spectral features. This paper will describe the physical characteristics and characterization of the experimental device. Design trades and test data will be presented.

Hyperspectral imaging has important benefits in remote sensing and material identification. This paper describes a class of hyperspectral imaging systems which utilize a novel optical processor that provides video-rate hyperspectral datacubes. These systems have no moving parts and do not operate by scanning in either the spatial or spectral dimension. They are capable of recording a full three-dimensional (two spatial, one spectral) hyperspectral datacube with each video frame, ideal for recording data on transient events, or from unstabilized platforms. We will present the results of laboratory and field-tests for several of these imagers operating in the visible, near-infrared, mid-wavelength infrared (MWIR) and long-wavelength infrared (LWIR) regions.

Advanced Coherent Technologies, LLC has demonstrated the use of multi-channel imaging systems in a variety of applications. These systems are composed of multiple cameras or 'channels', each of which can be coupled with spectral filters, polarization analyzers, or unique optics (e.g. for field of view (FOV) or aperture adjustments). The channel content is designed specifically for the extraction of information and/or the detection of targets. Airborne data collects have been made over forest and maritime environments for the detection of various targets. The results of these collects are discussed and analyzed. Of particular concern is how channel content is chosen in each environment and for each target.

A novel multi-object spectrometer (MOS) is being explored for use as an adaptive performance-driven sensor that tracks moving targets. Developed originally for astronomical applications, the instrument utilizes an array of micromirrors to reflect light to a panchromatic imaging array. When an object of interest is detected the individual micromirrors imaging the object are tilted to reflect the light to a spectrometer to collect a full spectrum. This paper will present example sensor performance from empirical data collected in laboratory experiments, as well as our approach in designing optical and radiometric models of the MOS channels and the micromirror array. Simulation of moving vehicles in a highfidelity, hyperspectral scene is used to generate a dynamic video input for the adaptive sensor. Performance-driven algorithms for feature-aided target tracking and modality selection exploit multiple electromagnetic observables to track moving vehicle targets.

We present a proof-of-principle demonstration using Sony's IBM Cell processor-based PlayStation 3 (PS3) to run-in near real-time-a hyperspectral anomaly detection algorithm (HADA) on real hyperspectral (HS) long-wave infrared imagery. The PS3 console proved to be ideal for doing precisely the kind of heavy computational lifting HS based algorithms require, and the fact that it is a relatively open platform makes programming scientific applications feasible. The PS3 HADA is a unique parallel-random sampling based anomaly detection approach that does not require prior spectra of the clutter background. The PS3 HADA is designed to handle known underlying difficulties (e.g., target shape/scale uncertainties) often ignored in the development of autonomous anomaly detection algorithms. The effort is part of an ongoing cooperative contribution between the Army Research Laboratory and the Army's Armament, Research, Development and Engineering Center, which aims at demonstrating performance of innovative algorithmic approaches for applications requiring autonomous anomaly detection using passive sensors.

Recent research in hyperspectral change detection has focused on the use of a single reference scene to identify anomalous changes that may be present in the current test scene by suppressing stationary background using various predictive algorithms. This paper extends such algorithms to the use of multiple reference scenes in an attempt to improve change detection performance in hyperspectral images in what is called multi-temporal change detection (MTCD). Often, an airborne hyperspectral sensor performs multiple reconnaissance passes over a specific spatial area. Consequently, a multi-temporal hyperspectral data set may exist for a spatial area of interest and the potential arises to improve clutter suppression and increase change detection performance by using multiple references. This multi-temporal change detection method, of course, requires precise co-registration of all the scenes being used as pixel level changes are being sought. Ground-based hyperspectral data collected using an imaging spectrometer mounted on a pan and tilt is used to perform this study. This method of collection helps ensure precise co-registration between scenes. The scenes are collected over a period of many months and consequently have different illumination and vegetation conditions present. These natural variations between scenes are not considered anomalous changes and should not be targeted by algorithms. A detection scene is collected as well with an anomalous change introduced to allow for testing. Various multi-temporal change detection approaches are derived using single-reference change detection techniques and simple signal processing knowledge. These methods are discussed and compared using ROC analysis on the data available.

An architecture and implementation is presented regarding persistent, hyperspectral, adaptive, multi-modal, feature-aided tracking within the urban context. A novel remote-sensing imager has been designed which employs a micro-mirror array at the focal plane for per-pixel adaptation. A suite of end-to-end synthetic experiments have been conducted, which include high-fidelity moving-target urban vignettes, DIRSIG hyperspectral rendering, and full image-chain treatment of the prototype adaptive sensor. Corresponding algorithm development has focused on: motion segmentation, spectral feature modeling, classification, fused kinematic/spectral association, and adaptive sensor feedback/control.

This study examines normalizing the imagery and the optimization metrics to enhance anomaly and change detection, respectively. The RX algorithm, the standard anomaly detector for hyperspectral imagery, more successfully extracts bright rather than dark man-made objects when applied to visible hyperspectral imagery. However, normalizing the imagery prior to applying the anomaly detector can help detect some of the problematic dark objects, but can also miss some bright objects. This study jointly fuses images of RX applied to normalized and unnormalized imagery and has a single decision surface. The technique was tested using imagery of commercial vehicles in urban environment gathered by a hyperspectral visible/near IR sensor mounted in an airborne platform. Combining detections first requires converting the detector output to a target probability. The observed anomaly detections were fitted with a linear combination of chi square distributions and these weights were used to help compute the target probability. Receiver Operator Characteristic (ROC) quantitatively assessed the target detection performance. The target detection performance is highly variable depending on the relative number of candidate bright and dark targets and false alarms and controlled in this study by using vegetation and street line masks. The joint Boolean OR and AND operations also generate variable performance depending on the scene. The joint SUM operation provides a reasonable compromise between OR and AND operations and has good target detection performance. In addition, new transforms based on normalizing correlation coefficient and least squares generate new transforms related to canonical correlation analysis (CCA) and a normalized image regression (NIR). Transforms based on CCA and NIR performed better than the standard approaches. Only RX detection of the unnormalized of the difference imagery in change detection provides adequate change detection performance.

This paper presents an algorithm for automated extraction of interest points (IPs)in multispectral and hyperspectral images. Interest points are features of the image that capture information from its neighbours and they are distinctive and stable under transformations such as translation and rotation. Interest-point operators for monochromatic images were proposed more than a decade ago and have since been studied extensively. IPs have been applied to diverse problems in computer vision, including image matching, recognition, registration, 3D reconstruction, change detection, and content-based image retrieval. Interest points are helpful in data reduction, and reduce the computational burden of various algorithms (like registration, object detection, 3D reconstruction etc) by replacing an exhaustive search over the entire image domain by a probe into a concise set of highly informative points. An interest operator seeks out points in an image that are structurally distinct, invariant to imaging conditions, stable under geometric transformation, and interpretable which are good candidates for interest points. Our approach extends ideas from Lowe's keypoint operator that uses local extrema of Difference of Gaussian (DoG) operator at multiple scales to detect interest point in gray level images. The proposed approach extends Lowe's method by direct conversion of scalar operations such as scale-space generation, and extreme point detection into operations that take the vector nature of the image into consideration. Experimental results with RGB and hyperspectral images which demonstrate the potential of the method for this application and the potential improvements of a fully vectorial approach over band-by-band approaches described in the literature.

We describe the development of a simulation framework for anomalous change detection that considers both the spatial and spectral aspects of the imagery. A purely spectral framework has previously been introduced, but the extension to spatio-spectral requires attention to a variety of new issues, and requires more careful modeling of the anomalous changes. Using this extended framework, we evaluate the utility of spatial image processing operators to enhance change detection sensitivity in (simulated) remote sensing imagery.

Hyperspectral images have textured regions and in many cases there are not sufficient samples to train classifiers. By simulating more samples that are self-similar to the original texture efficient classifiers can be trained and used for classification. The first step toward this goal is to develop a hyperspectral texture synthesis algorithm that efficiently combines both the spectral information and the spatial variability in the original image. We propose a method for texture region synthesis using neighboring pixel information, as well as interband correlation. The synthesis is done by a neighborhood search of a multiresolution pyramid constructed from the original sample that encodes the spectrum and spatial intensity gradients. Gaussian decomposition is used for the multiresolution decomposition. The pyramid is a representation of the hyperspectral image as a compact code. Starting from a seed image, the neighborhood is synthesized by sampling of the pyramid for the nearest neighbors. Results of the synthesis are presented using synthetic and remote sensing hyperspectral images. Discriminant analysis of texture properties of the synthesized texture are compared with the original texture and classification results are presented with different hyperspectral scenarios.

Image synthesis tools provide a means for generating hyperspectral image data without the expense of data collection. An important use of these tools is to provide data for the assessment of image exploitation algorithms. However, the detailed spectral/spatial structure of synthetic images is typically not sufficiently realistic to support the prediction of algorithm performance on real data. In this paper, we develop a new method for hyperspectral texture synthesis that accurately simulates the spectral/spatial structure of real hyperspectral image data. We demonstrate the utility of the new technique by presenting real and synthesized images and by analyzing spectral angle deviation from the mean curves that describe spectral properties. We also demonstrate that a signaturebased detection algorithm has similar performance against real and synthesized hyperspectral backgrounds.

Data from hyperspectral sensors calibrated to spectral radiance allow the prediction of spectra to be measured under different conditions through the use of physical models. Predicted spectra used in material detection algorithms, for example, are typically computed over a wide range of environmental conditions since environmental conditions are often unknown in advance of the measurements, and sometimes even during the data collections proper. A radiative transfer (RT) code such as MODTRAN® is commonly used to generate such predicted spectra. This requires the specification of a set of environmental conditions. We have developed an automated method for generating environmental conditions to obtain a uniform sampling of spectra in the VNIR/SWIR radiance domain. Our technique accounts for the nonlinear relationship between the environmental inputs and the RT spectral functions that are the outputs of the RT codes. An initial set of bounding conditions is used to define the bounding edges of the radiance domain. Based on the edges, the bounding surfaces are determined. Then, based on the bounding surfaces, the volume is populated. Variations of the sun angle, vertical-path-integrated water vapor density, and aerosol or cloud scatterer density and type are considered. Our approach to atmospheric modeling is demonstrated in signature comparisons using hyperspectral imagery acquired by the Hyperion VNIR/SWIR sensor.

Traditional approaches to hyperspectral target detection involve the application of detection algorithms to atmospherically compensated imagery. Rather than compensate the imagery, a more recent approach uses physical models to generate radiance signature spaces. The signature space is actually a representation of what the target might look like to the sensor as the reflectance propagates through the atmosphere. The model takes into account atmospherics, illumination conditions and target reflectivity. It is well known that the directional characteristics of reflectance vary considerably fromspecular to ideally diffuse (i.e., Lambertian). The current physical models assume the world is Lambertian. However, the reflectance properties are a function of wavelength, illumination angle, and viewing angle. The bidirectional reflectance distribution function (BRDF), which is actually a scattering function analogous to the angular scattering coefficient, describes the bidirectional reflectance values of all combinations of input-output angles and wavelength. This paper examines the impact of using the Lambertian assumption as it relates to physics bases material detection. The bidirectional reflectance studied is parameterized, based on laboratory measurements, using the Beard- Maxwell model. This parameterized reflectance is then coupled to the physics-based sensor-reaching radiance model to generate signature spaces. The signature spaces, along with hyperspectral imagery, are used in a target detection scheme where results are assessed through visual analysis.

The WorldView-2 sensor, to be launched mid-2009, will have 8 MSI bands - 4 standard MSI spectral channels and an additional 4 non-traditional bands. Hyperspectral data from the AURORA sensor (from the former Advanced Power Technologies, Inc. (APTI)) was used to simulate the spectral response of the WorldView-2 Sensor and DigitalGlobe's 4- band QuickBird system. A bandpass filter method was used to simulate the spectral response of the sensors. The resulting simulated images were analyzed to determine possible uses of the additional bands available with the WorldView-2 sensor. Particular attention is given to littoral (shallow water) applications. The overall classification accuracy for the simulated QuickBird scene was 89%, and 94% for the simulated WorldView-2 scene.

Identification of differing vegetation species has been a lauded ability of hyperspectral imagery and analysis but continues to be a challenging problem. Hyperspectral imagery has been used for years in applications such as vegetation analysis and delineation, terrain categorization, explosive mine detection, environmental impacts and effects, and agriculture and crop evaluation. Unlike applications which focus on detection of specific targets with constant spectral signatures, vegetation signatures continually vary across their growth cycle. In order to identify various vegetation species, either large collections of time-varying reference signatures are required, or ground truth/training data is needed. These are not always viable options and in many cases only in-scene data can be used. In this study we compare the performance of various spectral matching methods in separating vegetation at the species level. Parametric, non-parametric, derivative techniques, and other methods are compared. These methods are applied to a complex scene, the National Arboretum in Washington DC, which was imaged by an airborne hyperspectral sensor in August, 2008. This survey assesses performance of spectral matching methods for vegetation species delineation and makes recommendations for its application in hyperspectral data analysis.

In recent years hyper-spectral data has been used to retrieve information about vegetative canopies such as leaf area index and canopy water content. For the environmental scientist these two parameters are valuable, but there is potentially more information to be gained as high spatial resolution data becomes available. We developed an Amoeba (Nelder-Mead or Simplex) based program to invert a vegetative canopy radiosity model coupled with a leaf (PROSPECT5) reflectance model and modeled for the background reflectance (e.g. soil, water, leaf litter) to a measured reflectance spectrum. The PROSPECT5 leaf model has five parameters: leaf structure parameter Nstru, chlorophyll a+b concentration Cab, carotenoids content Car, equivalent water thickness Cw and dry matter content Cm. The canopy model has two parameters: total leaf area index (LAI) and number of layers. The background reflectance model is either a single reflectance spectrum from a spectral library() derived from a bare area pixel on an image or a linear mixture of soil spectra. We summarize the radiosity model of a layered canopy and give references to the leaf/needle models. The method is then tested on simulated and measured data. We investigate the uniqueness, limitations and accuracy of the retrieved parameters on canopy parameters (low, medium and high leaf area index) spectral resolution (32 to 211 band hyperspectral), sensor noise and initial conditions.

A single hyperspectral image can easily be hundreds of megabytes or even several gigabytes in size. For spectral processing, this is not an issue, as each pixel is processed indepedently. Additionally, many standard image processing algorithms can be readily adapted to process a few lines at a time (possibly with multiple passes over a file). Thus, clustering algorithms like k-means and dimension reduction methods such as PCA have also become staples of hyperspectral processing. More recently, however, new algorithms such as locally linear embedding (LLE) or non-local means have been shown to offer substantial performance increases, at least in theory. However, incremental processing is not feasible in many of these newer algorithms, and large amounts of processing power and memory are required to process large images. For this reason, mechanisms to efficiently reduce the size of large hyperspectral images while maintaining important information are desired. In this paper, we investigate a segmentation algorithm as a means to this end. We show that the amount of data that needs to be processed can be reduced by over an order of magnitude while maintaining the spectral purity of the data. This will be illustrated by showing classification accuracy before and after segmentation.

Clouds and cloud fields introduce important backscattering, obscuration, shadowing and radiative trapping effects in visible-NIR(near-infrared)-SWIR(short-wavelength infrared) hyperspectral imagery of the ground, especially in off-nadir (slant) viewing geometries where cloud thickness effects reduce the cloud-free line of sight (CFLOS). An investigation of these effects was conducted using monochromatic, multispectral and hyperspectral scene simulations performed with the Spectral Sciences, Inc. MCScene Monte Carlo code. Cloud fields were obtained from the Cloud Scene Simulation Model (CSSM) of Cianciolo and Raffensberger. The simulations took advantage of a data-fusion-based noise-removal method that enabled a dramatic reduction in computation time. Illumination levels at the sunlit ground showed enhancements of up to ~50% due to cloud scattering. Illumination in the cloud shadows was 20% of the full solar illumination or greater, with cloud optical depths of up to 10. Most of this illumination arises from solar scattering off the cloud tops and sides; however, a significant part can be ascribed to radiative trapping between the ground and the clouds, as represented by a local atmospheric spherical albedo. A simulation of a hyperspectral scene with cloud shadows was found to reproduce shadowing effects found in real data. Deeper shadowing is observed with increasing wavelength and in water-band regions, consistent with a previous analysis of cloud shadows in real imagery. The MCScene calculations also predict shadow enhancements of column water vapor retrievals from atmospheric correction/compensation codes, also in accord with field observations. CFLOS fractions were calculated as a function of off-nadir viewing angle and were found to be very accurately represented by a semi-empirical analytical function of both angle and cloud cover.

A set of 26 clear-sky, co-located, infrared data from NASA's space-based, downward looking Atmospheric Infrared Sounder (AIRS)¹ and its Clouds and the Earth's Radiant Energy System (CERES)² have been paired with the DOE's Southern Great Plains (SGP)³ ground-based, upward looking Atmospheric Emitted Radiance Interferometer (AERI)⁴. These data sets have then been simulated using the MODTRAN®5⁵ (MOD5) radiative transfer code with standard auxiliary 'truth' data as input. Of particular interest is the impact, if any, of the large Ground Sampling Distance (GSD) of AIRS and CERES (minimum radii of approximately 13 and 26 km, respectively) vs. the soda-straw up-looking mode of AERI. The smaller Moderate Resolution Imaging Spectroradiometer (MODIS)⁶ GSD (<1 km) provides estimated distributions of land-type and albedo within the larger footprints. The SGP's coincident vertical profile sondes and Aeronet⁷ retrievals, along with other satellite data [Ozone Monitoring Instrument (OMI)⁸] constrain the surface type, column ozone and aerosol optical depth that existed during the measurement events. Initial MOD5 calculations, using these ancillary data as input, have replicated the AERI and AIRS measurements to within an average difference of ~1% over their entire spectral range. Using MODIS visible albedo⁹ retrievals to extend the albedo into the long wave (LW) only slightly improved the statistical comparison between the CERES and MOD5 broadband LW radiances agreement, from 3% to ~1.5%, while increasing the variance. While these results are not sufficient for specific instrument inversion algorithms, they suggest some confidence in the generic use of MODTRAN®5 to help integrate and spectrally extend assorted data sets for sensitivity studies of Climate Change, where the estimated required sensitivity is <1%.

Irregular illumination across a hyperspectral image makes it difficult to detect targets in shadows, perform change detection, and segment the contents of the scene. To correct for the data in shadow, we first convert the data from Cartesian space to a hyperspherical coordinate system. Each N-dimensional spectral vector is converted to N-1 spectral angles and a magnitude representing the illumination value of the spectra. Similar materials will have similar angles and the differences in illumination will be described mostly by the magnitude. In the data analyzed, we found that the distribution of illumination values is well approximated by the sum of two- Gaussian distributions, one for shadow and one for non-shadow. The Levenberg-Marquardt algorithm is used to fit the empirical illumination distribution to the theoretical Gaussian sum. The LM algorithm is an iterative technique that locates the minimum of a multivariate function that is expressed as the sum of squares of non-linear real-valued functions. Once the shadow and non-shadow distributions have been modeled, we find the optimal point to be one standard deviation out on the shadow distribution, allowing for the selection of about 84% of the shadows. This point is then used as a threshold to decide if the pixel is shadow or not. Corrections are made to the shadow regions and a spectral matched filter is applied to the image to test target detection in shadow regions. Results show a signal-to-noise gain over other illumination suppression techniques.

The LSpec site in Nevada was established in late 2006 to enable unattended vicarious calibration of visible and nearinfrared remote sensing instruments. The site was selected because the large, relatively undisturbed, uniform, dry lakebed is suitable as a reflectance target. LSpec data include autonomous measurements of the playa taken at near-nadir viewing conditions every five minutes throughout the day. Thus, LSpec data can be retrieved for a calibration overflight time which will account for non-zenith solar angles. However, if the sensor under test views the playa from an off-nadir geometry, the LSpec reflectance data must be corrected for the playa's bi-directional reflectance factor. A ground-based experiment was performed on July 23, 2008 to collect reflective data from the playa over a series of solar zeniths, solar azimuths, and off-nadir collection angles. The collection conditions were managed to richly sample the space of zenith and azimuth differences between the sun and sensor. Data reduction shows that it is possible to experimentally derive a bi-directional reflectance distribution from field measurements suitable to use for calibration corrections. Additionally, analysis of the experimental data demonstrates consistent results with the data collected concurrently from the LSpec insitu sensors.

We describe a compressive projection algorithm and experimentally assess its performance when used with a Hyperspectral Image Projector (HIP). The HIP is being developed by NIST for system-level performance testing of hyperspectral and multispectral imagers. It projects a two-dimensional image into the unit under test (UUT), whereby each pixel can have an independently programmable arbitrary spectrum. To efficiently project a single frame of dynamic realistic hyperspectral imagery through the collimator into the UUT, a compression algorithm has been developed whereby the series of abundance images and corresponding endmember spectra that comprise the image cube of that frame are first computed using an automated endmember-finding algorithm such as the Sequential Maximum Angle Convex Cone (SMACC) endmember model. Then these endmember spectra are projected sequentially on the HIP spectral engine in sync with the projection of the abundance images on the HIP spatial engine, during the singleframe exposure time of the UUT. The integrated spatial image captured by the UUT is the endmember-weighted sum of the abundance images, which results in the formation of a datacube for that frame. Compressive projection enables a much smaller set of broadband spectra to be projected than monochromatic projection, and thus utilizes the inherent multiplex advantage of the HIP spectral engine. As a result, radiometric brightness and projection frame rate are enhanced. In this paper, we use a visible breadboard HIP to experimentally assess the compressive projection algorithm performance.

Improving the understanding of the optical scene components associated with coral reef imagery will advance the ability to map and monitor coral reefs using remote sensing. One tool that can aid in understanding the components in these scenes is the NIST Hyperspectral Image Projector (HIP). In this paper a hyperspectral scene is reformatted for projection using the HIP by first unmixing image spectra into endmembers. The abundance images representing each of the endmembers are then projected using the NIST HIP and collected by a hyperspectral imager. Since the scene is from a digital source, it can be used repeatedly without concern for changing measurement conditions. This work represents one of the first steps in developing scene projection capabilities that can be used for sensor characterization, algorithm testing or to have optical components changed independently in order to better understand the overall effects on the total observed scene.

A technique has been developed to quantitatively assess the relative radiometric accuracy of each detector at the focal plane of a pushbroom hyperspectral sensor. The approach requires only a calibrated radiance data cube as input. The results are visualized as a 2-D error image, analyzed statistically, and presented graphically as 2-D probability distribution function plots. We present results comparing the performance of the MaRS, COMPASS, HYCAS, and SpecTIR sensors. We also describe the algorithm basis, assumptions, and implementation details.

While able to measure the red, green, and blue channels, color imagers are not true spectral imagers capable of spectral measurements. In this paper we present a processing technique for estimating low resolution visible spectra from (RGB) imagery components. Such a methodology will find application in separating reflectivity components from source illumination as performed in Retinex or Linear Models. Numerical processing is discussed and formulated within non-regulated iterative restoration methodology. Numerical stability and convergence are considered relative to speed and accuracy as well as calibration data effects on overall results. Implementation is demonstrated together with several suggestions for real time processing. Imager calibration methodology is presented together with several results. Limitations of the method are discussed as well as directions for further investigations.

Hyperspectral images are customarily stored and transferred as radiance values. Image analysis may benefit from additional sensor-related information such as signal-dependent noise levels. This paper discusses representations of hyperspectral image data in forms which are intermediate between raw data and radiance data. The intermediate-form data can be processed directly, or they can be readily converted into radiance values and estimates of signal-dependent noise. The metadata needed for this data transformation constitutes an informative first-order description of the sensor, as an added benefit for the data user. One of the proposed data formats has already been adopted in commercial hyperspectral sensors. The proposed representations can be stored in a more compact data format than radiance values without loss of information, under reasonable assumptions about the sensor properties. In particular, it will be shown that a square-root transformation of the data leads to a representation which approaches the information-theoretic lower limit for storing light samples. The use of noise estimates derived from sensor physics is likely to be useful in hyperspectral image processing and image compression.

Image registration is the key component in image fusion where multiple imagery can be collectively exploited to provide much more useful information than that done with a single image. Image registration for images taken at different times, on different flight paths, by different sensors, is a hard problem to solve because of the change in image texture between the images. The changes between the images, mainly caused by change in look angles, platform positions and sensor characteristics, can be manifested into changes in scales, perspectives and intensities; are certain to break an image registration algorithm that is based on correlation of image data alone. Furthermore, the passive nature of an electro-optic (EO) sensor is known to be not very accurate in geolocation, especially with consideration of terrain variance. The availability of orthorectified reference data, such as Controlled Image Base (CIB) and Digital Terrain Elevation Data (DTED) can be used as a gateway to connect the two images to be registered by geolocation. Our approach to the problem is both image and feature based. It is assumed that Infrared (IR) bands be available in the night-time image which allows the separation of transient signals, for exmple man-made hot and bright spots, from the more permanent and natrual signals. The process developed is fully automated, and requires only a handful of system parameters. A successful execution of the process will produce a pair of registered images, with improved geolocation.

From an operational standpoint, road extraction remains largely a manual process despite the existence of several commercially available automation tools. The problem of automated feature extraction (AFE) in general is a challenging task as it involves the recognition, delineation, and attribution of image features. The efficacy of AFE algorithms in operational settings is difficult to measure due to the inherent subjectivity involved. Ultimately, the most meaningful measures of an automation method are its effect on productivity and actual utility. Several quantitative and qualitative factors go into these measures including spatial accuracy and timed comparisons of extraction, different user training levels, and human-computer interface issues. In this paper we investigate methodologies for evaluating automated road extraction in different operational modes. Interactive and batch extraction modes of automation are considered. The specific algorithms investigated are the GeoEye Interactive Road Tracker®(IRT) and the GeoEye Automated Road Tracker®(ART) respectively. Both are commercially available from GeoEye. Analysis metrics collected are derived from timed comparisons and spatial delineation accuracy. Spatial delineation accuracy is measured by comparing algorithm output against a manually derived image reference. The effect of object-level fusion of multiple imaging modalities is also considered. The goal is to gain insight into measuring an automation algorithm's utility on feature extraction productivity. Findings show sufficient evidence to demonstrate a potential gain in productivity when using an automation method when the situation is warranted. Fusion of feature layers from multiple images also demonstrates a potential for increased productivity compared to single or pair-wise combinations of feature layers.

This paper describes a new nonlinear joint fusion and anomaly detection technique for mine detection applications using two different types of sensor data (synthetic aperture radar (SAR) and Hyperspectral sensor (HS) data). A well-known anomaly detector so called the RX algorithm is first extended to perform fusion and detection simultaneously at the pixel level by appropriately concatenating the information from the two sensors. This approach is then extended to its nonlinear version. The nonlinear fusion-detection approach is based on the statistical kernel learning theory which implicitly exploits the higher order dependencies (nonlinear relationships) between the two sensor data through an appropriate kernel. Experimental results for detecting anomalies (mines) in hyperspectral imagery are presented for linear and nonlinear joint fusion and detection for a co-registered SAR and HS imagery. The result show that the nonlinear techniques outperform linear versions.

Fusion of Light Detection and Ranging (LiDAR) and Hyperspectral Imagery (HSI) products is useful for geological analysis, particularly for visualization of geomorphology and hydrology. In early 2007, coincident hyperspectral imagery and LiDAR were acquired over Cuprite, Nevada. The data were analyzed with ENVI and the ENVI LiDAR Toolkit. Results of the analysis of these data suggest, for some surfaces, a correlation between mineral content and surface roughness. However, the LiDAR resolution (~1 meter ground sampling distance) is likely too coarse to extract surface texture properties of clay minerals in some of the alluvial fans captured in the imagery. Though not demonstrated in this particular experiment (but a goal of the research), the relation between surface roughness and mineral composition may provide valuable information about the mechanical properties of the surface cover-in addition to generating another variable useful for material characterization, image classification, and scene segmentation. Future mission planning should include consideration of determining optimal ground sampling to be used by LiDAR and HSI systems. The fusion of LiDAR elevation data and multi- and hyperspectral classification results is, in and of itself, a valuable tool for imagery analysis and should be explored further.

Traditionally, only a narrow band of the Near-Infrared (NIR) spectrum (700-900nm) is utilized for iris recognition since this alleviates any physical discomfort from illumination, reduces specular reflections and increases the amount of texture captured for some iris colors. However, previous research has shown that matching performance is not invariant to iris color and can be improved by imaging outside of the NIR spectrum. Building on this research, we demonstrate that iris texture increases with the frequency of the illumination for lighter colored sections of the iris and decreases for darker sections. Using registered visible light and NIR iris images captured using a single-lens multispectral camera, we illustrate how physiological properties of the iris (e.g., the amount and distribution of melanin) impact the transmission, absorbance, and reflectance of different portions of the electromagnetic spectrum and consequently affect the quality of the imaged iris texture. We introduce a novel iris code, Multispectral Enhanced irisCode (MEC), which uses pixel-level fusion algorithms to exploit texture variations elicited by illuminating the iris at different frequencies, to improve iris matcher performance and reduce Failure-To-Enroll (FTE) rates. Finally, we present a model for approximating an NIR iris image using features derived from the color and structure of a visible light iris image. The simulated NIR images generated by this model are designed to improve the interoperability between legacy NIR iris images and those acquired under visible light by enabling cross wavelength matching of NIR and visible light iris images.

There has been significant research on pan-sharpening multispectral imagery with a high resolution image, but there has been little work extending the procedure to high dimensional hyperspectral imagery. We present a wavelet-based variational method for fusing a high resolution image and a hyperspectral image with an arbitrary number of bands. To ensure that the fused image can be used for tasks such as classification and detection, we explicitly enforce spectral coherence in the fusion process. This procedure produces images with both high spatial and spectral quality. We demonstrate this procedure on several AVIRIS and HYDICE images.

The methods used to evaluate automation tools are a critical part of the development process. In general, the most meaningful measure of an automation method from an operational standpoint is its effect on productivity. Both timed comparison between manual and automation based-extraction, as well as measures of spatial accuracy are needed. In this paper, we introduce the notion of correspondence to evaluate spatial accuracy of an automated update method. Over time, existing vector data becomes outdated because 1) land cover changes occur, or 2) more accurate overhead images are acquired, and/or vector data resolution requirements by the user may increase. Therefore, an automated vector data updating process has the potential to significantly increase productivity, particularly as existing worldwide vector database holdings increase in size, and become outdated more quickly. In this paper we apply the proposed evaluation methodology specifically to the process of automated updating of existing road centerline vectors. The operational scenario assumes that the accuracy of the existing vector data is in effect outdated with respect to newly acquired imagery. Whether the particular approach used is referred to as 1) vector-to-image registration, or 2) vector data updating-based automated feature extraction (AFE), it is open to interpretation of the application and bias of the developer or user. The objective of this paper is to present a quantitative and meaningful evaluation methodology of spatial accuracy for automated vector data updating methods.

In this paper, we propose spatio-spectral processing techniques for the detection of dust storms and automatically finding its transport direction in 5-band NOAA-AVHRR imagery. Previous methods that use simple band math analysis have produced promising results but have drawbacks in producing consistent results when low signal to noise ratio (SNR) images are used. Moreover, in seeking to automate the dust storm detection, the presence of clouds in the vicinity of the dust storm creates a challenge in being able to distinguish these two types of image texture. This paper not only addresses the detection of the dust storm in the imagery, it also attempts to find the transport direction and the location of the sources of the dust storm. We propose a spatio-spectral processing approach with two components: visualization and automation. Both approaches are based on digital image processing techniques including directional analysis and filtering. The visualization technique is intended to enhance the image in order to locate the dust sources. The automation technique is proposed to detect the transport direction of the dust storm. These techniques can be used in a system to provide timely warnings of dust storms or hazard assessments for transportation, aviation, environmental safety, and public health.

The current sea surface temperature (SST) algorithms were derived empirically using a large amount of in-situ observation data. The algorithms derived had no guarantee to be used for the different regions and time. Large amount of in-situ data was required for the algorithm regression analysis. The new algorithm did not require a large amount of in-situ data. The algorithm requires additional input of transmittance and emissivity values. The transmittance depends on atmospheric profiles. The standard profile was used. The input sensor zenith angle and water vapor contents were changed within a certain range. The data were then simulated by MODTRAN to obtain the transmittance. The derived equation of sea surface emissivity as a function of sensor zenith was used. Brightness temperature, sea surface emissivity and transmittance values were use to calculate the sea surface temperature of each cloud free water pixel using an image processing software. The results show that the new algorithm produce a comparable the R²=0.6569 and RMSE= 1.24 K. The new algorithm did not require the large amount of in-situ SST data, but still can give the SST data in moderately high accuracy.

Many researches showed that pollution particles are associated with adverse health effects. These particles are 10 micrometers in diameter or smaller (PM₁₀). Now the challenge is to develop a better technique for monitoring the concentration of these particles. In this study, we intend to use a Single Lens Reflex (SLR) camera to determine particle concentration in the air using our own developed algorithm. This algorithm was developed based on the regression analysis of relationship between the measured reflectance and the reflected components from a surface material and the atmosphere. This algorithm is used to convert multispectral image pixel values acquired by this camera into quantitative values of the concentrations of PM₁₀. These computed PM₁₀ concentrations were compared to other standard values measured by a DustTrak^TM meter. The validated result of the newly develop algorithm produced a high degree of accuracy as indicated by high correlation coefficient (R²) and low root-mean-square-error (RMS).

The effective dimensionality (ED) of hyperspectral images is often viewed as the dimensionality of an affine subspace defined by linear combinations of spectra of materials present in the image. That affine subspace is expected to give an acceptable approximation to all pixels. At this point, there is no precise definition of ED. In an effort to assess ED, a notion of virtual dimensionality (VD) has been developed, and it is being used in many papers including those published in TGARS. The ever- spreading use of VD warrants its thorough investigation. In this paper, we investigate properties of VD, and we show that VD largely depends on the average value of all spectra rather than on ED. We show specific examples when VD would give entirely misleading results. We also explain fallacies associated with justifications for VD.

The inherent dimension of hyperspectral data may be a useful metric for discriminating between the presence of manmade and natural materials in a scene without reliance on spectral signatures take from libraries. Previously, a simple geometric method for approximating the inherent dimension was introduced along with results from application to single material clusters. This method uses an estimate of the slope from a graph based on the point density estimation in the spectral space. Other information can be gathered from the plot which may aid in the discrimination between manmade and natural materials. In order to use these measures to differentiate between the two material types, the effect of the inclusion of manmade pixels on the phenomenology of the background distribution must be evaluated. Here, a procedure for injecting manmade pixels into a natural region of a scene is discussed. The results of dimension estimation on natural scenes with varying amounts of manmade pixels injected are presented here, indicating that these metrics can be sensitive to the presence of manmade phenomenology in an image.

Pixel Purity Index (PPI) has been widely used in endmember extraction. While it is available in ENVI software there are several interesting issues arising in its implementation. This paper re-invents the wheel by re-visiting the design rationale of the PPI and re-designing algorithms to implement PPI. More specifically, it develops the so-called Causal PPI (CPPI) which implements the PPI in a causal manner in the sense that the information used for data processing is only up to the data sample currently being visited. If the time required for computer processing is negligible, the CPPI actually becomes a real time PPI. The proposed CPPI can be implemented automatically and resolves the main issue of requiring human intervention to determine parameters.

Because hyperspectral imagery is generally low resolution, it is possible for one pixel in the image to contain several materials. The process of determining the abundance of representative materials in a single pixel is called spectral unmixing. We discuss the L1 unmixing model and fast computational approaches based on Bregman iteration. We then use the unmixing information and Total Variation (TV) minimization to produce a higher resolution hyperspectral image in which each pixel is driven towards a "pure" material. This method produces images with higher visual quality and can be used to indicate the subpixel location of features.