The design, operation, and performance of the fourth generation of Science and Technology International's Advanced Airborne Hyperspectral Imaging Sensors (AAHIS) are described. These imaging spectrometers have a variable bandwidth ranging from 390-840 nm. A three-axis image stabilization provides spatially and spectrally coherent imagery by damping most of the airborne platform's random motion. A wide 40-degree field of view coupled with sub-pixel detection allows for a large area coverage rate. A software controlled variable aperture, spectral shaping filters, and high quantum efficiency, back-illuminated CCD's contribute to the excellent sensitivity of the sensors. AAHIS sensors have been operated on a variety of fixed and rotary wing platforms, achieving ground-sampling distances ranging from 6.5 cm to 2 m. While these sensors have been primarily designed for use over littoral zones, they are able to operate over both land and water. AAHIS has been used for detecting and locating submarines, mines, tanks, divers, camouflage and disturbed earth. Civilian applications include search and rescue on land and at sea, agricultural analysis, environmental time-series, coral reef assessment, effluent plume detection, coastal mapping, damage assessment, and seasonal whale population monitoring

The Dual Etalon Cross Tilt Order Sorted Spectrometer (DECTOSS) uses relatively inexpensive off the shelf components in a small and simple package to provide ultra high spectral resolution over a limited spectral range. For example, the modest first try laboratory test setup DECTOSS we describe in this presentation achieves resolving power ~ 10⁵ on a spectral range of about 1 nm centered near 760 nm. This ultra high spectral resolution facilitates some important atmospheric remote sensing applications including profiling cirrus and/or aerosol above bright reflective surfaces in the O₂ A-band and the column measurements of CO and CO₂ utilizing solar reflectance spectra. We show details of the how the use of ultra high spectral resolution in the O₂ A-band improves the profiling of cirrus and aerosol. The DECTOSS utilizes a Narrow Band Spectral Filter (NBSF), a Low Resolution Etalon (LRE) and a High Resolution Etalon (HRE). Light passing through these elements is focused on to a 2 Dimensional Array Detector (2DAD). Off the shelf, solid etalons with airgap or solid spacer gap are used in this application. In its simplest application this setup utilizes a spatially uniform extended source so that spatial and spectral structure are not confused. In this presentation we'll show 2D spectral data obtained in a desktop test configuration, and in the first try laboratory test setup. These were obtained by illuminating a Lambertian screen with (1) monochromatic light, and (2) with atmospheric absorption spectra in the oxygen (O₂) A-band. Extracting the 1D spectra from these data is a work in progress and we show preliminary results compared with (1) solar absorption data obtained with a large Echelle grating spectrometer, and (2) theoretical spectra. We point out areas for improvement in our laboratory test setup, and general improvements in spectral range and sensitivity that are planned for our next generation field test setup.

In order to assist Rescue and Recovery personnel after 11 September 2001, Night Vision and Electronic Sensors Directorate was requested to collect a variety of airborne electro-optic data of the WTC site. The immediate objective was to provide FDNY with geo-rectified high-resolution and solar reflective hyperspectral data to help map the debris-field. Later data collections included calibrated MWIR data. This thermal data provided accurate temperature profiles, which could be warped to the high-resolution data. This paper will describe the assets and software used to help provide the FDNY data products, which were incorporated into their GIS database.

Stand-off identification in the field using thermal infrared spectrometers (hyperspectral) is a maturing technique for gases and aerosols. However, capabilities to identify solid-phase materials on the surface lag substantially, particularly for identification in the field without benefit of ground truth (e.g. for "denied areas"). Spectral signatures of solid phase materials vary in complex and non-intuitive ways, including non-linear variations with surface texture, particle size, and intimate mixing. Also, in contrast to airborne or satellite measurements, reflected downwelling radiance strongly affects the signature measured by field spectrometers. These complex issues can confound interpretations or cause a misidentification in the field. Problems that remain particularly obstinate are (1) low ambiguity identification when there is no accompanying ground truth (e.g. measurements of denied areas, or Mars surface by the 2003 Mars lander spectrometer); (2) real- or near real-time identification, especially when a low ambiguity answer is critical; (3) identification of intimate mixtures (e.g. two fine powders mixed together) and targets composed of very small particles (e.g. aerosol fallout dust, some tailings); and (4) identification of non-diffuse targets (e.g. smooth coatings such as paint and desert varnish), particularly when measured at a high emission angle. In most studies that focus on gas phase targets or specific manmade targets, the solid phase background signatures are called "clutter" and are thrown out. Here we discuss our field spectrometer images measured of test targets that were selected to include a range of particle sizes, diffuse, non-diffuse, high, and low reflectance materials. This study was designed to identify and improve understanding of the issues that complicate stand-off identification in the field, with a focus on developing identification capabilities to proceed without benefit of ground truth. This information allows both improved measurement protocols and identification quality. The Aerospace Corporation has a mature program for field hyperspectral measurements using van-mounted thermal-infrared spectrometers that raster-scan images. Aerospace is a non-profit Federally Funded Research and Development Center (FFRDC), managed by the Department of Defense. The precisely controlled viewing geometery, imaging capabilities, and sensitivity of the spectrometers used are critical to identifying and studying issues that can confound interpretations or cause a misidentification. We have released a portion of this data set publicly, and encourage researchers interested in the data set to contact us. More information is at www.lpi.usra.edu/science/kirkland.

Data from moderate resolution ocean color sensors, such as SeaWiFS and MODIS have greatly enhanced our understanding of the open ocean and shelf waters. However, the spatial and spectral complexity of the near coastal ocean, require higher resolution systems for the littoral zone. Recent experiments with aircraft imaging spectrometers have demonstrated their potential to be powerful tools for the characterization of the coastal ocan. Using the continuous spectral signature it is possible to measure shallow water bathymetry and bottom characteristics, and to gain insight into the distribution of phytoplankton and ot her optically active constituents. To demonstrate this I present recent results using AVIRIS and PHILLS data from the coastal environment. To obtain large area, repeated coverage of the coastal ocean two spaceborne hyperspectral imagers are planned. The Navy has joined in a partnership with industry to build and fly the Naval EarthMap Observer (NEMO). The NEMO spacecraft has the Coastal Ocean Imaging Spectrometer (COIS) a hyperspectral imager with adequate spectral and spatial resolution and a high signal-to-noise ratio to provide long-term monitoring and real-time characterization of the coastal environment. Additionally, the Integrated Program Office is considering a Coastal Ocean Imager (COI) as part of an Ocean Observer Spacecraft. COI is a hyperspectral imager in the visible with a two band thermal-IR imager for sea surface temperature. COI would provide 100 m resolution imagery over a 150 km wide swath of the coastal ocean.

The MTI Experiments Committee accepted a Tasking Proposal in 2001 for the Multispectral Thermal Imager (MTI) to make observations of Yellowstone National Park. Initial goals for this work were twofold. The first goal was to investigate Yellowstone National Park as a vicarious calibration site. The second goal was to gain experience in working with relatively high spatial resolution multispectral data whose near simultaneously imaged bands span the spectrum from the visible through the mid wave (roughly 3 μm to 5 μm) and long wave (roughly 8 μm to 12 μm) infrared atmospheric windows. Various sites in the Park were observed at various times, starting in June, 2001 and on-going through January of 2002. This paper reports on the progress of this work, summarizes the observations, provides a snapshot of preliminary forays into the analysis, and suggests directions for further exploitation of the data. Plans for the acquisition of more imagery in the spring and early summer of 2002 are in place.

Quantitative and robust metrics are required to objectively compare the performance of algorithms within a general functional class. This is especially true for classification algorithms that cluster and label data using spectral features. This is because spectral algorithms are usually based on a finite set of assumptions about the radiative transfer phenomenology. Thus, a suite of algorithms is needed to achieve a generalized and robust processing chain that performs well under all operational scenarios of interest. An adaptive processing chain that automatically selects the optimal combination of algorithms to generate a product of prescribed quality provides a framework for operational applications. To this end, we have developed Measures of Effectiveness (MOE's) and Figures of Merit (FOM's) that can quantitatively and objectively select the appropriate algorithm automatically. The FOM's are a weighted sum of MOE's, which are performance metrics such as the tightness and dissimilarity of clusters. We have also defined scene and sensor parameters that quantify a subset of factors that affect algorithm performance. Functional relationships between FOM's and MOE's and between the FOM's/MOE's and the scene/sensor parameters were also established. These functional relationships allow users to predict the expected classification product quality given a specific operational scenario based on a performance model that also automates the processing chain. Initial results of an application of this approach to hyperspectral data indicate that FOM's can be predicted with high accuracy with choices made correctly as high as 89% of the time depending on the FOM definitions. The results were obtained over a wide range of operational scenarios.

The performance of hyperspectral exploitation algorithms depends on the quality of the hyperspectral data processed. Some algorithms may perform better or be better suited to certain types or quality of data than other algorithms. To improve the hyperspectral exploitation production process, the dependencies of different types of algorithms to the quality of the hyperspectral data needs to be understood. A framework for predicting algorithm performance based on data parameters and metrics is presented. Figures of merit are defined for classes of algorithms which can be used to select between different algorithms to process a particular dataset. A training set of data is used to determine the dependence of each algorithm being tested in the class. Multiple regression is then applied to determine the dependence of the algorithm results on the different parameters and metrics. The performance on datasets not in the training set can then be predicted using the results of the regression analysis. Analysis of the regression results provides insight into the dependence of different types of algorithms on parameters of the data. In addition, the results provide insight into the data quality needed to provide quality exploitation products that meet minimum requirements. The technique is presented along with preliminary results for some basic algorithms in the atmospheric compensation and material identification categories.

In the framework of the APEX (Airborne Prism Experiment) pushbroom imaging spectrometer, a complete processing and archiving facility (PAF) is developed. The PAF not only includes imaging spectrometer data processing up to physical units, but also geometric and atmospheric correction for each scene, as well as calibration data input. The PAF software includes an Internet based web-server and provides interfaces to data users as well as instrument operators and programmers. The software design, the tools and its life cycle is discussed as well. Further we will discuss particular instrument requirements (resampling, bad pixel treatment, etc.) in view of the operation of the PAF as well as their consequences on the product quality. Finally we will discuss a combined approach for geometric and atmospheric correction including BRDF (or view angle) related effects.

The atmospheric correction of thermal infrared (TIR) imagery involves the combined tasks of separation of atmospheric transmittance, downwelling flux and upwelling radiance from the surface material spectral emissivity and temperature. The problem is ill posed and is thus hampered by spectral ambiguity among several possible feasible combinations of atmospheric temperature, constituent profiles, and surface material emissivities and temperatures. For many materials, their reflectance spectra in the Vis-SWIR provide a means of identification or at least classification into generic material types, vegetation, soil, etc. If Vis-SWIR data can be registered to TIR data or collected simultaneously as in sensors like the MASTER sensor, then the additional information on material type can be utilized to help lower the ambiguities in the TIR data. If the Vis-SWIR and TIR are collected simultaneously the water column amounts obtained form the atmospheric correction of the Vis-SWIR can also be utilized in reducing the ambiguity in the atmospheric quantities. The TIR atmospheric correction involves expansions in atmospheric and material emissivity basis sets. The method can be applied to hyperspectral and ultraspectral data, however it is particularly useful for multispectral TIR, where spectral smoothness techniques cannot be readily applied. The algorithm is described, and the approach applied to a MASTER sensor data set.

Model-based atmospheric correction of multi-spectral and hyperspectral imagery (MSI/HSI) typically involves searching through a look-up table (LUT) of potential atmospheric representations for a best fit, based on some fit criterion. These representations are generated using a radiation transport model such as MODTRAN. The parameter space covered by the LUT is defined to cover the likely atmospheric conditions encountered by the sensor that affect observed radiance over the spectral region covered by the sensor. For instance, aerosols play an important role in the visible through SWIR (450-2500 nm) but a minor role in the thermal IR, where water column content and atmospheric temperature are critical. We investigate the sampling and representation of the atmospheric parameter space in the thermal IR as it effects retrieval of the atmosphere. Using the SMACC convex projection technique we evaluate selection of significant basis members from a broadly-based LUT. We apply SMACC selected endmembers to solve for an arbitrary atmosphere.

The spectral signature of a material is the curve of power density vs. wavelength (λ) obtained from measurements of reflected light. It is used, among other things, for the identification of targets in remotely acquired images. Sometimes, however, unpredictable distortions may prevent this. In only a few cases have such distortions been explained. We propose some reasonable arguments that in a significant number of circumstances, atmospheric turbulence may contribute to such spectral signature distortion. We propose, based on this model, what appears to be one method that could combat such distortion.

A method for the determination of aerosol optical properties from imaging spectrometer data on a local scale is investigated, making use of the continuous spectral coverage, high spatial resolution, and the well-calibrated radiometry of such data. The method (correlated spectral unmixing) is based on the decomposition of the sensor signal in the short-wave infrared using spectrum database ground spectra, the reconstruction of image ground spectra in the visible, and forward modelling with a radiative transfer code. The sensitivity of the imaging spectrometer signal to different atmospheric condititions is explored, as well as the correlation of spectral reflectances in the visible and short-wave infrared for a variety of surfaces. The potential of the presented method is demonstrated for a scene from the airborne visible and infrared imaging spectrometer AVIRIS over rugged heterogeneous coastal terrain in California, and comparisons to multispectral methods are made.

The Compact High Resolution Imaging Spectrometer CHRIS sensor was launched on board of PROBA (PROject for on Board Autonomy) the 22nd of October 2001. CHRIS will acquire sets of images over the Belgian coastal zone near Oostende. Within this context CASI (Compact Airborne Spectrographic Imager) images was used as prototype of CHRIS data. This is to assess the performance of an atmospheric correction algorithm for hyperspectral ocean-color sensors (i.e. CHRIS). This approach couples the atmospheric attenuation processes with the underlying physics of water inherent optical properties. The algorithm employed the 6S code (second simulation of the satellite signal in the solar spectrum) to simulate the atmospheric and surface reflectance and the gaseous transmittances. Relationships between the water leaving reflectances at 860 870 and 880 nm are proposed. This is to estimate the water signals at these near infrared (NIR) bands, hence the aerosol reflectances. These negligible water-leaving signals were found to be very important for a reliable atmospheric-correction-algorithm over turbid waters. The choice of this NIR part of the spectrum was to satisfy a certain condition related to the corresponding water absorption coefficient. A look-up-table of total gaseous transmittance has been generated for 42 values of column water vapour. This table was used with a two-band ratio technique to estimate the contribution of water vapour to the total gaseous transmittance. The aerosol optical thicknesses were estimated by fitting calculated atmospheric reflectances at the water vapour window (860-880 nm) to 20 candidates of maritime aerosol models. The performance of the atmospheric correction is being investigated with other sensors (DASI and ROSIS) and in situ measurements.

The Digital Imaging and Remote Sensing Synthetic Image Generation (DIRSIG) model uses a quantitative first principles approach to generate synthetic hyperspectral imagery. This paper presents the methods used to add modeling of polarization phenomenology. The radiative transfer equations were modified to use Stokes vectors for the radiance values and Mueller matrices for the energy-matter interactions. The use of Stokes vectors enables a full polarimetric characterization of the illumination and sensor reaching radiances. The bi-directional reflectance distribution function (BRDF) module was rewritten and modularized to accommodate a variety of polarized and unpolarized BRDF models. Two new BRDF models based on Torrance-Sparrow and Beard-Maxwell were added to provide polarized BRDF estimations. The sensor polarization characteristics are modeled using Mueller matrix transformations on a per pixel basis. All polarized radiative transfer calculations are performed spectrally to preserve the hyperspectral capabilities of DIRSIG. Integration over sensor bandpasses is handled by the sensor module.

The Digital Imaging and Remote Sensing Image Generation (DIRSIG) model is a synthetic imagery generation model developed at the Center for Imaging Science (CIS) at the Rochester Institute of Technology (RIT). It is a quantitative first principle based model that calculates the sensor reaching radiance from the visible through to the long wave infrared on a spectral basis. DIRSIG generates a very accurate representation of what a sensor would see by modeling all the processes involved in the imaging chain. Currently, DIRSIG only models passive sources such as the sun and blackbody radiation due to the temperature of an object. Active systems have the benefit of the user being able to control the illumination source and tailor it for specific applications. Remote sensing Laser Detection and Ranging (LADAR) systems that utilize a laser as the active source have been in existence for over 30 years. Recent advances in tunable lasers and infrared detectors have allowed much more sophisticated and accurate work to be done, but a comprehensive spectral LADAR model has yet to be developed. In order to provide a tool to assist in LADAR development, this research incorporates a first principle based elastic LADAR model into DIRSIG. It calculates the irradiance onto the focal plane on a spectral basis for both the atmospheric and topographic return, based on the system characteristics and the assumed atmosphere. The geometrical form factor, a measure of the overlap between the sensor and receiver field-of-view, is carefully accounted for in both the monostatic and bistatic cases. The model includes the effect of multiple bounces from topographical targets. Currently, only direct detection systems will be modeled. Several sources of noise are extensively modeled, such as speckle from rough surfaces. Additionally, atmospheric turbulence effects including scintillation, beam effects, and image effects are accounted for. To allow for future growth, the model and coding are modular and anticipate the inclusion of advanced sensor modules and inelastic scattering.

A spectral signatures database (SSDB) has been developed to support the use of spectral sensing for remote sensing applications. A broad range of data from multiple regions of the electromagnetic spectrum is supported, including ultraviolet, visible, near-infrared, thermal infrared, and fluorescence. Future plans include support of hemispherical reflectance data. A priority in the database development was schema flexibility. Data can be archived with minimum or detailed sets of attributes. Pre-defined, community-standard attributes are included, but custom attributes may be added to meet the specific project or data requirements. The database incorporates all sources of signature data, including laboratory equipment, field radiometers, and imaging spectrometers. The database can also incorporate and reference other metadata such as project history, personnel information, spatial information, temporal information, equipment parameters, reference documents, photos, and imagery. The database was developed with entirely off-the-shelf products and exists in both stand-alone and web-based versions. Searching and filtering utilities have been included to allow a user to quickly locate and extract signatures of interest. Currently available application tools include two- and three-dimensional visualization, signature statistics, and surface matching and comparison. Data exporting is also available, which includes the creation of commercial image processing spectral libraries.

Hyperspectral imagery (HSI) of the ocean-land interface, known as the littoral zone (LZ) can provide a valuable source of information for identification of underwater objects and materials, determination of water depth, and retrieval of water composition. The first step in the analysis is removal of atmospheric effects, resulting in surface reflectance spectra. The atmospheric removal is accomplished with a new version of the MODTRAN-based FLAASH correction code. When available, infrared wavelengths are used to retrieve water vapor and aerosol parameters for the correction and to remove foam and glitter components to yield water-leaving reflectance. A visible-only spectral unmixing technique for foam and glitter removal has also been developed. Bathymetry algorithms that use the 500-700 nm region were developed based on Monte Carlo-simulated "ground truth" spectra. The end-to-end data analysis process has been demonstrated with publicly available AVIRIS imagery.

The shallow water and surf zone (SZ) regions are one of the more difficult environments currently being addressed in littoral mine counter measure (MCM) strategies, yet they are also critical regions for MCM with respect to military breaching tactics. The difficulties in optical remote sensing of the SZ lie mostly in the problem of clutter, which includes transient wave glint, foam patches, turbidity, and detritus. The problem is compounded by the refractive distortion of the small targets (mines and barriers) in these shallow waters. We have adopted several strategies for dealing with clutter rejection in the SZ. The first is a strictly statistical approach to clutter rejection, which is computationally efficient and mathematically simple. The second of these leverages hyperspectral algorithms used for the detection of submerged targets in deep water, wherein the glint is subtracted from the scene prior to image segmentation and anomaly detection. The second method, while more mathematically mature, does not appreciably increase the computation time and provides startlingly better results.

There is increasing interest in using wide-area standoff airborne hyperspectral sensors to detect potential targets at large oblique viewing angles. Under such conditions, the intervening atmosphere between the targets and the imager can attenuate and alter the detected signal. To help compensate for the reduced signal for long range viewing, recent efforts have focused on using hyperspectral sensors to collect imagery derived from short wave radiation SWIR (1-2.5 μm) rather than the more standard visible-near infrared radiation Vis-NIR (0.4-1.0 μm). However, unlike imagery collected using Vis-NIR, there is currently a relative dearth of analytical and classification algorithms that only use SWIR. To enhance the ability to detect spectral features confined to the SWIR regime, this study has examined extracting vegetation features in the SWIR. Visible-near infrared hyperspectral imagery has successfully extracted vegetation within a scene through computation of the normalized difference vegetation index (NDVI). The Visible NDVI computes a normalized difference of two bands corresponding to the chlorophyll absorption (0.67 μm) and IR edge (0.80 μm). This work extended and examined schemes for extracting vegetation within SWIR imagery. Specifically, this study examined the HYDICE data collect (0.4-2.5 μm) and the visible NDVI was used as the standard for determining the vegetation within the scene. A SWIR derived NDVI was generated using pairs of SWIR bands (1.08, 1.46 μm), (1.08, 1.57 μm), (1.08, 1.66 μm), and (1,08, 2.18 μm). All SWIR paired bands exhibited large (> 0.92) correlation coefficients with the Vis-NIR NDVI. Vis-NIR NDVI preferentially detects the greenest vegetation but the SWIR NDVI tends to favor vegetation residing in shadows. Water has large SWIR NDVI but has low reflectance throughout the SWIR. By setting a threshold, water can be eliminated from consideration and only vegetation is detected. In addition, minimizing the mean squared error between the visible and SWIR imagery can generate a suitable linear combination of all suitable bands involving SWIR wavelengths that exclude the atmospheric absorption. Using the high number of SWIR bands approach yields even larger correlation (>0.99) with visible NDVI. However, the specific coefficients used in the linear combination approach, varies from scene to scene. Therefore, using a fixed set of coefficients yields small correlation coefficient with visible NDVI.

In this paper, we describe the use of linear unmixing algorithms to spatially and spectrally separate fluorescence emission signals from fluorophores having highly overlapping emission spectra. Hyperspectral image data for mixtures of Nile Blue and HIDC Iodide in a methanol/polymer matrix were obtained using the Information-efficient Spectral Imaging sensor (ISIS) operated in its Hadamard Transform mode. The data were analyzed with a combination of Principal Components Analysis (PCA), orthogonal rotation, and equality and non-negativity constrained least squares methods. The analysis provided estimates of the pure-component fluorescence emission spectra and the spatial distributions of the fluorophores. In addition, spatially varying interferences from the background and laser excitation were identified and separated. A major finding resulting from this work is that the pure-component spectral estimates are very insensitive to the initial estimates supplied to the alternating least squares procedures. In fact, random number starting points reliably gave solutions that were effectively equivalent to those obtained when measured pure-component spectra were used as the initial estimates. While our proximate application is evaluating the possibility of multivariate quantitation of DNA microarrays, the results of this study should be generally applicable to hyperspectral imagery typical of remote sensing spectrometers.

A new matched filter-based algorithm has been developed for detecting and approximately correcting for shadows or other illumination variations in spectral imagery. Initial evaluations have been conducted with a handful of data cubes, including AVIRIS data. The de-shadowed images have a generally realistic appearance and reveal a wealth of previously hidden surface details.

The ability to detect man-made materials with known spectral signatures in hyperspectral images has many important applications. In this paper, we present and compare two, linear mixing model-based, algorithms for the detection of low probability of occurrence targets with known spectral signatures. One involves the estimation of the target abundance in each pixel, a form of spectral unmixing, and the other binary hypothesis testing using the generalized likelihood ratio test (GLRT). In an effort to improve detection, we investigate the effects of placing the Sum-to-One (STO) constraint on the abundances of the materials present in each pixel. Both theoretical and experimental results will be presented such that the benefits of the STO constraint can be directly compared. We shall demonstrate that, in theory, the enforcement of STO constraint improves detection performance. For abundance estimation based detectors, the constraint reduces the variance of the estimate. For GLRT detectors, the STO constraint increases the signal to interference plus noise ratio (SINR). Unfortunately, we do not see the same improvements with real data. In fact, enforcing the constraint leads to a performance degradation, in most cases we have investigated. It turns out that the abundance estimation based detector moves the full pixels, subpixels, and background pixels closer to each other; which makes reliable detection more difficult. With regard to the constrained GLRT detector, there is an introduction of bias to the background pixels, which naturally results to a deterioration in detection performance.

Automating the detection and identification of significant threats using multispectral (MS) imagery is a critical issue in remote sensing. Unlike previous multispectral target recognition approaches, we utilize a three-stage process that not only takes into account the spectral content, but also the spatial information. The first stage applies a matched filter to the calibrated MS data. Here, the matched filter is tuned to the spectral components of a given target and produces an image intensity map of where the best matches occur. The second stage represents a novel detection algorithm, known as the focus of attention (FOA) stage. The FOA performs an initial screening of the data based on intensity and size checks on the matched filter output. Next, using the target's pure components, the third stage performs constrained linear unmixing on MS pixels within the FOA detected regions. Knowledge sources derived from this process are combined using a sequential probability ratio test (SPRT). The SPRT can fuse contaminated, uncertain and disparate information from multiple sources. We demonstrate our approach on identifying a specific target using actual data collected in ideal conditions and also use approximately 35 square kilometers of urban clutter as false alarm data.

Two techniques for detecting point targets in hyperspectral imagery are described. The first technique is based on the principal component analysis of hyperspectral data. We combine the information of the first two principal component analysis images; the result is a single image display "summary" of the data cube. The summary frame is used to define image segments. The statistics, means and variances, of each segment for the principal component images is calculated and a covariance matrix is constructed. The local pixel statistics and the segment statistics are then used to evaluate the extent to which each pixel differs from its surroundings. Point target pixels will have abnormally high values. The second technique operates on each band of the hypercube. A local anti-median of each pixel is taken and is weighted by the standard deviation of the local neighborhood. The results of each band are then combined. Results will be shown for visible, SWIR, and MWIR hyperspectral imagery.

ASSET is based on physical first principles and was developed using synthetic data. The method treats each pixel independently, assumes homogeneous, isothermal pixels and requires the following inputs: 1) Hyperspectral LWIR radiance imagery, 2) Atmospheric parameters (downwelling irradiance, upwelling radiance, and transmissivity), and 3) A library of material emissivities. For each pixel, the method determines the most appropriate material from the emissivity library. The method computes the pixel temperature assuming pure pixels. Then, the pixel temperature is used to determine the emissivity. Note that the computed emissivity may differ from that of the selected library material due to a variety of factors such as noise, mixed pixels, natural spectral variability, and inadequate atmospheric compensation. The synthetic data used to develop ASSET were constructed by computing the thermally emitted radiances of a set of materials with specificed emissivities at a range of temperatures. A given set of atmospheric parameters was then applied to the radiances to obtain at-aperature radiance. Random additive gaussian noise was applied to the data. ASSET was run using the synthetic data, as well as additional materials. The initial results from ASSET are promising. With a signal-to-noise ratio (SNR) of 500, the material was correctly classified 100% of the time. The mean absolute temperature error for this case was 0.02 K with a standard deviation of 0.02. The maximum absolute temperature error was 0.12 K. With a SNR of 300, the material was correctly classified more than 99% of the time. The mean absolute temperature error for this case was 0.04 K with a standard deviation of 0.03. The maximum absolute temperature error was 1.07 K. We present results from a simple synthetic data set as well as results from applying ASSET to more sophisticated synthetic DIRSIG LWIR imagery.

Further refinements are presented on a simple and fast way to cluster/segment hyperspectral imagery. In earlier work, it was shown that, starting with the first 2 principal component images, one could form a 2-dimensional histogram and cluster all pixels on the basis of the proximity to the peaks. Issues that we analyzed this year are the proper weighting of the different principal components as a function of the peak shape and automatic methods based on an entropy measure to control the number of clusters and the segmentation of the data to produce the most meaningful results. Examples from both visible and infrared hyperspectral data will be shown.

In this paper we discuss the implementation of a multi-scale Bayesian classifier that operates on hyperspectral data in both the spatial as well as the spectral domain, the Sequential Maximum A Posteriori (SMAP) classifier. Class assignments are modeled as a Markov random process in multi-resolution scale. For applications such as terrain categorization, the SMAP algorithm results in an improved classification that is less noisy than spectral-only based techniques. In addition, for highly overlapping classes, the SMAP significantly outperforms conventional discriminant function approaches. We present the results of the SMAP classifier on several hyperspectral datasets and discuss an extension of the algorithm to perform shading and sub-pixel analyses.

Hyperspectral imagery data sets present an interesting challenge to feature extraction algorithm developers. Beyond the immediate problem of dealing with the sheer amount of spectral information per pixel in a hyperspectral image, the remote sensing scientist must explore a complex algorithm space in which both spatial and spectral signatures may be required to identify a feature of interest. Rather than carry out this algorithm exploration by hand, we are interested in developing learning systems that can evolve these algorithms. We describe a genetic programming/supervised classifier software system, called GENIE, which evolves image processing tools for remotely sensed imagery. Our primary application has been land-cover classification from satellite imagery. GENIE was developed to evolve classification algorithms for multispectral imagery, and the extension to hyperspectral imagery presents a chance to test a genetic programming system by greatly increasing the complexity of the data under analysis, as well as a chance to find interesting spatio-spectral algorithms for hyperspectral imagery. We demonstrate our system on publicly available imagery from the new Hyperion imaging spectrometer onboard the NASA Earth Observing-1 (EO-1) satellite.

This paper presents the results of creating a Data Model by using neural equation networks of high order polynomials to achieve 100% correct classification of the Jacoby stellar spectra. The Jacoby set is a challenging group of 161 spectra spanning the full range of temperature, sub-temperature and luminosity groupings of standard star types. To achieve full learning, the development of a cascaded decision architecture linking an extensive network of polynomial decision equations was required. The two dominant features were extracted, and complex decision maps generated. Also, the sensitivity of the equation architecture to misclassification due to measurement noise was analyzed.

This paper presents the research results that demonstrates hyperspectral imaging could be used effectively for detecting feces (from duodenum, ceca, and colon) and ingesta on the surface of poultry carcasses, and potential application for real-time, on-line processing of poultry for automatic safety inspection. The hyperspectral imaging system included a line scan camera with prism-grating-prism spectrograph, fiber optic line lighting, motorized lens control, and hyperspectral image processing software. Hyperspectral image processing algorithms, specifically band ratio of dual-wavelength (565/517) images and thresholding were effective on the identification of fecal and ingesta contamination of poultry carcasses. A multispectral imaging system including a common aperture camera with three optical trim filters (515.4 nm with 8.6- nm FWHM), 566.4 nm with 8.8-nm FWHM, and 631 nm with 10.2-nm FWHM), which were selected and validated by a hyperspectral imaging system, was developed for a real-time, on-line application. A total image processing time required to perform the current multispectral images captured by a common aperture camera was approximately 251 msec or 3.99 frames/sec. A preliminary test shows that the accuracy of real-time multispectral imaging system to detect feces and ingesta on corn/soybean fed poultry carcasses was 96%. However, many false positive spots that cause system errors were also detected.

Identification and separation of poultry carcasses contaminated by feces and/or crop ingesta are very important to protect the consumer from a potential source of food poisoning. A transportable hyperspectral imaging system was developed to detect fecal and ingesta contamination on the surface of poultry carcasses. Detection algorithms used with the imaging system were developed from visible/near infrared monochromator spectra and with contaminates from birds fed a corn/soybean meal diet. The objectives of this study were to investigate using regions of interest reflectance spectra from hyperspectral images to determine optimal wavelengths for fecal detection algorithms from images of birds fed corn, wheat and milo diets. Spectral and spatial data between 400 and 900 nm with a 1.0 nm spectral resolution were acquired from uncontaminated and fecal and ingesta contaminated poultry carcasses. Regions of interest (ROIs) were defined for fecal and ingesta contaminated and uncontaminated skin (i.e. breast, thigh, and wing). Average reflectance spectra of the ROIs were extracted for analysis. Reflectance spectra of contaminants and uncontaminated skin differed. Spectral data pre-processing treatments with a single-term, linear regression program to select wavelengths for optimum calibration coefficients to detect contamination were developed. Fecal and ingesta detection models, specifically a quotient of 2 and/or 3-wavelengths was 100% successful in classification of contaminates.

Hyperspectral imaging (HSI) shows great promise for the detection and classification of several diseases, particularly in the fields of "optical biopsy" as applied to oncology, and functional retinal imaging in ophthalmology. In this paper, we discuss the application of HSI to the detection of retinal diseases and technological solutions that address some of the fundamental difficulties of spectral imaging within the eye. HSI of the retina offers a route to non-invasively deduce biochemical and metabolic processes within the retina. For example it shows promise for the mapping of retinal blood perfusion using spectral signatures of oxygenated and deoxygenated hemoglobin. Compared with other techniques using just a few spectral measurements, it offers improved classification in the presence of spectral cross-contamination by pigments and other components within the retina. There are potential applications for this imaging technique in the investigation and treatment of the eye complications of diabetes, and other diseases involving disturbances to the retinal, or optic-nerve-head circulation. It is well known that high-performance HSI requires high signal-to-noise ratios (SNR) whereas the application of any imaging technique within the eye must cope with the twin limitations of the small numerical aperture provided by the entrance pupil to the eye and the limit on the radiant power at the retina. We advocate the use of spectrally-multiplexed spectral imaging techniques (the traditional filter wheel is a traditional example). These approaches enable a flexible approach to spectral imaging, with wider spectral range, higher SNRs and lower light intensity at the retina than could be achieved using a Fourier-transform (FT) approach. We report the use of spectral imaging to provide calibrated spectral albedo images of healthy and diseased retinas and the use of this data for screening purposes. These images clearly demonstrate the ability to distinguish between oxygenated and deoxygenated hemoglobin using spectral imaging and this shows promise for the early detection of various retinopathies.

We present a method for separating a sensor radiance spectrum into a reflectance spectrum and an illumination spectrum. The method is based on the use of subspace models for both reflectance and illumination spectra. The method exploits the fact that reflectance and illumination spectra typically lie in distinct subspaces. The separation algorithm finds the best reflectance and illumination spectra within their respective subspaces. We have applied the algorithm to simulated VNIR radiance spectra using a large database of reflectance and illumination spectra. We have also examined the use of the recovered reflectance spectra for material identification over a database of materials.

The linear mixing model (LMM) is a well-known and useful method for decomposing spectra in a hyperspectral image into the sum of their constituents, or endmembers. Mathematically, if the spectra are represented as n-dimensional vectors, then the LMM implies that the set of endmembers defines a basis or coordinate system for the set of spectra. Because the endmembers themselves are generally not orthogonal, the geometry (distances, difference angles, etc.) is changed by moving from band space to endmember space. We explore some of the differences between the two coordinate systems, and show in particular that the difference in angle measurements leads to an improved method for subpixel target detection.

Hyperspectral data rarely spans the full band space because of factors such as sensor noise, numerical round-off, sparse sampling, and band correlation introduced by data processing. Standard exploitation of data, which often does not consider the possibility of a reduced band space, leads to reduced detection performance. Spectral signature detection performance can be improved by estimating the covariance on a subset of the band space components. The decision about how to limit the band space can be determined by factors such as in-scene estimation of noise. In-scene estimation of noise can be used to optimize spectral signature detection when spectral filtering methods based on covariance inverses are used. We present here a method for determining instrument noise and a new method of covariance inverse regularization which increases spectral filtering performance.

The traditional approach for achieving hyperspectral imaging over a broad range of infrared (IR) wavelengths involves multiple focal plane arrays (FPAs), dispersive elements, and optical beamsplitters. It has been shown that the current state-of-the-art in dual-band infrared focal plane array (IRFPA) technology allows spectral imagery to be obtained in two wavebands simultaneously with a single FPA, therefore reducing cryo-cooler and power requirements. The new approach described here advances the capabilities of the current state-of-the-art one step further and achieves a spectrometer concept based on a tri-band IRFPA. The tri-band spectrometer concept would lead to spectral imagery collected simultaneously in the SWIR, MWIR, and LWIR spectral regions with high efficiency. To achieve this a unique characteristic of a dispersive grating, that of overlapping spectral orders, would be exploited to allow simultaneous focusing of three spectral bands onto the multi-waveband FPA, thereby creating co-registered spectral images. The capabilities of a multi-waveband FPA then allow integration of spectra independently in the various orders. In addition, spectral images would be perfectly registered both spatially and spectrally, a difficult prospect for the traditional approach. By providing hyperspectral imagery in the SWIR, MWIR, and LWIR spectral regions, we capture the bulk of reflected and thermally emissive target and background phenomenology, within the constraints of atmospheric transmission. The characteristics of a suitable tri-band FPA are derived on the basis of our modeling efforts.

A familiar concept in imaging spectrometry is that of the three dimensional data cube, with one spectral and two spatial dimensions. However, available detectors have at most two dimensions, which generally leads to the introduction of either scanning or multiplexing techniques for imaging spectrometers. For situations in which noise increases less rapidly than as the square root of the signal, multiplexing techniques have the potential to provide superior signal-to-noise ratios. This paper presents a theoretical description and numerical simulations for a new and simple type of Hadamard transform multiplexed imaging spectrometer. Compared to previous types of spatially encoded imaging spectrometers, it increases etendue by eliminating the need for anamorphically compressed re-imaging onto the entrance aperture of a monochromator or spectrophotometer. Compared to previous types of spectrally encoded imaging spectrometers, it increases end-to-end transmittance by eliminating the need for spectral re-combining optics. These simplifications are attained by treating the pixels of a digital mirror array as virtual entrance slits and the pixels of a 2-D array detector as virtual exit slits of an imaging spectrometer, and by applying a novel signal processing technique.

Although the throughput and multiplex advantages of Fourier transform spectrometry were established in the early 1950's (by Jacquinot and Fellgett , respectively) confusion and debate arise when these advantages are cited in reference to imaging spectrometry. In non-imaging spectrometry the terms throughput and spectral bandwidth clearly refer to the throughput of the entire field-of-view (FOV), and the spectral bandwidth of the entire FOV, but in imaging spectrometry these terms may refer to either the entire FOV or to a single element in the FOV. The continued development of new and fundamentally different types of imaging spectrometers also adds to the complexity of predictions of signal and comparisons of signal collection abilities. Imaging spectrometers used for remote sensing may be divided into classes according to how they relate the object space coordinates of cross-track position, along-track position, and wavelength (or wavenumber) to the image space coordinates of column number, row number, and exposure number for the detector array. This transformation must be taken into account when predicting the signal or comparing the signal collection abilities of different classes of imaging spectrometer. The invariance of radiance in an imaging system allows the calculation of signal to be performed at any space in the system, from the object space to the final image space. Our calculations of signal - performed at several different spaces in several different classes of imaging spectrometer - show an interesting result: regardless of the plane in which the calculation is performed, interferometric (Fourier transform) spectrometers have a dramatic advantage in signal, but the term in the signal equation from which the advantage results depends upon the space in which the calculation is performed. In image space, the advantage results from the spectral term in the signal equation, suggesting that this could be referred to as the multiplex (Fellgett) advantage. In an intermediate image plane the advantage results from a difference in a spatial term, while for the exit pupil plane it results from the angular term, both of which suggest the throughput (Jacquinot) advantage. When the calculation is performed in object coordinates the advantage results from differences in the temporal term.

We have previously shown those circumstances for which the multiplex advantage of temporally scanned Fourier transform imaging spectrometers enables higher signal-to-noise ratios than other techniques. Unfortunately, for many real-life applications, such as aerial reconnaissance, deployment of FT instruments based on traditional moving-mirror interferometers is problematic due to their inherent sensitivity to vibration. We will describe a new type of Fourier transform imaging spectrometer, employing moving birefringent prisms to create the necessary path difference modulations. This new system retains the accepted sensitivity advantages of traditional Fourier transform devices, but because it employs common-path interferometry and because path differences are introduced within a single optical element, the system is inherently very robust. Furthermore, the precision of the movement can be typically two orders of magnitude lower than for a traditional two-beam interferometer, resulting in a simpler instrument. Experimental results will be presented.

We present a new approach to hyperspectral imaging that is inspired by biological imaging systems, such as human vision, which employ high spectral and spatial discrimination only in a small central patch. This foveal technique addresses several problems of conventional approaches to HSI: they cannot provide snapshot, high spectral-resolution imagery in a two dimensional format. The ability to provide the data in a single snapshot removes temporal mis-registration issues. High signal to noise ratios naturally result from the absence of any multiplexing technique and the corresponding loss of light. Other reported snapshot techniques are either low spectral resolution or provide only a one-dimensional field of view. A high-spectral-resolution imager with a wide field of view could produce giga-sample data rates, which would make real-time data processing problematic. By gathering hyperspectral data from only a selected portion of the scene, we reduce the data processing rates to manageable levels. For many applications only a small field of view is required, but needs to be cued for situational awareness. In our system, this is provided for by a wide field of view, panchromatic imager, which fills a similar role to peripheral vision in the biological systems mentioned above. Our technique images the selected region onto a coherent fibre bundle, which reformats the input into a line array constituting the input to a dispersive hyperspectral imager. Computer processing reformats the dispersed one-dimensional output into a rectangular image and applies calibration routines to produce a high spectral resolution, small hyperspectral image. This is combined with a high-spatial-resolution panchromatic image. Experimental results will be presented.

In this paper, a novel filter-based greedy modular subspace (GMS)technique is proposed to improve the accuracy of high-dimensional remote sensing image supervisor classification. The approach initially divides the whole set of high-dimensional features into several arbitrary number of highly correlated subgroup by performing a greedy correlation matrix reordering transformation for each class. These GMS can be regarded as a unique feature for each distinguishable class in high-dimensional data sets. The similarity measures are next calculated by projecting the samples into different modular feature subspaces. Finally, a supervised multi-class classifer which is implemented based on positive Boolean function (PBF) schemes is adopted to build a non-linear optimal classifer. A PBF is exactly one sum-of-product form without any negative components. The PBF possesses the well-known threshold decomposition and stacking properties. The classification errors can be calculated from the summation of the absolute errors incurred at each level. The optimal PBF are found and designed as a classifer by minimize the classification error rate among the training samples. Experimental results demonstrate that the proposed GMS feature extraction method suits the PBF classifer best as a classification preprocess. It signifcantly improves the precision of image classification compared with conventional feature extraction schemes. Moreover, a practicable and convenient "vague" boundary sampling property of PBF is introduced to visually select training samples from high-dimensional data sets more effciently.

Imaging systems such as the Computed Tomographic Imaging Spectrometer (CTIS) are modeled by the matrix equation g = Hf, which is the discretized form of the general imaging integral equation.. The matrix H describes the contribution to each element of the image g from each element of the hyperspectral object cube f. The vector g is the image of the spatial/spectral projections of f on a focal plane array (FPA). The matrix H is enormous, sparse and rectangular. It is extremely difficult to discretize the integral operator to obtain the matrix operator H. Normally H is constructed empirically from a series of monochromatic calibration images, which is a time consuming process. However we have been able to synthetically construct H by numerically modeling how the optical and diffractive elements in the CTIS project monochromatic point source data onto the FPA. We can evaluate a CTIS system by solving the imaging equation for f using both the empirical and synthetic H from some test data g. Comparison between the two results provides a means to evaluate and improve CTIS system calibration procedures noting that the synthetic system matrix H represents a baseline ideal system.

The development of an imaging spectrometer to acquire 3 spatial dimensions and hyper-spectral information simultaneously is detailed. The spectrometer is based on the Computed Tomographic Imaging Spectrometer (CTIS) developed at the Optical Science Center and the Scannerless Laser Radar (LADAR) architecture developed at Sandia National Labs. The new 4-D imager, called the Spectral LADAR System (SLS), operates in the visible to near-infrared portion of the spectrum (600-900 nm). The system has 30 spectral intervals (10 nm bands), 1024 range samples, and approximately 80 x 80 spatial sampling. CTIS and LADAR are discussed, as well as preliminary results of the SLS.