In this paper we evaluate the ability of the Matched Subspace Detector (MSD), Matched Filter Detector (MFD) and Orthogonal Subspace Projection (OSP) to discriminate material types in laboratory samples of intimately mixed bidirectional reflectance data. The analysis consists of a series of experiments where bidirectional reflectance spectra of intimate mixtures of enstatite-olivine and anorthite-olivine in various proportions are converted to single scattering albedo (SSA) using Hapke's model for bidirectional reflectance. The linearized SSA spectra are used as inputs to the various detectors and the output for each is evaluated as a function of the proportion of target- to-interference. Results are presented as a series of figures that show overall the MSD has a higher target-to- background separation (i.e., better class separation) than either the MFD or OSP. This target-to-background separation results in fewer false alarms for the MSD than either of the other two detectors.

We present an algorithm for subpixel material identification that is invariant to the illumination and atmospheric conditions. The target material spectral reflectance is the only prior information required by the algorithm. A target material subspace model is constructed from the reflectance using a physical model and a background subspace model is estimated directly from the image. These two subspace models are used to compute maximum likelihood estimates for the target material component and the background component at each image pixel. These estimates form the basis of a generalized likelihood ratio test for subpixel material identification. We present experimental results using HYDICE imagery that demonstrate the utility of the algorithm for subpixel material identification under varying illumination and atmospheric conditions.

The increasing levels of greenhouse gases into the atmosphere is of rising concern to the world community. The formulation of treaties to monitor such emissions levels must be developed upon the ability to detect these gases in the differing levels of concentration found in the real world. This in turn is dependent upon a technical understanding of the capability to reliably and accurately measure their spectral signatures through the atmosphere. The effects of gas concentrations, atmosphere, electro- optical sensing system performance and ground processing can be quantified using statistical measures in order to assess sensing capability.

We present two approaches for removing ocean surface (and subsurface) clutter in multispectral and hyperspectral imagery. The first approach is a computationally simple, physics-based algorithm that models surface clutter as a two-component mixture of surface-reflected light (glint) and scattered upwelling light. The model exploits the difference in the spectral content of the two components in order to differentiate glint from subsurface scattered light. The second approach is a statistics-based algorithm that simultaneously models the local spectral and spatial correlation structure with a linear predictive filter that spatially adapts to the statistical properties of the image on a region-by-region basis as determined by a spectrally generated segmentation map. The filter coefficients for each image segment are estimated via autoregression using a multichannel, multidimensional formulation of the Yule- Walker equations. The combined spatial-spectral processing allows the filter predictor to remove subsurface background clutter, as well as glint. In both algorithms, decluttered residuals are obtained by subtracting the background estimate from the input image data. In a test of effectiveness, the data conditioning provided by decluttering on a regional basis improved the performance of our set of matched-filter detectors by an average of 3 dB. The performance enhancement demonstrates the need for regional estimates of background when the assumption of spatially stationary data is no longer valid. Examples of both decluttering and matched-filter detection processing are presented using data collected by the AAHIS sensor.

High-quality, multispectral thermal infrared sensors can, under certain conditions, be used to measure more than one surface temperature in a single pixel. Surface temperature retrieval in general is a difficult task, because even for a single unknown surface, the problem is under-determined. For the example of an N-band sensor, a pixel with two materials at two temperatures will, in principle, have 2(N + 1) unknowns (N emissivities and one temperature for each of two materials). In addition, the upwelling path and reflected downwelling radiances must be considered. Split-window (two or more bands) and multi-look (two or more images of the same scene) techniques provide additional information that can be used to reduce the uncertainties in temperature retrieval. Further reduction in the uncertainties is made if the emissivities are known, either a priori (e.g., for water) or by ancillary measurements. Ultimately, if the number of unknowns is reduced sufficiently, the performance of the sensor will determine the achievable temperature sensitivity. This paper will explore the temperature sensitivity for a pixel with two temperatures that can be obtained under various assumptions of sensor performance, atmospheric conditions, number of bands, number of looks, surface emissivity knowledge, and surface composition. Results on synthetic data sets will be presented.

Algorithms have been developed for estimating temperature and emissivity of terrestrial targets from airborne LWIR hyperspectral radiance data. This report documents a performance comparison of these algorithms. Data collected at the Department of Energy's Atmospheric Radiation Measurement Site by the Spatially-Enhanced Broadband Array Spectrograph System (SEBASS) was used. SEBASS is a 128-band hyperspectral instrument covering a wavelength range of 7.3 to 13.6 micrometers. Spectra of natural and man-made objects were processed by Thermal Log Residual, Alpha Log Residual, Normalized Emissivity, Inverse Wave, Graybody, EOS/ASTER, N- Temperature, and Max-Min Difference methods. Temperature and emissivity estimates are compared to ground truth measurements. Sample results and performance statistics are presented.

With the advent of multi-spectral thermal imagers such as EOS's ASTER high spatial resolution thermal imagery of the Earth's surface will soon be a reality. Previous high resolution sensors such as Landsat 5 had only one spectral channel in the thermal infrared and its utility to determine absolute sea surface temperatures was limited to 6 - 8 K for water warmer than 25 deg C. This inaccuracy resulted from insufficient knowledge of the atmospheric temperature and water vapor, inaccurate sensor calibration, and cooling effects of thin high cirrus clouds. We will present two studies of algorithms and compare their performance. The first algorithm we call `robust' since it retrieves sea surface temperatures accurately over a fairly wide range of atmospheric conditions using linear combinations of nadir and off-nadir brightness temperatures. The second we call `physics-based' because it relies on physics-based models of the atmosphere. It attempts to come up with a unique sea surface temperature which fits one set of atmospheric parameters.

Spectral features are often extracted from multispectral/hyperspectral data using a multiresolutional decomposition known as the spectral fingerprint. While the spectral fingerprint method has proven to be quite powerful, it has also shown several shortcomings: (1) its implementation requires multiple convolutions with Laplacian-of-Gaussian filters which are computationally expensive, (2) it requires a truncation of the filter impulse response which can cause spurious errors, and (3) it provides information about the sizes and areas of radiance features but not the shapes. It is proposed that a wavelet- based spectral fingerprint can overcome these shortcomings while maintaining the advantages of the traditional method. In this study, we investigate the use of the wavelet transform modulus-maximus method to generate a wavelet-based spectral fingerprint. The computation of the wavelet-based fingerprint is based on recent fast wavelet algorithms. The analyses consists of two parts: (1) the computational expense of the new method is compared with the computational costs of current methods, and (2) the outputs of the wavelet-based methods are compared with those of current methods to determine any practical differences in the resulting spectral fingerprints.

In this paper, we present a practical and potential useful approach to spatial spectral feature extraction of hyperspectral imagery. Many new hyperspectral imaging sensors have collected hyperspectral data cubes in optical- infrared wavebands. The data cubes have two spatial dimensions and one spectral dimension providing hundreds of images of the same scene in different wavebands. Radiance clutter may interfere with the detection of specific target signals in the data cubes. But, target signals and background sources generally have different spatial features in different spectral bands. For example, target signals may have a higher contrast than the local background in one band but not in a different band. We exploit these band differences to detect the target signals and extract spatial and spectral features. In our analysis we use real image cube data from sensors such as the Fourier Transform Hyperspectral Imager. We combine traditional spatial processing, such as frame differencing and adaptive filters, but apply them to different image band instead of different images of the same scene obtained at different times. We compute the probabilities of detection and false alarms for targets of a given strength against the measured optical clutter. We compare target detection algorithms using only one band with those using multiple bands.

This paper presents a wavelet-based hyperspectral image coder that is optimized for transmission over the binary symmetric channel. The proposed coder uses a robust channel- optimized trellis-coded quantization stage that is designed to optimize the image coding based on the channel characteristics. This optimization is performed only at the level of the source encoder, and does not include any channel coding for error protection. The robust nature of the coder increases the security level of the encoded bit stream and provides a much more visually pleasing rendition of the decoded image. In the absence of channel noise, the proposed coder is shown to achieve a compression ratio greater than 70:1, with an average PSNR of the coded hyperspectral sequence exceeding 40 dB. Additionally, the coder is shown to exhibit graceful degradation with increasing channel errors.

A new method of automatically determining control points for registration and fusion of multispectral images is presented. This method was motivated by the question of whether it is possible to computationally assess the quality of optical flow estimates at various points throughout the image without knowing the true flow field. Somewhat surprisingly the answer is yes and is determined by the norm of the least squares operator associated with the (windowed) optical flow equations. This approach has several advantages. First it shows the danger of using the condition number of the optical flow equations to measure the reliability of the computed flow. Second the method isolates points in the image corresponding to maximum reliability. These points in turn can be used as control points for registration and fusion without actually computing the optical flow and indeed only require a single frame for computation. Since this computation only requires a few operations per pixel it is very fast. The control points are defined as the minima of the norm of the least squares operator and as such enjoy a great deal of invariance with respect to the regional intensity changes seen in multispectral images. For this reason they are ideal for multispectral registration. A multiscale version of this method has been developed that allows a coarse to fine control point decomposition for suppressing the negative effects of noise and clutter. Various applications are presented demonstrating the utility of this approach for real world images including multispectral satellite images and dual spectral IR image sequences. For the latter sequence we were able to obtain subpixel motion estimates that were accurate to within one percent of the true motion.

Automated material characterization and identification from airborne imagery is an important capability for many applications including target recognition and geospatial database construction. Hyperspectral imagery provides a rich source of information for this purpose but utilization is complicated by the variability in a material's observed spectral signature due to the ambient conditions and the scene geometry. In this paper, we present a method that uses a single spectral radiance function measured from a material under unknown conditions to synthesize a comprehensive set of radiance spectra that corresponds to that material over a wide range of conditions. This set of radiance spectra can be used to build a hyperspectral subspace representation that can be used for material identification over a wide range of circumstances. We demonstrate the use of these algorithms for model synthesis and material mapping using HYDICE imagery acquired at Fort Hood, Texas. The method correctly maps several classes of roofing materials, roads, and vegetation over significant spectral changes due to variation in surface orientation. We show that the approach outperforms methods based on direct spectral comparison.

The use of hyperspectral imagery for remote sensing detection applications has received attention recently due to the ability of the hyperspectral sensor to provide registered information in both space and frequency. In this paper we extend our work on the development of an efficient implementation of the Maximum Likelihood (ML) detector in which we use a 3D Gauss-Markov random field to model the clutter background in the hyperspectral data. We review the details of the optimal ML estimation approach for obtaining the Markov parameters and discuss a gradient based optimization scheme for obtaining these estimates. To improve the computational efficiency of the overall detection algorithm, we develop an estimation method based on some simple mathematical approximations that allows us to explicitly solve for the Markov parameters. In addition, we use a stochastic, rather than deterministic target model by implementing a single hypothesis test in place of the more traditional binary hypothesis paradigm. We compare the detection performance and the computational requirements of our updated model and detector implementation to the benchmark RX detection algorithm for hyperspectral imagery.

Recent developments of more sophisticated sensors enable the measurement of radiation in many more spectral intervals at a higher spectral resolution than previously possible. As the number of bands in high spectral resolution data increases, the capability to detect more objects and the detection accuracy should increase as well. Most of the detection techniques presently used in hyperspectral data require the use of spectral libraries that contain information on specific objects to be detected. An example of one technique used for detection purposes in hyperspectral imagery is the spectral angle approach based on the Euclidean inner product of the spectral signatures. This method has good performance on objects that have sufficient differences between their spectral signatures. This paper presents a partially supervised detection approach that uses previously measured spectral responses as inputs and is capable of differentiating objects that have similar spectral signatures. Two versions will be presented: one that is based on Statistical Pattern Recognition and other based on Fuzzy Pattern Recognition. The detection mechanisms are tested with objects of very similar spectral signatures and the detection results are compared with those from the spectral angle approach.

Deriving information about the Earth's surface requires atmospheric corrections of the measured top-of-the- atmosphere radiances. One possible path is to use atmospheric radiative transfer codes to predict how the radiance leaving the ground is affected by the scattering and attenuation. In practice the atmosphere is usually not well known and thus it is necessary to use more practical methods. We will describe how to find dark surfaces, estimate the atmospheric optical depth, estimate path radiance and identify thick clouds using thresholds on reflectance and NDVI and columnar water vapor. We describe a simple method to correct a visible channel contaminated by a thin cirrus clouds.

An airborne hyperspectral line scanner is used to image the ground as the aircraft moves on a single trajectory. In reality, it may be difficult for the aircraft to maintain a perfectly steady course causing distortions in the imagery. So, special subsystems including stabilizers are used to maintain the hyperspectral line scanner on the proper course. If the subsystems of an airborne hyperspectral line scanner are malfunctioning or if the proper stabilizers are not available, then a technique is needed to restore the imagery. It no stabilizers are used on the airborne line scanner, but if aircraft navigation information is available including yaw, pitch and roll, then the restoration may be automated. However, if the stabilizers are malfunctioning or if the navigation information is corrupted or unavailable, then a technique is needed to restore the imagery. This paper introduces an automated technique for restoring hyperspectral images that was used on some images obtained for the Dynamic Data Base program sponsored by the Defense Advanced Research Projects Agency. The automated approach is based on image flow vectors obtained from the unstable image. The approach is introduced along with results that demonstrate how successful the restoration is at the feature level.

Imaging spectrometers must be calibrated and characterized in order to accurately interpret collected data, especially for sharp spectral features in the target spectra. Two key aspects of this characterization are the spectral bandwidth and spectral registration. Calibration and characterization methodologies are being developed, with the casi (Compact Airborne Spectrographic Imager) as a test instrument.

This presentation describes the major performance characteristics of the developed camera and demonstrates some results of target contrast enhancement in laboratory and field experiments.

Nonlinear predictors based on feedforward artificial neural networks are investigated for use in lossless compression of AVHRR Imagery. The FNN predictors are designed and compared to the optimum nonlinear Mean Square Error predictor, and to the linear predictor. The predictors are compared based on the first order entropy of the predictor error, on run time, and memory requirements. The FNN predictors can be designed to have a wide range of performance with a trade off between first order entropy error, and memory and run time. There is little difference in prediction errors between the linear and FNN predictors for large sample sizes, when the image is segmented into large areas. The difference can be greater for smaller sample sizes, when the image is segmented into smaller areas such as the typical 8 X 8 pixel size. The results indicate there is no advantage to using nonlinear predictors when compression and run time requirements are taken into account.