A multiprocessor version of the ORASIS hyperspectral analysis program has been implemented in support of the ASRP. In brief, the long-term technical objectives of the ASRP are to demonstrate the feasibility and military utility of real-time target detection from uncrewed air vehicles using hyperspectral data. This paper presents a preliminary assessment of ORASIS performance and describes the ORASIS development effort designed to meet the ASRP goals. Real-time performance of the analysis program and its potential effectiveness as a target detection method are demonstrated.

The release of harmful effluents into the earth's atmosphere is an increasing world-wide concern. Technical feasibility to detect and localize such releases is a necessary first step before sound policy decisions can be established to regulate such releases. This paper examines key system parameters by quantifying their effects in terms of Receiver Operating Characteristics (ROC) curves. It establishes upper performance bounds based on perfect apriori information of the atmospheric state and thus can be used to gauge measured effectiveness of candidate detection algorithms. In the first part of this paper, a theoretical discussion is presented on the development of the probability density functions (pdfs) required to perform the ROC analysis. These pdfs are associated with 1₂ measures of spectral vector magnitude lengths, a measurable quantity for effluent detection. It will be shown that these functions are non-Gaussian and are functions of the number of spectral bands in the data. Generation of these functions are also discussed in this paper both analytical and through Monte Carlo methods. In the second half of this paper, ROC performance curves are generated for various sensor and source parameters. These curves are generated for various sensor noise and spectral aggregation conditions constrained by the assumption of shot-limited performance. Impacts of spectral aggregation and plume-to- ground temperature differentials are also examined and related to ROC performance. Parametric evaluations are confined to the long-wave infrared (LWIR) spectral regime where sensor resolution is systematically varied between 1.0 and 8.0 cm^-1. System models, analysis methodology, and ROC results are presented.

This paper presents (1) trade-off studies of detection performance versus the number of bands using reflective hyperspectral imagery; (2) the quantitative detection performance of various approaches used in automatic target detection. The trade-off studies of detection performance versus the number of bands are based on the Adaptive Real-Time Endmember Selection and Clutter Suppression (ARES) algorithm. The ARES algorithm presents a new concept and approach for spectral-spatial aided/automatic target detection based on the unique characteristics of the spectral signatures produced by the hyperspectral imaging system for remote sensing surveillance and reconnaissance applications. This paper compares the quantitative detection performance based on the ARES algorithm with other automatic target detection approaches. This paper uses the Forest Radiance I database collected with the HYDICE hyperspectral sensor at Aberdeen U.S. Army Proving Ground in Maryland, including scenarios such as targets in the open, with footprint of 1 meter, and at different times of day.

In this work, we generate ROC curves on real and synthetic scenes and develop scoring methods to evaluate the performance of the ORASIS hyperspectral algorithm. The goal of this effort is to improve the overall performance of ORASIS, focusing on the endmember selection methods. ROC curve evaluations have been performed on hyperspectral data sets from different scenes. We have scored by target and by target pixel. A scene generator has been developed allowing many features: combination of real or synthetic background and multiple, distinct targets; user-defined angle of target spectrum to background subspace; and user-specified non-uniform target/background transparency.

Principal Components Analysis is very effective at compressing information in multivariate data sets by computing orthogonal projections that maximize the amount of signal variance. Unfortunately, information content in hyperspectral images does not always coincide with such projections. We propose the application of Projection Pursuit, which seeks to find a set of orthogonal projections that are 'interesting' in the sense that they deviate from the Gaussian distribution assumption. Once these projections are obtained, they can be used for image compression, segmentation, or enhancement for visual analysis. To find these projections we follow a 2-step iterative process where we first search for a projection that maximizes a projection index based on the divergence of the projection's estimated probability distribution from the Gaussian distribution, and then reduce the rank by projecting the data onto the subspace orthogonal to the previous projections. To find the projection that maximizes the index, a novel approach is taken which does not use an optimization algorithm, but rather searches for a solution by obtaining a set of candidate projections from the data and choosing the one with the highest projection index. This method is shown to work with simulated examples as well as data for the Hyperspectral Digital Imagery Collection Experiment.

The measured spectral radiance signature for a material can vary significantly due to atmospheric conditions and scene geometry. We show using a statistical analysis of a comprehensive physical model that the variation in a material's spectral signature lies in a low-dimensional space. The spectral radiance model includes reflected solar and sky radiation as well as path radiance. Signature variability is introduced by effects such as solar occlusion and variation in the concentrations of atmospheric gases aerosols. The MODTRAN 3.5 code was employed for computing radiative transfer aspects of the model. Using the new model, we develop a maximum likelihood algorithm for automatic material identification that is invariant to atmospheric conditions and scene geometry. We demonstrate the algorithm for the identification of exposed and concealed material samples in HYDICE imagery.

A system is presented for compression of hyperspectral imagery. Specifically, DPCM is used for spectral decorrelation, while a robust 2-D discrete wavelet coding scheme is used for spatial decorrelation. Trellis-coded quantization is used to encode the wavelet coefficients. Side information and rate allocation strategies are discussed. Fixed-rate codebooks are designed using a modified version of the generalized Lloyd algorithm. This system achieves a compression ratio of greater than 70:1, with an average PSNR of the coded hyperspectral sequence exceeding 40 dB.

This paper presents an image quality improvement approach for extraction of map features based on single and multisensor image enhancement techniques. The approach relies upon the evaluation of sensor imagery for extraction of map features based on its sensor characteristics. The approach will be illustrated by evaluation results for RADARSAT using two visual metrics. Samples of single sensor and multiband enhancement results will be presented for a range of map features. Features will include those essential for shoreline categorization and delineation for support of environmental applications.

Work has been on-going at the Rochester Institute of Technology's Center for Imaging Science in the spatial resolution enhancement ofthermal infrared imagery. Based upon the most recent work by the Digital Imaging and Remote Sensing group, several areas of study were carried out using M7 and Landsat data sets. All investigations undertaken were confmed to cases that ensure radiometric fidelity across image processing operations, since derivation of accurate temperature or emissivity maps necessitate this requirement. Given a low spatial resolution thermal band, these methods produced a high resolution estimate thereofbased on enhancement using: (1) a single panchromatic band; and (2) a high resolution class-map derived from multi-spectral bands. The single band method relies on thresholding of grey-levels within a local window. Local processing is also the key to the class-map, multi-band approach. A variety of factors related to output quality, such as resolution ratio and the number of local windows, were examined. M7 scenes were used to develop/test the algorithm by degrading thermal data and restoring it to the original resolution. The completed algorithms have also been applied to Landsat imagery with the reflective channels used to enhance the 120 m thermal band. An overview ofthe background equations is given, workings ofthe algorithm are presented and resulting output is shown and compared to truth.

This paper describes our approach and presents measured results of the extension of multispectral sharpening techniques to hyperspectral imagery. Our approach produce high spatial resolution spectral imagery using a least squares estimator. The estimator is based on the underlying physics of the spectral imaging process. The intent of the process it to produce high spatial resolutions with the best possible spectral fidelity. The results on multiple test cases demonstrate sharpened imagery within 5% of the true high resolution hyperspectral values.

High resolution panchromatic imagery can be used to increase the spatial resolution of low resolution spectral imagery through spatial/spectral sharpening techniques. Recently, sharpening techniques have been presented that use a multiresolution analysis by manipulating the images at different resolution scales that use a Laplacian pyramid or wavelet transform. This paper presents a model for sharpening multispectral images (MRA/MTF) that uses multiresolution analysis (MRA) to extract the high frequency information from the panchromatic image and matches the spatial response between imagery using a modulation transfer function (MTF) correction. When carefully executed, the MRA/MTF model is shown to provide a sharpened image with minimum spectral distortion and visually pleasing results. Multispectral data was used to evaluate the algorithm for sharpening 30 meter Landsat data with 1 meter aerial photography with comparison to other sharpening algorithms available within ERDAS Imagine and with ERIM's sharpening algorithm called Sparkle. The algorithms were tested for spectral distortion by comparing the covariance between bands before and after sharpening.

Classification of earth terrain from satellite radar imagery represents an important and continually developing application of microwave remote sensing. The basic objective of this paper is to derive more information, through combining, than is present in any individual element of input data. Multispectral data has been used to provide complementary information so as to utilize a single SAR data for the purpose of land-cover classification. More recently neural networks have been applied to a number of image classification problems and have shown considerable success in exceeding the performance of conventional algorithms. In this work, a comparison study has been carried out between a conventional Maximum Likelihood (ML) classifier and a neural network (back-error-propagation) classifier in terms of classification accuracy. The results reveal that the combination of SAR and MSS data of the same scene produced better classification accuracy than either alone and the neural network classification has an edge over the conventional classification scheme.

Multispectral sensing with acoustic-optical tunable filters (AOTFs) offers several advantages for the automated recognition of targets and classification of terrain features. High spectral resolution, real-time selection of the wavelength, and dual polarization are among the chief advantages. AOTFs employed in an imaging sensor usually involve a shifting of the spatial image on the focal plane as the wavelength is sampled. This results in a misregistered data hypercube where selected spectral images are not aligned spatially. This can severely limit the sensor's application if not accounted for and rectified. An edged-based routine operating on data taken with the Real-Time Multispectral Sensor (RTMS), an imaging AOTF sensor produced by the Jet Propulsion Laboratory, will be described and demonstrated in this paper. The method is completely general and is capable of removing misregistration for any reason (e.g. platform jitter) not only for AOTF-induced misregistration. This paper will also provide examples of image classification within several scenes collected by RTMS during tower data collection. The basis for the classification is spectral and/or polarization characteristics of the targets and scenes.

This paper discusses the nonuniform illumination of individual pixels in an array that is intrinsic to the scene viewed, as opposed to turbulence or platform motion as an error source in quantitative imagery. It describes two classes of algorithms to treat this type of problem. It points out that this problem can be viewed as a type of inverse problem with a corresponding integral equation unlike those commonly treated in the literature. One class allows estimation of the spatial variation of radiance within pixels using the single digital number irradiances produced by the measurements of the detectors within their instantaneous-fields-of-view (IFOVs). Usually it is assumed without discussion that the intrapixel radiance distribution is constant. Results are presented showing the improvements obtained by the methods discussed.

The calculation of ground reflectance imagery from satellite scenes acquired over mountainous terrain depends on a number of factors. Two factors are considered here: the spatial resolution of the Digital Elevation Model (DEM) and bidirectional reflectance (BRDF) effects. Due to the large range of local solar incidence angles in a rugged terrain a strong departure from the isotropic reflectance behavior is often apparent in the imagery. Simple empirical functions are offered to reduce the BRDF influence for the reflectance image product. DEM errors, an inadequate spatial resolution or a small sub-pixel misregistration between an image pixel and the DEM resolution cell lead to reflectance errors. The magnitude of this error is wavelength-dependent. Some typical configurations are investigated to assess the DEM influence.

In this paper we address the problem of detecting targets in hyperspectral images when the target signature is buried in random noise and interference (from other materials in the same pixel). We assume that the hyperspectral pixel measurement is a linear combination of the target and interference signatures observed in additive noise. The linear mixing assumption leads to a linear vector space interpretation of the measurement vector, which can be decomposed into a noise-only subspace and a target-plus- interference subspace. While it is true that the target and interference subspaces are orthogonal to the noise-only subspace, the target subspace and interference subspace are, in general, not orthogonal. The non-orthogonality between the target and interference subspaces results in leakage of interference signals into the output of matched filters resulting in false detections (i.e., higher false alarm rates). In this paper, we replace the Matched Filer Detector (MFD), which is based on orthogonal projections, with a Matched Subspace Detector (MSD), which is built on non- orthogonal or oblique projections. The advantage of oblique projections is that they eliminate the leakage of interference signals into the detector, thereby making detectors based on oblique projections invariant to the amount of interference. Furthermore, under Gaussian assumptions for the additive noise, it has been shown that the MSD is Uniformly Most Powerful (higher probability of detect for a fixed probability of false alarm) among all detectors that share this invariance to interference power. In this paper we evaluate the ability of two versions of the MSD to detect targets in HYDICE data collected over sites A and B located at the U.S. Army Yuma proving grounds. We compute data derived receiver operating characteristics (ROC) curves and show that the MSD out- performs the MFD.

Hyperspectral sensors collect hundreds of images in contiguous and narrowly spaced spectral bands. They have the potential to simultaneously provide high spatial and spectral resolution of targets of interest in Automatic Target Detection and Recognition (ATD/R). The price to be paid is the need to process and store an extremely large amount of data in an effective and timely manner. We develop a new implementation of the maximum-likelihood (ML) detector which is both practical and efficient. Our detection is based on a Gauss- Markov Random Field (GMRF) model for the data which avoids the inversion of large data covariance matrices usually encountered in ML-detectors. The paper presents two algorithms to fit the GMRF to the hyperspectral sensor data: an optimal ML estimation algorithm and a suboptimal Least Squares (LS) estimation algorithm. Using the LS-algorithm, we develop the structure of the detector and present estimation results from a real hyperspectral data set.

We present a hierarchical classification technique that discriminates broad categories of surface materials in terms of ground true features, such as water, vegetation, and soils from spectral information. Subsequently, we further discriminate these materials and extract finer ground features, like chemistries, peculiar to each. The interaction at various scales of the 3D spatial and the spectral domains is decomposed by using wavelet tools to address scale dependencies in the spatial domain, a robust spectral unmixing technique, called Hierarchical Foreground Background Analysis (HFBA) along the spectral axis. HFBA sequentially derives a series of weighting vectors discriminating features at different levels of detection: (1) constituent materials, (2) types within constituents, and (3) chemistries peculiar to each type. Our goal is two-fold. First, we present the combination of HFBA and wavelets as a supervised classification technique validating the categories imposed by the supervised classification, and manifesting clusters which can refine the classification at different scales. Second, we identify spectral redundancies between hyperspectral and multispectral information, studying mixture at different spatial/spectral resolutions and assess whether targeted features may be extracted as efficiently from multispectral data as they could be from hyperspectral data. Results on AVIRIS and simulated MODIS data illustrate the robustness and effectivity of the technique.

Face recognition has potential applications in surveillance and security systems. Most algorithms proceed by constructing a feature vector, which represents those aspects of the face which vary most between individuals. In general these algorithms locate critical features such as the eyes, nose and mouth, and use a mixture of direct comparison, and spatial distribution of features to identify those candidate images in a large database which match the image currently under analysis. Since pixel intensity comparisons for regions around the eyes are often used to compute a measure of similarity between two images lighting can play a major part in the success of recognition techniques. By using a combination of infrared and white light image information we hope to normalize pixel intensities, reducing the error margin due to variations in lighting conditions. However, the most significant problem with recognition systems is the processing delay required to find a match. Since a high degree of certainty is required in security applications, this increases the level of processing required leading to even greater delays. This paper proposes and evaluates a parallel algorithm for face recognition. The general performance characteristics of the algorithm are analyzed using formal methods to obtain the expected temporal and spatial complexities of the algorithm, and to estimate the speedup that can be achieved. Confirmation of these predictions is provided by implementing the parallel algorithm on an eight processor Sequent Balance shared memory multiprocessor and PRISM/DiST, a multiprocessor simulator.

In this paper, we propose an optimal feature extraction method for normally distributed data. The feature extraction algorithm is optimal in the sense that we search the whole feature space to find a set of features which give the smallest classification error for the Gaussian ML classifier. Initially, we start with an arbitrary feature vector. Assuming that the feature vector is used for classification, we compute the classification error. Then we move the feature vector slightly in the direction so that the classification error decreases most rapidly. This can be done by taking gradient. We propose two search methods, sequential search and global search. In the sequential search, if more features are needed, we try to find an additional feature which gives the best classification accuracy with the already chosen features. In the global search, we are not restricted to use the already chosen features. Experiment results show that the proposed method outperforms the conventional feature extraction algorithms.

We used a 3-D wavelet-based denoising method to reduce the noise from multispectral imagery. In our approach, we compared denoising of different bands of a multispectral image using a 2-D denoising technique, by which the wavelet coefficients corresponding to each band were denoised independent of each band, and a 3-D denoising technique by which the wavleet coefficients were denoised by involving all bands in thresholding the wavelet coefficients. Due to the high correlation of the multispectral imagery data along the wavelength axis, the noise can be easily reduced by applying the wavelet transform along the wavelength direction. Our results showed that the 3-D denoising approach improved the overall SNR of a noisy multispectral imagery over the 2-D denoising approach, due to the correlation between the different bands.

We use an algorithm based on the natural immune system for classification of aerial multispectral imagery. Our artificial immune system works by maintaining a population of detectors that remove undesired patterns, but pass a specified training set of positive examples. Any detectors reacting with input patterns are optimized to remove as many of them as possible while not removing ones similar to the training examples. This paper consists of an introduction to the natural and artificial immune systems (AIS), explanation of the AIS algorithm, results of forest and water classification using multispectral data, and discussion of sources of error and possible improvements.