A novel automatic change detection procedure has been developed using multi-spectral imagery for a number of important applications including the surveillance of enemy military installations, detection of military vehicle movements, battle damage assessments and monitoring of environmental changes. The proposed approach consists of key algorithm components that includes data normalization, image registration and distance measurement. The algorithm is selected based on performance and computational considerations for near real-time implementation. Limited study on real multi-spectral data has shown that the performance of our proposed change detection approach is far superior to that of standard techniques. Example results using M7 imagery are presented to illustrate the performance improvements of this approach over the other techniques.

We used a 3D wavelet denoising method to reduce noise from multispectral imagery so that small objects may be more readily detected. Our approach exploits the correlation between bands typically present in multispectral imagery. Using our approach, the resulting image generally consists of a weighted sum of both spectral bands and spatial frequencies. We found that we could generally increase the SNR of a multispectral image more than if the spectral bands wee processed independently.

This paper describes a novel multi-spectral algorithm that combines linear and nonlinear processes to detect and identify materials with known spectral signatures. The nonlinear multi-spectral process is an anomaly detector that applies geometric whitening filters. It has demonstrated good detection and false alarm rejection performances without the knowledge of a prior target spectral information. In some instances, it achieved performance equivalent to material identification just by proper selection of spectral bands. This capability, i.e. material identification, was greatly enhanced by the incorporation of a priori target statistics.

Multivariate correlation techniques can be used to enhance target contrast in spectroscopic imagery. But in most cases the detectability of dim targets remains limited by residual background clutter. If, however, multiple-time measurements can be made, detection performance can be markedly enhanced by an integrated spectral/temporal technique that exploits the correlated nature of background spectral trajectories. We demonstrate the detection of extreme subpixel objects, such as is required by long-range remote sensing systems. We also show that the time intervals between data collections can be long. The confusing effects of natural background evolution-in temperature distribution or illumination-can be distinguished from anomalous changes. Data collected with longwave infrared point- and imaging-spectrometers have validated the concept.

Soccer is a very popular sport all over the world, Coaches and sport commentators need accurate information about soccer games, especially about the players behavior. These information can be gathered by inspectors who watch the soccer match and report manually the actions of the players involved in the principal phases of the game. Generally, these inspectors focus their attention on the few players standing near the ball and don't report about the motion of all the other players. So it seems desirable to design a system which automatically tracks all the players in real- time. That's why we propose to automatically track each player through the successive color images of the sequences acquired by a fixed color camera. Each player which is present in the image, is modelized by an active contour model or snake. When, during the soccer match, a player is hidden by another, the snakes which track these two players merge. So, it becomes impossible to track the players, except if the snakes are interactively re-initialized. Fortunately, in most cases, the two players don't belong to the same team. That is why we present an algorithm which recognizes the teams of the players by pixels representing the soccer ground which must be withdrawn before considering the players themselves. To eliminate these pixels, the color characteristics of the ground are determined interactively. In a second step, dealing with windows containing only one player of one team, the color features which yield the best discrimination between the two teams are selected. Thanks to these color features, the pixels associated to the players of the two teams form two separated clusters into a color space. In fact, there are many color representation systems and it's interesting to evaluate the features which provide the best separation between the two classes of pixels according to the players soccer suit. Finally, the classification process for image segmentation is based on the three most discriminating color features which define the coordinates of each pixel in an 'hybrid color space.' Thanks to this hybrid color representation, each pixel can be assigned to one of the two classes by a minimum distance classification.

A technique has been developed to estimate bounds on the spectra of major constituents of multispectral images. The bounds are two distinct sets of spectra, one in which the spectra are maximally independent from one another and another set in which the spectra a re minimally independent. Both sets and their corresponding estimated abundance maps satisfy feasibility constraints for both spectral elements and fractional abundances. The actual spectra will have an independence measure between the minimal and maximum sets. An approach to mapping the feasibility region for all intermediate independence measures is described. In general, for a given level of independence here is an infinity of rotation axes about which small rotations of the spectra leads to another feasible set. In our approach, the selected rotation axes is the one which takes the maximally independent basis into the minimally independent basis. The effects of noise and low levels of additional components are expected to have a larger effect on altering the spectra than the modifications due to small arbitrary rotations of feasible spectra. The technique is illustrated by application to a computer generated multispectral data array.

Hyperspectral images are frequently analyzed in terms of the linear mixing model, which assumes that observed pixel radiances are generated by linear combinations of a relatively small number of spectral constituent signatures. The constituents are generally modeled as deterministic points in color space whose locations can in principle be found by exploiting the convex geometry of the mixture simplex. This paper presents an alternative stochastic mixing model (SMM) that associates scene constituents with distinct probability distributions,the parameters of which are estimated from observed data using statistical clustering methods. By defining distributions corresponding to both constituent and mixed pixel classes, the SMM can often be used to generate physically meaningful classification maps of spectrally-heterogeneous scenes. However, the most significant application of the stochastic approach is to hyperspectral target detection based on statistical decision theory concepts. A SMM can provide accurate parametric estimates of the spectral distributions for mixed scenes, thereby improving the performance of hypothesis testing procedures that utilize replacement targets with spectral signature uncertainty. SMM principles and applications are illustrated using hyperspectral imagery collected by the LIFTIRS and HYDICE instruments.

The detection of objects of subpixel size in high-spectral- resolution imagery is a subject of much current interest. In this paper, we discuss this subject from a practical perspective and then develop a mathematical foundation for the detection of subpixel objects for which the spectral signature is either known or unknown. We capitalize on techniques from linear algebra that theoretically allow the projection of each pixel in the image onto a subspace orthogonal to the subspace defined by the interfering spectra. Using the properties of projection operators, a technique is developed for estimating an orthogonal-subspace projection operator using the singular value decomposition when interfering background spectra are unknown. We then show how a vector operator of interest. Finally, we discuss algorithm performance characterization and present some preliminary results based on Monte Carlo simulations.

Band sharpening involving multi-sensor and multi-resolution imagery is an excellent means of utilizing the complementary nature of various data types. The synergistic use of these imagery types can provide additional information that is not independently available in each source. In the case of band sharpening, a higher spatial resolution panchromatic image is fused with a lower spatial resolution multispectral image. This fusion creates a product with the spectral characteristics of the multispectral image and a spatial resolution approaching that of the panchromatic image. The goal of this project was to evaluate MSI band sharpening in four research areas. The first area explored the 'effective ground sample distance' and relative utility of multispectral imagery sharpened with panchromatic imagery. The second area examined interactions between data compression and the band sharpening process. The third area determined the effectiveness of band sharpening using a pair of high resolution sharpening bands covering different regions of the electromagnetic spectrum. The fourth area determined the effect of band sharpening on the accuracy of automated exploitation algorithms such as terrain categorization and normalized difference vegetation index.

The multi-sensor multi-resolution technique (MMT) is applied to unmix a TM/LANDSAT-5 thermal image of a typical agricultural scene using higher-resolution images in the reflective TM channels. The technique allows to retrieve the mean thermal radiance for the multispectral classes which can be recognized in the higher-resolution reflective images. As a result, the unmixed thermal image can be restored with the pixel size of 30 m and merged with the reflective images for combined data analysis. Moving-window processing, as well as low-pass correction are used to reduce the effect of mixing the thermal features which can not be recognized in the reflective images. The accuracy of the technique is tested by comparing the unmixed TM thermal image with the airborne thermal images of the same scene, which were obtained by the DAIS-7915 imaging spectrometer shortly after the LANDSAT-5 fly-by, as well as with on- ground temperature measurements. The technique can be applied for unmixing thermal images of multi-resolution sensors in the near-future spaceborne Earth observation missions.

We present a new approach for estimating land surface fluxes using remote sensing optical and thermal IR observations. We employ an artificial neural network and train it with a radiosity reflectance model. We then apply the network without retraining to extract geometrical view factors from AVIRIS imagery. We use the retrieved view factors and in- situ meteorological data to drive a surface energy balance model. Theoretical directional view factors were retrieved with an average absolute error of 15 percent. Hemispherical view factors were retrieved with a root mean square error of 6 percent. Surface net radiation estimated using the AVIRIS imagery and the surface energy balance model varied form 520 W m^-2 to 650 W m^-2 and are consistent with tower measurements. The retrieved view factors may also be used to model mixed pixel response for directional thermal IR data.

A novel feed forward neural network is used to classify hyperspectral data from the AVIRIS sensor. The network applies an alternating direction singular value decomposition technique to achieve rapid training times. Very few samples are required for training. 100 percent accurate classification is obtained using test data sets. The methodology combines this rapid training neural network together with data reduction and maximal feature separation techniques such as principal component analysis and simultaneous diagonalization of covariance matrices, for rapid and accurate classification of large hyperspectral images. The results are compared to those of standard statistical classifiers.

Background clutter and target signatures have traditionally been described by parameters derived form measurements of spatial structure and spectral ratios derived from fixed spectral bandpass images. The advent of hyperspectral imagery requires descriptions of background clutter in a mixed wavelength-spatial or Fourier-transformed (FT) spectral - spatial framework because the data stream may contain simultaneous spatial - spectral, or FT spectral - spatial clutter components. We have developed and tested analytical routines for characterizing the background clutter and target signatures observed by Fourier-transform instruments, without requiring production of a hyperspectral data cube having spectra wavelength and 2D spatial image dimensions. The Kestrel Fourier-transform hyper spectral imager, a Sagnac format interferometer produces a data stream consisting of the Fourier spectra of the background in the in-track focal plane dimension and the spatial information int he cross-track dimension. The temporal data stream thus consists of a time series of frames in FT- spectral vs. spatial dimensions. Spectral wavelength filtering and guard-band subtraction can be accomplished in FT space by binary shift and add algorithms without prior transformation of the data into a hyperspectral data cube. Spatial filtering in the cross track dimension can similarly benefit from efficient binary operations. This paper summarizes some of the target and background clutter characterization algorithms developed and their evaluation against an example atmospheric gas detection scenario.

The SWIR full spectrum imager (SFSI), an imaging spectrometer covering the short-wave IR (SWIR) from 1220 to 2420 nm, has been developed at the Canada Centre for Remote Sensing for use on an airborne platform. The sensor has ben designed to acquire simultaneously a full spectrum resolution and a full image swath at high spatial resolution. The sensor was test flown in Nevada in June 1995. Data from this mission are analyzed on the Imaging Spectrometer Data Analysis System. A look-up-table driven atmospheric correction procedure is used to retrieve surface reflectances. An image cube of a site near Virginia City is processed via spectral unmixing using reflectance spectra extracted from the image at ground sample sites and spectra of mineral samples acquired by a ground-based field spectrometer. The resulting end member abundance maps indicate that SFSI data can be analyzed successfully in this way for mineral identification.

For many years Hughes Danbury Optical Systems has ben developing algorithms for detecting trace gases in the atmosphere using hyperspectral data processing techniques. We have shown in the past that our orthogonal background suppression (OBS) algorithms are effective for measuring the column density-thermal radiance contrast product of a gas plume in the atmosphere at some distance from a passive thermal-IR emission spectrometer. The algorithm facilitates the detection of the target signal in the presence of low signal to spectral clutter ratio.Our current work shows that using the non-linear absorption features of a target gases' spectral signature, coupled with our OBS algorithm, we can separate column density-thermal radiance contrast product and obtain absolute plume column density and plume temperature. The OBS algorithms are straight forward and allow detection near theoretical random noise limits. The efficacy of our novel technique is demonstrated using simulations and field data.

The advanced airborne hyperspectral imaging system (AAHIS) is an operational, high signal-to-noise ratio, high resolution, integrated hyperspectral imaging spectrometer. The compact, lightweight and portable AAHIS system is normally flown in Piper Aztec aircraft. AAHIS collect 'push- broom' data with 385 spatial channels and 288 simultaneous spectral channels from 433 nm to 832 nm, recording at 12 bits up to 55 frames/second. Typical operation incorporates on-chip pixel binning of four pixels spectrally and two pixels spatially, increasing the signal-to-noise ratio and reducing data rate. When binned, the spectral resolution is 5.5 nm and the instantaneous field-of-view is 1 mrad, resulting in a ground sample distance of 0.5 m from 500 m altitude. The sensor is optimized for littoral region remote sensing for a variety of civilian and defense applications including ecosystem surveying and inventory, detection and monitoring of environmental pollution, infrastructure mapping, and surveillance. Since August 1994, AAHIS has acquired over 120 GB of hyperspectral image data of littoral, urban, desert and tropical scenes. System upgrades include real-time spectral image processing, integrated flight navigation and 3-axis image stabilization. A description of the sensor system, its performance characteristics, and several processed images demonstrating material discrimination are presented. The remote assessment, characterization, and mapping of coral reef health and species identification and floral species at Nu'upia Ponds, are shown and compared to extensive ground truthing in and around Kaneohe Bay, Oahu, Hawaii. SETS emphasizes providing georegistered, GIS-integrated, value- added data products for customers to help them solve real- world problems.

For a dispersive hyperspectral imaging sensor, frames are continuously being generated with spatially continuous rows of differing spectral wavelengths. As the sensor advances in the direction of travel, a hyperspectral data cube can be constructed from adjacent frames. A hyperspectral sensor residing on a satellite would require either an extremely large bandwidth for the downlink or onboard data compression to transmit the majority of the data. This paper presents a compression algorithm and the implementation of the algorithm on a real-time computational architecture. The compression algorithm sues wavelet subband coding, and universal trellis code quantization. The full implementation algorithm might include differential pulse code modulation between spectral images. The computational implementation of the algorithm uses a real-time operating system and a single general-propose microprocessor upon a VME backplane. Tradeoffs between algorithm performance and computational burden are discussed. Performance of the algorithm is presented in terms root-mean-squared error and execution time. Quantitative results for the implementation of the algorithm are provided.

Images obtained from satellite and airborne multispectral collection platforms exhibit a high degree of spatial and spectral correlations that must be properly exploited in any multispectral bandwidth compression scheme. Removing the inherent spectral correlation in the data results in a significant compaction of data to be coded. Discrete approximate trigonometric expansions have previously been proposed for exploiting spatial correlation in 1D signals and images for the purpose of coding.In this paper, we apply the approximate trigonometric expansions to multispectral data, and explore their capability of spectral decorrelation across bands. We show that the compression algorithms employing approximate trigonometric expansions to multispectral imagery provide fast implementation and some how better spectral decorrelation efficiency than discrete cosine transform. For comparison purposes, the results are compared with the techniques employing the discrete cosine transform. Computer simulation results are presented.

A system is presented for compression of hyperspectral imagery. Specifically, DPCM is used for spectral decorrelation, while an adaptive 2D discrete cosine transform coding scheme is used for spatial decorrelation. Trellis coded quantization is used to encode the transform coefficients. Side information and rate allocation strategies are discussed. Entropy-constrained codebooks are designed using a modified version of the generalized Lloyd algorithm. This entropy constrained system achieves a compression ratio of greater than 70:1 with an average PSNR of the coded hyperspectral sequence approaching 41 dB.

A method for the combined correction of atmospheric and topographic effects has been developed. It accounts for horizontally varying atmospheric conditions and also includes the height dependence of the atmospheric radiance and transmittance functions to simulate the simplified properties of a 3D atmosphere. A digital elevation model is used to obtain information about surface elevation, slope, and orientation. Based on the Lambertian assumption the surface reflectance in rugged terrain is calculated. The method is restricted to high spatial resolution satellite sensors like Landsat TM and SPOT HRV, since some simplifying assumptions are being made to reduce the required image processing time. The algorithm can be applied to multispectral and panchromatic imagery.

In the past various authors pointed out, that the value of imaging spectrometer data is closely related to the accuracy with which the data are calibrated to represent physical parameters. the AVIRIS team at JPL gave good examples on how the calibration can be performed in the laboratory and how its accuracy can be evaluated independently by means of an in-flight calibration/validation experiment. The first part of this paper presents the laboratory instrumentation and measurements that were brought into place at the German Aerospace Research Establishment (DLR) to calibrate the DAIS 7915 sensor. Some estimates of the accuracy of these measurements are given to allow the derivation of an overall precision of the laboratory calibration. It is the purpose of an in-flight calibration and validation campaign to check the validity of the laboratory calibration for data acquired under in-flight conditions. In a joint experiment of DLR and the Remote Sensing Laboratories of the University of Zurich the DAIS instrument flew a standard test site in the center of Switzerland in summer 1996. In parallel to this overflight a number of ground reference measurements are acquired. The influence of the atmosphere is accounted for using the MODTRAN radiative transfer code. Sample spectra for different in-flight calibration targets are displayed.

This paper presents the results of ongoing research aimed at reducing the size of a hyperspectral data set without significant loss in information prior to classifying a scene. We are using an atmospherically corrected 614 by 397 pixel subset of a 1994 AVIRIS image of the Lunar Crater Volcanic Field, where a great diversity of cover types can be found. An artificial neural network (ANN) has already ben sued to distinguish over twenty different surface units some of which exhibit very subtle spectral differences. This ANN classification utilized the entire 224 spectral bands obtained at each pixel. We test the hypothesis that a discrete wavelet transform of these spectral data vectors can be used to reduce their length prior to classifying the scene. This is possible because the spectra can be relatively sparse in the wavelet domain after removal of the smallest wavelet components. Since the transform is linear, spectral information is preserved and pixel classification can be based on the smaller data vectors. An ANN is being used as a sensitive tool to test this hypothesis and determine relative loss of information due to the wavelet compression. A substantial amount of ground truth from past extensive research by us and others is also being used in support of our analysis. If successful, wavelet compression could significantly increase the efficiency of a classification.

In December 1995 an inflight calibration campaign was conducted in India for the LISS-2 cameras onboard the IRS-1B satellite. For this purpose three test sites were selected where ground reflectance measurements were performed simultaneously with overpasses of the IRS-1B and Landsat-5 satellites. Due to weather conditions, only the data of 8 December 1995 was appropriate for the evaluation of the LISS-2 calibration coefficients. Ground truth data of several reference areas in ICRISAT near Hyderabad was jointly collected by DLR, ISRO, and GFZ using two field spectrometers and a 4-band radiometer. Weather data was recorded at a local meteorological station. The ATCOR2 model, based on the MODTRAN 2 radiative transfer code, was employed to calculate the calibration coefficients for the LISS-2B sensor. The derived inflight calibration coefficients agree within 5 percent with the preflight coefficients. The offset coefficients were not evaluated since no low reflectance target was available at this time.

Helicopter-based spectral radiance measurements over several homogeneous ground targets were performed with a portable calibrated spectrometer covering the wavelength region 400- 1000 nm. The flight altitude was varied from 200 to 1000 m and radiance spectra were collected in altitude steps of 200 m. Up to 20 spectra were collected for each flight level to evaluate average and standard deviation for homogeneity. For the grass and concrete target, ground reflectance measurements were performed shortly before and after the airborne radiance spectra were acquired. Meteorological data from a local DLR weather station is used as input for the MODTRAN radiative transfer code. The measurements of airborne radiance spectra are intended to support the inflight calibration of the hyperspectral DAIS sensor. Results of measured and modeled radiance spectra are presented and discussed.