Hyperspectral sensors produce large quantities of data when operating on uninhabited aerial vehicles (UAV) that can overwhelm available data links. Technical Research Associates, Inc. designed, developed, and implemented a data compression approach that is capable of reducing this data volume by a factor of 100 or more with no loss in the tactical utility of the data. This algorithm, Full Spectrum Wavelet, combines efficient coding of the spectral dimension with a wavelet transformation of the spatial dimension. The approach has been tested on a wide variety of reflection band and thermal band hyperspectral data sets. In addition to such traditional measures as the error introduced by the compression, the performance of the algorithm was evaluated using application-oriented measures such as Receiver Operating Curves (ROC) and terrain categorization maps. Comparisons between these products showed little or no degradation of performance out to compression factors of 100. The evaluation procedure provided results directly relevant to tactical users of the data
Proc. SPIE. 8048, Algorithms and Technologies for Multispectral, Hyperspectral, and Ultraspectral Imagery XVII
KEYWORDS: Principal component analysis, Independent component analysis, Video acceleration, Detection and tracking algorithms, Visualization, Image processing, Video, Data processing, Video processing, C++
With the advent of the commercial 3D video card in the mid 1990s, we have seen an order of magnitude performance
increase with each generation of new video cards. While these cards were designed primarily for visualization and video
games, it became apparent after a short while that they could be used for scientific purposes. These Graphical Processing
Units (GPUs) are rapidly being incorporated into data processing tasks usually reserved for general purpose computers.
It has been found that many image processing problems scale well to modern GPU systems. We have implemented four
popular hyperspectral processing algorithms (N-FINDR, linear unmixing, Principal Components, and the RX anomaly
detection algorithm). These algorithms show an across the board speedup of at least a factor of 10, with some special
cases showing extreme speedups of a hundred times or more.
Hyperspectral imagers tend to have lower spatial resolution than multispectral ones. This often results in a (sometimes
difficult) trade-off between spectral and spatial resolution. We have developed a technique, called CRISP, that combines
low-resolution hyperspectral data and high-resolution multispectral data to produce high quality, high-resolution
hyperspectral data. This technique shows good quantitative performance when applied to realistic applications such as
land cover estimation and anomaly detection. As a test of this technique, we have performed an experiment using
HyMap hyperspectral data and multispectral instruments over the coast waters of Oahu, Hawaii. The accuracy of the
CRISP sharpening approach when used for coastal applications such as depth mapping is assessed.
We have developed a new and innovative technique for combining a high-spatial-resolution multispectral image with a
lower-spatial-resolution hyperspectral image. The approach, called CRISP, compares the spectral information present
in the multispectral image to the spectral content in the hyperspectral image and derives a set of equations to
approximately transform the multispectral image into a synthetic hyperspectral image. This synthetic hyperspectral
image is then recombined with the original low-spatial-resolution hyperspectral image to produce a sharpened product.
The result is a product that has the spectral properties of the hyperspectral image at a spatial resolution approaching
that of the multispectral image. To test the accuracy of the CRISP method, we applied the method to synthetic data
generated from hyperspectral images acquired with an airborne sensor. These high-spatial-resolution images were used
to generate both a lower-spatial-resolution hyperspectral data set and a four-band multispectral data set. With this
method, it is possible to compare the output of the CRISP process to the 'truth data' (the original scene). In all of these
controlled tests, the CRISP product showed both good spectral and visual fidelity, with an RMS error less than one
percent when compared to the 'truth' image. We then applied the method to real world imagery collected by the
Hyperion sensor on EO-1 as part of the Hurricane Katrina support effort. In addition to multiple Hyperion data sets,
both Ikonos and QuickBird data were also acquired over the New Orleans area. Following registration of the data sets,
multiple high-spatial-resolution CRISP-generated hyperspectral data sets were created. In this paper, we present the
results of this study that shows the utility of the CRISP-sharpened products to form material classification maps at four-meter
resolution from space-based hyperspectral data. These products are compared to the equivalent products
generated from the source 30m resolution Hyperion data.
When analyzing a hyperspectral image using the linear mixture model, one makes a variety of assumptions relating to the distribution of error and the underlying mixture model. In order to test the validity of these assumptions, a simple model of hyperspectral data is examined. Generally, simple linear unmixing is performed assuming that sensor error rates are the same for each band. This assumption is violated quite easily when unmixing reflectance data. Assuming a perfect sensor, image data that perfectly obeys the linear mixture model, and perfectly known end-member spectra, the error rate for least squares linear unmixing is determinable using a simple formula. When data is transformed into reflectance, the error rates for the unmixed image increases by a significant factor due to the poor statistical normalization of the resulting data. As a means of mitigating error in unmixed imagery, two alternative unmixing methods are examined: non-negative least squares, and total least squares. Non-negative least squares can be shown to significantly outperform simple least squares, while total least squares behaves pathologically. Unmixing hyperspectral images inherently transfers error from the original hyperspectral image to the unmixed fraction plane image. Care should be taken when unmixing, so that this error is known and minimized.
Hyperspectral imagers tend to have lower spatial resolution than multispectral ones. This often results in a (sometimes difficult) trade-off between spectral and spatial resolution. One means of addressing this spatial/spectral resolution trade-off is to acquire both multispectral and hyperspectral data simultaneously, and then combine the two to produce a hyperspectral image with the high spatial resolution of the multispectral image. This process, called 'sharpening', results in a product that fuses the rich spectral content of a hyperspectral image with the high spatial content of the multispectral image. The approach we have been investigating compares the spectral information present in the multispectral image to the spectral content in the hyperspectral image and derives a set of equations to approximately transform the multispectral image into a synthetic hyperspectral image. This synthetic hyperspectral image is then recombined with the original low-spatial-resolution hyperspectral image to produce a sharpened product. We have evaluated this technique against several types of data for terrain classification and it has demonstrated good performance across all data sets. The spectra predicted by the sharpening algorithm match truth spectra in synthetic image tests, and performance with detection algorithms show little, if any, degradation of detection performance.
Multispectral sharpening of hyperspectral imagery fuses the spectral content of a hyperspectral image with the spatial and spectral content of the multispectral image. The approach we have been investigating compares the spectral information present in the multispectral image to the spectral content in the hyperspectral image and derives a set of equations to approximately transform the multispectral image into a synthetic hyperspectral image. This synthetic hyperspectral image is then recombined with the original low-resolution hyperspectral image to produce a sharpened product. We evaluate this technique against several types of data, showing good performance across with all data sets. Recent improvements in the algorithm allow target detection to be performed without loss of performance even at extreme sharpening ratios.
N-FINDR, an automated end-member detection and unmixing algorithm, was first proposed four years ago. Since then, the algorithm has been used successfully in a number of situations. The apparent success of the N-FINDR algorithm is a strong motivator for a complete review of its approach, from its assumptions to its implementation details. This paper reviews the approach used in N-FINDR, and makes a theoretical argument that the algorithm works. The algorithm can be proven to work perfectly on theoretically perfect data. Moreover, N-FINDR can be shown to have good (although imperfect) convergence properties with non-ideal data.
The University of Hawaii’s Airborne Hyperspectral Imager (AHI) consists of a long-wave infrared pushbroom hyperspectral imager and a boresighted 3-color visible high resolution CCD linescan camera. A new data system was added to the AHI in a recent upgrade of the sensor, resulting in the ability to collect data at full resolution in 256 spectral channels. This upgrade motivated the design of a new calibration procedure that removes image distortion and bad pixels from the produced imagery. The approach used is a novel method using a runtime-calculated transform. This transform describes the means of converting the distorted AHI focal plane into a corrected “virtual” AHI focal plane. The transform is formulated using several spatial-statistical assumptions as to the way information varies on the focal plane, and is based on geostatistical interpolation techniques. This transform removes the distortion present in the AHI imager and delivers high quality imagery.
The University of Hawaii AHI LWIR hyperspectral sensor has been in active use for several years. Since previous publications the sensor characteristics have evolved, and new applications have been encountered. This paper reviews the current status of the sensor and its characteristics, reviews a gas detection experiment conducted using natural sulfur dioxide emitted from a Hawaiian volcano, and test images from a hyperspectral polarization upgrade.
Unmixing hyperspectral images inherently transfers error from the original hyperspectral image to the unmixed fraction plane image. In essence by reducing the entire information content of an image down to a handful of representative spectra a significant amount of information is lost. In an image with low spectral diversity that obeys the linear mixture model (such as a simple geologic scene), this loss is negligible. However there exist inherent problems in unmixing a hyperspectral image where the actual number of spectrally distinct items in the image exceeds the resolving ability of an unmixing algorithm given sensor noise. This process is demonstrated here with a simple statistical analysis. Stepwise unmixing, where a subset of end-members is used to unmix each pixel provides a means of mitigating this error. The simplest case of stepwise unmixing, constrained unmixing, is statistically examined here. This approach provides a significant reduction in unmixed image error with a corresponding increase in goodness of fit. Some suggestions for future algorithms are presented.
Hyperspectral images can be conveniently and quickly interpreted by detecting spectral endmembers present in the image and unmixing the image in terms of those endmembers. However, spectral diversity common in hyperspectral images leads to high errors in the unmixing process by increasing the likelihood that spectral anomalies will be detected as endmembers. We have developed an algorithm to detect target-like spectral anomalies in the image which are likely to detrimentally interfere with the endmember detection process. The hyperspectral image is preprocessed by detecting target-like spectra and masking them from the subsequent endmember detection analysis. By partitioning target-like spectra from the scene, a set of spectral endmembers is detected which can be used to more accurately unmix the image. The vast majority of data in the original image can be interpreted in terms of these detected spectral endmembers. The few spectra which represent the bulk of the spectral diversity in the scene can then be interpreted individually.
In this paper we examine how the projection of hyperspectral data into smaller dimensional subspaces can effect the propagation of error. In particular, we show that the nonorthogonality of endmembers in the linear mixing model can cause small changes in band space (as, for example, from the addition of noise) to lead to relatively large changes in the estimated abundance coefficients. We also show that increasing the number of endmembers can actually lead to an increase in the amount of possible error.
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.
The University of Hawaii's Efficient Materials Mapping program aims to automatically and rapidly produce material maps from hyperspectral scenes. The program combines an end- member determination algorithm and a material identification algorithm to produce context maps in real time without user intervention. The material identification algorithm is a combination of a spectral databse and analytic code; each spectrum in the library augmented with computer readable diagnostic instructions. At present, the material library consists of over three hundred different spectra, generally geological materials from the USGS digital spectral library, however selected spectra from other libraries have been incorporated. Our method has been applied to an AVIRIS sceme taken over Kaneohe Bay, Hawaii. This scene contains large expanses of ocean, developed and undeveloped land, thus providing a good test bed for the program. The results of applying this methodolgy were verified by ground truth where possible by team equipped with hand held spectrometer. Algorithm derived archetypical en-member locations were well matched well by the material identification database, however the end-member determination itself operated sub- optimally on this scene. These results will guid progress with respect to the continued development of this program.
There has been considerable interest in the application of real-time processing techniques to the problem of hyperspectral scene analysis. Recent satellite and aircraft systems can produce data at a rate far faster than the data can be analyzed by interactive computer procedures. Automated and fast procedures for preparing the data for analyst inspection are required for even laboratory use of the large quantities of data. In addition, there are several real-time applications where the data must be processed as it is being acquired. A typical application is a computing system on-board an airplane for operator analysis of the scene as the hyperspectral sensor collects data. In this paper the possible tradeoffs fore rapid analysis are discussed, including choice of algorithm, possible dimensionality reduction, and reduced display level. A real time anomaly detection processing system based on the N- FINDR algorithm has been designed and implemented for the Night Vision Imaging Spectrometer (NVIS). The N-FINDR algorithm is a linear unmixing based algorithm that automatically finds spectral endmembers. The algorithm works by inflating a simplex inside the data, beginning with a random set of pixels. Once these endmember spectra have been found, the image cube can be unmixed using a least-squares approach into a map of fractional abundances of each endmember material in each pixel. In addition to the N-FINDR algorithm, the real-time processing system performs calibration, bad pixel removal, and display of selected fraction planes. The real-time processor is implemented in a commercial Pentium IV computer.
Hyperspectral imaging systems are assuming a greater importance for a wide variety of commercial and military systems. The reason for this increased interest is the fact that a hyperspectral sensor of a give4n spatial resolution or pixel sized will reveal information on the scene that are not obtainable by single band or multi-spectral sensors. There have been several approaches to using a single higher spatial resolution band to improve the spatial resolution fo the hyperspectral data. In this paper, a new technique for improving the spatial resolution of hyperspectral image data will be presented. This technique, called Joint End-member Determination and Unmixing, combines a high-resolution image with a lower spatial resolution hyperspectral image to produce a product that has the spectral properties of the hyperspectral image at a spatial resolution approaching that of the panchromatic image. Instead of using statistical methods to sharpen hyperspectral imagery, a physical model is used where the data present in both the hyperspectral and high-resolution data are assumed to follow linear mixing model. In this paper, the new mixture model based resolution enhancement approach will be compared to the statistical approach using data from NASA/JPL AVIRIS hyperspectral sensor.
The AHI sensor consists of a long-wave infrared pushbroom hyperspectral imager and a boresighted 3-color visible high resolution CCD linescan camera. The system used a background suppression system to achieve good noise characteristics (less than 1(mu) fl NESR). Work with AHI has shown the utility of the long-wave infrared a variety of applications. The AHI system has been used successfully in the detection of buried land mines using infrared absorption features of disturbed soil. Recently, the AHI has been used to examine the feasibility active and passive hyperspectral imaging under outdoor and laboratory conditions at three ranges. In addition, the AHI was flown over a coral reef ecosystem on the Hawaiian island of Molokai to study fresh water intrusion into coral reef ecosystems. Theoretical calculations have been done propose extensions to the AHI design in order to produce an instrument with a higher signal to noise ratio.
The MUlti Sensor Trial 2000 experiment was a multi-platform remote sensing deployment in Cairns Australia. Included in the deployment were both visible and infrared airborne hyperspectral images. The University of Hawaii's Airborne Hyperspectral Imager represented the thermal infrared portion of the data collect. The ability to discriminate various targets using the thermal infrared was explored. Consequent data processing involved separating targets from clutter using matched filters. In addition, a preliminary atmospheric correction algorithm was developed based on the ISIS algorithm used in SEBASS.
Hyperspectral data rates and volumes challenge analysis approaches that are not highly automated and efficient. Derived products from hyperspectral data, which are presented in units that are physically meaningful, have added value to analysts who are not spectral or statistical experts. The Efficient Materials Mapping project involves developing an approach that is both efficient in terms of processing time and analyzed data volume and produces outputs in terms of surface chemical or material composition. Our approach will exploit the typical redundancy inherent in hyperspectral data of natural scenes to reduce data volume. This data volume reduction is combined with an automated approach to extract chemical information from spectral data. The results will be a method to produce maps of chemical quantities that can be readily interpreted by analysts specializing in characteristics of terrains and targets rather than photons and spectra.
While reflection band hyperspectral instruments have been in use for over a decade, only recently has data from airborne thermal IR hyperspectral instruments become available. One such instrument is the Airborne Hyperspectral Imager (AHI). AHI is a pushbroom sensor developed by the University of Hawaii that spans the 8 to 11.5 micrometer spectral band with 32 spectral bands and 256 simultaneous spatial channels. While many analysis techniques used for reflection band hyperspectral processing can be applied to the thermal band, new procedures had to be developed. In particular, sensor noise and sensor non-linearity induced spectral artifacts are a greater problem than for the VNIR and SWIR. The process begins with calibration, with different calibration files being used to optimize the reduction of sensor artifacts such as shading and striping. Once the data has been calibrated to radiance units, the absorption and path radiance effects of the atmosphere can be removed, if atmospheric truth is available. Following this step, the apparent emissivity is calculated for every pixel in each band. The data is now in a form that is analogous to the apparent reflectance images developed for reflection band data. At this point spectral analysis techniques can be applied to classify the image. The procedure used here was to use an automated endmember determination algorithm, N- FINDR, to determine spectral endmembers and unmix the data cube into fractional abundances. Since some endmembers are likely to result from residual sensor and cultural artifacts, the automated endmember determination and unmixing procedure is performed interactively to optimize results. Both the fractional abundance planes and the endmember spectra themselves are then reviewed for artifacts. Selected abundance planes that correspond to real minerals can then be combined into a classification map. In this paper, AHI data collected for two applications: the detection of buried land mine application and a geological remote sensing application will be presented using similar processing steps.
The AHI sensor consists of a long-wave infrared pushbroom hyperspectral imager and a boresighted 3- color visible high resolution CCD linescan camera. The system used a background suppression system to achieve good noise characteristics (less than 1µfl NESR). Work with AHI has shown the utility of the longwave infrared a variety of applications. The AHI system has been used successfully in the detection of buried land mines using infrared absorption features of disturbed soil. Gas detection was also shown feasible, with gas absorption being clearly visible in the thermal IR. This allowed the mapping of a gas release using a matched filter. Geological mapping using AHI can be performed using the thermal band absorption features of different minerals. A large-scale geological map was obtained over a dry lake area in California using a mosaic of AHI flightlines, including mineral spectra and relative abundance maps.
The Marsokhod Field Experiment performed in Silver Lake, California, was designed to test the use of a “robotic geologist” for future unmanned Mars missions. The University of Hawaii’s Airborne Hyperspectral Imager was included in the experiment to provide geologic context information. The AHI sensor was flown over a 3 by 3 km area, imaging in the long-wave infrared. The hyperspectral data was then processed with the N-FINDR algorithm to produce estimates of constituent material spectra and mineral abundance maps. The derived mineral spectra were identified by comparison to library spectra and found generally consistent with the geology of the area.
Recently, new hyperspectral sensors have become available that provide both high spatial resolution and high spectral resolution. These characteristics combined with high signal to noise ratio allow the differentiation of vegetation or mineral types based upon the spectra of small patches of the surface. In this paper, automated endmember determination methods are applied to high spatial and spectral resolution data from two new sensors, TRWIS III and NVIS. Both of these sensors are high quality low noise pushbroom imaging spectrometers that acquire data at 5 to 6 nm resolution from 400 to 2450 nm. The data sets collected will be used for two different applications of the automated determination of endmembers: scene material classification and the detection of spectral anomalies. The NVIS hyperspectral data was collected from approximately 6000 ft above ground level over Cuprite, Nevada, resulting in a footprint of approximately two meters. The TRWIS III data was collected from 1500 meters altitude over mixed agriculture backgrounds in Ventura County, California, a largely agricultural area about 100 km from Los Angeles. After calibration and other preprocessing steps, the data in each case was processed using the N-FINDR algorithm, which extracts endmembers based upon the geometry of convex sets. Once these endmember spectra are found, the image cube can be "unmixed" into fractional abundances of each material in each pixel. The results of processing this high spatial and spectral resolution data for these two different applications will be presented.
The Airborne Hyperspectral Imager (AHI) system is a long- wave infrared imaging spectrometer originally designed to detect the presence of buried land mines. Subsequent work with AHI has shown the utility of the long-wave infrared for other applications. The AHI system has been used successfully in the detection of buried land mines using infrared absorption features of disturbed soil. Gas detection was also shown to be feasible, with gas absorption being clearly visible in the thermal IR. This allowed the mapping of a gas release using a matched filter. Geological mapping using AHI can be performed using the thermal band absorption features of different minerals. A large-scale geological map was obtained over a dry lake area in California using a mosaic of AHI flightlines, including mineral spectra and relative abundance maps.
The use of hyperspectral sensors for geological, agricultural and other remote sensing applications is continually increasing. In addition to airborne sensors, there are now at least four hyperspectral satellite sensors under development. These sensors will be producing a near continual stream of high dimensional data, leading to an obvious analysis bottleneck. Much of the planned analysis of hyperspectral image cubes requires the determination of certain basis spectra called 'end-members.' Once these spectra are found, the image cube can be 'unmixed' into fractional abundances of each material in each pixel. There exist several techniques for accomplishing the determination of these end-members, most of which require the intervention of a trained geologist. This process and the associated computations are often time- consuming. There is a need for automated techniques to allow the quick review of data collected by the sensors. Several different approaches to finding end-members in data will be reviewed, including the Pixel Purity Index, Orasis, and the Iterative Error Estimation methods. A new method, called N- FINDR, which extracts end-members based upon the geometry of convex sets, will be discussed in detail. End-member spectra and abundance maps will be compared to USGS results on AVIRIS data. Data examples from AVIRIS will also be used to compare several of the algorithms.
The analysis of hyperspectral data sets requires the determination of certain basis spectra called 'end-members.' Once these spectra are found, the image cube can be 'unmixed' into the fractional abundance of each material in each pixel. There exist several techniques for accomplishing the determination of the end-members, most of which involve the intervention of a trained geologist. Often these-end-members are assumed to be present in the image, in the form of pure, or unmixed, pixels. In this paper a method based upon the geometry of convex sets is proposed to find a unique set of purest pixels in an image. The technique is based on the fact that in N spectral dimensions, the N-volume contained by a simplex formed of the purest pixels is larger than any other volume formed from any other combination of pixels. The algorithm works by 'inflating' a simplex inside the data, beginning with a random set of pixels. For each pixel and each end-member, the end-member is replaced with the spectrum of the pixel and the volume is recalculated. If it increases, the spectrum of the new pixel replaces that end-member. This procedure is repeated until no more replacements are done. This algorithm successfully derives end-members in a synthetic data set, and appears robust with less than perfect data. Spectral end-members have been extracted for the AVIRIS Cuprite data set which closely match reference spectra, and resulting abundance maps match published mineral maps.