This pdf file contains the Front Matter associated with SPIE Proceedings Volume 7695, including the Title page, Copyright information, Table of Contents, Conference Committee listing, and Introduction (if any).

A new formalism has been developed for producing detection algorithms in model-based problems, in which one or more parameter values is unknown. Continuum fusion can be used to generate different flavors of algorithm for any composite hypothesis testing problem. The method is defined by a fusion logic that can be translated into a set of partial differential relations. However, some important hyperspectral models can be solved with purely geometrical means, one of which is described. The new methodology affords a new interpretation of the orthodox solution, the generalized likelihood ratio test, which is revealed as but one point in a spectrum of fusion methods.

In this paper, we propose a new sparsity-based algorithm for automatic target detection in hyperspectral images (HSI). This algorithm is based on the concept that a pixel in HSI lies in a low-dimensional subspace and thus can be represented by a sparse linear combination of the training samples. The sparse representation (a sparse vector representing the selected training samples) of the test sample can be recovered by solving an 0-norm minimization problem. With the recent development of Compressed Sensing theories, the minimization problem can be recast as a linear programming or solved efficiently by a greedy pursuit algorithm. Once the sparse vector is obtained, the class of the test sample can be directly determined by the behavior of the vector on reconstruction. In addition to the constraints on sparsity and reconstruction accuracy, we also exploit the fact that HSI are usually smooth in that neighboring pixels have a similar spectral characteristic. In our proposed algorithm, a smoothness constraint is also imposed by forcing the Laplacian of the reconstructed image to be zero in the minimization process. The proposed sparsity-based algorithm is applied to several hyperspectral images to detect targets of interest. Simulation results show that our algorithm outperforms the other HSI target detection algorithms, including the popular spectral matched filters, matched subspace detectors, and adaptive subspace detectors.

The most widespread methods of anomaly detection in hyperspectral imagery (HSI) are the RX algorithm and its variants (e.g. Subspace RX). RX is optimal for any unimodal elliptically contoured distribution (ECD), and in certain data sets, it misinterprets any deviations from this model as true anomalies. Singleton outliers are by definition anomalous, but other RX detections can arise from less severe departures from the ECD, in the form of spectral "prominences." We describe a method that mitigates such persistent false alarms by augmenting RX in a recursive process with truncated versions of the Adaptive Cosine Estimator (ACE). ACE is applied to RX exceedances that arise from prominences, bulges appearing in the whitened clutter distribution that indicate anisotropy. The ACE-augmented RX decision surface resembles a sea urchin.

Hyperspectral anomaly detection algorithms are well developed however their ability to account for illumination conditions is limited only to mild variations. We propose an approach specifically designed to handle shadows and poorly illuminated regions present in otherwise well-illuminated imagery without making any assumptions about shaded backgrounds or object signature evolution. The algorithm, Segmentation Adaptive RX (SARX), relies on panchromatic segmentation of the data into dark and bright clusters based on the illumination level. Bright cluster detection employs standard subspace RX and dark cluster detection subspace is limited by only few higher variance spectral dimensions to reflect diminished signal-to-noise ratio in shadows. Anomaly detection near the geographical border between the clusters utilizes Stochastic Mixing Model. We demonstrate experimentally superior ability of SARX to detect anomalous objects in variety of illumination conditions.

An outlier detection algorithm for hyperspectral imaging based on likelihood ratio test is presented in this article. The null hypothesis tests if a test pixel is from the conditional distribution of the pixel given the background subspace and the alternative hypothesis tests if a test pixel is from the conditional distribution of the pixel given the target subspace. Using principal components for the complementary subspaces, a practical outlier detector is developed and is compared to conventional outlier detectors using a VNIR hyperspectral imagery.

We use a nonlinear dimensionality reduction technique to improve anomaly detection in a hyperspectral imaging application. A nonlinear transformation, diffusion map, is used to map pixels from the high-dimensional spectral space to a (possibly) lower-dimensional manifold. The transformation is designed to retain a measure of distance between the selected pixels. This lower-dimensional manifold represents the background of the scene with high probability and selecting a subset of points reduces the computational overhead associated with diffusion map. The remaining pixels are mapped to the manifold by means of a Nystr¨om extension. A distance measure is computed for each new pixel and those that do not reside near the background manifold, as determined by a threshold, are identified as anomalous. We compare our results with the RX and subspace RX methods of anomaly detection.

The Landsat Data Continuity Mission (LDCM) is the follow-on mission to Landsat 7 and will be the eighth mission in the Landsat series. The mission is in development via an interagency partnership between the National Aeronautics and Space Administration (NASA) and the Department of Interior (DOI) / United States Geological Survey (USGS). The LDCM satellite will carry two earth-observing sensors, the Operational Land Imager (OLI) to collect image data for nine spectral bands in the reflective portion of the spectrum and the Thermal Infrared Sensor (TIRS) to collect coincident image data for two thermal spectral bands. The LDCM ground segment will control the satellite and will receive, process, archive, and distribute the science data collected by the OLI and TIRS instruments. The USGS Earth Resources Observation & Science Center (EROS) will distribute LDCM data products at no cost to requestors. The mission objective is to continues the Landsat program's collection, archive, and distribution of multispectral imagery affording global, synoptic, and repetitive coverage of the Earth's land surfaces at a scale where natural and human-induced changes can be detected, differentiated, characterized, and monitored over time. The LDCM launch readiness date is currently December, 2012.

A compact SWIR/MWIR infrared hyperspectral imaging polarimeter (IHIP) is currently under development at the Optical Detection Lab at the University of Arizona. The sensor uses a pair of sapphire Wollaston prisms and high order retarders to form an imaging birefringent Fourier transform spectropolarimeter. Polarimetric data are acquired through the use of channeled spectropolarimetry to modulate the spectrum with the Stokes parameter information. The two dimensional interferogram is Fourier filtered and reconstructed to recover the complete Stokes vector data across the image. The IHIP operates over a +/-5° field of view and will use a dual-scan false signature reduction technique to suppress polarimetric aliasing artifacts. We present current instrument development progress, initial laboratory results, and our plan for future work.

An airborne system for hyperspectral target detection is described. The main sensor is a HySpex pushbroom hyperspectral imager for the visible and near-infrared spectral range with 1600 pixels across track, supplemented by a panchromatic line imager. An optional third sensor can be added, either a SWIR hyperspectral camera or a thermal camera. In real time, the system performs radiometric calibration and georeferencing of the images, followed by image processing for target detection and visualization. The current version of the system implements only spectral anomaly detection, based on normal mixture models. Image processing runs on a PC with a multicore Intel processor and an Nvidia graphics processing unit (GPU). The processing runs in a software framework optimized for large sustained data rates. The platform is a Cessna 172 aircraft based close to FFI, modified with a camera port in the floor.

Remote sensing by infrared spectroscopy allows identification and quantification of atmospheric gases as well as airborne pollutants. An application of the method that has gained increased interest in recent years is remote sensing of hazardous gases. If hazardous compounds are released into the atmosphere, for example in the case of a chemical accident, emergency response forces require information about the released compounds immediately in order to take appropriate measures to protect workers, residents, and the environment. A hyperspectral sensor allows identification and visualisation of hazardous clouds in the atmosphere from long distances. The image of a cloud allows an assessment of the dimensions and the dispersion of a cloud. In addition, the source of a cloud may be located. A hyperspectral sensor based on an imaging Fourier-transform spectrometer with a focal plane array detector is currently being developed for this application. Compared to conventional spectrometers, hyperspectral systems allow the use of spatial information in addition to spectral information. In addition to the application of remote sensing of hazardous gases, the system may be applied in other fields of research such as the detection of liquids and atmospheric measurements. In this work, the HYGAS system and first results of measurements are presented.

In the past we have presented a framework for deriving a set of intrinsic manifold coordinates that directly parameterize high-dimensional data, such as that found in hyperspectral imagery^1234567.¹⁰ In these previous works, we have described the potential utility of these representations for such diverse problems as land-cover mapping and in-water retrievals such as bathymetry.¹⁰ Because the manifold coordinates are intrinsic, they offer the potential for significant compression of the data, and are furthermore very useful for displaying data structure that can not be seen by linear image processing representations when the data is inherently nonlinear. This is especially true, for example, when the data are known to contain strong nonlinearities, such as in the reflectance data obtained from hyperspectral imaging sensors over the water, where the medium itself is attenuating²³⁵.⁷ These representations are also potentially useful in such applications as anomaly finding².³ A number of other researchers have looked at different aspects of the manifold coordinate representations such as the best way to exploit these representations through the backend classifier,¹⁵ while others have examined alternative manifold coordinate models.¹⁴ In this paper, we provide an overview of our scalable algorithm for deriving manifold coordinate representations of high-dimensional data such as hyperspectral imagery, describe some of our recent work to improve the local estimation of spectral neighborhood size, and demonstrate the benefits for problems such as anomaly finding.

Working with the Naval Research Laboratory, Celestech has implemented advanced non-linear hyperspectral image (HSI) processing algorithms optimized for Graphics Processing Units (GPU). These algorithms have demonstrated performance improvements of nearly 2 orders of magnitude over optimal CPU-based implementations. The paper briefly covers the architecture of the NIVIDIA GPU to provide a basis for discussing GPU optimization challenges and strategies. The paper then covers optimization approaches employed to extract performance from the GPU implementation of Dr. Bachmann's algorithms including memory utilization and process thread optimization considerations. The paper goes on to discuss strategies for deploying GPU-enabled servers into enterprise service oriented architectures. Also discussed are Celestech's on-going work in the area of middleware frameworks to provide an optimized multi-GPU utilization and scheduling approach that supports both multiple GPUs in a single computer as well as across multiple computers. This paper is a complementary work to the paper submitted by Dr. Charles Bachmann entitled "A Scalable Approach to Modeling Nonlinear Structure in Hyperspectral Imagery and Other High-Dimensional Data Using Manifold Coordinate Representations". Dr. Bachmann's paper covers the algorithmic and theoretical basis for the HSI processing approach.

Dimensionality reduction is a key step in hyperspectral image processing. Recent investigations are looking into nonlinear manifold learning for dimensionality reduction in hyperspectral imagery. Nonlinear manifold learning methods such as Isomap, and Local Linear Embedding are use to recover the low dimensional representation of an unknown nonlinear manifold for high dimensional data where it is important to retain the neighborhood structure of the manifold. Although these algorithms use different philosophies for recovering the nonlinear manifold, they all incorporate neighborhood information from each data point to construct a weighted graph having the data points as vertices. Thus the performance of these methods highly depends on how well these neighborhoods are selected, since all subsequent steps rely on it. The k-NN algorithm is the most widely used technique for neighbor selection in manifold learning. However, it can result in a disconnected graph and it does not fully exploit spatial neighborhood information from the image in selecting points to form the neighborhoods. In this paper, recently proposed methods for constructing the weighted graph in manifold learning are studied: k-VC, k-EC and k-MST, which have the advantage of creating connected graphs and have performed well in artificial data sets. Spatial information of the hyperspectral images is included in the manifold learning process by using spatial coherence. Experiments are conducted with artificial data and hyperspectral images. For hyperspectral images, classification accuracy was used as an indirect measure to understand how the low dimensional embedding of the data reduces dimensionality while still maintaining good discrimination performance.

We develop a method for the recognition of textures with three-dimensional structure in hyperspectral images. Properties of a texture are captured by a feature vector that is generated by a bank of spectral/spatial Gabor filters. Variation in the illumination and atmospheric conditions is modeled using a subspace of the feature vectors. Since a large bank of filters is used, we develop methods for reducing the dimension of the feature vector that is used to represent a texture. The goal of the dimension-reduction process is to optimize the discriminability of a set of textures. We demonstrate the utility of the approach using experiments with hyperspectral textures of three-dimensional objects that are generated by DIRSIG over a range of conditions.

Remote sensing of shallow submerged marine ecosystems presents a challenging environment for information extraction algorithms, where physically based solutions commonly require complex, computationally intensive algorithms. The inherent variations in water depth, water properties, and surface waves all impact the measured remote sensing signal, and the strong absorption of light in water also limits the effective range of wavelengths available for analysis. An algorithm has been developed to address this multifaceted problem. The algorithm uses a two-stage inverse semianalytical optimization model and spectral unmixing scheme to derive water column properties, water depth and habitat composition from imaging spectroscopy data. In addition to testing and validation studies, work on this algorithm has included improving its efficiency using the computing power of graphical processor units (GPUs). This improvement provides accelerated execution of the algorithm, and by leveraging more robust optimization routines, also facilitates increased accuracy in algorithm output. Initial results from implementing the algorithm on a single GPU using a conservative optimization strategy indicate substantial improvement in performance can be achieved using this technology. We present an overview of the algorithm, provide example output, discuss the GPU parallelization approach, and illustrate the performance achievements that have been obtained using GPU technology.

Canopy cover is a significant factor in assessing the performance of target detection algorithms in forested environments. This is true of electro-optical (EO), radar frequency (RF), light detection and ranging (LIDAR), multi/hyperspectral (MSI/HSI), and other remote sensing methods. This research compares traditional ground based methods of estimating canopy closure with estimates of canopy cover via spectral detection methods applied to VNIR/SWIR hyperspectral imagery. This paper uses canopy cover and canopy closure as defined by Jennings, et al. [1]. In the Summer of 2009, a pushbroom VNIR/SWIR hyperspectral sensor collected data over a forested region of the Naval Surface Warfare Center, Dahlgren Division, Virginia. This forested region can be best described as single canopy cover with multiple tree species. Hyperspectral imagery was collected over multiple days and at multiple altitudes in August and September, 2009. On the ground, densiometer measurements and hemispherical photography were used to estimate canopy closure at 10 meter intervals across a 2500 m² grid. Several spectral detection methods including vegetation indices, matched filtering, linear un-mixing, and distance measures, are used to calculate canopy coverage at varying ground sample distances and across multiple days. These multiple estimates are compared to the ground based measurements of canopy closure. Results indicate that estimates of canopy coverage via VNIR/SWIR hyperspectral imagery compare well to the ground based canopy closure estimates for this single canopy region. This would lead to the conclusion that it is possible to use airborne VNIR/SWIR hyperspectral alone to provide an accurate estimate of canopy cover.

This paper introduces a simple approach for object tracking using hyperspectral (HS) spectral features. The approach addresses the object tracking problem using a small object sample size. For a particular application, the key challenges are: (i) Offline training cannot be utilized; (ii) motor vehicles of interest (targets) have a small sample size (e.g., less than 9); and (iii) kinematic states of targets cannot be used for tracking, since stationary targets are also of interest. Using HS imagery, this paper introduces a method that exploits the mean and median averages spectra to estimate higher moments of the underlying (and unknown) probability distribution function of spectra; in particular, skew tendency and sign. Tracking HS targets is then possible using this algorithm to test a sequence of HS imagery, given that target spectra are initially cued by the user. The approach was implemented into a commercially off the shelf workstation, featuring the IBM Cell Processor and GA-180 Add in Board. Preliminary results are promising using a challenging HS data cube.

The hyperspectral image cube can be modeled as a three dimensional array. Tensors and the tools of multilinear algebra provide a natural framework to deal with this type of mathematical object. Singular value decomposition (SVD) and its variants have been used by the HSI community for denoising of hyperspectral imagery. Denoising of HSI using SVD is achieved by finding a low rank approximation of a matrix representation of the hyperspectral image cube. This paper investigates similar concepts in hyperspectral denoising by using a low multilinear rank approximation the given HSI tensor representation. The Best Multilinear Rank Approximation (BMRA) of a given tensor A is to find a lower multilinear rank tensor B that is as close as possible to A in the Frobenius norm. Different numerical methods to compute the BMRA using Alternating Least Square (ALS) method and Newton's Methods over product of Grassmann manifolds are presented. The effect of the multilinear rank, the numerical method used to compute the BMRA, and different parameter choices in those methods are studied. Results show that comparable results are achievable with both ALS and Newton type methods. Also, classification results using the filtered tensor are better than those obtained either with denoising using SVD or MNF.

The Spectral and Polarimetric Imagery Collection Experiment (SPICE) is a collaborative effort between the US Army ARDEC and ARL for the collection of mid-wave and long-wave infrared imagery using hyperspectral, polarimetric, and broadband sensors. The objective of the program is to collect a comprehensive database of the different modalities over the course of 1 to 2 years to capture sensor performance over a wide variety of adverse weather conditions, diurnal, and seasonal changes inherent to Picatinny's northern New Jersey location. Using the Precision Armament Laboratory (PAL) tower at Picatinny Arsenal, the sensors will autonomously collect the desired data around the clock at different ranges where surrogate 2S3 Self-Propelled Howitzer targets are positioned at different viewing perspectives at 549 and 1280m from the sensor location. The collected database will allow for: 1) Understand of signature variability under the different weather conditions; 2) Development of robust algorithms; 3) Development of new sensors; 4) Evaluation of hyperspectral and polarimetric technologies; and 5) Evaluation of fusing the different sensor modalities. In this paper, we will present the SPICE data collection objectives, the ongoing effort, the sensors that are currently deployed, and how this work will assist researches on the development and evaluation of sensors, algorithms, and fusion applications.

We present a new hyperspectral data set that FOI will keep publicly available. The hyperspectral data set was collected in an airborne measurement over the countryside. The spectral resolution was about 10 nm which allowed registrations in 60 spectral bands in the visual and near infrared range (390-960 nm). Objects with various signature properties were placed in three areas: the edge of a wood, an open field and a rough open terrain. Several overflights were performed over the areas. Between the overflights some of the objects were moved, representing different scenarios. Our interest is primarily in anomaly detection of man-made objects placed in nature where no such objects are expected. The objects in the trial were military and civilian vehicles, boards of different size and a camouflage net. The size of the boards range from multipixel to subpixel size. Due to wind and cloud conditions the stability and the flight height of the airplane vary between the overflights, which makes the analysis extra challenging.

A pixel represents the limit of spatial knowledge that can be represented in an image. It is represented as a single (perhaps spectral) digital count value that represents the energy propagating from a spatial portion of a scene. In any captured image, that single value is the result of many factors including the composition of scene optical properties within the projected pixel, the characteristic point spread function (or, equivalently, modulation transfer function) of the system, and the sensitivity of the detector element itself. This presentation examines the importance of sub-pixel variability in the context of generating synthetic imagery for remote sensing applications. The study was performed using the Digital Imaging and Remote Sensing Image Generation (DIRSIG) tool, an established ray-tracing based synthetic modeling system whose approach to sub-pixel computations was updated during this study. The paper examines three aspects of sub-pixel variability of interest to the remote sensing community. The first study simply looks at sampling frequency relative to structural frequency in a scene and the effects of aliasing on an image. The second considers the task of modeling a sub-pixel target whose signature would be mixed with background clutter, such as a small, hot target in a thermal image. The final study looks at capturing the inherent spectral variation in a single class of material, such as grass in hyperspectral imagery. Through each study we demonstrate in a quantitative fashion, the improved capabilities of DIRSIG's sub-pixel rendering algorithms.

The MCScene code, a high fidelity model for full optical spectrum (UV to LWIR) spectral image simulation, will be discussed and its features illustrated with sample calculations. The MCScene simulation is based on a Direct Simulation Monte Carlo approach for modeling 3D atmospheric radiative transport, as well as spatially inhomogeneous surfaces including surface BRDF effects. The model includes treatment of land and ocean surfaces, 3D terrain, 3D surface objects, and effects of finite clouds with surface shadowing. This paper will review the more recent upgrades to the model including the development of an approach for incorporating direct and scattered thermal emission predictions into the MCScene simulations. Sample calculations presented in the paper include a full optical spectrum simulation from the visible to the LWIR for a desert scene under a broken cloud field. This scene was derived from an AVIRIS visible to SWIR spectral imaging data collect over the Virgin Mountains in Nevada. The data has been extrapolated to the thermal IR. Other calculations include complex 3D clouds over urban and rural terrain.

The characterization of material reflectance properties is important in the analysis of hyperspectral and polarization imagery as well as accurate simulation of such images. This paper merges the results of empirical reflectance property (spectral pBRDF) measurements with detailed model based simulations. The empirical data are collected with a laboratory spectroradiometer as well as an RIT-developed spectro-polarimetric imaging goniometer. The modeling uses an adaptation of RIT's Digital Imaging and Remote Sensing Image Generation (DIRSIG) model to capture the radiative transfer in rough surfaces with micron-scale features. Measurements and model results for several man-made materials under various conditions are presented.

The U.S. Army Corps of Engineers (USACE) began developing airborne lidar bathymetry systems for coastal mapping applications in 1986, and fielded its first system in 1994. In the ensuing years, the Scanning Hydrographic Operational Airborne Lidar Survey research and development program led to the creation of the Joint Airborne Lidar Bathymetry Technical Center of Expertise (JALBTCX), a robust federal government partnership in airborne lidar bathymetry; the USACE National Coastal Mapping Program, a program of mapping built around airborne lidar bathymetry and complementary airborne remote sensing technologies; and a healthy commercial field of airborne lidar bathymeter manufacturers and service providers. The Coastal Zone Mapping and Imaging Lidar (CZMIL) is a new USACE sensor development effort and a partnership among the JALBTCX, Optech International, and The University of Southern Mississippi. The goal of CZMIL is to produce an integrated lidar and imagery sensor suite and software package designed for highly automated generation of physical and environmental information products for the coastal zone. CZMIL is an opportunity to revamp existing hardware and software to address the turbidity and shallow water limitations of existing systems; improve environmental applications of the data; take advantage of advances in laser, scanner, and receiver technology, and in signal processing and data fusion algorithms; while maintaining accurate depth measurement capability. The CZMIL program has been underway since 2006, resulting in a detailed design of the CZMIL software and hardware. CZMIL fabrication will be complete in 2010 and fielded in USACE operations in 2011.

CZMIL is a new airborne mapping and imaging system designed to simultaneously produce high resolution 3D images of the beach and shallow water seafloor, and to achieve benthic classification and water column characterization. It is designed to have high performance in shallow, turbid waters. The Data Acquisition System (DAS) is composed of a new bathymetric lidar integrated with a commercial imaging spectrometer and digital metric camera. The Data Processing System (DPS) employs new algorithms and software designed to automatically produce environmental image products by combining data from the three sensors within a data fusion paradigm. CZMIL is specifically designed to meet the requirements of the USACE Coastal Mapping Program, and is scheduled to enter field trials in the spring of 2011.

Integration of a bathymetric lidar and imaging spectrometer in CHARTS presented the challenge of developing new algorithms and software for combining these two types of data. To support this development, we conducted several field campaigns to collect airborne and in-situ data of the water column and seafloor. This work, sponsored by the Office of Naval Research (ONR) led to development of the Rapid Environmental Assessment (REA) processor. REA can be used to produce seafloor reflectance images from both sensors, and classification maps of the seafloor. Ultimately, REA became the prototype software for CZMIL, and the CZMIL Data Processing System (DPS) has been produced as a continuous refinement of REA. Here, we describe the datasets collected and illustrate results achieved with the REA software.

Upon the completion of the prototype CZMIL by Optech International, the University of Southern Mississippi will conduct a field validation campaign of the system in various environments to evaluate the accuracy of the sensors and the products derived from the data of those sensors, individually or combined. In addition to the IHO order 1 vertical and horizontal accuracy assessment of the CZMIL, the effectiveness of the DPS algorithms to map environmental parameters will be tested and estimated. This paper presents a preliminary campaign plan to realize the positional and thematic accuracy assessments. To support these efforts, an array of in-situ sensors will be deployed on land and at sea, prior to and during the CZMIL survey, to obtain a full environmental characterization of the survey site. This program will identify the materials and methodology for each selected environment to create a standardized field survey procedure.

The CZMIL system is a new generation airborne bathymetric and topographic remote sensing platform composed of an active lidar, passive hyperspectral imager, high resolution frame camera, navigation system, and storage media running on a linux-based Gigabit Ethernet network. The lidar is a hybrid scanned-flash system employing a 10 KHz green laser and novel circular scanner, with a large aperture receiver (0.20m) having multiple channels. A PMT-based segmented detector is used on one channel to support simultaneous topographic and bathymetric data collection, and multiple fields-of- view are measured to support bathymetric measurements. The measured laser returns are digitized at 1 GHz to produce the waveforms required for ranging measurements, and unique data compression and storage techniques are used to address the large data volume. Simulated results demonstrate CZMIL's capability to discriminate bottom and surface returns in very shallow water conditions without compromising performance in deep water. Simulated waveforms are compared with measured data from the SHOALS system and show promising expected results. The system's prototype is expected to be completed by end of 2010, and ready for initial calibration tests in the spring of 2010.

The Coastal Zone Mapping and Imaging Lidar (CZMIL) system¹ aims to expand bathymetric lidar performance in shallow water conditions, where high spatial and temporal data resolutions are required. This must be achieved without sacrificing performance in deep water. Very high frequency data collection (as high as 70KHz) utilizing a hybrid flash-point lidar approach is utilized, which places high demands on the transmitted laser beam, namely high pulse power with short duration with very high fidelity beam characteristics. To meet these requirements, we have developed a compact, high-power, short pulse length, high repetition rate laser with beam quality acceptable for use with the hybrid lidar approach employed in CZMIL. The short pulse length is achieved using a CW-pumped Q-switched oscillator, and the required power is achieved by using multiple amplifiers employing fiber coupled, CW end-pumped Nd:YVO₄ crystals prior to doubling the 1064 nm beam. The resulting output power is ~ 30W at 10 kHz with a pulse length of < 2.5 ns FWHM at 532 nm and ~ 20W of residual power at 1064 nm.

In pursuit of increased reliability and improved signal levels, we employ a scanning approach that supports the high frequency acquisition requirements of the system while utilizing a large aperture receiver. A continuously rotating circular scanner was developed using direct inductance drives to rotate a Fresnel prism, acting as synchronized transmitter-receiver optical element with a fixed incident angle to the sea surface. Circular scanning introduces challenges in evaluating the system's coverage. By its nature, a circular scanning pattern introduces non-uniformity. Unlike galvo-based scanners roll and pitch compensation is practically impossible for a large mechanism, therefore careful planning is required to ensure continuous coverage under realistic operational conditions. Selection of optimal operation parameters is discussed as well as different ways to evaluate surface coverage in different operation modes.

The Coastal Zone Mapping and Imaging Lidar (CZMIL) is a multi-sensor airborne system with dedicated data fusion software producing 3D images and maps of environmental parameters describing the beach, seafloor and water column. To reduce overall program development risk, a commercial off-the-shelf (COTS) imaging spectrometer and digital metric camera are used. These imagers are installed on the same mounting plate as the lidar so as to share navigation data from a single inertial measurement unit (IMU). In this paper, we discuss the capabilities of the passive imagers as they relate to spatial and spectral requirements of the U.S. Army Corps of Engineers (USACE) mission, and illustrate the anticipated data coverage based on the expected deployment mode.

The Coastal Zone Mapping and Imaging Lidar (CZMIL)¹ system is a new generation airborne remote sensing system. The multi-sensor system integrates a lidar system, hyperspectral camera, digital camera, and a positioning system together to seamlessly record data over topographic and bathymetric environments. The recorded data is post-processed using data fusion algorithms to generate image products. In order to amplify the amount of information contained within the data fed to the algorithms, the architecture of the lidar receiver is designed using numerous techniques to maximize the quality of the recorded data. The receiver architecture employs commercial photomultiplier tubes operated in a logarithmic mode coupled to 10bit, 1GHz analog to digital converters. This architecture achieves sufficient dynamic range to support operating the system in both terrestrial and ocean environments. The multi-channel design of the lidar system requires the digitization of 9 channels of optical return signal data. The resulting large data volume necessitated design of a novel data reduction strategy. These important aspects of the lidar receiver's design are presented. The strategies illustrate how the receiver's architecture is designed to optimize the fidelity of the recorded data.

Range measurements in CZMIL^1,2 are accomplished with signal processing techniques applied to green lidar waveforms. In the design phase of the project, we developed software to simulate waveforms for CZMIL, and have used these simulated waveforms to design ranging algorithms, and test their accuracies. Our results indicate the topographic ranging accuracy to hard targets should be on the order of 2cm. In this paper, we discuss the simulations, algorithms, and results.

The Data Processing System (DPS) of the Coastal Zone Mapping and Imaging Lidar (CZMIL) has been designed to automatically produce a number of novel environmental products through the fusion of Lidar, spectrometer, and camera data in a single software package. These new products significantly transcend use of the system as a bathymeter, and support use of CZMIL as a complete coastal and benthic mapping tool. The DPS provides a spinning globe capability for accessing data files; automated generation of combined topographic and bathymetric point clouds; a fully-integrated manual editor and data analysis tool; automated generation of orthophoto mosaics; automated generation of reflectance data cubes from the imaging spectrometer; a coupled air-ocean spectral optimization model producing images of chlorophyll and CDOM concentrations; and a fusion based capability to produce images and classifications of the shallow water seafloor. Adopting a multitasking approach, we expect to achieve computation of the point clouds, DEMs, and reflectance images at a 1:1 processing to acquisition ratio.

A significant challenge in the CZMIL program was to develop a topographic/bathymetric lidar delivering high spatial resolution 3D data in shallow, turbid waters, without sacrificing performance in deeper waters. To support analysis of the trade space inherent in the design process, we developed a waveform simulator capable of predicting CZMIL waveforms by varying parameters of the physical design and environmental properties of the seafloor and water column. Here, we describe the predicted performance of the proposed hardware and algorithms for generating seafloor point clouds in a number of simulated environments.

The University of Southern Mississippi's Center of Higher Learning has developed a Waveform Viewer, Attribute Viewer, and a 3D Editor for use in the CZMIL Point Cloud Manual Editor (CME). The Waveform Viewer displays various channels of CZMIL waveforms within the 2D/3D editor interface of CME. This module provides the user an interactive tool set consisting of a cross sectioning mechanism for the intensity time-bin relationship, waveform file output, and zooming capabilities. The Attribute Viewer provides the data analyst with information to analyze various environmental and spatial parameters that might contribute to errors in the measured points. The 3D Editor offers the benefits of capturing depth outliers; an intuitive visual connectivity with the 2D editor; and the implementation of volumetric directional slice isolation of data outliers.

We have developed a combined atmospheric-oceanographic spectral optimization solution decomposing measured airborne radiance data from the passive spectrometer into environmental parameters of interest. In this model, we hold depth measurements from the lidar as fixed constraints, thereby gaining a degree of freedom in the solution, and extending the solution into deeper waters than achieved with passive data alone. In this paper, we illustrate results of the data processing procedure and assess the accuracy of estimated IOPs (Inherent Optical Properties) parameters through comparison to in-situ measurements.

Algorithms for estimation of the concentrations of suspended particulate matter (SPM) are being developed in order to fulfill USACE design requirements for the Coastal Zone Mapping and Imaging Lidar (CZMIL). The lidar sensor will be used to characterize bottom depth as well as water column reflectance properties that can be related to SPM concentrations. Because lidar uses an artificial light source (i.e., the laser), it is considered an active system. Additionally, CZMIL will make observations of the water column with a hyperspectral imaging sensor, which acquires images of the magnitude of water leaving radiance at multiple wavelengths. Since the hyperspectral sensor relies solely on ambient light, it is considered a passive system. Current research includes the testing and validation of published algorithms for estimating SPM from passive spectral data produced by ship-borne, airborne and satellite-based sensors, as well as the development of a new active-passive data fusion algorithm. The new algorithm will combine observations from CZMIL's lidar and hyperspectral sensors. Data which are being collected in the northern Gulf of Mexico as part of an NSF-funded project will be applied to this research. These data come from ship-borne hyperspectral radiometers, as well as in situ SPM and optical observations. These observations will be used to validate the applicability of published SPM algorithms to the project site. This paper explores the relationship between SPM and two optical parameters used in algorithm development, remote sensing reflectance and the backscattering coefficient.

CZMIL will simultaneously acquire lidar and passive spectral data. These data will be fused to produce enhanced seafloor reflectance images from each sensor, and combined at a higher level to achieve seafloor classification. In the DPS software, the lidar data will first be processed to solve for depth, attenuation, and reflectance. The depth measurements will then be used to constrain the spectral optimization of the passive spectral data, and the resulting water column estimates will be used recursively to improve the estimates of seafloor reflectance from the lidar. Finally, the resulting seafloor reflectance cube will be combined with texture metrics estimated from the seafloor topography to produce classifications of the seafloor.

The Telops Hyper-Cam midwave (InSb 1.5-5.μm) imaging Fourier-transform spectrometer observed the plume from a coal-burning power plant smokestack. From a distance of 600 meters, the plume was captured on a 128×64 pixel sub-window of the focal-plane array with each pixel imaging a 19.5×19.5cm² region. Asymmetric interferograms were collected with long side and short side maximal optical path differences of 2.4cm and 0.9cm, respectively. Interferograms were recorded for each scan direction. The plume was strongly emissive across 1800-3000cm^-1, and raw spectra revealed emissions from CO₂, CO, H₂O, NO, SO₂, and HCl. A complete description of the instrument calibration and lineshape modeling is presented, including a simple and computationally efficient method of averaging spectra from forward- and reverse-scan interferograms that avoids the need to model a complex instrument lineshape. A simple radiative transfer model is developed to interpret the spectrum between 2565 ≤ ^~ν ≤ 3000cm^-1. Examination of the HCl spectrum demonstrates exceptional agreement between the data and an ideal instrument lineshape. For a pixel immediately above the stack exit, the plume temperature is estimated to be 399.6±0.6K with an SO₂ concentration of 376±10ppmv, and these values agree well with in situ measurements of 407.0±0.2K and 383±2ppmv, respectively.

Models for material reflectance properties, atmospheric effects and hyperspectral sensor properties in the VNIR/SWIR spectral region can be used to predict the detailed spectral structure of spectra to be measured in future field data collections. It is useful to have databases of predicted spectra that span a wide range of possible atmospheric conditions, given that the ambient environmental conditions for a measurement are rarely known in advance. A radiative transfer (RT) code such as MODTRAN® can be used to model the effects of atmospheric propagation in the predicted spectra as measured by a remotely deployed sensor. We are generating a database of optimized atmospheres that considers variations in solar angle, vertical-path-integrated water vapor density, and aerosol and cloud scatterer types and densities. This database provides a large set of discrete atmospheres. To improve the accuracy of the atmospheric spectral functions for a given hyperspectral image, we have developed an atmospheric perturbation model that is based on a rigorous physics-based approach to the variation of the spectral functions that are contained in the database. We will demonstrate how this new approach to atmospheric modeling improves the accuracy of models for VNIR/SWIR spectra collected with a space-based hyperspectral sensor.

Automated registration has been pursued for decades within academia, government, and commercial sectors as a fundamental enabling technology to support improved positioning, automated change detection, target recognition, and multi-source fusion. The focus of previous and current research has largely been on automated image-to-image registration tools. Comparatively little attention has been paid to automated registration of non-raster data (e.g., vector) stored within Geographical Information Systems (GIS) or other types of databases. The Vision Inspired Spatial Engine (VISE) is an innovative approach to automated registration. Rather than focusing on automated registration of a specific data source such as imagery, VISE uses a novel object-matching paradigm which is independent of data source. VISE assesses the fuzzy spatial similarity between two or more object patterns that can be of different shape or size by use of a top-down multiple resolution approach that simultaneously optimizes both edge and area match between vector-represented spatial features. As a by-product of the VISE pattern-matching process, object-to-object and object-to-image registration between different data sources are possible. This paper demonstrates the VISE technology applied toward the automated registration and object-level correlation of Hyperspectral (HSI), LIDAR and Electro-Optical (EO) Imagery and derived objects, and other GIS data sources.

Even with the most extensive precautions and careful planning, space based imagers will inevitably experience problems resulting in partial data corruption and possible loss. Such a loss occurs, for example, when individual image detectors are damaged. For a scanning imager this results in missing lines in the image. Images with missing lines can wreak havoc since algorithms not typically designed to handle missing pixels. Currently the metadata stores the locations of missing data, and naive spatial interpolation is used to fill it in. Naive interpolation methods can create image artifacts and even statistically or physically implausible image values. We present a general method, which uses non-linear statistical regression to estimate the values of the missing data in a principled manner. A statistically based estimate is desirable because it will preserve the statistical structure of the uncorrupted data and avoid the artifacts of naive interpolation. It also means that the restored images are suitable as input for higher-level statistical products. Previous methods replaced the missing values with those of a single closely related band, by applying a function or lookup table. We propose to use the redundant information in multiple bands to restore the lost information. The estimator we present in this paper uses values in a neighborhood of the pixel to be estimated, and propose a value based on training data from the uncorrupted pixels. Since we use the spatial variations in other channels, we avoid the blurring inherent spatial interpolation, which have implicit smoothness priors.

Nonnegative Matrix Factorization (NMF) and its variants have recently been successfully used as dimensionality reduction techniques for identification of the materials present in hyperspectral images. In this paper, we present a new variant of NMF called Nonnegative Matrix Underapproximation (NMU): it is based on the introduction of underapproximation constraints which enables one to extract features in a recursive way, like PCA, but preserving nonnegativity. Moreover, we explain why these additional constraints make NMU particularly wellsuited to achieve a parts-based and sparse representation of the data, enabling it to recover the constitutive elements in hyperspectral data. We experimentally show the efficiency of this new strategy on hyperspectral images associated with space object material identification, and on HYDICE and related remote sensing images.

Dimensionality reduction (DR) using tensor structures in morphological scale-space decomposition (MSSD) for HSI has been investigated in order to incorporate spatial information in DR.We present results of a comprehensive investigation of two issues underlying DR in MSSD. Firstly, information contained in MSSD is reduced using HOSVD but its nonconvex formulation implicates that in some cases a large number of local solutions can be found. For all experiments, HOSVD always reach an unique global solution in the parameter region suitable to practical applications. Secondly, scale parameters in MSSD are presented in relation to connected components size and the influence of scale parameters in DR and subsequent classification is studied.

Information extraction from high spatial resolution imagery is sometimes hampered by the limited number of spectral channels available from these systems. Standard supervised classification algorithms found in commercial software packages may misclassify different features with similar spectral characteristics; leading to a high occurrence of false positives. An additional step in the information extraction process was developed incorporating the concept of object geometry. Objects are defined as a contiguous group of pixels identified as belonging to a single class in the spectral classification. Using results from the spectral classification, a supervised approach was developed using genetic programming to select and mathematically combine feature-specific shape descriptors from a larger set of shape descriptors, to form a new classifier. This investigation focused on extraction of residential housing from QuickBird and IKONOS imagery of the Mississippi Gulf Coast before and immediately after hurricane Katrina. Use of genetic programming significantly reduced false positives caused by asphalt pavement and isolated roofing material scattered throughout the image.

This study presents image complexity measures applied to hyperspectral images and their relation to the percentage of correct classification (PCC). Specifically, it studies the relationship between these metrics and the PCC for Maximum Likelihood and Angle Detection classifiers. First, many complexity measures were studied to determine if there was a relation between the measure and the PCC. Results showed a correlation of above 0.7 between complexity measures based on entropy and uncertainty and the PCC of the classifiers mentioned above. Once the relation was established, PCC estimators based on the metrics using simple and multiple regression models were designed. This design was performed using data from both synthetic and real images. The real images were from two hyperspectral sensors, the space based AISA and a portable SOC 700 hyperspectral sensor, and include scenes from the Enrique Reef in La Parguera Puerto Rico. The models were then tested with real data. Results show that confidence intervals on the PCC can be reliably obtained for real images.

Hyperspectral images have traditionally been analyzed by pixel based methods. Invariant methods that consider surface and shape geometry have not been used with these images. However, there is a need for such methods due to the spectral and spatial variability present in these images. In this paper, we develop a method for classifying these images invariant to translation and rotation. The method is based on developing shape descriptors using spherical harmonics. These orthogonal functions have been widely used as a powerful tool for 3D shape recognition and are better suited for hyperspectral images due to its inherent dimensionality. A spherical function defined on the surface of a shape extracts rotation invariant features. In this case, the hyperspectral image is converted to spherical coordinates, decomposed as a sum of its harmonics and then converted to Cartesian coordinates. A classifier is trained with spherical harmonic descriptors computed from training samples. Support vector machines and Maximum Likelihood are considered for classification. The method is tested with hyperspectral image from AISA, AVIRIS and HYDICE sensors. The results show that the descriptors are effective in improving the accuracy of classification.

Stochastic watershed is a robust method to estimate the probability density function (pdf) of contours of a multi-variate image using MonteCarlo simulations of watersheds from random markers. The aim of this paper is to propose a stochastic watershed-based algorithm for segmenting hyperspectral images using a semi-supervised approach. Starting from a training dataset consisting in a selection of representative pixel vectors of each spectral class of the image, the algorithm calculate for each class a membership probability map (MPM). Then, the MPM of class k is considered as a regionalized density function which is used to simulate the random markers for the MonteCarlo estimation of the pdf of contours of the corresponding class k. This pdf favours the spatial regions of the image spectrally close to the class k. After applying the same technique to each class, a series of pdf are obtained for a single image. Finally, the pdf's can be segmented hierarchically either separately for each class or after combination, as a single pdf function. In the results, besides the generic spatial-spectral segmentation of hyperspectral images, the interest of the approach is also illustrated for target segmentation.

The majority of pixel-level hyperspectral change detection algorithms have risen out of probabilistic models developed for the data. These algorithms typically operate in two stages. In the first stage, the illumination differences and other changes due to atmospheric and environmental conditions between the two scenes are removed. In the second stage, a hypothesis test is performed on the difference between these normalized pixels. These particular change detection methods often suffer due to local variability within the data. As an alternative to these statistical-based change detection algorithms, this paper examines the use of a parametric physical model towards change detection. For a single hyperspectral data set, the number of unknown parameters in the model is greater than the number of measurements. However, if a second data set exists and the underlying material reflectance of each pixel is assumed to remain constant between the two, one can develop a problem for which the number of measurements is greater than the number of unknowns allowing for application of standard constrained optimization methods for parameter estimation. Assuming the validity of the physical model used, any residual error remaining after obtaining the optimal parameter estimates must result from noise or a violation of the reflectance assumption made, i.e., a change in material reflectance from time-1 to time-2. Accordingly, the fit error for each pixel is an indicator of reflectance change. Additionally, the proposed framework allows for incorporating spatial information at some later point. This paper provides a preliminary look at the proposed change detection method and associated challenges.

A family of subtraction-based anomalous change detection algorithms is derived from a total least squares (TLSQ) framework. This provides an alternative to the well-known chronochrome algorithm, which is derived from ordinary least squares. In both cases, the most anomalous changes are identified with the pixels that exhibit the largest residuals with respect to the regression of the two images against each other. The family of TLSQbased anomalous change detectors is shown to be equivalent to the subspace RX formulation for straight anomaly detection, but applied to the stacked space. However, this family is not invariant to linear coordinate transforms. On the other hand, whitened TLSQ is coordinate invariant, and special cases of it are equivalent to canonical correlation analysis and optimized covariance equalization. What whitened TLSQ offers is a generalization of these algorithms with the potential for better performance.

Historically in change detection, statistically based methods have been used. However, as the spatial resolution of spectral images improves, the data no longer maintain a Gaussian distribution, and some assumptions about the data - and subsequently all algorithms based upon those statistical assumptions - fail. Here we present the Simplex Volume Estimation algorithm (SVE), which avoids these potential hindrances by taking a geometrical approach. In particular, we employ the linear mixture model to approximate the convex hull enclosing the data through identification of the simplex vertices (known as endmembers). SVE begins by processing an image and tiling it into squares. Next, SVE iterates through the tiles and for each set of pixels it identifies the number of corners (as vectors) that define the simplex of that set of data. For each tile, it then iterates through the increasing dimensionality, or number of endmembers, while every time calculating the volume of the simplex that is defined by that number of endmembers. When the volume is calculated in a dimension that is higher than that of the inherent dimensionality of the data, the volume will theoretically drop to zero. This value is indicative of the inherent dimensionality of the data as represented by the convex hull. Further, the volume of the simplex will fluctuate when a new material is introduced to the dataset, indicating a change in the image. The algorithm then analyzes the volume function associated with each tile and assigns the tile a metric value based on that function. The values of these metrics will be compared by using hyperspectral imagery collected from different platforms over experimental setups with known changes between flights. Results from these tests will be presented along with a path forward for future research.

A new method for change detection of two large area scenes based on the point density of the pixel distribution in the hyperspace is presented. This method is derived from the point density approach to hyperspectral analysis, originally developed for material discrimination based on inherent dimension estimation. In this method, two registered large area scenes are tiled for individual scoring and comparison. The point density tail length is estimated for each tile in both scenes. The difference between this value for corresponding tiles indicates whether change has likely occurred in a tile and how significant the change is relative to other changes in the image. The method does not identify changes in individual pixels, but uses a tiling approach to identify changes in small sub-regions of the image. Preliminary results of this methodology are presented for multiple images and changing scene phenomenology.

Urban areas are highly variable in remote sensing data, thus it can be difficult to detect changes over time caused by development or by movement of specific targets. This research is a first attempt at exploration of repeat flights of 20m spatial resolution imaging spectrometer data to identify and characterize sources and the nature of spectral variability in urban/rural environments. Data for two dates were atmospherically corrected using a MODTRAN-based model to independently correct each dataset for atmospheric effects. The data were geocorrected and co-registered then cropped to a common spatial coverage. Spectra for individual pixels and for ROIs were extracted from the two datasets and visual and statistical comparisons were made between spectra to assess the effects of collection parameters and atmospheric corrections. Spectral endmembers were determined for each flightline, mapped using spectral matching methods, and compared across flightlines. This approach allowed determination of relatively spectrally invariant areas versus areas with significant spectral change. The combined datasets were used to develop thematic layers and evaluate spectral variability versus changes principally due to urban development.

A hyperspectral endmember detection and spectral unmixing algorithm based on an l₁ norm factorization of the input hyperspectral data is developed and compared to a method based on l₂ norm factorization. Both algorithms, the L1-Endmembers algorithm based on the l₁ norm and the SPICE algorithm based on the l₂ norm, simultaneously and autonomously estimate endmember spectra, abundance values and the number of endmembers needed for a hyperspectral image. The l₁ norm factorization of the hyperspectral data is approximated through the use of the Huber M-estimator. Results showing the stability of the L1-Endmembers algorithm in terms of the number of endmembers estimated with noise and outliers are presented. Results indicate that the proposed algorithm is more consistent in estimating the correct number of endmembers over SPICE. However, when both algorithms determine the correct number of endmembers, SPICE results provide a better estimate of endmembers and a lower variance of endmember estimates over many runs with random initialization.

Endmember extraction in Hyperspectral Images (HSI) is a critical step for target detection and abundance estimation. In this paper, we propose a new approach to endmember extraction, which takes advantage of the sparsity property of the linear representation of HSI's spectral vector. Sparsity is measured by the l0 norm of the abundance vector. It is also well known that l1 norm well resembles l0 in boosting sparsity while keeping the minimization problem convex and tractable. By adding the l1 norm term to the objective function, we result in a constrained quadratic programming which can be solved effectively using the Linear Complementary Programming (LCP). Unlike existing methods which require expensive computations in each iteration, LCP only requires pivoting steps, which are extremely simple and efficient for the un-mixing problem, since the number of signatures in the reconstructing basis is reasonably small. Preliminary experiments of the proposed methods for both supervised and unsupervised abundance decomposition showed competitive results as compared to LS-based method like Fully Constrained Least Square (FCLS). Furthermore, combination of our unsupervised decomposition with anomaly detection makes a decent target detection algorithm as compared to methods which require prior information of target and background signatures.

A new endmember finder and spectral unmixing algorithm based on the LARS/Lasso method for linear regression is developed. The endmember finder is sequential; a single endmember is identified at first and further endmembers which depend on the previous ones are found. The process terminates once a pre-determined number of endmembers have been found, or when the modeling error has attained the noise floor. The unmixing algorithm is a straightforward procedure that expresses each pixel as a linear combination of endmembers in a physically meaningful way. This algorithm successfully unmixes simulated data, and shows promising results on real hyperspectral images as well.

The linear mixture model for hyperspectral images assumes that all the image spectra lie on a high-dimensional simplex with corners called endmembers. Given the set of endmembers, one typically calculates fractional abundances for each pixel using constrained least squares. This method likely reconstructs the spectra as combinations of most, if not all, the endmembers. We instead assume that pixels are combinations of only a few of the endmembers, yielding sparse abundance vectors. We introduce a new method, similar to Matching Pursuit (MP) from the signal processing literature, to calculate these sparse abundances. We combine this sparse demixing algorithm with dictionary learning methods to automatically calculate endmembers for a provided set of spectra. We apply our method to an AVIRIS image of Cuprite, NV, for which we compare our endmembers with spectral signatures from the USGS spectral library.

Linear mixture model (LMM) is widely used in remote sensing image analysis. In the original LMM, it is assumed that each pixel is a linear combination of all the endmember signatures. For an image scene covering a large geospatial area, the number of endmembers is quite large but only a small subset of these endmembers may actually participate in the composition of a specific pixel. If all the endmembers are used in the spectral unmixing process, the result may contain additional estimation error. The Multiple Endmember Spectral Mixture Analysis (MESMA) approach has been proposed which allows the endmember subset to be varied from pixel to pixel. We have developed approaches that can automatically search for an endmember subset for each pixel from a large number of endmembers for the entire image scene, without the limitation on the number of endmembers to be included. Due to high computational complexity, in this paper, we will develop a fast searching algorithm by considering the neighborhood relationship among pixels. We will show that the resulting endmember sets can reduce pixel reconstruction error and improve the quality of estimated abundances (i.e., all the pixels have nonnegative abundances and most of pixels have sum-to-one abundances).

N-FINDR suffers from several issues in its practical implementation. One is the search region which is usually the entire data space. Another related issue is its excessive computation. A third issue is the use of random initial conditions which causes inconsistency in final results that can not be reproducible. This paper develops two ways to speed up the N-FINDR in computation. One is to narrow down the search region for the N-FINDR to a feasible range, called region of interest (ROI) where data sphering/thresholding and the well-known pixel purity index (PPI) are used as a preprocessing to find a desire ROI. The other is to simplify the simplex volume computation where three methods are proposed for this purpose to reduce computational complexity of matrix determinant. In addition, in order to further reduce computational complexity two sequential N-FINDR algorithms are developed which implement the N-FINDR by finding one endmember after another in sequence so that the information provided by previously found endmembers can be used to reduce computational complexity. The conducted experiments demonstrate that while the proposed fast algorithms can greatly reduce computational complexity, their performance remains as good as the N-FINDR is and is not compromised by reduction of the search region to an ROI and simplified matrix determinant.

Unmixing is an ill-posed inverse problem and as such the solution computed with different unmixing algorithms depends on the underlying assumptions for the inverse problem. Ideally one would expect similar solutions for unmixing a hyperspectral image of different spatial resolutions of the same scene. In this paper, we study the results of unmixing different images of the same area at different spatial resolution using different unmixing algorithms. We also compare the estimation of the number of endmembers using the rank of a scaled correlation matrix against the positive rank estimated with the fitting error of a positive matrix factorization. The results show that algorithms that assume the pure pixels in the image given consistent results in the same scale and are limited to the number of endmembers determined from the rank of the scaled correlation matrix while algorithms that do not assume pure pixels are consistent across spatial scales and the number of endmembers is better estimated by the positive rank. One and four meter data collected with the AISA sensor over southwestern Puerto Rico is used for the study.

The clutter-matched filter (CMF) and the Adaptive Cosine Estimator (ACE) have become established metrics for detecting chemical vapor plumes from hyperspectral infrared imagery. Both metrics follow from the presumption of a linear additive signal model. However, examination of the underlying radiative transfer equation (RTE) indicates that while the use of a linear additive signal model is a reasonable approximation when considering an optically-thin plume viewed against blackbody background the RTE is in fact nonlinear. Unfortunately, presumption of a linear additive signal model can significantly degrade plume detection statistics and results in significant errors in estimated chemical vapor column density when plumes are not optically-thin or are viewed against spectrally-complex backgrounds. This paper describes a nonlinear estimation approach which integrates a parameterized signal model based on the RTE with a statistical model for the infrared background. We show results obtained by applying the nonlinear estimation approach to background-only hyperspectral imagery augmented with synthetic chemical vapor plumes and compare them with results obtained presuming a linear additive signal model. As plumes become optically-thick, nonlinear estimation yields significantly more accurate estimates of chemical vapor column density and significantly more favorable plume detection statistics than clutter-matched-filter-based and adaptive-subspace-detector-based plume characterization and detection.

An ensemble learning approach using a number of weak classifiers, each classifier conducting learning based on a random subset of spectral features (bands) of the training sample, is used to detect/identify a specific chemical plume. The support vector machine (SVM) is used as the weak classifier. The detection results of the multiple SVMs are combined to generate a final decision on a pixel's class membership. Due to the multiple learning processes conducted in the randomly selected spectral subspaces, the proposed ensemble learning can improve solution generality. This work uses a two-class approach, using samples taken from hyper-spectral image (HSI) cubes collected during a release of the test chemical. Performance results in the form of receiver operator characteristic curves, show similar performance when compared to a single SVM using the full spectrum. Initial results were obtained by training with samples taken from a single HSI cube. These results are compared to results that are more recent from training with sample data from 28 HSI cubes. Performance of algorithms trained with high concentration spectra show very strong responses when scored only on high concentration data. However, performance drops substantially when low concentration pixels are scored as well. Training with the low concentration pixels along with the high concentration pixels can improve over all solution generality and shows the strength of the ensemble approach. However, it appears that careful selection of the training data and the number of examples can have a significant impact on performance.

Recently, a SVM-based ensemble learning technique has been introduced by the authors for hyperspectral plume detection/classification. The SVM-based ensemble learning consists of a number of SVM classifiers and the decisions from these sub-classifiers are combined to generate a final ensemble decision. The SVM-based ensemble technique first randomly selects spectral feature subspaces from the input data. Each individual classifier then independently conducts its own learning within its corresponding spectral feature space. Each classifier constitutes a weak classifier. These weak classifiers are combined to make an ensemble decision. The ensemble learning technique provides better performance than the conventional single SVM in terms of error rate. Various aggregating techniques like bagging, boosting, majority voting and weighted averaging were used to combine the weak classifiers, of which majority voting was found to be most robust. Yet, the ensemble of SVMs is suboptimal. Techniques that optimally weight the individual decisions from the sub-classifiers are strongly desirable to improve ensemble learning performance. In the proposed work, a recently introduced kernel learning technique called Multiple Kernel Learning (MKL) is used to optimally weight the kernel matrices of the sub-SVM classifiers. MKL basically iteratively performs l2 optimization on the Euclidian norm of the normal vector of the separating hyperplane between the classes (background and chemical plume) defined by the weighted kernel matrix followed by gradient descent optimization on the l1 regularized weighting coefficients of the individual kernel matrices. Due to l1 regularization on the weighting coefficients, the optimized weighting coefficients become sparse. The proposed work utilizes the sparse weighting coefficients to combine decision results of the SVM-based ensemble technique. A performance comparison between the aggregating techniques - MKL and majority voting as applied to hyperspectral chemical plume detection is presented in the paper.

A novel approach to VNIR hyperspectral target identification is presented based on the Least-Angle Regression (LARS) variable selection and model building algorithm. The problem to be solved is that of accurately identifying a target's primary signature component given a sub-pixel observation. Traditional matched detectors (MF, ACE, etc.) perform well at discriminating a target from a random cluttered background, but do not perform so well at unambiguously matching an observation with its counterpart in a large spectral library containing thousands of signatures. The LARS model-building algorithm efficiently selects a parsimonious subset of a large ensemble of model terms to optimally describe a particular target observation. The LARS solution technique is a recent addition to the family of model selection algorithms that includes Stepwise Regression, Forward Selection, and Backward Elimination. LARS is particularly well-suited to this problem as it is easily modified to enforce material abundance constraints: positive coefficients that sum to unity. Other approaches generally enforce such constraints in an ad-hoc fashion or use computationally demanding nonlinear programming solution techniques. LARS enforces these constraints as an inherent property of the model while remaining as computationally efficient as traditional sequential linear least-squares solvers. We demonstrate and quantify sub-pixel material identification performance using simulated target observations tested against large signature libraries.

One of the fundamental challenges for a hyperspectral imaging surveillance system is the detection of sub-pixel objects in background clutter. The background surrounding the object, which acts as interference, provides the major obstacle to successful detection. Two additional limiting factors are the spectral variabilities of the background and the object to be detected. In this paper, we evaluate the performance of detection algorithms for sub-pixel objects using a replacement signal model, where the spectral variability is modeled by multivariate normal distributions. The detection algorithms considered are the classical matched filter, the matched filter with false alarm mitigation, the mixture tuned matched filter and the finite target matched filter. These algorithms are compared using simulated and actual hyperspectral imaging data.

We produced a one-dimensional model of the signal to noise performance of Fourier Transform hyperspectral imagers using microbolometer array detectors. Our prior work could predict the peak response of such sensors, but not a measured fall-off in SNR at short and long wavelengths. We improved the fidelity of the model by including necessary wavelength dependent terms. We then tested the model using the measured performance of an FTIR, and critical measurements of the relative wavelength responsivity and NEDT of the microbolometer camera, and the wavelength response of the beamsplitter modulation. The model agrees with measurement with 15%, and the difference is consistent with the lack of spatial nonuniformity in the model. We also present similar analysis using the interferometer and a cooled mercury cadmium telluride camera showing that high performance can be obtained if dark currents can be reduced relative to commercial broadband IR cooled cameras.

Improved understanding of midwave infrared (1-5μm) spectral emissions from detonation fireballs is needed to develop a battle space optical forensics capability. While Fourier-transform spectrometers (FTS) are an attractive tool, interferometer-based spectroscopic measurements can be corrupted when the observed scene intensity systematically varies during the measurement time. Approximating a detonation fireball as a blackbody radiator with a time-varying temperature T and modified by atmospheric attenuation τ(^~ν), double-sided interferograms from an ideal FTS were calculated and converted to measured spectra L_m(^~ν) to characterize the nature and magnitude of scene-change artifacts. T(x) decreased exponentially with optical path difference x, -x_m ≤ x ≤ x_m, at various rates relative to the Michelson mirror speed so that changing scene spectra could be simulated on 1700 ≤ ^~ν ≤ 7900cm^-1 at δ ^~ν = 3.64cm^-1 resolution (x_m = 0.25cm, Hamming apodization). The real part of L_m(^~ν), Re{_Lm(^~ν)}, is well approximated by the instantaneous spectrum at zero path difference, L(^~ν,x = 0). In regions where τ(^~ν) is highly structured, both the imaginary component Im{L_m(^~ν)} and the differences between Re{L_m(^~ν)} and L(^~ν,0) exhibit spectral features, and in general |Im{L_m(^~ν)}|>>|Re{L_m(^~ν)}-L(^~ν,0)|. In a region of highly structured absorption, 2800 ≤^~ν ≤ 3500cm^-1, a 600K decrease in temperature produced RMS values of 62 and 5μW/(cm² • sr •cm^-1) in Im{L_m(^~ν)} and Re{L_m( ^~ν)-L(^~ν,0)}, respectively, compared with an RMS value of 1924μW/ (cm² • sr • cm^-1) in Re{L_m(^~ν)}. A method based on theoretical expressions developed by Kick et al. is devised to interpret L_m(^~ν) and provide estimates of the temporal evolution T(x) when its functional formis not known a priori.

A hyperspectral image projector (HIP) is introduced that is built with liquid crystal based spatial light modulators (SLM) as opposed to micromirror arrays. The use of an SLM as a broadband intensity modulator presents several benefits to this application. With slight modifications to the SLM design, SLMs can be built for a wide range of spectral regimes, ranging from the ultraviolet (UV) to the long-wavelength infrared (LWIR). SLMs can have a large pixel pitch, significantly reducing diffraction in the mid-wavelength infrared (MWIR) and LWIR. Liquid crystal based devices offer direct analog intensity modulation, thus eliminating flicker from time sequential drive schemes. SLMs allow for an on-axis configuration, enabling a simple and compact optical layout. The design of the HIP system is broken into two parts consisting of a spectral and spatial engine. In the spectral engine a diffraction grating is used to disperse a broadband source into spectral components, where an SLM modulates the relative intensity of the components to dynamically generate complex spectra. The recombined output is fed to the spatial engine which is used to construct two-dimensional scenes. The system is used to simulate a broad range of real world environments, and will be delivered to the National Institute of Standards and Technology as an enabling tool for the development of calibration standards and performance testing techniques for multispectral and hyperspectral imagers. The focus of this paper is on a visible-band HIP system; however, related work is presented with regard to SLM use in the MWIR and LWIR.

The ROC curve is the most frequently used performance measure for detection methods and the underlying sensor configuration. Common problems are that the ROC curve does not present a single number that can be compared to other systems and that no discrimination between sensor performance and algorithm performance is done. To address the first problem, a number of measures are used in practice, like detection rate at a specific false alarm rate, or area-under-curve. For the second problem, we proposed in a previous paper¹ an information theoretic method for measuring sensor performance. We now relate the method to the ROC curve, show that it is equivalent to selecting a certain point on the ROC curve, and that this point is easily determined. Our scope is hyperspectral data, studying discrimination between single pixels.

The development and testing of algorithms for unresolved target detection in hyperspectral imagery requires the availability of empirical imagery with adequate ground truth. However, target deployment and collection of imagery can be expensive, and the resulting data often have limited distribution due to concerns of a security or propriety nature. When data are made available, it is usually with full ground truth leading to the possibility of analysts "tuning" their algorithm and reporting optimistic results. There exists an ongoing need for widely available, well ground truthed, and independent data for the community. This paper provides an overview and introduction to such a standard blind test data set. Airborne hyperspectral imagery is provided together with spectral reflectance signatures of several fabric panels and vehicles in the scene. A self-test image is accompanied by pixel locations in the image for the targets of interest for algorithm development. A blind test image has additional targets in different locations and is provided without pixel truth for independent performance assessment. Since publicizing the data set in 2008, over 150 researchers from around the world have downloaded the data for testing. Further details on the data, the project, and the results of participants are presented.

In this paper, a hybrid reflective hyperspectral target detection algorithm based on the Constrained Energy Minimization (CEM) and Kelly's detection algorithms is presented. Detection performance is evaluated using the dataset and scoring methodology of the Rochester Institute of Technology (RIT) Target Detection Blind Test project. Results show that the proposed hybrid algorithm enables the detection of subpixel targets at a false alarm rate lower than either CEM or Kelly's algorithms.

The authors have recently introduced the Generalized Linear Mixing Model (GLMM), which extends the traditional Linear Mixing Model by generalizing the concept of an endmember vector to an endmember subspace. This generalization allows us to model the spectral variability that is present in a given class. The model also naturally includes the use of 'target spaces', which have been previously developed to model the variability of at-sensor radiance for a given library spectrum due to atmospheric and illumination uncertainty. In this paper, we apply the GLMM / target space approach to detecting targets in the recently released RIT test data set. In particular, we give a brief description of the underlying model, and then present our results of applying this model to the RIT data set.

In this paper, we evaluate several popular and simple matched-filter-based target detection algorithms using the online accessible HyMap data and ground truth provided by Digital Imaging and Remote Sensing (DIRS) Laboratory at Rochester Institute of Technology, and compare with their regularized versions. Based on both self and blind tests, the performance of these widely used detectors is objectively and fairly compared. Based on our experience, some recommendations for possible system improvements are provided.

In this paper we investigate the use of anomaly detection to identify pixels to be removed prior to covariance computation. The resulting covariance matrix provides a better model of the image background and is less likely to be tainted by target spectra. In our tests, this method results in robust improvement in target detection performance for quadratic detection algorithms. Tests are conducted using imagery and targets freely available online. The imagery was acquired over Cooke City, Montana, a small town near Yellowstone Park, using the HyMap V/NIR/SWIR sensor with 126 spectral bands. There are three vehicle and four fabric targets located in the town and surrounding area.

This work was motivated by the availability of a new ground truthed hyperspectral data set, freely accessible to the scientific community for target detection algorithm testing. In our research, we are interested in physics-based approaches to target detection, i.e. those techniques aimed at modeling the radiation transfer within the atmosphere in order to account for atmospheric/viewing/illumination effects. This is a crucial aspect in target detection applications, where the available information resides in the sensor-acquired radiance image and field-measured spectral reflectances of the targets. Properly backing out the aforementioned effects allows detection to be performed in either of the two domains, i.e. radiance or reflectance. As part of our research into the use of physics-based radiative transfer modeling (RTM) for target detection with these new data, it was important to accurately analyze the available a priori information concerning data acquisition, and investigate the value of enhancing this information by making use of freely accessible meteorological and environmental data. In this work, the characterization procedure of the RTM parameters applied to these data is described, and the corresponding RTM parameters thus obtained are reported. A range of variation for some of these parameters were determined as well, in order to allow for a certain degree of variability around nominal conditions (e.g. spatial variability within the scene, non-perfect acquisition condition knowledge, etc.). Target detection results obtained by adopting the RTM parameters attained by the characterization procedure show similar performance in both the radiance and the reflectance domains.

As hyperspectral sensors and exploitation methods have evolved, the accuracy of conventional background models has become a limiting factor for high confidence and low false alarm detection of mixed pixel targets. Many common target detection algorithms, such as the Adaptive Coherence/Cosine Estimator, implicitly use a global background model that assumes the background can be modeled by a single, multivariate Gaussian random variable with additive independent and identically distributed Gaussian noise. In order to improve the accuracy of the Gaussianity assumptions, a local background model, which models the background as multiple, disjoint Gaussian clusters, has also been widely considered. This paper introduces an improved variant of the local background model, as well as a novel stochastic background model that is free from distributional assumptions and accounts for the spectral variability of the background on a pixel-by-pixel basis. The performance of the global, local, and stochastic background models is evaluated on a controlled data set and the tradeoffs associated with each method are discussed.

We address the problem of estimating atmosphere parameters (temperature, water vapour content) from data captured by an airborne thermal hyperspectral imager, and propose a method based on direct optimization. The method also involves the estimation of object parameters (temperature and emissivity) under the restriction that the emissivity is constant for all wavelengths. Certain sensor parameters can be estimated as well in the same process. The method is analyzed with respect to sensitivity to noise and number of spectral bands. Simulations with synthetic signatures are performed to validate the analysis, showing that estimation can be performed with as few as 10-20 spectral bands at moderate noise levels. More than 20 bands does not improve the estimates. The proposed method is also extended to incorporate additional knowledge, for example measurements of atmospheric parameters and sensor noise.

The in-scene atmospheric compensation (ISAC) algorithm of Young et al. (2002) [14] (and as implemented in the ENVI® software system [16] as 'Thermal Atm Correction') is commonly applied to thermal infrared multi- and hyperspectral imagery (MSI and HSI, respectively). ISAC estimates atmospheric transmissivity and upwelling radiance using only the scene data. The ISAC-derived transmissivity and upwelling radiance are compared to those derived from the emissive empirical line method (EELM), another in-scene atmospheric compensation algorithm for thermal infrared MSI and HSI data. EELM is based on the presence of calibration targets (e.g., panels, water pools) captured in the spectral image data for which the emissivity and temperature are well known at the moment of MSI/HSI data acquisition. EELM is similar in concept to the empirical line method (ELM) algorithm commonly applied to visible/near-infrared to shortwave infrared (VNIR/SWIR) spectral imagery and is implemented as a custom ENVI® plugin application. Both ISAC and EELM are in-scene methods and do not require radiative transfer modeling. ISAC and EELM have been applied to airborne longwave infrared (LWIR; ~7.5 μm to ~13.5 μm) HSI data. Captured in the imagery are calibration panels and/or water pools maintained at different temperatures facilitating the application of EELM. Overall, the atmospheric compensation parameters derived from the two methods are in close agreement: the EELM-derived ground-leaving radiance spectra generally contain fewer residual atmospheric spectral features, although ISAC sometimes produces smoother ground-leaving radiance spectra. Nonetheless, the agreement is viewed as validation of ISAC. ISAC is an effective atmospheric compensation algorithm that is readily available to the remote sensing community in the ENVI® software system. Thus studies such as the present testing and comparing ISAC to other methods are important. The ISAC and EELM algorithms are discussed as are the airborne LWIR and simulated HSI data to which they are applied. Also presented are analyses and comparisons of the retrieved transmissivity and upwelling radiance terms.

The inherent dimension of hyperspectral data is commonly estimated for the purpose of dimension reduction. However, the dimension estimate itself may be a useful measure for extracting information about hyperspectral data, including scene content, complexity, and clutter. There are many ways to estimate the inherent dimension of data, each measuring the data in a different way. This paper compares a group of dimension estimation metrics on a variety of data, both full scene and individual material regions, to determine the relationship between the different estimates and what features each method is measuring when applied to complex data.

We present an efficient method for facial recognition using hyperspectral imaging and orthogonal subspaces. Projecting the data into orthogonal subspaces has the advantage of compactness and reduction of redundancy. We focus on two approaches: Principal Component Analysis and Orthogonal Subspace Projection. Our work is separated in three stages. First, we designed an experimental setup that allowed us to create a hyperspectral image database of 17 subjects under different facial expressions and viewing angles. Second, we investigated approaches to employ spectral information for the generation of fused grayscale images. Third, we designed and tested a recognition system based on the methods described above. The experimental results show that spectral fusion leads to improvement of recognition accuracy when compared to regular imaging. The work expands on previous band extraction research and has the distinct advantage of being one of the first that combines spatial information (i.e. face characteristics) with spectral information. In addition, the techniques are general enough to accommodate differences in skin spectra.

There has been significant research on specific image sharpening algorithms, but there has been less work incorporating them into a robust and validated tool. We present the Image Sharpening Toolkit (ISTK), a collection of four validated algorithms that provide error propagation and greater spectral or spatial fidelity, dependent on the user's needs. The algorithms exploit the physics of a sensor, including knowledge of its Tasseled Cap coefficients and Spectral Response Function. In addition, the HSV color model is used to fuse two single-band images and produce a color image. We demonstrate the tool's performance on QuickBird and IKONOS imagery.

This paper presents a novel method for anomaly detection based on a cluster unmixing approach. Several algorithms for endmember extraction and unmixing have been reported in literature. Endmember extraction algorithms search for pure materials which constitute the significant structure of the environment. For abundance estimation in hyperspectral imagery, various physically motivated least squares methods are considered. In real hyperspectral data, signatures of each pure material vary with physical texture and perspective. In this work, clustering of data is performed and normal distributions - instead of constant signatures - are used to represent the endmembers. This representation allows determination of class membership by means of unmixing. Furthermore, a parameter optimization is performed. Using only endmembers in a focal window around each pixel better fits the physical model. As result of this local approach, the residual of the reconstruction indicates the magnitude of anomalies. The results obtained with the new approach is called 'Cluster Mixing' (CM). The performance of Cluster Mixing is illustrated by a comparison with other anomaly detection algorithms.

Target detection is an essential component for defense, security and medical applications of hyperspectral imagery. Structured and unstructured models are used to model variability of spectral signatures, for the design of information extraction algorithms. In structured models, spectral variability is modeled using different geometric representations. In linear approaches, the spectral signatures are assumed to be generated by the linear combination of basis vectors. The nature of the basis vectors, and its allowable linear combinations, define different structural models such as vector subspaces, polyhedral cones, and convex hulls. In this paper, we investigate the use of these models to describe background of hyperspectral images, and study the performance of target detection algorithms based on these models. We also study the effect of the model order in the performance of target detection algorithms based on these models. Results show that model order is critical to algorithm performance. Underfitting or overfitting result in poor performance. Models based on subspace are of lower order than those based on polyhedral cones or convex hulls. With good target to background contrast all models perform well.

To accurately assess the area of land cover in hill land, we integrated DEM data and remote sensing image in Lihe River Valley, China. Firstly, the DEM data was combined into decision tree to increase the accuracy of land cover classification. Secondly, a slope corrected model was built to transfer the projected area to surface area by DEM data. At last, the area of different land cover was calculated and the dynamic of land cover in Lihe River Valley were analyzed from 1998 to 2003. The results show that: the area of forestland increased more than 10% by the slope corrected model, that indicates the area correcting is very important for hill land; the accuracy of classification especially for forestland and garden plot is enhanced by integrating of DEM data. It can be greater than 85%. The indexes of land use extent were 266.2 in 1998, 273.1 in 2001, and 276.7 in 2003. The change rates of land use extent were 2.59 during 1998 to 2001 and 1.34 during 2001 to 2003.

The usefulness of orthogonal subspace projection (OSP) has been demonstrated in many applications. Automatic Target Generation Process (ATGP) was previously developed for automatic target recognition for Hyperspectral imagery by implementing a successive OSP. However, ATGP itself does not provide a stopping rule to determine how many signal sources present and need to be extracted in the image. This paper presents a new application of ATGP in determining the number of signal sources and finding these signal sources in the image at the same time. The idea is to categorize signal sources into target classes and background classes in terms of their inter-sample spectral correlation (ISSC). Two separate algorithms, unsupervised target sample generation (UTSG) and unsupervised background sample generation (UBSG) are developed for this purpose. The UTSG implements a sequence of successive OSP in the sphered hyperspectral data to determine the number of target signal sources whose ISSC are characterized by high order statistics (HOS) and find the target signal sources to at the same time. It is then followed by the UBSG which operates the ATGP on a space orthogonal to the subspace generated by the target samples to determine and find background signal sources. Both UTSG and UBSG are terminated by an effective stopping rule which can be used to estimate the virtual dimensionality (VD). Two data sets, synthetic image data and real image scenes are used for experiments. Experimental results demonstrate that the UTSG and UBSG are effective in extracting signal sources in various applications.