This PDF file contains the front matter associated with SPIE Proceedings Volume 8390, including the Title Page, Copyright information, Table of Contents, Introduction, and Conference Committee listing.

The continuum fusion (CF) methodology for producing detection algorithms is generalized to include a new class of theoretical problems not considered previously by any published methods. The current CF formalism distinguishes two types of epistemic unknowns for binary composite hypothesis (CH) testing problems. One type is associated with a target (hypothesis H₁), the other with the clutter (hypothesis H₀). Older methods, including the GLR and invariant methods, treat the two types of parameters as independent of each other. Here a new type of parameter is introduced, which is shared jointly with both hypotheses. In many common applications, this is a distinction imposed by the physical models that generate the hypothesis test. Furthermore, it is a distinction not recognized by traditional methods, but is treated naturally in the CF formalism. Examples are described where models with such shared parameters produce optimal detectors, while the older methods perform at extremely degraded levels.

This paper discusses issues related to the inherent ambiguity of the composite hypothesis testing problem, a problem that is central to the detection of target signals in cluttered backgrounds. In particular, the paper examines the recently proposed method of continuum fusion (which, because it combines an ensemble of clairvoyant detectors, might also be called clairvoyant fusion), and its relationship to other strategies for composite hypothesis testing. A specific example involving the affine subspace model adds to the confusion by illustrating irreconcilable differences between Bayesian and non-Bayesian approaches to target detection.

Anomaly detectors based on subspace models have the dimension of the clutter subspace as the parameter with a large range of values. An anomaly detector that has a different parameter with fewer values is proposed. The known pixel from a hyperspectral image is predicted with a linear transformation of the unknown variables from the clutter subspace and the coefficients of the linear transformation are unknown. The dimension of the clutter subspace can vary from one spectral component of the pixel to another. The anomaly detector is the Mahalanobis distance of the error. The experimental results show that the parameter in the anomaly detector has a significantly reduced number of possible values in comparison with the conventional anomaly detectors.

A nonlinear kernel-based classifier for hyperspectral target detection is proposed in this paper. The proposed approach relies on sparsely representing a test sample in terms of all the training samples in a high-dimensional feature space induced by a kernel function. Specifically, the feature representation of a test pixel is assumed to be compactly expressed as a sparse linear combination of few atoms from a training dictionary consisting of both background and target training samples in the same feature space. The sparse representation vector is obtained by decomposing the test pixel over the training dictionary via a kernelized greedy algorithm, which uses the kernel trick to avoid explicit evaluations of the data in the feature space. The class label is then determined by comparing the reconstruction accuracy with respect to the background and target sub-dictionaries using the recovered sparse vector. Designing the classifier in a high-dimensional feature subspace will implicitly exploit the higher-order structure (correlations) within the data which cannot be captured by a linear model. Therefore, by projecting the pixels into a kernel feature space and kernelizing the linear sparse representation model, the data separability between the background and target classes will be shown to be improved, leading to a more accurate detection performance. The effectiveness of the proposed kernel sparsity model for target detection is demonstrated by experimental results on real hyperspectral images.

Hyperspectral Image (HSI) anomaly detectors typically employ local background modeling techniques to facilitate target detection from surrounding clutter. Global background modeling has been challenging due to the multi-modal content that must be automatically modeled to enable target/background separation. We have previously developed a support vector based anomaly detector that does not impose an a priori parametric model on the data and enables multi-modal modeling of large background regions with inhomogeneous content. Effective application of this support vector approach requires the setting of a kernel parameter that controls the tightness of the model fit to the background data. Estimation of the kernel parameter has typically considered Type I / false-positive error optimization due to the availability of background samples, but this approach has not proven effective for general application since these methods only control the false alarm level, without any optimization for maximizing detection. Parameter optimization with respect to Type II / false-negative error has remained elusive due to the lack of sufficient target training exemplars. We present an approach that optimizes parameter selection based on both Type I and Type II error criteria by introducing outliers based on existing hypercube content to guide parameter estimation. The approach has been applied to hyperspectral imagery and has demonstrated automatic estimation of parameters consistent with those that were found to be optimal, thereby providing an automated method for general anomaly detection applications.

This paper talks about the problem of nding targets in shadows. It discusses, through example and empirical analysis, why shadowed targets look dierent to a sensor. A forward modeling approach is used to describe how ground materials (i.e., targets) manifest themselves through the atmosphere and appear to the sensor in the radiance domain. Changes in illumination can be obtained by processing co-registered LiDAR point cloud data to obtain solar and sky-loading scaling factors. These scaling factors are then used in the forward model to better estimate varying illuminated targets in the scene. A target detection application was applied and showed that the modied or dynamic forward model was able to detect targets in both the open and hard shadow.

We extend the data fusion pixel level to the more semantically meaningful blob level, using the mean-shift algorithm to form labeled blobs having high similarity in the feature domain, and connectivity in the spatial domain. We have also developed Bhattacharyya Distance (BD) and rule-based classifiers, and have implemented these higher-level data fusion algorithms into the CZMIL Data Processing System. Applying these new algorithms to recent SHOALS and CASI data at Plymouth Harbor, Massachusetts, we achieved improved benthic classification accuracies over those produced with either single sensor, or pixel-level fusion strategies. These results appear to validate the hypothesis that classification accuracy may be generally improved by adopting higher spatial and semantic levels of fusion.

Recent terrorist attacks have sprung a need for a large scale explosive detector. Our group has developed differential reflection spectroscopy which can detect explosive residue on surfaces such as parcel, cargo and luggage. In short, broad band ultra-violet and visible light is shone onto a material (such as a parcel) moving on a conveyor belt. Upon reflection off the surface, the light intensity is recorded with a spectrograph (spectrometer in combination with a CCD camera). This reflected light intensity is then subtracted and normalized with the next data point collected, resulting in differential reflection spectra in the 200-500 nm range. Explosives show spectral finger-prints at specific wavelengths, for example, the spectrum of 2,4,6, trinitrotoluene (TNT) shows an absorption edge at 420 nm. Additionally, we have developed an automated software which detects the characteristic features of explosives. One of the biggest challenges for the algorithm is to reach a practical limit of detection. In this study, we introduce our automatic detection software which is a combination of principal component analysis and support vector machines. Finally we present the sensitivity and selectivity response of our algorithm as a function of the amount of explosive detected on a given surface.

Remote sensing can be used to rapidly generate land use maps for assisting emergency response personnel with resource deployment decisions and impact assessments. In this study we focus on constructing accurate land cover maps to map the impacted area in the case of a nuclear material release. The proposed methodology involves integration of results from two different approaches to increase classification accuracy. The data used included RapidEye scenes over Nine Mile Point Nuclear Power Station (Oswego, NY). The first step was building a coarse-scale land cover map from freely available, high temporal resolution, MODIS data using a time-series approach. In the case of a nuclear accident, high spatial resolution commercial satellites such as RapidEye or IKONOS can acquire images of the affected area. Land use maps from the two image sources were integrated using a probability-based approach. Classification results were obtained for four land classes - forest, urban, water and vegetation - using Euclidean and Mahalanobis distances as metrics. Despite the coarse resolution of MODIS pixels, acceptable accuracies were obtained using time series features. The overall accuracies using the fusion based approach were in the neighborhood of 80%, when compared with GIS data sets from New York State. The classifications were augmented using this fused approach, with few supplementary advantages such as correction for cloud cover and independence from time of year. We concluded that this method would generate highly accurate land maps, using coarse spatial resolution time series satellite imagery and a single date, high spatial resolution, multi-spectral image.

In this paper, we consider detecting man-made objects in natural images. We segment the image into tiles; we consider a variety of statistical metrics and correlate them to the presence of man-made targets. To quantify the metric, we apply a method of implanting targets and evaluating the resulting ROC (Receiver Operating Characteristic) curves. We rank previously reported algorithms and develop new ones in this paper.

A passive remote sensing technique for accurately monitoring the combustion efficiency (CE) of petrochemical flares is greatly desired. A Phase II DOE-funded SBIR lead by Spectral Sciences, Inc. is underway to develop such a method. This paper presents an overview of the current progress of the Air Force Institute of Technology's (AFIT) contribution to this effort. A Telops Hyper-Cam Mid-wave infrared (MWIR 1800-6667cm^-1 or 1.5-5.5μm) imaging Fourier-transform spectrometer (IFTS) is used to examine a laminar calibration flame produced by a Hencken burner. Ethylene fuel (C₂H₄) was burned at four different equivalency ratios Φ=0.80, 0.91, 1.0 and 1.25. This work focuses on the qualitative spectrally-resolved visualization of a Hencken burner flame and the spatial distribution of combustion by-products. A simple radiative transfer model is then developed and fit to a single-pixel spectrum. The flame spectra were characterized by structured emissions from CO₂, H₂O and CO. For the Φ = 0.91 flame, the spectrally-estimated temperature was T = 2172±28K at a height 10mm above the burner, a favorable result compared to OH-PLIF measurements (T = 2226±112K) made on an identical flame. H₂O and CO₂ mole fractions across the flame at the same height of 10 mmwere measured to be 13.7±0.6% and 15.5±0.8%, respectively.

In this paper we consider methods for estimating forest tree stem volumes by species using images taken from light unmanned aircraft systems (UAS). Instead of using LiDAR and additional multiband imagery a color infrared camera mounted to a light UAS is used to acquire both imagery and the DSM of target area. The goal of this study is to accurately estimate tree stem volumes in three classes. The status of the ongoing work is described and an initial method for delineating and classifying treetops is presented.

The Digital Imaging and Remote Sensing Image Generation (DIRSIG) model has been developed at the Rochester Institute of Technology (RIT) for over two decades. The last major update of the model, DIRSIG 4, built on an established, first-principles, multi- and hyper-spectral scene simulation tool. It introduced a modern and flexible software architecture to support new sensor modalities and more complex and dynamic scenes. Since that time, the needs of the user community have grown and diversified in tandem with the computational capabilities of modern hardware. Faced with a desire to model more complex, multi-component systems that are beyond the original intent and capabilities of an aging software design, a new version of DIRSIG, version 5, is being introduced to the community. This paper describes the core of DIRSIG 5 that is responsible for linking the disparate sensor, scene, and environmental models together, spatially, temporally, and parametrically. The spatial relationships are governed by a planet-centric universe model encompassing a whole globe digital elevation and optical property model, the scene model(s), globally varying atmospheric models, and a space model. Temporal relationships are driven by a formal modeling and simulation architecture based on approaches used in engineering and biological sciences to model highly dynamic and interactive systems. Finally, the parametric interfaces are described by a universal data model that facilitates scripting, inter-dependent properties and user interface construction. The design of these components will be presented along with specific module implementation details. These simulation tools will be used to demonstrate some of the new capabilities and applications of DIRSIG 5.

The Digital Imaging and Remote Sensing Image Generation (DIRSIG) tool is a rst principles-based synthetic image generation model, developed at the Rochester Institute of Technology (RIT) over the past 20+ years. By calculating the sensor reaching radiance between the bandpass 0.2 to 20mm, it produces multi or hyperspectral remote sensing images. By integrating independent rst principles based sub-models, such as MODTRAN, DIRSIG generates a representation of what a sensor would see with high radiometric delity. In order to detect temporal changes in a process within the scene, currently the eort is devoted to enhance the capacity of DIRSIG by incorporating process models. The parking lot process model is interesting to many applications. Therefore, this paper builds a parking lot process model PARKVIEW based on the statistical description of the parking lot which includes parking lot occupancy, parking duration and parking spot preference. The output of PARKVIEW could then be fed into DIRSIG to enhance the scene simulation capacity of DIRSIG in terms of including temporal information of the parking lot. In order to show an accurate and ecient way of extracting the statistical description of the parking lot, an experiment is set up to record the distribution of cars in several parking lots on the RIT campus during one weekday by taking photos every ve minutes. The image data are processed to extract the parking spot status of the parking lot for each frame taken from the experiment. The parking spot status information is then described in a statistical way.

A methodology is proposed for automatically extracting primitive models of buildings in a scene from a three-dimensional point cloud derived from multi-view depth extraction techniques. By exploring the information provided by the two-dimensional images and the three-dimensional point cloud and the relationship between the two, automated methods for extraction are presented. Using the inertial measurement unit (IMU) and global positioning system (GPS) data that accompanies the aerial imagery, the geometry is derived in a world-coordinate system so the model can be used with GIS software. This work uses imagery collected by the Rochester Institute of Technology's Digital Imaging and Remote Sensing Laboratory's WASP sensor platform. The data used was collected over downtown Rochester, New York. Multiple target buildings have their primitive three-dimensional model geometry extracted using modern point-cloud processing techniques.

Phototourism is a burgeoning field that uses collections of ground-based photographs to construct a three-dimensional model of a tourist site, using computer vision techniques. These techniques capitalize on the extensive overlap generated by the various visitor-acquired images from which a three-dimensional point cloud can be generated. From there, a facetized version of the structure can be created. Remotely sensed data tends to focus on nadir or near nadir imagery while trying to minimize overlap in order to achieve the greatest ground coverage possible during a data collection. A workflow is being developed at Digital Imaging and Remote Sensing (DIRS) Group at the Rochester Institute of Technology (RIT) that utilizes these phototourism techniques, which typically use dense coverage of a small object or region, and applies them to remotely sensed imagery, which involves sparse data coverage of a large area. In addition to this, RIT has planned and executed a high-overlap image collection, using the RIT WASP system, to study the requirements needed for such three-dimensional reconstruction efforts. While the collection was extensive, the intention was to find the minimum number of images and frame overlap needed to generate quality point clouds. This paper will discuss the image data collection effort and what it means to generate and evaluate a quality point cloud for reconstruction purposes.

Over the last decade DigitalGlobe (DG) has built and launched a series of remote sensing satellites with steadily increasing capabilities: QuickBird, WorldView-1 (WV-1), and WorldView-2 (WV-2). Today, this constellation acquires over 2.5 million km² of imagery on a daily basis. This paper presents the configuration and performance capabilities of each of these satellites, with emphasis on the unique spatial and spectral capabilities of WV-2. WV-2 employs high-precision star tracker and inertial measurement units to achieve a geolocation accuracy of 5 m Circular Error, 90% confidence (CE90). The native resolution of WV-2 is 0.5 m GSD in the panchromatic band and 2 m GSD in 8 multispectral bands. Four of the multispectral bands match those of the Landsat series of satellites; four new bands enable novel and expanded applications. We are rapidly establishing and refreshing a global database of very high resolution (VHR) 8-band multispectral imagery. Control moment gyroscopes (CMGs) on both WV-1 and WV-2 improve collection capacity and provide the agility to capture multi-angle sequences in rapid succession. These capabilities result in a rich combination of image features that can be exploited to develop enhanced monitoring solutions. Algorithms for interpretation and analysis can leverage: 1) broader and more continuous spectral coverage at 2 m resolution; 2) textural and morphological information from the 0.5 m panchromatic band; 3) ancillary information from stereo and multi-angle collects, including high precision digital elevation models; 4) frequent revisits and time-series collects; and 5) the global reference image archives. We introduce the topic of creative fusion of image attributes, as this provides a unifying theme for many of the papers in this WV-2 Special Session.

The enumeration of the population remains a critical task in the management of refugee/IDP camps. Analysis of very high spatial resolution satellite data proofed to be an efficient and secure approach for the estimation of dwellings and the monitoring of the camp over time. In this paper we propose a new methodology for the automated extraction of features based on differential morphological decomposition segmentation for feature extraction and interactive training sample selection from the max-tree and min-tree structures. This feature extraction methodology is tested on a WorldView-2 scene of an IDP camp in Darfur Sudan. Special emphasis is given to the additional available bands of the WorldView-2 sensor. The results obtained show that the interactive image information tool is performing very well by tuning the feature extraction to the local conditions. The analysis of different spectral subsets shows that it is possible to obtain good results already with an RGB combination, but by increasing the number of spectral bands the detection of dwellings becomes more accurate. Best results were obtained using all eight bands of WorldView-2 satellite.

Multispectral imagery (MSI) provides information to support decision making across a growing number of private and industrial applications. Among them, land mapping, terrain classification and feature extraction rank highly in the interest of those who analyze the data to produce information, reports, and intelligence products. The 8 nominal band centers of WorldView-2 allow us to use non-traditional means of measuring the differences which exist in the features, artifacts, and surface materials in the data, and we can determine the most effective method for processing this information by exploiting the unique response values within those wavelength channels. The difference in responses across select bands can be sought using normalized difference index ratios to measure moisture content, indicate vegetation health, and distinguish natural features from man-made objects. The focus of this effort is to develop an approach to measure, identify and threshold these differences in order to establish an effective land mapping and feature extraction process germane to WorldView-2 imagery.

There are clear indications that densification of built-up areas within cities and new developments in their outskirts, in conjunction with urban population activities, are at the origin of climate changes at the local level and have a direct impact on air and water quality. Densification of the vegetation cover is often mentioned as one of the most important means to mitigate the impacts of climate changes and to improve the quality of the urban environment. Decision making on vegetation cover densification presupposes that urban planners and managers know exactly the actual situation in terms of vegetation location, types and biomass. However, in many cities, inventories of vegetation cover are usually absent. This study examines the feasibility of an automatic system for vegetation cover inventory and mapping in urban areas based on WorldView-2 imagery. The city of Laval, Canada, was chosen as the experimental site. The principal conclusions are as follows: a) conversion of digital counts to ground reflectances is a crucial step in order to fully exploit the potential of WV-2 multispectral images for mapping vegetation cover and recognizing vegetation classes; b) the combined use of NDVIs computed using the three infrared available bands and the red band provides an accurate means of differentiating vegetation cover from other land covers; and c) it is possible to separate trees from other vegetation types and to identify tree species even in dense urban areas using spectral signature characteristics and segmentation algorithms.

Nearshore depths for Waimanalo Beach, HI, are extracted from optical imagery, taken by the WorldView-2 (WV-2) satellite on 31 March 2011, by means of automated Wave Kinematics Bathymetry (WKB). Two sets of three sequential images taken at intervals of about 10 seconds are used for the analyses herein. Water depths are calculated using a computer program that registers the images, estimates the currents, and then uses the linear dispersion relationship for surface gravity waves to estimate depth. Depths are generated from close to shore out to about 20 meters depth. Comparisons with SHOALS Light Detection and Ranging (LiDAR) bathymetry values show WKB depths are accurate to about half a meter, with R² values of 90%, and are frequently in the range of 10 to 20 percent relative error for depths ranging from 2 to 16 meters.

LWIR hyperspectral imaging has a wide range of civil and military applications with its ability to sense chemical compositions at standoff ranges. Most recent implementations of this technology use spectrographs employing varying degrees of cryogenic cooling to reduce sensor self-emission that can severely limit sensitivity. We have taken an interferometric approach that promises to reduce the need for cooling while preserving high resolution. Reduced cooling has multiple benefits including faster system readiness from a power off state, lower mass, and potentially lower cost owing to lower system complexity. We coupled an uncooled Sagnac interferometer with a 256x320 mercury cadmium telluride array with an 11 micron cutoff to produce a spatial interferometric LWIR hyperspectral imaging system operating from 7.5 to 11 microns. The sensor was tested in ground-ground applications, and from a small aircraft producing spectral imagery including detection of gas emission from high vapor pressure liquids.

A prototype long wave infrared Fourier transform spectral imaging system using a wedged Fabry-Perot interferometer and a microbolometer array was designed and built. The instrument can be used at both short (cm) and long standoff ranges (infinity focus). Signal to noise ratios are in the several hundred range for 30 C targets. The sensor is compact, fitting in a volume about 12 x12 x 4 inches.

The development and demonstration of a new snapshot hyperspectral sensor is described. The system is a significant extension of the four dimensional imaging spectrometer (4DIS) concept, which resolves all four dimensions of hyperspectral imaging data (2D spatial, spectral, and temporal) in real-time. The new sensor, dubbed "4×4DIS" uses a single fiber optic reformatter that feeds into four separate, miniature visible to near-infrared (VNIR) imaging spectrometers, providing significantly better spatial resolution than previous systems. Full data cubes are captured in each frame period without scanning, i.e., "HyperVideo". The current system operates up to 30 Hz (i.e., 30 cubes/s), has 300 spectral bands from 400 to 1100 nm (~2.4 nm resolution), and a spatial resolution of 44×40 pixels. An additional 1.4 Megapixel video camera provides scene context and effectively sharpens the spatial resolution of the hyperspectral data. Essentially, the 4×4DIS provides a 2D spatially resolved grid of 44×40 = 1760 separate spectral measurements every 33 ms, which is overlaid on the detailed spatial information provided by the context camera. The system can use a wide range of off-the-shelf lenses and can either be operated so that the fields of view match, or in a "spectral fovea" mode, in which the 4×4DIS system uses narrow field of view optics, and is cued by a wider field of view context camera. Unlike other hyperspectral snapshot schemes, which require intensive computations to deconvolve the data (e.g., Computed Tomographic Imaging Spectrometer), the 4×4DIS requires only a linear remapping, enabling real-time display and analysis. The system concept has a range of applications including biomedical imaging, missile defense, infrared counter measure (IRCM) threat characterization, and ground based remote sensing.

One of the drawbacks of slit-based dispersive imaging spectrometers is the need for an inertial measurement unit and a photogrammetric software to register the line images. On the contrary, since they perform instantaneous 2D imaging, static Fourier transform imaging spectrometers allow image registration from the images themselves. We propose to merge these two spectral imagers, by cutting the slit in a mirror that reflects the light toward the static Fourier transform spectral imager. Thus, the two spectrometers can be exactly co-registered. We present the preliminary design of such a dual-band hyperspectral imager (visible and near infrared), and an evaluation of the expected geometric and radiometric performances.

Land and ocean data product generation from visible-through-shortwave-infrared multispectral and hyperspectral imagery requires atmospheric correction or compensation, that is, the removal of atmospheric absorption and scattering effects that contaminate the measured spectra. We have recently developed a prototype software system for automated, low-latency, high-accuracy atmospheric correction based on a C++-language version of the Spectral Sciences, Inc. FLAASH™ code. In this system, pre-calculated look-up tables replace on-the-fly MODTRAN® radiative transfer calculations, while the portable C++ code enables parallel processing on multicore/multiprocessor computer systems. The initial software has been installed on the Sensor Web at NASA Goddard Space Flight Center, where it is currently atmospherically correcting new data from the EO-1 Hyperion and ALI sensors. Computation time is around 10 s per data cube per processor. Further development will be conducted to implement the new atmospheric correction software on board the upcoming HyspIRI mission's Intelligent Payload Module, where it would generate data products in nearreal time for Direct Broadcast to the ground. The rapid turn-around of data products made possible by this software would benefit a broad range of applications in areas of emergency response, environmental monitoring and national defense.

In the present work, the WorldView-2 (WV2) capability for retrieving Case 2 water components is analyzed. The WV2 sensor characteristics, such as a 11-bit quantization, 8 bands in the VNIR (visible and near infrared) region and high Signal-to-Noise Ratios (SNR) make WV2 potentially suitable for a retrieval process. In the Case 2 water problem, the sensor-reaching signal due to water is very small when compared to the signal due to the atmospheric eects. Therefore, adequate atmospheric compensation becomes an important rst step to accurately retrieve water parameters. The problem becomes more dicult when using multispectral imagery as there are typically only a handful of bands suitable for performing atmospheric compensation. In this work, we test atmospheric compensation techniques for the WV2 satellite, enabling it to be used for water constituent retrieval in both deep and shallow water. A look-up-table (LUT) methodology is implemented to retrieve the water parameters chlorophyll, suspended materials, colored dissolved organic matter, bathymetry, bottom type and water clarity for a simulated case study. The in-water radiative transfer code HydroLight is used to simulate re ectance data in this study while the MODTRAN code is used to simulate atmospheric eects. The resulting modeled sensor-reaching radiance data can be sampled to a WV2 sensor model to simulate WV2 image data. This data is used to test the proposed methodology. Finally, a sensitivity analysis is performed to evaluate how sensitive the constituent retrieval process is to adequate atmospheric compensation.

In the atmospheric correction of the CASI hyperspectral image, we found that the biggest factor is the downward scattering of the direct solar beam in the exact direction to be reflected at the water surface to be detected by the sensor. The downward scattering angle was calculated using navigation data, viewing geometry, and solar ephemeris. One benefit of this approach is that it is now possible to avoid the limitations posed by the dark-pixel method. Since the scattering angle is computed using geometry only, it is completely free from the possible trouble met by the dark-pixel approach. In this paper, we illustrate the computational procedure and show examples of marine remote sensing data.

Here, we present a new prototype algorithm for the simultaneous retrieval of the atmospheric profiles (temperature, humidity, ozone and aerosol) and the surface reflectance from hyperspectral radiance measurements obtained from air/space-borne, hyperspectral imagers such as the 'Airborne Visible/Infrared Imager (AVIRIS) or Hyperion on board of the Earth Observatory 1. The new scheme, proposed here, consists of a fast radiative transfer code, based on empirical orthogonal functions (EOFs), in conjunction with a 1D-Var retrieval scheme. The inclusion of an 'exact' scattering code based on spherical harmonics, allows for an accurate treatment of Rayleigh scattering and scattering by aerosols, water droplets and ice-crystals, thus making it possible to also retrieve cloud and aerosol optical properties, although here we will concentrate on non-cloudy scenes. We successfully tested this new approach using two hyperspectral images taken by AVIRIS, a whiskbroom imaging spectrometer operated by the NASA Jet Propulsion Laboratory.

Field spectral signatures are commonly collected in conjunction with remote sensing campaigns. Unfortunately, the lack of sufficient metadata associated with campaign-specific spectral signatures often makes it difficult for others to utilize them for their own applications. The first step in improving the utility of field collected spectral signatures is achieved by establishing fully documented procedures that minimize controllable error sources. A major source of error when collecting field spectral signatures is the variability of solar illumination. By periodically monitoring a static reference panel it is possible to both characterize the variance in solar illumination during collection as well as to correct collected spectra. In addition, recent advances in instrument sensitivity greatly reduce the time required to collect high quality spectra that in turn reduces the magnitude of potential errors associated with changes in solar illumination. Since libraries of field spectral signatures are commonly used to analyze remotely sensed imagery, it is important that field collection is performed at a relevant spatial scale and with illumination and view geometry that is similar to that for the image collection. This is particularly true of vegetation since the observed spectral signature is the result of the complex interaction of multiple illumination sources (i.e. direct sunlight, sky illumination and light scattered off other elements in the scene), canopy architecture and the reflectance properties of the individual elements within the canopy. Suggested field collection approaches that minimize these sources of error are presented.

Fractional abundance maps have been produced from Hyperion hyperspectral data over Oaxaca, Mexico, by applying a new spatially adaptive spectral unmixing algorithm. The goal of this research is to produce land-use maps for aiding archaeologists studying the Zapotec civilization. However, to correlate the fractional abundance maps generated from the HSI image processing, a relationship between the known materials located in Oaxaca, Mexico, and the spectral profiles of these materials must be established. A field campaign during December 2011, (the dry season in Oaxaca) took place for the explicit task of obtaining spectral profiles of the most common materials found in the region. Ground-truth information was collected for three Oaxaca valleys (Tlacolula, Yanhuitlan, and Ycuitla). Common materials and associated regions were recorded and material samples were collected at many of these locations. Laboratory reflectance spectral profiles of the collected material samples are measured after the field campaign using a FieldSpec Pro. Wavelength ranges of the FieldSpec Pro spanned 350-2500nm matching that of the hyperspectral imagery collected from the Hyperion sensor on board the EO-1 satellite. GIS maps of the three valleys in Oaxaca, Mexico, are used to identify where these samples were collected and correspond to the laboratory measured material samples. The spectral library entries obtained correspond to bare soils, senescent agricultural vegetation, senescent natural vegetation, and terra cotta tile.

The Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) measures reflected solar radiation in the shortwave infrared and has been used to map methane (CH4) using both a radiative transfer technique [1] and a band ratio method [2]. However, these methods are best suited to water bodies with high sunglint and are not well suited for terrestrial scenes. In this study, a cluster-tuned matched filter algorithm originally developed by Funk et al. [3] for synthetic thermal infrared data was used for gas plume detection over more heterogeneous backgrounds. This approach permits mapping of CH₄, CO₂ (carbon dioxide), and N₂O (nitrous oxide) trace gas emissions in multiple AVIRIS scenes for terrestrial and marine targets. At the Coal Oil Point marine seeps offshore of Santa Barbara, CA, strong CH₄ anomalies were detected that closely resemble results obtained using the band ratio index. CO₂ anomalies were mapped for a fossil-fuel power plant, while multiple N₂O and CH₄ anomalies were present at the Hyperion wastewater treatment facility in Los Angeles, CA. Nearby, smaller CH₄ anomalies were also detected immediately downwind of hydrocarbon storage tanks and centered on a flaring stack at the Inglewood Gas Plant. Improving these detection methods might permit gas detection over large search areas, e.g. identifying fugitive CH₄ emissions from damaged natural gas pipelines or hydraulic fracturing. Further, this technique could be applied to other trace gasses with distinct absorption features and to data from planned instruments such as AVIRISng, the NEON Airborne Observation Platform (AOP), and the visible-shortwave infrared (VSWIR) sensor on the proposed HyspIRI satellite.

Identifying materials by measuring and analyzing their reflectance spectra has been an important procedure in analytical chemistry for decades. Airborne and space-based imaging spectrometers allow materials to be mapped across the landscape. With many existing airborne sensors and new satellite-borne sensors planned for the future, robust methods are needed to fully exploit the information content of hyperspectral remote sensing data. A method of identifying and mapping materials using spectral feature analyses of reflectance data in an expert-system framework called MICA (Material Identification and Characterization Algorithm) is described. MICA is a module of the PRISM (Processing Routines in IDL for Spectroscopic Measurements) software, available to the public from the U.S. Geological Survey (USGS) at http://pubs.usgs.gov/of/2011/1155/. The core concepts of MICA include continuum removal and linear regression to compare key diagnostic absorption features in reference laboratory/field spectra and the spectra being analyzed. The reference spectra, diagnostic features, and threshold constraints are defined within a user-developed MICA command file (MCF). Building on several decades of experience in mineral mapping, a broadly-applicable MCF was developed to detect a set of minerals frequently occurring on the Earth's surface and applied to map minerals in the country-wide coverage of the 2007 Afghanistan HyMap data set. MICA has also been applied to detect sub-pixel oil contamination in marshes impacted by the Deepwater Horizon incident by discriminating the C-H absorption features in oil residues from background vegetation. These two recent examples demonstrate the utility of a spectroscopic approach to remote sensing for identifying and mapping the distributions of materials in imaging spectrometer data.

We present a method for partitioning a multispectral image into fixed size look-up tables (LUTs) that are dynamically updated for presence or absence of simple distribution characterizing features of the sub-frames they represent. If the features have been previously observed, the sub-frame is recognized and no update occurs, if not the table is updated and a suitable anomaly reported. Our method enables dynamic change detection to occur at multiple wavelengths independently by creating suitable LUTs for each wavelength band. Details of our approach are presented.

Various hyperspectral change detection methods exist in the literature. Here prediction-based methods, such as chronochrome and covariance equalization, are reviewed and compared with a more recently developed model-based approach. These methods are typically applied for anomalous change detection. Several methods for extending these algorithms to achieve matched change detection are discussed. The algorithms are then applied to airborne visible to near infrared hyperspectral data collected recently over Rochester, New York.

The increasing volume of data produced by hyperspectral image sensors have forced researches and developers to seek out new and more ecient ways of analyzing the data as quick as possible. Medical, scientic, and military applications present performance requirements for tools that perform operations on hyperspectral sensor data. By providing a hyperspectral image analysis library, we aim to accelerate hyperspectral image application development. Development of a cross-platform library, Libdect, with GPU support for hyperspectral image analysis is presented. Coupling library development with ecient hyperspectral algorithms escalates into a signicant time invest- ment in many projects or prototypes. Provided a solution to these issues, developers can implement hyperspectral image analysis applications in less time. Developers will not be focused on implementing target detection code and potential issues related to platform or GPU architecture dierences. Libdect's development team counts with previously implemented detection algorithms. By utilizing proven tools, such as CMake and CTest, to develop Libdect's infrastructure, we were able to develop and test a prototype library that provides target detection code with GPU support on Linux platforms. As a whole, Libdect is an early prototype of an open and documented example of Software Engineering practices and tools. They are put together in an eort to increase developer productivity and encourage new developers into the eld of hyperspectral image application development.

Automated hyperspectral image processing enables rapid detection and identification of important military targets from hyperspectral surveillance and reconnaissance images. The majority of this processing is done using ground-based CPUs on hyperspectral data after it has been manually exfiltrated from the mobile sensor platform. However, by utilizing high-performance, on-board processing hardware, the data can be immediately processed, and the exploitation results can be distributed over a low-bandwidth downlink, allowing rapid responses to situations as they unfold. Additionally, transitioning to higher-performance and more-compact processing architectures such as GPUs, DSPs, and FPGAs will allow the size, weight, and power (SWaP) demands of the system to be reduced. This will allow the next generation of hyperspectral imaging and processing systems to be deployed on a much wider range of smaller manned and unmanned vehicles. In this paper, we present results on the development of an automated, near-real-time hyperspectral processing system using a commercially available NVIDIA® Telsa™ GPU. The processing chain utilizes GPU-optimized implementations of well-known atmospheric-correction, anomaly-detection, and target-detection algorithms in order to identify targetmaterial spectra from a hyperspectral image. We demonstrate that the system can return target-detection results for HYDICE data with 308×1280 pixels and 145 bands against 30 target spectra in less than four seconds.

We propose a new architecture to accomplish lossy ultraspectral data compression where we particularly focus on AIRS (Atmospheric Infrared Sounder) images. In general AIRS images are good candidates for compression as they include more than two thousand spectral bands that account for over 40MB of data per single data cube. In our proposed compression technique the input image is first preprocessed by means of spatial subband decomposition followed by a spectral band ordering stage which is applied in order to increase the correlation between contiguous spectral bands. The resulting image is segmented on a spectral band basis in such a way that spectral bands are scanned to generate a speech-like signal that exhibits a higher spectral interband than intraband correlation and therefore can be modeled as an AR (autoregressive) process. As final step the data is processed through a compression stage involving short and long term forward linear prediction that produces an error signal that is encoded using a CELP (Code Excited Linear Prediction) scheme. The forward linear prediction filter order and the resolution of the CELP codebooks are adjusted depending on the spatial subband that originates the signal being predicted. By manipulating several parameters of both the preprocessing and compression stages different rate-distortion curves are obtained and highly efficient compression is achieved.

Compressed sensing (CS) has attracted a lot of attention in recent years as a promising signal processing technique that exploits a signal's sparsity to reduce its size. It allows for simple compression that does not require a lot of additional computational power, and would allow physical implementation at the sensor using spatial light multiplexers using Texas Instruments (TI) digital micro-mirror device (DMD). The DMD can be used as a random measurement matrix, reflecting the image off the DMD is the equivalent of an inner product between the images individual pixels and the measurement matrix. CS however is asymmetrical, meaning that the signals recovery or reconstruction from the measurements does require a higher level of computation. This makes the prospect of working with the compressed version of the signal in implementations such as detection or classification much more efficient. If an initial analysis shows nothing of interest, the signal need not be reconstructed. Many hyper-spectral image applications are precisely focused on these areas, and would greatly benefit from a compression technique like CS that could help minimize the light sensor down to a single pixel, lowering costs associated with the cameras while reducing the large amounts of data generated by all the bands. The present paper will show an implementation of CS using a single pixel hyper-spectral sensor, and compare the reconstructed images to those obtained through the use of a regular sensor.

A computational skin re ectance model is used here to provide the re ectance, absorption, scattering, and transmittance based on the constitutive biological components that make up the layers of the skin. The changes in re ectance are mapped back to deviations in model parameters, which include melanosome level, collagen level and blood oxygenation. The computational model implemented in this work is based on the Kubelka- Munk multi-layer re ectance model and the Fresnel Equations that describe a generic N-layer model structure. This assumes the skin as a multi-layered material, with each layer consisting of specic absorption, scattering coecients, re ectance spectra and transmittance based on the model parameters. These model parameters include melanosome level, collagen level, blood oxygenation, blood level, dermal depth, and subcutaneous tissue re ectance. We use this model, coupled with support vector machine based regression (SVR), to predict the biological parameters that make up the layers of the skin. In the proposed approach, the physics-based forward mapping is used to generate a large set of training exemplars. The samples in this dataset are then used as training inputs for the SVR algorithm to learn the inverse mapping. This approach was tested on VIS-range hyperspectral data. Performance validation of the proposed approach was performed by measuring the prediction error on the skin constitutive parameters and exhibited very promising results.

In the hyperspectral imaging community, there is a need for relevant and valid data for testing scientific principles. Several sets of data exist which support satellite and airborne HSI research applications. However, there is a need for data sets to support research looking at detecting and classifying smaller targets such as pedestrians. While many studies capture data on pedestrians, the data sets are often only related to the specific study being conducted. As a result, these types of data sets do not contain the necessary documentation or ground truth needed to apply the data in other contexts. This paper reports on a fully ground truthed HSI data set which was captured in June 2011 over an urban scene with pedestrians present. The imagery was collected using a modified airborne hyperspectral imager suited for ground level imaging in the 450 - 2500 nm wavelength region. The data captured are described along with the ground truth information and documentation which are part of the data package. Preliminary results from an initial study on material spectral separability using the data are included to demonstrate the utility of this particular data set.

In June 2011, a multi-sensor airborne remote sensing campaign was flown at the Virginia Coast Reserve Long Term Ecological Research site with coordinated ground and water calibration and validation (cal/val) measurements. Remote sensing imagery acquired during the ten day exercise included hyperspectral imagery (CASI-1500), topographic LiDAR, and thermal infra-red imagery, all simultaneously from the same aircraft. Airborne synthetic aperture radar (SAR) data acquisition for a smaller subset of sites occurred in September 2011 (VCR'11). Focus areas for VCR'11 were properties of beaches and tidal flats and barrier island vegetation and, in the water column, shallow water bathymetry. On land, cal/val emphasized tidal flat and beach grain size distributions, density, moisture content, and other geotechnical properties such as shear and bearing strength (dynamic deflection modulus), which were related to hyperspectral BRDF measurements taken with the new NRL Goniometer for Outdoor Portable Hyperspectral Earth Reflectance (GOPHER). This builds on our earlier work at this site in 2007 related to beach properties and shallow water bathymetry. A priority for VCR'11 was to collect and model relationships between hyperspectral imagery, acquired from the aircraft at a variety of different phase angles, and geotechnical properties of beaches and tidal flats. One aspect of this effort was a demonstration that sand density differences are observable and consistent in reflectance spectra from GOPHER data, in CASI hyperspectral imagery, as well as in hyperspectral goniometer measurements conducted in our laboratory after VCR'11.

This research demonstrates the application of spectral-feature-based analysis to identifying and mapping Earth-surface materials using spectral libraries and imaging spectrometer data. Feature extraction utilizing a continuum-removal and local minimum detection approach was tested for analysis of both reflectance and emissivity spectral libraries by extracting and characterizing spectral features of rocks, soils, minerals, and man-made materials. Library-derived information was then used to illustrate both reflectance- and emissivity-feature-based spectral mapping using imaging spectrometer data (AVIRIS and SEBASS). An additional spectral library of emission spectra from selected nocturnal lighting types was used to develop a database of key spectral features that allowed mapping and characterization of night lights from ProSpecTIR-VS imaging spectrometer data. Results from these case histories demonstrate that the spectralfeature- based approach can be used with either reflectance or emission spectra and applied to a wide variety of imaging spectrometer data types for extraction of key surface composition information.

For the flooding seasons of 2011-2012 multiple space assets were used in a "sensorweb" to track major flooding in Thailand. Worldview-2 multispectral data was used in this effort and provided extremely high spatial resolution (2m / pixel) multispectral (8 bands at 0.45-1.05 μ m spectra) data from which mostly automated workflows derived surface water extent and volumetric water information for use by a range of NGO and national authorities. We first describe how Worldview-2 and its data was integrated into the overall flood tracking sensorweb. We next describe the use of Support Vector Machine learning techniques that were used to derive surface water extent classifiers. Then we describe the fusion of surface water extent and digital elevation map (DEM) data to derive volumetric water calculations. Finally we discuss key future work such as speeding up the workflows and automating the data registration process (the only portion of the workflow requiring human input).

We explore the use of machine learning, computer vision, and pattern recognition techniques to automatically identify volcanic ash plumes and plume shadows, in WorldView-2 imagery. Using information of the relative position of the sun and spacecraft and terrain information in the form of a digital elevation map, classification, the height of the ash plume can also be inferred. We present the results from applying this approach to six scenes acquired on two separate days in April and May of 2010 of the Eyjafjallajökull eruption in Iceland. These results show rough agreement with ash plume height estimates from visual and radar based measurements.

Multispectral imaging (MSI) data acquired at different view angles provide an analyst with a unique view into shallow water. Observations from DigitalGlobe's WorldView-2 (WV-2) sensor, acquired in 39 images in one orbital pass on 30 July 2011, are being analyzed to determine bathymetry along the windward side of the Oahu coastline. Satellite azimuth and elevation range from 18.8 to 185.8 degrees, and 24.9 (forward-looking) to 24.9 (backward-looking) degrees with 90 degrees representing a nadir view (respectively). WV-2's eight multispectral bands provide depth information (especially using the Blue, Green, and Yellow bands), as well as information about bottom type and surface glint (using the Red and NIR bands). Bathymetric analyses of the optical data will be compared to LiDAR-derived bathymetry in future work. This research shows the impact of varying view angle on inferred bathymetry and discusses the differences between angle acquisitions.

The capability of the WorldView-2 (WV02) satellite is analyzed for bathymetric applications in shallow coastal waters. We use an Optimal Estimation method, which provides lower bounds on retrievals errors for an idealized sensor and idealized model of the environment. Retrieval performance is studied over different substrates and column water properties. We also study effects of increased signal to noise ratio. This analysis is supported by numerical inversion of imagery, using a variant of a least-square optimization. Results from 4 different study areas collected across a few sites in clear Case 1 and Case 2 waters show that the water depth can be realistically measured on a pixel-by-pixel basis with 10% standard deviation of the error, down to nearly 20 meters depth in Case 1 waters over bright sandy bottom. The same accuracy over dark sea grass or coral is valid down to 10 meters, assuming that reliable a priori substrate albedo is available. Water turbidity has an important effect on retrievals - Case 2 water with small concentrations of suspended solids allows for 10% accuracy down to 10 meters over bright targets. The retrieval accuracy is likely to improve with tighter a priori constrains, constraints obtained from the context of entire image, or independent information from multiple stereo pairs collected in a single WV02 pass.

To facilitate locating archaeological sites before they are compromised or destroyed, we are developing approaches for generating maps of probable archaeological sites, through detecting subtle anomalies in vegetative cover, soil chemistry, and soil moisture by analyzing remotely sensed data from multiple sources. We previously reported some success in this eort with a statistical analysis of slope, radar, and Ikonos data (including tasseled cap and NDVI transforms) with Student's t-test. We report here on new developments in our work, performing an analysis of 8-band multispectral Worldview-2 data. The Worldview-2 analysis begins by computing medians and median absolute deviations for the pixels in various annuli around each site of interest on the 28 band dierence ratios. We then use principle components analysis followed by linear discriminant analysis to train a classier which assigns a posterior probability that a location is an archaeological site. We tested the procedure using leave-one-out cross validation with a second leave-one-out step to choose parameters on a 9,859x23,000 subset of the WorldView-2 data over the western portion of Ft. Irwin, CA, USA. We used 100 known non-sites and trained one classier for lithic sites (n=33) and one classier for habitation sites (n=16). We then analyzed convex combinations of scores from the Archaeological Predictive Model (APM) and our scores. We found that that the combined scores had a higher area under the ROC curve than either individual method, indicating that including WorldView-2 data in analysis improved the predictive power of the provided APM.

A method of incorporating macroscopic and microscopic reflectance models into hyperspectral pixel unmixing is presented and discussed. A vast majority of hyperspectral unmixing methods rely on the linear mixture model to describe pixel spectra resulting from mixtures of endmembers. Methods exist to unmix hyperspectral pixels using nonlinear models, but rely on severely limiting assumptions or estimations of the nonlinearity. This paper will present a hyperspectral pixel unmixing method that utilizes the bidirectional reflectance distribution function to model microscopic mixtures. Using this model, along with the linear mixture model to incorporate macroscopic mixtures, this method is able to accurately unmix hyperspectral images composed of both macroscopic and microscopic mixtures. The mixtures are estimated directly from the hyperspectral data without the need for a priori knowledge of the mixture types. Results are presented using synthetic datasets, of macroscopic and microscopic mixtures, to demonstrate the increased accuracy in unmixing using this new physics-based method over linear methods. In addition, results are presented using a well-known laboratory dataset. Using these results, and other published results from this dataset, increased accuracy in unmixing over other nonlinear methods is shown.

Nonnegative matrix factorization and its variants are powerful techniques for the analysis of hyperspectral images (HSI). Nonnegative matrix underapproximation (NMU) is a recent closely related model that uses additional underapproximation constraints enabling the extraction of features (e.g., abundance maps in HSI) in a recursive way while preserving nonnegativity. We propose to further improve NMU by using the spatial information: we incorporate into the model the fact that neighboring pixels are likely to contain the same materials. This approach thus incorporates structural and textural information from neighboring pixels. We use an ℓ₁-norm penalty term more suitable to preserving sharp changes, and solve the corresponding optimization problem using iteratively reweighted least squares. The effectiveness of the approach is illustrated with analysis of the real-world cuprite dataset.

In hyperspectral imaging, the radiation represented by a single pixel rarely comes from the interaction with a single homogeneous material. However, the high spectral resolution of imaging spectrometers enables the detection, identification, and classification of subpixel objects from their contribution to the measured spectral signal. Unmixing is a hyperspectral image processing approach where the measured spectral signature is decomposed into a collection of constituent spectra, or endmembers, and a set of corresponding fractions or abundances which correspond to the fractional area occupied by the particular endmember in that pixel. The use of a single spectrum to represent an endmember class does not take into account the variability of spectral signatures caused by natural factors. Simple spectral mixture analysis can, by itself, provide suitable accuracies in some relatively homogeneous environments, but because of the spectral complexity of many landscapes, the use of fixed endmember spectra may results in inaccurate unmixing analysis for complex regions over large landscapes. This paper addresses the question of how to perform unsupervised unmixing where local information is used to extract local endmember information and merged at a global level to extract endmembers classes for developing an accurate description of the scene under study using the nonnegative matrix factorization. Preliminary results using AVIRIS data are presented. Results show that this approach better captures local structures that are not possible with global unmixing approach. Furthermore, they show that spatial information allows the identification of more spectral endmembers than is it possible with just spectral-only methods.

Automated unmixing consists of finding the number of endmembers, their spectral signatures and their abundances from a hyperspectral image. Most unmixing techniques are pixel-to-pixel procedures that do not take advantage of spatial information provided by hyperspectral sensor. This paper explores a new approach for unmixing analysis of hyperspectral imagery based on a multiscale representation for the joint estimation of the number of endmember and their spectral signatures. Experimental results using an AVIRIS image is presented.