Technical Innovation
Mixed Material Classifier - Detection of sub-visual bank erosion features
For more than 25 years, AAI has been focused on developing new approaches and technologies for retrieving more detailed and quantitatively defensible information from spectral imagery, and doing it automatically. The technologies were developed as part of an ongoing R&D program to meet our customers' requirements for increasingly automated, detailed and quantitatively defensible information products from spectral imagery. The following is a brief description of some of the key advanced hyperspectral and multispectral image analysis technologies developed and used by AAI to generate its truly unique line of automated image-derived information products.
iCee™ Atmospheric Correction - automaticatic atmospheric (and sensor) correction to material reflectance. This product addresses one of the key requirements for retrieving detailed and quantitatively defensible information from spectral imagery. The atmosphere and sensor distort the image-reported spectral properties of the material on the ground, and these effects need to be suppressed in order to retrieve actual material reflectance properties. This is critical to enabling accurate materials identification and quantitatively defensible assessments. iCee™ Atmospheric Correction is unique in the industry with respect to accuracy, repeatability, and full automation. It is the critical first enabling step for all of AAI's other spectral image analysis technologies and the generation of its information products. In addition to its role as a core enabling technology for AAI's line of image analysis products, iCee™ automated atmospheric correction is also available as a product on its own. Click here to learn more.
Mixed Material Classifier - automatic subpixel detection and classification of up to three material components per pixel. This technology addresses another key requirement for retrieving detailed and quantitatively defensible information from spectral imagery. Image pixels almost always contain mixtures of materials regardless of pixel dimensions, and the spectral properties of this mixture are different than those of any of the individual component materials by themselves. This can significantly impact detection and identification accuracies of materials of interest, even when only small amounts of the other materials are present. Mixed Material Classifier not only addresses the critical "mixed pixel problem" in a reliable, repeatable, adaptive, and automated non-interactive way. It also classifies pixels based on the retrieved combinations of materials present in them. This allows higher level actionable classifications, such as standard beach classes, submerged coastal and riverine bottom material classes, and fire fuels classes, among others, to be retrieved directly from the imagery.
Material Identifier - automatic identification of materials in images calibrated to reflectance by iCee™. Material Identifier ability to calibrate image pixels to material reflectance with iCee™ and to transform a library of reflectance spectra for known materials so as to match the spectral band characteristics of the image pixels using AAI's Universal Sensor core technology (see below) makes possible the automatic identification of materials in the image. Pixel spectra are initially placed into generalized material identification classes (e.g., sparse, medium, or dense vegetation, water, wet soil, etc.), and then selectively refined based on the goodness of match to the library spectra. Material Identifier works best with the component material reflectance images produced by Mixed Material Classifier to suppress the "mixed pixel" effect, which can mask the spectra and the correct identification of individual components.
Universal Sensor - automatically adapts AAI spectral exploitation technologies for use with any spectral sensor of opportunity (specified by a simple text file list of band centers and widths). Universal Sensor is incorporated in each of AAI's technologies, allowing them to be used with images from any multispectral, hyperspectral, or ultraspectral sensor operating within the 350 - 2600nm wavelength range.
Signature Adapter - reports the individual spectra of subpixel-scale materials detected in an image using a spectral signature, and develops one or more refined signatures based on those image-retrieved spectra. The process is fully automatic, and addresses the pervasive problem of having signatures that are unrepresentative of materials that have natural variation.
Signature Transformer - automatic transformation of spectral signatures so they can be reused with data from another sensor, including sensors with a different number of spectral bands. This not only allows unique signatures of opportunity to be reused with image data that may only be available from other sensors. It also allows signatures from traditional spectral libraries to be meaningfully transformed for use with imagery from a spectral imager of opportunity, and vice versa.
BANDS - characterization of material-diagnostic features in hyperspectral image pixel spectra. The process automatically derives and reports the wavelength positions, widths, and strengths of optical absorption and emission features in image pixel spectra. The features can be used for material identification using standard spectroscopic methods. They can also be used for data compression, and/or classification. The process tolerates significant spectral noise and feature overlap. BANDS is a core component of AAI's Hyperspectral Distiller application, as well as a principal engine for those classes of hyperspectral image analysis products requiring detection of specific discrete absorption or emission features.
Autonomous Mission-Directed Technologies
Shore Zone Analyzer (SHORZAN) / Land-Water Interface - automatic identification and segmentation of water from non-water (land, clouds, etc.) cover materials in images, and shape file delineation of waterway boundaries. The process first corrects imagery to units of material reflectance using the iCee™ atmospheric correction core technology (described above). It next spectrally identifies water, including subpixel occurrences (e.g., from narrow or partially obscured waterways), using the Mixed Material Classifer and Material Identifier core technologies (see above). It then converts the raster layer into a shape file. The process automatically estimates an Accuracy Confidence value for the results, and displays a graphic symbol on the GUI when the processing is complete, ranking the result as either Good, Fair, or Poor. An illustration is shown above. The raster water layer is shown to the left (blue), and the corresponding shape file is shown to the right (cyan).
Glint Remover - Automatically identifies and removes surface reflection components (solar glint and sky reflections) from water pixel spectra. Glint Remover uses the iCee™ Atmospheric Correction technology to correct the image to units of material reflectance so that the water and surface reflection components can be accurately identified. The Mixed Pixel Classifier core technology separates the pixel components, and Material Identifier identifies the water components. The process replaces the pixel spectrum with the water component, thereby effectively removing the surface reflection contribution. An illustration is shown above showing a zoomed-in portion of a QuickBird image of a US Naval rescue operation in the Harbor at Meloubah, Indonesia following the 2004 Tsunami. The image, corrected to units of material reflectance, is shown on the left. On the right is the image corrected for surface reflections using Glint Remover. With the glint removed, the image reveals that the water in the harbor contains a thin laminar surface layer of sediment-laden water from the river to the immediate north. Note, for example, how the sediment-laden water (red) is being diverted by the hull of the ship in the lower half of the image (to the right of the arrows), revealing clearer seawater underneath. Note also the debris (navigational hazards) floating in the water, indicated by the arrows. These features are all masked by surface reflections and glint in the image to the left.
Supervised Anomaly Finder (SAF) - Automatically identifies spectral anomalies, i.e. scene features that are spectrally out of context with their surroundings. The SAF technology provides a core technology for Disturbed Soil Finder (DSF) and Spoil Pile Detector (SPD), but for those applications it is restricted to rock and soil anomalies. Developed initially to search for anomalies associated with human activity along waterway corridors, the SAF application has more general utility and can be used to retrieve the spectral properties of anomalous subpixel-scale features for identification across a wide range of terrain and water settings. The above illustration shows the results of a search for subpixel-scale structures (rusty metal rooftops) in a Landsat Thematic Mapper of a South American jungle setting. Two of the detected structures were confirmed using independent collateral information. The third is an unconfirmed structure.
Hyperspectral Distiller - Fully automatic identification of the most discriminating spectral bands (image planes) in a hyperspectral image, and creation of multispectral subset image consisting only of those image planes. Hyperspectral Distiller is a core technology for those classes of image analysis products not requiring detection of specific discrete absorption or emission features. The process can operate in any of three modes, identifying the most discriminating bands for either: 1) general land cover classification; 2) discriminating a specific material of interest from the principal land cover materials in a scene; or 3) discriminating a specific material of interest from false alarm materials in the scene. For some image analysis products all three sets of bands have been used. The distilled images not only have an obvious compression advantage. They are also typically more discriminating than a full-resolution hyperspectral image because of elimination of the “superfluous” and redundant bands. Furthermore, the images are brought into the realm of multispectral image analysis with its myriad of available analytical processes and tools. In the above illustration, Hyperspectral Distiller selected the most discriminating bands for discrimination of plywood from the other materials in the scene (Mode 2). It successfully discriminated the plywood against the other materials in the scene (above left), detecting all three pieces of plywood (detections are red). It also detected a particle board panel, however, which is spectrally similar to the plywood. Next, Hyperspectral Distiller operated on the Mode 2 detections with a different set of bands designed to discriminate the plywood from particle board. The Mode 3 results (above right) successfully discriminated the two sets of materials.
AAI's suite of advanced hyperspectral and multispectral image analysis technologies have been leveraged to create autonomous mission-directed spectral image exploitation technologies specific to our nation's needs. DAM is a collection of these applications (see below) that AAI licenses to DoD users. Support and periodic revisions are available for a modest annual fee. The applications enable the entry-level spectral analyst to automatically identify features of interest within an image, such as excavation spoil piles, disturbed soil, and user-specified colors, with little or no interaction. DAM is a plug-in available for ERDAS IMAGINE 9.3 and SOCET GXP version 2.3.1. It is also available on GeoServer® through our industry partner Photon Research Associates, a Raytheon Company. DAM currently contains the following applications:
Disturbed Soil Finder (DSF) - automatically identifies and characterizes soil features that are spectrally out of context with their immediate surroundings. Features can differ at subpixel scales and not be visible to the unaided eye. Each feature is characterized by 1) its spectral properties; 2) pixel fraction of the anomalous component; and 3) its area, elongation, and geographical coordinates. It was developed primarily to perform automated searches for improvised explosive devices (IEDs), buried weapons caches and hazards, mass graves, and other buried objects. DSF can be used with both airborne and satellite spectral imagery.
Disturbed Soil Finder example
Detection of buried roadside IED hazards and scarring disturbances likely associated with their emplacement
 |
"Known" buried hazard locations with indicated ground truth error of uncertainty |
 |
Detected(≤ 1 meter diameter) disturbances associated with buried hazards |
 |
Scarring disturbances associated with vehicular and egress activity |
Spoil Pile Detector (SPD) - Automatically identifies tunnel excavation spoil, and distinguishes it from other soil occurrences. SPD spectrally characterizes the spoil and uses that characterization to further identify areas in the image that are candidate tunnel locations. Outputs include spoil pile thematic images and shapefiles, as well as tunnel locator thematic images and shapefiles.
Spoil Pile Detector: Detected tunnel excavation spoil piles at Natanz, Iran tunnel construction site
Spoil Pile Detector: Spoil pile source material (possible tunnel locations) at Natanz, Iran tunnel construction site
Object-Based Change Detector (OBCD)- Automatic registration and comparison of "before" and "after" images containing identified specific objects of interest, such as disturbances from DSF. OBCD determines which objects are pre-existing and which are new. The process autonomously identifies and uses discrete natural image features as control points for derivation of the image-to-image co-registration transformation. The transformation is applied only to the coordinates of the objects of interest, rather than to the pixel coordinates. Consequently, images need not be registered and they can be from airborne or satellite sensors having very different pixel sizes, numbers of spectral bands, and flight paths.
| 5 August 2007 | 18 August 2007 |
 |
|
OBCD uses shrubs in the images as ground control points (shown in cyan) for coordinate transformation
The coordinate transformation is applied to disturbance coordinates in the earlier image, and the transformed coordinates of the disturbances in the earlier image are compared to the disturbance coordinates in the later image to determine which disturbances are persistent (cyan in image below) and which are new (dark blue). Most of the disturbances at this location
QuickBird image of the Euphrates River in Syria, 18 August 2007, with a cluster of roadside soil disturbances detected by Disturbed Soil Finder. Most of the disturbances had developed over the 18-day period since 5 August 2007. Only a few had persisted over that period.
Specialized Detector (SPEC-D)- identifies pixels having spectral characteristics consistent with a specific primary color. Spec-D was developed to identify specific man-made features.