Hyperspectral Band Reduction Ground Perspective Applications, 09-R8081Printer Friendly Version
Inclusive Dates: 7/7/09 – 7/7/10
Background: This project investigated hyperspectral sensors and hyperspectral processing technology for use in ground-based systems. Hyperspectral imaging is a passive imaging technology that captures the reflected or emitted energy from objects across different narrow wavelengths of the electromagnetic spectrum. This technique, also known as an "imaging spectrometer," acquires images in contiguous spectral bands to enable improved characterization and identification of targets. Commonly used in optical remote sensing, hyperspectral imaging is beneficial in applications such as agricultural health monitoring, mineral identification and camouflaged target detection. The sensors employed by such systems, for example, NASA's Airborne Visible/Infrared Imaging Spectrometer (AVIRIS), are typically designed for high resolution, with the intent being to process the collected data in a lab setting rather than in real time (AVIRIS has 224 spectral channels). The exploitation of hyperspectral data provides value in the ability to identify materials of interest based on their spectral signatures. On the other hand, the large amounts of collected data and requirement processing techniques make hyperspectral imaging inefficient and challenging to apply to many real-time problems.
Approach: The primary aim of this project was to transition hyperspectral technology toward ground-based applications by minimizing the spectral bands used to reduce data acquisition time, data processing time, and overall cost of hyperspectral systems as compared to wide spectrum imagers currently in use today, while still maintaining the ability to classify materials based on their spectral signatures. The project was divided into six main technical tasks. In task one, acquire/verify test hardware and data, the team selected a commercial-off-the-shelf (COTS) hyperspectral imager and used it to acquire a large number of representative hyperspectral data cubes of ground-based scenes. In task two, model-based sampling optimization, the team selected a training set of signatures from the collected material database and developed a feature selection algorithm for automatically identifying effective multispectral channels from a set of COTS filters. An optimized material identification algorithm was then developed that used simulated reduced band inputs, rather than the full hyperspectral inputs, to classify materials by their spectral signatures. In task three, segmentation analysis, the team developed a spatial/spectral segmentation algorithm to cluster image based regions according to surrounding identified material signatures. In the fourth and fifth tasks, fuse algorithms and test final system, the team integrated the segmentation algorithm with a material classifier to form an effective overall system for identifying materials from multi-spectral samples. In task six, generate hardware requirements, the team developed specifications and selected COTS hardware that could be used in place of the hyperspectral imager to collect reduced spectral channel images while maintaining the spectral variability.
Accomplishments: The results showed that a small number of optimally selected, commercially available filters captured most of the variance as compared to the full set of channels that were available in the hyperspectral imager. The results allowed a high rate of correct classification of our selected materials on a pixel by pixel level.