Algorithms Improve Spectral Imaging

An imaging spectrometer in an airplane or satellite can picture the Earth's surface while simultaneously detecting the broad range of radiation wavelengths that emanate from each pixel. More than 40 models of imaging spectrometers are in use today, and they all have a common feature: Each rapidly produces a plethora of data that is a challenge to process in real time.

Glenn Healey's algorithm can identify low-contrast target materials in both sun and shade in a forest environment with a false alarm rate of about 1 pixel in 10⁶. The six targets present in the upper image are detected by the algorithm and labeled in red in the lower image. Courtesy of the University of California at Irvine.

Glenn Healey, director of the Computer Vision Laboratory at the University of California, is one of those researchers. To find small, partially hidden or camouflaged objects on the ground, he has applied computer algorithms and reflectance models to airborne imaging spectrometer data. He developed a computational system that detects spectral target profiles in variable illumination and weather conditions, even when the target is as small as 5 percent of a pixel's area and when the object is only partially visible.

Healey's algorithms have processed data from the Hyperspectral Digital Imagery Collection Experiment, which uses a 320 x 210-pixel indium antimonide sensor array. In a typical search, the device is mounted in an airplane flying 100 mph at 5000 ft. At this elevation, the resolution of a single pixel is 30 in. The array searches a square mile of ground in four minutes. The light that is detected by each pixel is separated into 210 discrete wavelengths from 400 to 2500 nm. The rate of digital data acquisition is 40 million measurements per second.

When they rely solely on spatial characteristics, target recognition systems are limited in their ability to detect objects in low contrast, or when targets have been partially hidden or camouflaged. Hyperspectral imagers process spectral as well as spatial data. As they seek the wavelength profiles that are unique to a target, identification algorithms analyze spectral data and ignore the surrounding background.

"The most important aspect of our work," Healey explained, "is that we are able to define target signatures that are invariant to scene properties, such as the illumination environment or the geometric arrangements of objects." Those invariant properties are used to build the search algorithm that processes the imagery data.

Because hyperspectral images can detect targets through dense foliage or across rugged terrain, potential applications include military reconnaissance, mineral exploration, crop surveys or searches for lost people or wreckage.