Transforming the Science of Remote Sensing

March 2013
Topics: Remote Sensing, Sensor Technology, Image Processing
Imagine the advances possible if we could see more of the invisible world.
scientists working in the field

"Sometimes it's good to think differently from how everyone else has been thinking for a long time," says Marin Halper, a MITRE signal-processing specialist. "A lot of the work that's gone into PRISM is about trying things differently from how people have tried them for the last 10 or 20 years."

PRISM (Probabilistic Identification of Solid Materials)——a new algorithm for hyperspectral imaging——promises to transform the science of remote sensing. While conventional techniques have made it possible to remotely detect materials on the Earth's surface, PRISM now enables analysts to go even further and to identify a specific material with confidence. Its applications cover a wide range of domains, including mineralogy, geology, forestry, agriculture, environmental monitoring, astronomy, search and rescue, and disaster relief.

Halper played a key role in developing PRISM. When he started working on PRISM in October 2010, he made a deliberate decision to try something new. "We were going into areas that nobody else in our field has gone before," he says. "There wasn't a lot of work to draw from —lessons learned, failures, or anything else."

Beyond Red, Green, and Blue

Remote sensing refers to the detection and identification of materials from a standoff distance. Since early attempts in the 19th century, which included primitive cameras mounted on hot-air balloons, remote sensing has continued to evolve. During the Cold War, remote sensing was largely associated with surveillance and reconnaissance missions. A decade later, the Landsat program gained prominence for its satellite imagery of Earth.

In the 1980s, remote sensing expanded to include hyperspectral imaging. Following Hurricane Katrina, the Environmental Protection Agency used hyperspectral data to map potentially hazardous chemicals in the Gulf of Mexico. The Rochester Institute of Technology also used hyperspectral data after the 2010 earthquake in Haiti to locate refugee camps.

Hyperspectral data is derived from a series of measurements taken simultaneously by sensors at hundreds of distinct wavelengths, many of them beyond the visible light spectrum. "In hyperspectral data, you'll have hundreds of bands of light——not just RGB [red, green, or blue] —but all these different wavelengths, and each one will have a corresponding measurement of reflected or emitted energy," says Halper.

The result is a digital image in which each data pixel consists of a vector —or spectral signature— of many tens or hundreds of energy measurements. Each signature represents a physical quantity linked by the underlying physics and chemistry of a particular material, or mixture of materials, within the sampled pixel. By comparing the signature of an image pixel to a library of reference of signatures, it's possible to assign the pixel a material label. This is the foundation of hyperspectral imaging.

The Missing Part

Halper arrived at MITRE in May 2002 at the end of his freshman year at Yale, where he was studying mathematics and economics. That summer he worked as an intern with MITRE's Nanotechnology Group, and he continued to return to MITRE between semesters and during summer breaks. Eventually he started working with Sherry Olson, a scientist who specializes in hyperspectral sensing, on her research.

Last year Halper, collaborating with a colleague from the Rochester Institute of Technology, formally began his work on PRISM. They had gone over an algorithm many times, according to Halper, but it still wasn't working. There was something missing. As they struggled to figure out the problem, Halper says he knew at a certain point that it was going to work.

"There was one specific thing we weren't calculating right," he says. "It took me a day of not thinking about it to come up with a way of solving that part."

Confusers Be Gone!

By converting raw hyperspectral data into pixels with geospatial attributes, the PRISM algorithm represents a transformational engineering solution for remote sensing. One of the biggest challenges involved separating targets from materials known as "confusers," which look spectrally similar to the targets and therefore result in a high percentage of false alarms during the identification process.

Conventional remote sensing techniques have been unable to address this challenge. Instead of using the single best model to estimate the confidence of a material identification, which is the process used by conventional remote sensing techniques, PRISM offers an improved method by averaging over a set of models based on Bayesian Model Averaging (see sidebar).

PRISM has enabled a decrease in the false alarm rate by an order of magnitude. In addition, results are available in a drastically reduced amount of time than previously. "PRISM has changed the way we— and others —use hyperspectral sensor data," says Olson. "What started as MITRE research has turned into a game changer."

"Let's Do Something Else"

PRISM is now widely considered the state of the art for automated material identification throughout the hyperspectral community. With assistance from Halper's colleague, Shannon Jordan, PRISM has been enhanced and used in multiple R&D tests and demonstrations. MITRE is currently in the process of patenting the new technology. "It's an enabling capability, which opens up a lot of doors and research opportunities that we can explore," says Halper. "The key focus is expanding it to more and more problem sets."

For example, PRISM may serve as a catalyst to enable the automated and accurate identification of a rich and diverse set of materials on the Earth's surface. "Maybe we can build a library of every known material in the world, similar to the Google Street View car that drives around and takes pictures of everything," says Halper. "Maybe we can take spectral signatures of everything and try to label every pixel on the ground automatically."

As researchers continue to look for new ways to apply PRISM, Halper reflects back on its evolution. "We said to ourselves: the way we've been doing things for the past ten years isn't going to work anymore," he recalls. "Let's do something else."

——by David Savold


Publication Search