Deep machine learning algorithms can help government agencies analyze satellite imagery.
WASHINGTON — For all the hype and promise around artificial intelligence and machine learning technologies in military applications, it always comes down to what specifically can be done with it.
The industry keeps rolling out new gee-whiz artificial intelligence tools but the defense and intelligence communities still are trying to figure out how to use them and whether they really work as promised.
According to a new study, there is one area where deep machine learning algorithms can definitely help the government, and that is to analyze satellite imagery.
Officials from the National Geospatial Intelligence Agency have called on the private sector to bring forth machine learning tools to automate repetitive and time-consuming image analysis tasks. They want to free up skilled analysts to spend more time on hard intelligence problems that can’t be turned over to a computer.
Researchers from the Center for Geospatial Intelligence at the University of Missouri used a deep learning neural network to assist human analysts in visual searches for surface-to-air missile sites over a large area in southeastern China. The results showed that the computer performed an average search time of only 42 minutes for an area of approximately 90,000 square kilometers. By comparison, North Korea is about 120,000 square kilometers.
“This was more than 80 times more efficient than a traditional human visual search,” the center’s director and University of Missouri electrical engineering and computer science professor Curt Davis told SpaceNews.
The software achieved the same overall statistical accuracy as human analysts — 90 percent — for correctly locating the missile sites.
“I’ve been doing this research for almost 20 years, and I do believe the application of deep machine learning technology to satellite imagery reconnaissance is revolutionary,” he said. “I never expected this type of performance that we’ve been able to see both in the lab and the study. The metrics we’re seeing, the applications to larger-scale data sets to me is revolutionary.”
Historically, machine learning algorithms haven’t performed well when they have been applied to large satellite imagery data sets, he said. The breakthroughs came in the last couple of years. The computer used in the study searched the 90,000 square kilometer area in less than an hour.
U.S. defense and intelligence agencies are drowning in high-resolution imagery they need to analyze every day to monitor events unfolding around the world. “There is simply not enough manpower to effectively analyze all the image data collected today, and the problem is only getting worse,” Davis said.
And the technology is only going to get better, he said. “The ultimate goal is to recognize dozens and hundreds of different types of objects very quickly,” said Davis. “I believe that goal is achievable in the near future.” Researchers will be training networks to search for things military analysts typically look for, including bunkers, aircraft shelters, radar sites, antennas, satellite dishes, launch pads and tank formations.
The study used commercially available remote sensing satellite imagery of one-meter resolution. With new generations of satellites soon to be launched by commercial firms, including some with sub-meter resolution, the data deluge will continue. “It has taken a while for the remote sensing community to evaluate these deep machine learning methods,” said Davis. “Most of the studies I’ve seen were only experiments against limited data sets,” he continued. “Now we’ve been able to apply deep learning models to a large data set.”
The research was published in the SPIE Journal of Applied Remote Sensing in a special issue on deep learning in remote sensing applications. Readers can search for Chinese surface-to-air missile sites on a demonstration website that uses the same high-resolution satellite imagery and deep learning algorithms used in the study.
If and when these tools start replacing human analysts remains to be seen. National Geospatial-Intelligence Agency Director Robert Cardillo said recently he wants to automate 75 percent of the repetitive tasks analysts perform so they can focus on the “25 percent that require the most attention.”
Deep learning methods can help do that, said Davis. The tough threat posed by North Korea is a case in point. The computer can find most of the fixed-site missiles but it takes human skills to track Pyongyang’s notoriously elusive mobile ballistic missile launchers. “That’s a harder problem. They can be hiding in a cave, pop out and launch a test missile.”
The Pentagon years ago identified machine learning and artificial intelligence as central elements to the military’s modernization strategy for weapons and information systems. Clearly the industry is progressing quickly, but the Defense Department has not moved as fast in applying the technology.
“One of the challenges DoD faces in this area is that we are too often in this position where we discuss something in an abstract or theoretical way,” said Shawn Steene, senior force planner for emerging technologies at the Defense Department. He spoke Oct. 19 at a CNA panel discussion on artificial intelligence.
In recognition of the growing role of these technologies in defense, CNA, a federally funded nonprofit think tank in Arlington, Va., announced the opening of a “Center for Artificial Intelligence and Autonomy.”
“To some degree we’re limited by our creativity in the application of these capabilities,” Steene said.
He recalled when the U.S. Geospatial Intelligence Foundation put out an open-source challenge offering a prize to whoever would come up with the best algorithms to take overhead imagery and identify the buildings in the picture. “The point was to remove from the analysts the first cut layer,” he said. “Just having that program to do that, having the machine doing the first layer, I can pass that to an analyst. And instead of spending time doing basic tasks, now they can do the ‘value added’ work.”
Using artificial intelligence for data mining also could help prioritize information so networks are not clogged by data that may not be valuable, said Steene. “Instead of needing a giant pipe, if I have some screening at the front end, I can constrain the data flow,” he said. This technology offers infinite applications but the Defense Department needs to define the problems it is trying to address and “we need to use more creativity.”