Just how many jets, cargo planes, and other military vehicles is Russia deploying to the border of Ukraine? Answering that sort of question is a labor-intensive project for analysts, requiring them to pore through satellite photos to find and classify specific objects. A new tool from data analysis firm Orbital Insight could change that.
The multi-class object detection algorithms can detect and classify a wide number of objects of relevance to the military, alerting analysts to events like buildups or unusual deployments anywhere that can be photographed by satellites. Orbital will announce the multi-class object detection algorithms today, part of the company’s GO platform.
Analysts “don’t have enough time to look at all the targets. [They] focus on all the top-tier ones. But the idea [with the new tool] is to see what’s also going on, in some of these second-tier and third-tier [ones] that may be really related,” Patrick Podejko, a geospatial analyst with Orbital Insight, said ahead of the public announcement. He estimated that the company tracks around 8,100 airfields worldwide, in addition to ports, testing facilities, and other locations.
After recognizing different types of aircraft, ships, or weapons, the algorithm can alert analysts to new or unusual activity at the sites, such as a large number of cargo planes arriving, or more than the usual bomber or fighter jet traffic.
That information becomes more useful in the context of other information. For example, is Russia just staging an exercise, or does other information point to something more sinister?
“It’s good to look not only at these components of the airfields, but supply depots or something further back,” Podejko said. “Start looking at the logistics chain. What’s going on in these other sites? We also have a truck-detector algorithm…Instead of the analysts going and counting all those trucks and all the supply depots or even something from up closer to the border, now you can start quantifying this and start quantifying what is their [Russia’s] ability to do an operation? Do they have the logistics in place to do it?”
That visual data can be combined with other pieces of collectable information, such as telephonic metadata from data-brokers or automatic identification system, or AIS, data from ships, to refine results, boost confidence in particular findings, or discover new events.
“We also developed a dark ship-tracking algorithm where, you know, if they turn their AIS on and off… we can see that history,” Podejko said.
Right now, it’s easier to train AIs to recognize some things than others; there are simply more pictures of things like jets on tarmacs than there are of mobile missile launchers. One of the solutions the intelligence community is considering to bridge that gap is what they call synthetic data, or taking the handful of available pictures of a rare object and then generating more pictures from different angles, under different conditions, etc. to train the algorithm to recognize it.
Dan Soller, Orbital’s senior advisor on national security programs, said those rarer objects, like missiles, are “going to require a lot more synthetic data to be able to train the model to be able to detect them in areas where they like to hide, whether that’s hills, fighting positions, forests,…covered areas…that’s where the real value of synthetic data is going to be.”
But, for synthetic data, he cautioned, it’s still the “early days.”
© 2021 Government Executive Media Group LLC
Distributed by Tribune Content Agency, LLC.