Activity at North Korea’s Pyongsan uranium mine appears to have increased from 2017 to 2020, though its output still lags the country’s uranium processing capabilities, according to Stanford researchers who used artificial intelligence software in their analysis.
Uranium from Pyongsan can be refined into low-enriched uranium, which is suitable for nuclear power reactors, or high-enriched uranium, suitable for weapons. No outside visitors or international monitors have seen the Pyongsan mine since the IAEA’s last visit in 1992.
Sulgiye Park, lead author of the paper that appeared in the journal Science and Global Security last month, said she and her Stanford researchers worked with geospatial intelligence company Orbital Insight to apply machine learning software to satellite photos of Pyongsan to understand what was happening at the mine.
The team ran photos of the site through landuse algorithm that can detect things like forests, grassland and other natural features. By extension, it can also detect the presence of elements that it can’t classify. Those areas get the label “other.” In the case of Pyongsan, that “other” material is a steady accumulation of mine waste, indicative of increased activity, an accumulation that grew at the expense of forests.
“A reduction of vegetation, including forests and grasslands, by 20% from 2017 to 2020, is concurrent to an almost four times increase in ‘others’ is likely a result of mine and wastes expansion over time. A major advantage of the application of machine learning to satellite images is efficiency. However, caution is needed to interpret such results, as vegetation change may not always indicate a mine expansion,” she writes.
That timing is significant because that suggests an increase in mining activity around the same time that the North Korean regime and the Trump administration were in talks to cease North Korean nuclear activity in exchange for sanctions relief.
The algorithm from Orbital Insight allowed her to sift through results and accelerate the analysis process by orders of magnitude, bringing a process that took tens of hours down to less than an hour, she told Defense One on Wednesday.
Other photos show increased activity in building uranium storage containers at the nearby Pyongsan uranium concentrate plant, about a half of a kilometer away. But, said Park, activity at the concentration plant is well below capacity, meaning that the North Koreans still have a lot of room to scale up production.
Park’s future research will look at other indicators of activity, such as the use of rail cars, since material will come into and out of the plant by rail. Different cars, of course, also bring different materials, such as pressurized containers for dangerous substances and coal cars to bring in fuel to run the facility.
This presents a different challenge in terms of satellite images. Pictures clearly indicating changes of land use are easier to come by, especially with the proliferation of low earth orbit satellite constellations that can visit a place once a day. But rail cars come and go, which means you need more images, spanning different times of day to catch them.
Says Park, AI could help analysts here as well. “The ongoing project will look at how many rail cars are coming in and out of this facility. And I can assure you in at least one out of every two images I look at, there are always railcars. Instead of me individually counting, ‘One, two, three railcars,’ the algorithm will help pick up all.”
Olivia Koski, solutions engineer at Orbital Insight, said the company is moving ahead with efforts to not just identify broad categories of objects like railcars, but also offer more granular and specific insights. The company already has a multi-class aircraft detector and they are working to release a multi-classification ship detector, which would alert analysts not only to the presence of ships but also the type of ship present.
Collecting intelligence data on North Korea is notoriously difficult as the country is so isolated. The regime keeps very close tabs on North Korean daily life, especially for any individual that might have information of use to Western powers. That leaves outside analysts with satellite photos. But the vast number of satellite photos available can make it difficult for analysts to know what to pay attention to or how to track changes over time.
One major challenge in training algorithms to detect specific types of hardware is a lack of good satellite photos of that hardware to train algorithms, since it might take hundreds of thousands of photos for an algorithm to learn to recognize different objects. Dan Soller, the senior advisor on national security programs at Orbital Insight, said the company is working now with a government client to develop synthetic data, essentially taking pictures of things that are rarely seen such as ground combat equipment, and making more versions from different angles to develop a dataset capable of training algorithms.
Said Soller, the increased mining activity may not be an indicator of an effort to accelerate the development of nuclear weapons, but the North Koreans are probably hoping that the West will make such an interpretation.
“Currently there’s a food shortage going on,” he said. “North Korea is definitely feeling the pinch and so when that occurs, as you can imagine, you see activities,” such as last month’s submarine ballistic missile launch. “With the case of developing yellowcake here, they want to continue that and make it look larger than it actually is, just to send the message that they are still pursuing their interests in order to gain concessions.”
___
© 2021 Government Executive Media Group LLC
Distributed by Tribune Content Agency, LLC.
