Planets in our solar system are frequently pelted by random space rocks as they orbit the Sun. Some of these impacts are catastrophic, but others leave relatively small craters that are tricky for even the most keen-eyed scientists to spot.
For the first time ever, a machine learning tool has discovered this type of subtle crater formation on Mars, according to NASA’s Jet Propulsion Laboratory (JPL). The AI tool, which is called “an automated fresh impact crater classifier,” found a previously unknown cluster of Martian craters in images captured by NASA’s Mars Reconnaissance Orbiter (MRO). The new detection indicates that there might be a whole host of craters on Mars that we currently can’t see.
The craters are only about four meters (13 feet) in diameter, and were likely created by debris from a meteor that exploded over the Noctis Fossae region of Mars sometime between 2010 and 2012. Though the feature has been in MRO’s dataset for years, it have been overlooked by humans until now.
“The goal here is really to be able to get a sense of where these [craters] appear on the surface,” said Kiri Wagstaff, a computer scientist at JPL, in a call. “The existing catalog that we have, of about 1,000 of them, was generated by manually reviewing images and it shows a very heterogeneous distribution across the surface of Mars.”
In other words, it “appears that they are concentrated in certain areas and we know that physically there is no reason for that,” she added. “Meteors should hit everywhere with equal probability, and that’s what’s giving us a sense that we are missing a whole lot of them and we’d love to get a more complete catalog.”
These hidden craters may be tougher to spot for a variety of reasons, but one of the most important is the nature of the Martian surface environment surrounding the impacts. Craters that form in dusty reflective regions of the planet are easier to pick out in manual reviews because of the contrast between the dark impact zone and the lighter surface.
“If the same impact happened on exposed bedrock where there’s not that bright dust, probably it’s just a lot harder to see because it would be dark-on-dark,” Wagstaff said.
In other words, there is likely a sampling bias of Martian craters that obscures the planet’s full impact history. The new AI classifier may have already expressed a special talent for evening out our lopsided observations of craters.
The tool is trained to scan through tens of thousands of low-resolution images taken by the MRO’s Context Camera, and to flag potential crater candidates so that humans can follow up using sharper images taken by the orbiter’s HiRISE instrument. So far, the classifier has been finding many candidate craters that were missed by human review over the past several years.
The “big question,” Wagstaff said, is if there is “something different about the ones the machine learning is finding that were not already known.”
“Are they smaller or fainter or do they have weird shapes or properties, or are they on darker surfaces?” she continued. “It could be that we are finding a subpopulation of these craters that was not previously known because it was overlooked for some reason.”
Wagstaff and her colleagues have already requested HiRISE observations of nearly 70 other potential craters that the classifier has earmarked in the Context Camera data. Those additional observations could help scientists figure out whether the AI tool is pinpointing a special class of Martian craters that has slipped under the radar until now.
In addition to creating a more expansive catalog of impacts on Mars, the project is an example of the complementary skill-sets that human researchers and AI tools can bring to planetary exploration.
“I like to be very cautious that people don’t think we’re saying: ‘Oh, AI is now going to make all the scientific discoveries all on its own,’” Wagstaff said. “That’s just not possible. This is definitely people working with the machines to make this happen.”