It is the first time that machine learning has been used to find previously unknown craters on the Red Planet.
At one point, between March 2010 and May 2012, a meteor crossed the Martian sky and broke into pieces, hitting the planet’s surface. The resulting craters were relatively small – only 13 feet (4 meters) in diameter. The smaller the features, the harder it is to spot them Mars orbiters. But in this case – and for the first time – scientists have spotted them with a little extra help: artificial intelligence (AI).
It is a milestone for planetary scientists and AI researchers NASAFrom the Jet Propulsion Laboratory in Southern California, they worked together to develop the machine learning tool that helped make the discovery. The result offers hope both to save time and to increase the volume of results.
Typically, scientists spend hours every day studying images captured by NASA’s Mars Reconnaissance Orbiter (MRO), looking for ever-changing surface phenomena such as dust devils, avalanches, and shifting dunes. In the 14 years of the Mars orbiter, scientists have relied on MRO data to find over 1,000 new craters. They are usually first detected with the spacecraft’s Context Camera, which takes low-resolution images spanning hundreds of miles at a time.
In these images, only the signs of the explosion around an impact stand out, not individual craters, so the next step is to take a closer look with the High-Resolution Imaging Science Experiment, or HiRISE. The tool is so powerful that it can see fine details like the traces left by the Curiosity Mars rover. (The HiRISE team allows anyone, including audience members, to request specific images via their HiWish page.)
The process takes patience, and it takes a researcher about 40 minutes to carefully scan a single Context Camera image. Save time, JPL The researchers created a tool – called the automatic fresh impact crater classifier – as part of a larger JPL effort called COSMIC (Capturing Onboard Summarization to Monitor Image Change) that develops technologies for future generations of Martian orbiters.
Learn the landscape
To train the crater classifier, the researchers fed 6,830 images of the Context Camera, including those of previously discovered impact locations that had already been confirmed via HiRISE. The tool was also fed with no new impact images to show the classifier what not to look for.
Once trained, the classifier was deployed across the entire repository of approximately 112,000 Context Camera images. Running on a supercomputer cluster at JPL made up of dozens of high-performance computers capable of working in concert with each other, a process that takes 40 minutes for a human, and the artificial intelligence tool requires in averages only five seconds.
One challenge was figuring out how to run up to 750 copies of the classifier across the cluster at one time, said JPL computer scientist Gary Doran. “It would not be possible to process over 112,000 images in a reasonable amount of time without spreading the work across many computers,” said Doran. “The strategy is to break the problem down into smaller parts that can be solved in parallel.”
But despite all that computing power, the classifier still requires a human to check his work.
“Artificial intelligence can’t do the kind of expert analysis a scientist can do,” said JPL computer scientist Kiri Wagstaff. “But tools like this new algorithm can be their assistants. This opens the way for an exciting symbiosis of human and artificial intelligence “researchers” working together to accelerate scientific discovery. “
On August 26, 2020, HiRISE confirmed that a dark spot detected by the classifier in a region called Noctis Fossae was actually the crater cluster. The team has already submitted more than 20 additional HiRISE candidates for verification.
Although this crater classifier works on Earth-bound computers, the ultimate goal is to develop similar classifiers tailored for use on board by future Martian orbiters. Right now, data sent to Earth requires scientists to sift through for interesting images, much like trying to find a needle in a haystack, said Michael Munje, a graduate student at Georgia Tech who worked on the classifier as an intern at JPL. .
“The hope is that in the future, artificial intelligence can prioritize orbital images that scientists are most likely to be interested in,” Munje said.
Ingrid Daubar, a JPL and Brown University scientist who was also involved in the work, hopes the new tool will offer a more complete picture of how often meteors hit Mars and reveal even small impacts in areas where they are not. been discovered before. The more craters found, the more scientists add the size, shape, and frequency of meteorite impacts to Mars to the body of knowledge.
“There are probably many other impacts that we haven’t found yet,” he said. “This advancement shows you what you can do with veteran missions like MRO using modern analytical techniques.”