For the past For 15 years, NASA’s Mars Reconnaissance Orbiter has been orbiting the Red Planet to study climate and geology. Every day, the orbiter sends back a wealth of images and other sensor data that NASA scientists have used to search for safe landing sites for rovers and to understand the spread of water ice on the planet. Of particular interest to scientists are the orbiter’s crater photos, which provide a glimpse into the planet’s deep history. NASA engineers are still on a mission to return samples from Mars; without the rocks that will help them remotely calibrate satellite data with the conditions on the surface, they have to do a lot of educated guesswork when it comes to determining the age and composition of each crater.
For now, they need other ways to disseminate that information. A proven method is to extrapolate the age of the oldest craters from the features of the planet’s newest. Because scientists can know the age of some recent impact sites within a few years – or even weeks -, they can use them as a basis to determine the age and composition of much older craters. The problem is finding them. Sifting through a planet’s image data for the telltale signs of new impact is tedious work, but it’s just the kind of problem an AI needed to solve.
Late last year, NASA researchers used a machine-learning algorithm to discover fresh Martian craters for the first time. The AI discovered dozens of them hiding in image data from the Mars Reconnaissance Orbiter and revealed a promising new way to study planets in our solar system. “From a scientific standpoint, that’s exciting because it increases our understanding of those functions,” said Kiri Wagstaff, a computer scientist at NASA’s Jet Propulsion Laboratory and one of the leaders of the research team. “The data was always there, but we hadn’t seen it ourselves.”
The Mars Reconnaissance Orbiter carries three cameras, but Wagstaff and her colleagues trained their AI using only images from the Context and HiRISE imagers. Context is a relatively low resolution grayscale camera, while HiRISE uses the largest mirror telescope ever sent into deep space to produce images with a resolution roughly three times higher than the images used on Google Maps.
First, the AI was given nearly 7,000 orbiter photos of Mars – some with previously discovered craters and others without – to teach the algorithm how to detect a new attack. After the classifier was able to accurately detect craters in the training set, Wagstaff and her team loaded the algorithm onto a Jet Propulsion Laboratory supercomputer and used it to comb through a database of more than 112,000 images from the orbiter.
“There is nothing new about the underlying machine learning technology,” Wagstaff says. “We used a fairly standard convolutional network to analyze the image data, but it is still a challenge to scale this out. That was one of the things we had to struggle with here. “
The most recent craters on Mars are small and may be only a few feet wide, which means they appear as dark grainy spots on context images. If the algorithm compares the image of the candidate crater with a previous photo from the same area and finds that the dark spot is missing, chances are a new crater has been found. The date of the earlier image also helps establish the timeline for when the impact occurred.
After the AI identified some promising candidates, NASA researchers were able to make some follow-up observations with the orbiter’s high-resolution camera to confirm that the craters really existed. Last August, the team got its first confirmation when the orbiter photographed a cluster of craters identified by the algorithm. It was the first time an AI had discovered a crater on another planet. “There was no guarantee that new things would come,” Wagstaff says. “But there were many, and one of our big questions is why are they harder to find?”