For decades, scientists have been trying to find out how many craters are present on the moon. This arduous task has recently found new impetus thanks to a powerful neural network, which has recently identified several thousand of craters which were until then unknown.
Researchers Ari Silburt of Penn State University (USA) and Mohamad Ali-Dib of the University of Toronto (Canada) led a team that developed an artificial intelligence that discovered between 6,000 and 7,000 lunar craters, which had until now escaped the observations of astronomers. Identifying and inventorying the craters of the Moon is a rather difficult mission due to their immense number!
The artificial intelligence that made it possible to find these unknown craters is nothing more than a network of convolutional artificial neurons whose structure is inspired by the visual cortex of animals. In addition, this neural network is based on both machine learning and deep learning.
A first step was to train this AI to identify lunar craters more than 5 km in diameter already known by the researchers, and this on a surface covering two thirds of the moon. The images submitted to the neural network came from stocks of photographs from the work of space probes such as the Lunar Reconnaissance Orbiter, launched in 2008 to map our satellite.
According to the results published in an article on the arXiv platform, the IA has managed to identify twice as many craters as a human beings, with 42% of these same craters being new ones! In addition, some of them have a diameter of less than 5 km, which is to say that the AI found craters smaller than those studied during its training.
Finally, the scientists explain that the AI has also identified craters from images of the planet Mercury, without training. While the surface of Mercury is very different from that of the Moon, this performance indicates that the AI seems to have really understood what was an impact crater in the broad sense of the term!