The Earth’s satellite is littered with craters of various sizes and ages. So far, however, there is no complete map of all moon craters – there are simply too many. That is why scientists have now turned to artificial intelligence to help. After training with marked recordings, the adaptive system independently scanned recordings of the two Chinese lunar probes Chang’e 1 and Chang’e 2. The result is a mapping that includes not only the known but also more than 109,000 previously unrecognized lunar craters. According to the researchers, such AI systems could also provide valuable support for mapping other celestial bodies in the future.
The lunar craters are among the most striking features of the earth’s satellite. Because its entire surface is littered with these relics of past impacts – and new ones are always being added. Because like all celestial bodies in the inner solar system, the moon is exposed to a constant bombardment of larger and smaller fragments from space. While the dense earth atmosphere only allows the largest meteorites to reach the earth’s surface and the craters are quickly eroded or overgrown by vegetation, this is different with the moon. Because it has no atmosphere, even small meteorite hits leave their traces and these can even be preserved for billions of years. “The craters are therefore the lunar equivalent of fossils, they conserve the history of the solar system,” explain Chen Yang from Jilin University in Changchun, China, and her colleagues.
Neural network as a mapping aid
But mapping the lunar craters is not easy – if only because of their sheer number. In addition, the lunar craters can be shaped very variably and, depending on age, are sometimes more or less strongly changed and covered by more recent impacts. In addition, they cover an enormous range of diameters, ranging from a few meters to impact basins hundreds of kilometers in diameter. All of this makes it difficult to clearly identify the craters and, above all, makes an automated evaluation of space probe images of the lunar surface very difficult. The existing databases of lunar craters therefore contain very different and sometimes contradicting information on the total number of lunar craters. For this reason, among other things, the International Astronomical Union (IAU) has only officially recognized 9137 lunar impact craters since 1919, as Yang and his colleagues explain.
In search of a way to improve and simplify lunar crater mapping, the researchers have now turned to artificial intelligence – an adaptive system based on a neural network. As training material, this system initially received a good 5600 images of the lunar surface in three different resolutions, which had been created by the Chinese lunar probes Chang’e 1 and Chang’e 2. The AI also received digital terrain models for the same areas. In the training phase, the craters were initially marked so that the system could learn which features characterize lunar craters. For the actual mapping, the program then received additional images, which it should now map independently.
109,000 newly identified craters
The AI system detected around 117,200 craters in the images of the lunar surface, the size of which ranged from 0.9 to 5323 kilometers. Of these lunar craters, 109,000 were previously unmapped. “That is almost 15 times more craters than previously identified,” report Yang and her team. “88.14 percent of them have a diameter of less than ten kilometers.” The comparison with existing databases showed a good match for the already known craters and confirmed the “better view” of the AI especially for the smaller lunar craters: with crater sizes between one and 20 kilometers, the new mapping was systematically on top of the older data. In a supplementary analysis, the adaptive system also succeeded in correctly estimating the age of almost 19,000 larger craters based on their characteristics.
According to the scientists, this system is therefore suitable for compiling a new, more comprehensive database and mapping of the lunar craters, especially the equatorial and mid-latitudes of the moon. “The principle could also be adapted to other celestial bodies in the solar system, such as Mars, Mercury, Venus, Vesta or Ceres, and extract more semantic information from the data than the usual manual analysis methods,” write Yang and her colleagues.
Source: Chen Yan (Jilin University, Changchun, China) et al., Nature Communications, doi: 10.1038 / s41467-020-20215-y