Counting craters
Machine learning could help researchers date planetary surfaces by tallying up their craters
Counting the number of craters in a region of a planetary surface gives scientists important information on how old the planet is. Like specks of raindrops accumulating on a pavement, the longer a surface has been exposed, the more craters it will have. Crater densities tell you the relative age of features like lava flows or wind deposits, and if we’ve also been able to analyse a sample from the surface (as we have with lunar rocks returned by the Apollo programme and then dated precisely by measuring radioactive isotopes, for example), we can then also calculate the absolute age of whole regions.
The problem is that crater counting has traditionally been an exceedingly slow and laborious exercise; up until now it’s been mostly done by human eye. The data we’ve been able to gather this way either covers large areas of a surface taking note of only the largest craters, or it includes the smaller craters but covers only a very specific, limited geographic region. Automated methods using computer algorithms have been developed, but they can often be confused by overlapping or degraded craters, variations in illumination or other landscape features, such as ridges.
Ari Silburt at the University of Toronto and his colleagues have been trying to change all of this. They have applied a new computer technique to the challenge based on Deep Learning, which uses artificial neural networks, where the computer emulates the functioning of a simple brain. As Silburt himself explains, “Similar to how a human learns to recognise a cat by seeing many different examples of cats, a computer can learn to recognise a cat via machine learning by receiving many examples of what is and isn’t a cat.”
Silburt and his team applied their neural network to landscape maps of the Moon’s surface created by the Lunar Reconnaissance Orbiter and Kaguya probes. The immediate advantage of starting with these ‘digital elevation maps’, rather than simply photographs, is that they are not affected by varying shadows from different angles of sunlight.
They first tested their neural network on images that had already been counted by scientists to check that it worked reliably. Their technique successfully located 92 per cent of the craters identified by the human experts. Crucially, it also found a large number of new craters – almost twice as many, in fact, and in particular scores of those very small craters that are often neglected in crater counts carried out with the human eye. Silburt then applied his neural network to the planet Mercury and found that his technique also worked very well on this completely different terrain.
While this is a very promising new approach, Silburt’s crater-counting computer code still had an error rate of about 11 per cent. This isn’t too bad, but probably not quite good enough just yet to be used for completely automating the process of building accurate crater catalogues. Silburt and his team are now working on tweaking and perfecting their system. Nevertheless, this represents a very promising approach for automating the laborious process of identifying the number and sizes of craters, and so ultimately improving our understanding of how different landscapes on planets in the Solar System formed.
LEWIS DARTNELL was reading… Lunar Crater Identification via Deep Learning by Ari Silburt et al Read it online at https://arxiv.org/abs/1803.02192
“Their technique located 92 per cent of the craters identified by experts. It also found a large number of new crater"