Machine learning is uncovering the entrances to underground caves on Mars
Planetary scientists get particularly excited about finding caves on Mars. Such subterranean structures offer very promising locations for establishing human habitats – their rock ceilings would provide protection from dust storms and micrometeoroid impacts, as well as shielding from cosmic radiation. They also represent enticing targets for searching for signs of simple martian life, as caves on Earth often sustain abundant microbial growth.
Lava tubes are one promising source of Martian caves. These tubes form when a stream of volcanic lava develops a hardened crust on top, but continues to flow within, then empties to leave behind a hollow tunnel beneath the surface. On Earth, they range from just a few metres long to the enormous 65km-long Kazumura Cave in Hawaii, and they can be up to 30 metres in diameter.
The main way that lava tubes are identified, and which also provides access down into them, is when a portion of the ceiling collapses to create a skylight. Due to the weaker gravity on Mars, these potential cave entrances (PCEs) can be significantly larger than on Earth, and the lack of covering vegetation makes spotting them in satellite imagery relatively easy.
Traditionally, though, identifying PCEs on Mars – and indeed any other surface feature – has been very slow and laborious, with researchers having to eyeball thousands of images. Thomas Watson and James Baldini, both in the Department of Earth Sciences at Durham University, have been working on automating this process using machine learning.
Tunnel vision
They developed a computer system known as an artificial neutral network, simulating a structure loosely based on mammalian brains. They then trained this system to recognise lava tube openings by showing it lots of examples of already-identified PCEs, and then used it to process other images of the martian surface. They focused on regions that have had lots of vol
