Martian cave detection via machine learning coupled with visible light imagery

Abstract

Cave entrances on Mars, formed from lava tube collapses, are key to the future exploration of the planet. They represent valuable exploration targets for signs of life and could provide shelter for prospective human endeavours. In this survey, a convolutional neural network (CNN), trained to identify potential cave entrances (PCEs) from images of the Martian surface, is used to locate new potential caves. Five regions were targeted for search, totalling around 1.1% of the Martian surface. Region A (latitude: -20o N to -8o N, longitude: -120o E to -132o E, centre: -14o N, −126o E), Region B (latitude: 0o N to -12o N, longitude: -108o E to -120o E, centre: -6o N, −114o E), Region C (latitude: 0o N to 12o N, longitude: -100o E to -112o E, centre: 6o N, −106o E), Region D (latitude: -28o N to -40o N, longitude: 88o E to 100o E, centre: -34o N, 94o E) and Region E (latitude: 20o N to 32o N, longitude: 140o E to 152o E, centre: 26o N, 146o E). Each region selected either contains a high abundance of previously identified PCEs or is known to contain volcanic surface features. The network identified 61 new and 24 previously identified PCEs out of 10,834 positive outputs (0.78%). This accounted for ∼7.0% of the 341 previously identified PCEs present in the five regions surveyed. Four newly identified PCEs are highlighted as promising candidates for future research, including a very large (∼700 m diameter) PCE (‘Marvin’; following previous convention, PCEs are informally named for ease of reference (Cushing et al., 2007)), as well as a PCE whose low altitude could enable exploration via remote controlled drone (‘Emily’). Twelve PCE-dense sub-regions capable of facilitating rapid exploration were also identified. Of these, sub-region B2 contains the largest number of PCEs suggesting the most promise for future research. Overall, the network's 0.78% success rate of PCE detection is approximately 37 times more effective than random selection of locations (estimated as a 0.02% chance of detection). This suggests that there is great potential for cave discovery with this method, although improvements in the size and quality of the training dataset are required prior to planet-wide application. Advancements in current exploration technologies are also necessary before confirming any PCE identified as an actual cave entrance.