Description
In this investigation a deep learning (DL) neural network was used to detect Transverse Aeolian Ridges (TARs) in High Resolution Imaging Science Experiment (HiRISE) images of the surface of Mars. TARs are decametre scale bedforms which are found ubiquitously on the surface of Mars. They consist of ridges aligned perpendicular to the direction of the prevailing peak wind. Because these features are mostly immobile in the present day they have significance as a geomorphic marker of past wind conditions.
In order to derive statistically significant data related to TAR morphometry, distribution, and orientation, it is necessary to segment them in large numbers. A HiRISE image where TARs are present will likely have 1000s to 10,000s of features. This makes manually segmentation challenging.
We developed a semi supervised pipeline for automatic retrieval of TAR features within a Geographic Information System (GIS) environment. A Mask R-CNN model segments candidate TARs. These are stored as a polygon feature class which is fed into a series of GIS tools which clean up the dataset, for example by filtering out features which are likely to be false positives based on morphometry characteristics not expected for TARs. Morphometry and orientation statistics are then computed.
The model was trained using a desktop workstation computer, using commercial off the shelf models. This demonstrates that machine learning is becoming increasingly accessible and will likely see increasing use in the field of Earth and Planetary Science.
