P008-06
Morphologic Analysis of Eolian Bedforms on Mars using Fully Convolutional Instance Segmentation Networks
Monday, 7 December 2020: 05:50
Virtual
Lior Rubanenko1, Mathieu Gaetan Andre Lapotre2, Joseph Schull1, Lori K Fenton3 and Ryan Ewing4, (1)Stanford University, Stanford, CA, United States, (2)Stanford University, Geological Sciences, Stanford, CA, United States, (3)Carl Sagan Center, SETI Inst., Mountain View, CA, United States, (4)Texas A&M, Department of Geology and Geophysics, College Station, TX, United States
Abstract:
The prevalence of eolian landforms on Mars is a testament to the past and present atmospheric conditions on the red planet. Deciphering the morphologic information encoded in these landforms could help characterize local and global wind circulation patterns and atmospheric history. Here we use an instance segmentation neural network to analyze the global distribution and morphometrics of two types of eolian features: Barchan dunes and Transverse Aeolian Ridges (TARs). Barchan dunes are crescent-shaped windblown bedforms that form when sand supply is limited and winds are approximately unidirectional. TARs are another ubiquitous eolian landform on Mars whose formation mechanism is currently poorly understood. Previously, the morphology of dunes mapped on Mars by orbiting spacecraft was used to estimate past and present-day wind patterns. However, due to the great effort in extracting and analyzing individual images, these surveys were mostly focused on localized regions.
For our analysis, we employ a fully convolutional instance segmentation neural network to map and outline barchan dunes and TARs on Mars on a global scale. The network was trained on samples derived from a new global catalog of dune fields on Mars, employing images obtained by the Mars Reconnaissance Orbiter Context Camera (MRO CTX). Unlike object-detection networks, instance segmentation allows to simultaneously predict the object class, position, and shape, and requires less extensive labeling compared with semantic segmentation networks. For barchans, our model detects approximately 80% of the dunes in a single field, with a false positive rate <5%. We expect this accuracy to increase as more training samples are labeled.
Unlike previous catalogs, which characterized dune fields, our database will include the geographic location of individual dunes, along with their orientation and typical wavelength. Using this dataset, we intend to infer local and global wind circulation patterns.
