Transverse Aeolian Ridges at the ExoMars Rover landing sites

Introduction

Apart from the Earth, no planetary body is mapped more extensively and to such fine resolution as Mars. The increasing volume of remote sensing data means we are better equipped than ever to answer the fundamental questions about the history of the planet. However, the volume of data grows much faster than the number of scientists who can use it. Machine Learning (ML) is a powerful tool for automating the analysis of ever-increasing volumes of remote sensing data.

Aeolian bedforms exhibit varied morphologies at different scales in remote sensing imagery, therefore, automated detection is a complicated problem. Linear dune fields have been successfully characterized at regional scales using edge detection on Titan from synthetic aperture radar images [1]. Within the field of Earth observation, an edge detection algorithm has been proposed that is optimized for recognizing linear dune fields in panchromatic Landsat 8 data and digital elevation models [2]. Fingerprint minutiae extraction software designed for forensic applications has also successfully detected dune crests and their bifurcations and terminations for linear dunes in the Namib Sand Sea and Strzelecki Desert, and for Transverse Aeolian Ridges (TARs) on Mars [3].

A method for mapping aeolian ripples has been demonstrated using HiRISE imagery from Gale crater [4]. Similarly to earlier studies, this uses a two-step algorithm that segments the bedforms from the surrounding terrain and then detects the crestlines [5]. This study uses the same approach but with a segmentation step that classifies bedforms according to scale and morphology as opposed to foreground-background.

The aim of this study is to create a more general bedform detector that can be applied over larger and more texturally diverse areas of Mars. Moreover, it should perform as well as classic methods employed by geologists such as manually mapping crestlines. This will be assessed in terms of orientations and crest line maps produced but also in terms of the inferred wind regime. The secondary goal of this study is to demonstrate how ML terrain classifications designed for rover navigation can be repurposed for science.

Method

A machine learning system called the Novelty or Anomaly Hunter – HiRISE (NOAH-H) has been developed to classify terrain in HiRISE images from Oxia Planum and Mawrth Vallis according to texture. It was designed to assess terrain for rover traversability but also demonstrates great potential to be used for science [6]. Each pixel of an input HiRISE image is assigned one of 14 classes. These classes represent every type of terrain that can be found at the Oxia Planum and Mawrth Vallis landing sites, summarized in table 1. Classes 8 through to 13 are the six types of ripple morphology that are recognized by NOAH-H.

1	Non-bedrock	Smooth, Featureless
2	Smooth, Lineated
3	Textured
4	Bedrock	Smooth
5	Textured
6	Rugged
7	Fractured
8	Large Ripples	Simple form, Continuous
9	Simple form, Isolated
10	Rectilinear form
11	Small Ripples	Continuous
12	Non-continuous, Bedrock substrate
13	Non-continuous, Non-bedrock substrate
14	Other Cover	Boulder fields

Table 1: Ontological classes used by NOAH-H. Large refers to decimeter scale features and small refers to meter scale features.

Class 9, "large simple form isolated ripples", corresponds to the larger-scale TARs in these regions and we use the NOAH-H output to segment the TARs from the surrounding terrain. Some of these classified regions contain more than one TAR, therefore the next step splits these into separate regions. Now that we can assume that every region corresponds to a single TAR, we calculate an orientation for each region using second order central image moments.

Planned Analysis

This method will be applied to HiRISE images already classified by NOAH-H in Oxia Planum. We will compare the spatial distribution and orientation of TARs using the proposed method with those measured from a study that measured 10,753 TARs by manually digitizing crestlines [8]. They will also be compared in terms of inferred wind regime and compared with a global climate model to see if they give the same conclusions. The next step to build on this work is to implement existing or new methods for the remaining 5 bedform classes detectable by NOAH-H, in order to make a more general bedform characterization method.

References: [1] Lucas A. et al. (2014) JGR, 41, 6093–6100. [2] Telfer M. W. et al. (2015) Aeolian Research, 19, 215-224. [3] Scuderi L. (2019) Aeolian Research, 39, 1-12. [4] Vaz D. A. and Silvestro S. (2014) Icarus, 230, 151-161. [5] Pina P. et al. (2004) LPS XXXV, Abstract #1621. [6] Barrett A. M. et al. (2022) Icarus, 371, 114701. [7] Canny J. (1986) IEEE TPAMI, PAMI-8, 6, 679-698. [8] Favaro E. A. et al. (2021) JGR Planets, 126, e2020JE006723.