Abstract:Introduction: Habitats with a taphonomic window are high priority targets for astrobiology missions to Mars [1]. However, if life existed on Mars it did not interact with its physical environment in a way that has been obvious to identify from orbit. Studying terrestrial Mars analogs can inform us of the types of features that serve as habitats with a taphonomic window on Earth, and by extension the types of features that should be sought on Mars. But what spatial resolutions are necessary to positively identify such targets? To address the problem of establishing thresholds of identification in spatial resolution, we are assessing the ability of a deep convolutional neural network (CNN) to perform semantic segmentation (classify each pixel) on images of the astrobiologically-relevant Mars-analog environment, Salar de Pajonales (SdP), in Chile (Fig. 1A). Our aim is to discover the resolutions at which the habitats at SdP are no longer identifiable by the CNN to understand if the same types of features could be identified in salt-encrusted basins on Mars. CNNs excel at many tasks related to image analysis, including image classification [2], with modern CNNs capable of outperforming human analysts [2, 3]. Our investigation requires a consistent analysis of the same terrain over many ground sampling distances (GSDs). A single human analyst would become biased after the first classification task, and many human analysts would vary in their ability to classify a scene. Therefore, a CNN is a natural choice to tackle this problem. Here we present initial results of the semantic segmentation of our highest resolution (3 cm/pixel) scene of SdP. In future work, we will evaluate the CNN performance at progressively lower GSDs to establish thresholds of identification for habitats at SdP. Geologic Background and Relevance to Mars: Salar de Pajonales (SdP) is a salt-encrusted basin in the Chilean Altiplano (25°08'29"S, 68°46'20"W, 3547m, Fig. 1A) [4]. Its polyextreme characteristics make it a suitable analog for a post-Noachian martian climate [5]. Our study site within SdP is a gypsum-dominated area that notably hosts ridges and tumuli [6] that serve as endolithic habitats [5], and a “patterned ground” unit that represents a biological soil crust [5] (Table 1, Fig. 1B). Data and Methods: We captured images of our

Mars terrain image classification using Cartesian genetic programming

Low-cost Mars Terrain Classification System Based on Coarse-grained Annotation

Self-Supervised Learning to Guide Scientifically Relevant Categorization of Martian Terrain Images

Automatic Data Processing for Space Robotics Machine Learning

Mapping "Brain Terrain" Regions on Mars using Deep Learning

Automated detection of geological landforms on Mars using Convolutional Neural Networks

Deep Learning based Semantic Segmentation for Mars Rover Terrain Classification

Rock Hunting With Martian Machine Vision

Terrain Classification Using Mars Raw Images Based on Deep Learning Algorithms with Application to Wheeled Planetary Rovers

CONVOLUTIONAL NEURAL NETWORK IMAGE CLASSIFICATION OF MARS-ANALOG TERRAIN: PRELIMINARY RESULTS AND IMPLICATIONS FOR THE SEARCH FOR LIFE ON MARS.

Mars Image Content Classification: Three Years of NASA Deployment and Recent Advances

Autonomous Martian rock image classification based on transfer deep learning methods

MARS MULTISPECTRAL IMAGE CLASSIFICATION USING MACHINE LEARNING TECHNIQUES

Deep Mars: CNN Classification of Mars Imagery for the PDS Imaging Atlas

Learning Surface Terrain Classifications from Ground Penetrating Radar

Towards Safer Planetary Exploration: A Hybrid Architecture for Terrain Traversability Analysis in Mars Rovers

Automatically Evolving Texture Image Descriptors Using the Multitree Representation in Genetic Programming Using Few Instances

Planetary Mapping Using Deep Learning: A Method to Evaluate Feature Identification Confidence Applied to Habitats in Mars-Analog Terrain

GPGM-SLAM: a Robust SLAM System for Unstructured Planetary Environments with Gaussian Process Gradient Maps

Mixed-domain Training Improves Multi-Mission Terrain Segmentation

Generalized Unsupervised Clustering of Hyperspectral Images of Geological Targets in the Near Infrared