SPACESeg: Automated Detection of Bed Junction Morphologies Indicating Signs of Life in Ediacaran Period
With Perseverance out looking for life on Mars, we identify the need to equip ourselves with automated techniques for remote assessment of geological information. The first step in this translational research is studying early signs of life on Earth. More specifically, we examine the Ediacaran sedimentological record of the Flinders region in Australia, whose unique bed ripple junction morphologies have been determined as the definite indicators of early life on Earth. We propose an automated technique, SPACESeg, that robustly detects the artifact-clouded, miniature ripple structures from cross-sectional views of the Ediacaran rocks. We demonstrate the efficacy of SPACESeg in precisely extracting the desired structures with high accuracy, outperforming many techniques. We also establish the robustness of this technique as it extracts desired biosignatures from drastically varying image conditions, even when the ripples comprise of <1% of the image around significant artifacts. We provide quantitative and qualitative analysis and compare our method against many unsupervised rule-based and supervised deep learning methods, outperforming them all.