Automatic crater shape retrieval using unsupervised and semi-supervised systems

Abstract

Impact craters are depressions formed due to impacts on the surface of planetary bodies. Recent deep learning-based crater detection methods assume craters as circular-shaped without much attention to extracting craters' shape and morphology. Craters' shape information (i.e., instance segmentation map for craters) can be helpful for many advanced analyses including crater formation and surface characteristics. However, publicly available ground truth catalogs for the lunar surface do not contain crater shape annotations and are also challenged by the missing craters problem. We attempt to solve these challenges by proposing a novel estimation and refinement-based approach using a combination of unsupervised and semi-supervised systems. Our method consists of (a) learning estimated crater segmentation (ECS) maps by a novel adaptive rim estimation algorithm (unsupervised system) using side information, (b) refining ECS by a cascade of Mask region-based convolutional neural networks (R-CNNs) to obtain refined crater segmentation (RCS) maps (semi-supervised system), and (c) combining RCS followed by predicting a highly accurate crater segmentation map. In the absence of any publicly available catalog for crater shape annotations, we conducted a ranking-based user study to compare against the state-of-the-art. The proposed method outperforms by achieving the best ranking for 63.53% crater images as compared to 9.67% for state-of-the-art. Further, the extracted shapes of the craters are utilized to improve the estimate of the craters' diameter, depth, and other morphological factors to be made publicly available: https://drive.google.com/drive/folders/1ghBf2FXNIJUEQkAM2GjLZNiIXKEhZMEB?usp=sharing.