Deep Learning for Boulder Detection on Planetary Surfaces
Abstract
On the surface of rocky planetary bodies, boulders form predominantly as rock fragments ejected upon meteor impact and deposited around impact craters. Some of those fragments are ejected with sufficient kinetic energy to create secondary craters, therefore increasing the density of impact craters and, in doing so, biasing surface ages estimated from crater counting. Thus, impact-generated boulders offer a unique opportunity to correct for such biases in planetary surface age determinations. However, hundreds of thousands of resolvable boulders are often associated with a single impact structure, rendering manual mapping of individual boulders highly impractical. As a result, automating boulder detection is key to obtaining statistically significant, high-quality boulder statistics over large and varied planetary surfaces. Here, we use convolutional neural networks (CNN) to automatically detect boulders in satellite imagery of the Moon and Mars. To train the algorithm, we have collected more than 35,000 outlines of boulders to date. Our labeling strategy involved mapping boulder populations around impact structures with different levels of crater degradation and terrain types to account for potential variations in boulder appearance. The preliminary trained model setup (Mask R-CNN with default "vanilla" hyperparameters) gives a mean average precision of 25%. Differences in boulder size, shape, and orientation distributions between automatically and manually outlined boulders are statistically insignificant, and therefore give promising results. At the conference, we will report on a comparison between different model architectures, extensive hyperparameters search, and the use of augmentation techniques and additional boulder labels on boulder detection. The best model will then be used to automatically detect and characterize boulder populations around impact structures on the lunar and martian surfaces as a first step towards a better understanding of impact-driven boulder ejection.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFM.P23A..02P