[ESAC Seminar] Max Filter

“Decoding Bedrock Erodibility from River Patterns”

Abstract: Landscape evolution models make it possible to explore the geologic processes that give rise to the variety of landscapes present on Earth and other terrestrial bodies. But, to accurately model the future evolution of an existing landscape, maps of bedrock erodibility are necessary. Currently, these maps are not generally available on Earth. In this work, we seek to develop a technique that is applicable not only to Earth, but also to other terrestrial bodies where accurate topography maps are not available. We hypothesize that binary river networks, which can be generated from elevation maps or satellite imagery, are sufficient for predicting bedrock erodibility of a landscape. River networks are often the most visible feature on a landscape and the patterns they take are determined by the erodibility of the underlying bedrock. We take a deep learning approach to learn a mapping between binary river networks and erodibility maps, combining a pre-trained DINOv2 backbone with a patch-level CNN classifier. Our training dataset consists of binary river network and erodibility image pairs generated by a landscape evolution model. We use the trained model to infer erodibility of the Tennessee River Basin and find that it identifies known contrasts in erodibility.

ESAC Student Seminar Series —

A forum for students and postdocs to share recent research, hone presentation skills, and build community among peers, sponsored by the EAPS Student Advisory Committee. Open to current EAPS graduate and undergraduate students and postdocs. Typically hosted on Thursdays during the semester, including pizza lunch.

Contact: esac.officers@gmail.com