2023 CSDMS meeting-110


(if you haven't already)

Log in (or create account for non-CSDMS members)
Forgot username? Search or email:CSDMSweb@colorado.edu

Browse  abstracts

A deep learning framework for linear dune mapping and pattern classification

Maike Nowatzki, University of Oxford Oxford , United Kingdom. maike.nowatzki@ouce.ox.ac.uk
David Thomas, University of Oxford Oxford , United Kingdom. david.thomas@ouce.ox.ac.uk
Richard Bailey, University of Oxford Oxford , United Kingdom. richard.bailey@ouce.ox.ac.uk

Morphological patterns reflect climatic and geomorphic influences throughout a dunefield’s history and can thus be a valuable source for information about past and present environmental conditions. The quantitative assessment and interpretation of such patterns requires precise information about dune locations and arrangements, i.e., dune maps. Mapping dunes based on satellite data has evolved from a simple tool for dryland research to a notable research area. Globally available datasets and the progression of computational infrastructure have facilitated the operation of increasingly elaborate automated algorithms to map spatially extensive areas where manual approaches would be inefficient.

We present a deep learning framework that employs semantic segmentation techniques on optical satellite imagery and medium resolution digital elevation models to map linear dunefields. The workflow includes the access and pre-processing of training and prediction data, with a Neural Network as the centrepiece that is trained and applied to identify dune crests. We conducted a preliminary case study to develop and evaluate the framework on the dunefields of the Kalahari Desert, producing promising results.

Our next step is to leverage the generated dune maps to classify different dune patterns and investigate their relationship with climate and topography. We hope to provide valuable insights into the complex interplay between dunefield morphology and its environmental and climatic drivers.