A new workflow for mapping dune features (outline, crestline and defects) combining deep learning and skeletonization from DEM-derived data

Dune morphology, dynamics, and spatial arrangement mainly depend on the characteristics of the moving fluid (e.g. wind regime) and of the sediment supply (e.g. grain size) modulated by surface conditions (e.g. vegetation). A comprehensive mapping of dunes could provide essential and quantitative information for characterizing, analyzing, and monitoring dunes, dune patterns, and dynamics. This would help in understanding the underlying processes, from individual dunes to extensive dune fields. Here, we present a new workflow for the automated mapping of three dune features (outlines, crestlines, and defects) in GIS, based on DEM derived data (Residual Relief and Volumetric Obscurance; Copernicus 30 m DEM) and suitable for morphometric analysis. We first adapted a U-Net convolutional neural network (CNN) to map dune outlines. We then used a skeletonization tool to map dune crestlines. Crestline networks were then used to refine the individualization of coalescent dune outlines from a region-growing algorithm, while dune defects were mapped using a network topological analysis tool. Accuracy assessments demonstrate the efficiency of outline mapping from CNN (recall = 87 %, precision = 90 %, quality = 69 %) and we identified the limitations related to i) input DEM resolution, ii) training samples definition, and iii) skeletonization threshold settings. Finally, the application of the workflow to the Rub'Al Khali sand sea produced a map of 122,522 dune outlines/crestline networks and 514,083 defects, which were used for morphometric mapping and analysis (mean dune height, mean dune volume, defect density, and crestline density). Preliminary results demonstrate the robustness of the workflow to produce a large cartographic database suitable for morphometric analysis through its consistency with previous investigations while providing further uniformly detailed information. This workflow offers promising prospects for quantifying and understanding driving factors and boundary conditions, monitoring dune dynamics, from single dune to dune fields, including superimposed dune patterns of different scales where very high-resolution DEM are available. [Display omitted] •AI workflow for dune outlines, crestlines and defects automated mapping from DEM•Deep learning outlining validated with 90 % recall, 87 % precision and 69 % quality•Workflow not limited by the significant variability in dune morphologies•Application to the Rub'Al Khali produced over 100,000 dunes and 5