Automatically dunes mapping and morphometric analysis using Artificial Intelligence

The surface of some planet’s present abundant periodic topographic forms at different scales (mm- km) and in different environments. They are called bedforms and develop at the interface between a moving fluid and a deformable and/or erodible material. Sand dunes are major bedforms in aeolian systems and play an important role in understanding how aeolian environments evolve. Generally grouped in dune fields, their morphological characteristics (e.g. shape, size, patterns of spatial organization) play a critical role understanding how aeolian environments evolve and interact with global changes. Detailed maps of these morphologies are produced by GIS manual digitizing from high-resolution satellite imagery and digital terrain models but this approach remains time-consuming. One way to override these locks is to use Artificial Intelligence (AI) to increase the computational speed, accuracy and reproducibility of mapping.We here present a GIS automated mapping protocol of aeolian bedforms contours and crestlines based on AI. First, we extract the Residual Relief in order to delete the regional topographic trend and map the different dune generations. Secondly, an unsupervised pixel-based classificator (Deep Learning, U-Net CNN) trained with Residual Relief samples of different dune forms is used to detect and map dunes independently of the bedrock. Thirdly, the dune crests are skeletonized from the identification of high inflection point of the dunes from a Volumetric Obscurance approach. The protocol is repeated for each dune order of magnitude to extract the different superimposed generations of dunes and their relationships.To illustrate its performance, the protocol was applied on a part of the Rub’Al Khali (220,000 km²) and Namib deserts (115,000 km²) to map the various dune forms and a first morphometric analysis is realized at the scale of the two sand seas. We produced a detailed map of the aeolian morphologies for each desert at two orders of magnitude (kilometer-scale and hectometer-scale dunes). For the Rub’Al Khali desert, more than 78,000 dunes (58,000 km²) and crestlines were mapped in 6 hours of processing and more than 17,000 dunes (12,000 km²) and crestlines for the Namib desert in the same processing time. The first morphometric parameters calculated from the previous results show a spatial variability of the length, width, height and crests orientation for Rub’Al Khali and Namib dunes. Thus, the protocol contributes to provide a digital atlas of the different dune generations (kilometer-scale and hectometer-scale dunes). All of these results allow to visualize the morphological dunes diversity and contribute to understand the relationships between forms and processes at a dune field scale.