Automatic Detection, Segmentation, and Tracking of Barchan Dunes in Satellite Optical Images Using Deep Learning

Barchan dunes are prevalent features in planetary systems. Compared to other wind-blown bedforms, they are distinguished by their crescent form and variable sizes. While they archive information on environmental conditions over considerable timescales, their isolated small sizes also make them responsive to changes in wind regimes, and vegetation cover. Observing their dynamics can be inform studies of heterogenous eolian deposits. The dune boundaries are used to measure different morphometric parameters and correlate to local environmental parameters, in addition to their position in the landscape. We use a pre-trained Mask R-CNN model to perform instance segmentation (i.e. detection and segmentation) on barchan dunes in satellite images of Qatar. Because the model segmentation is not precise enough for our study, we needed to find a method to improve the dune delineation. We first tried “active contours” which are techniques that iteratively deform the predicted dune polygons to fit the dune edges better. However, they fail to conform to the dune boundaries due to variable image illumination, dune edges with inconsistent gradient magnitudes and low contrast in radiance between the dunes and the bed. Therefore, we experiment with the recently released “Segment-Anything” model to improve the segmentation of the dunes. A test run on barchan dunes in Qatar and Saudi Arabia yielded fair agreement between dune dimensions measured manually and automatically. Tracking sand dunes in modern arid environments with different conditions would give clues about timescales of aggradation/lateral heterogeneity, especially in mixed fluvial-eolian deposits.