AI-enhanced categorisation and analysis of dune patterns in the south-western Kalahari

Understanding dune patterning on a landscape scale provides key insights into the controls shaping dune fields and requires robust methods that capture morphological patterns in a geomorphologically meaningful manner. This study presents a data-driven approach for the creation of dune pattern classification schemes and their mapping. We apply dimensionality reduction and clustering to 198 randomly distributed dune crest tiles to derive morphometric-based classification schemes for the planimetric dune patterns across the south-western Kalahari dune field. A convolutional neural network is trained on two newly generated and a foundational classification scheme to map pattern distribution across the study area. Maps based on the new schemes presented here reproduced some broad spatial trends from previous work, like the shift from straight, parallel ridges in the north to dendritic patterns in the south. We further characterise the defining morphometrics and morphometric signature of classification schemes and pattern types, finding that orientation range is the dominant discriminating morphometric. Based on model uncertainty, we locate transitional zones between patterns in the study area. We argue that these are the most geomorphologically meaningful product of the classification: their recurrence across schemes suggests that they may reflect changes in boundary conditions that have shaped dune morphology.