Abstract:
Barchan morphometric data have been used as proxies of meteorological and topographical data in environments where this data is lacking (such as other planetary bodies), gaining insights into barchan dune field dynamics such as barchan collision and sediment dynamics, and estimating migration speeds. However, manual extraction of this data is time-consuming which can impose limits on the spatial extent and temporal frequencies of observations. Combining remotely sensed big data with automated processing techniques such as convolutional neural networks (CNNs) can therefore increase the amount of data on barchan morphology. However, such techniques have not yet been applied to barchans and their efficacy remains unknown. This study addresses this issue by evaluating the classification performance (using the ACC, F 1 -score and MCC metrics) of CNNs on several different morphometric tasks: the side of horn elongation, the magnitude of elongation, the barchans a/c ratio, and a new metric, bilateral asymmetry, which takes a more holistic view of barchan asymmetry. Specifically, bilateral asymmetry offers a means by which the total points of variation on a barchan that is used in describing barchan morphology, can be expressed with a single measure. Twelve different CNN architectures, each with different hyperparameters, are trained and tested on a sample of 90 barchan dunes. Additionally, the potential of transfer learning is assessed using the VGG16 and ResNet50 architectures. The results show that the accuracy of the CNNs can exceed 80% in some cases and that “from scratch” CNNs can match the performance obtained using transfer learning approaches.
