Remotely sensed imagery is a valuable source of data for studying barchan morphology. However, manual methods of data extraction constrain both the spatial and temporal resolution of studies because they are time consuming to carry out. Therefore, to effectively use the increasing availability of remotely sensed imagery, novel methods need to be developed that can extract the desired data from imagery automatically. Convolutional Neural Networks (CNNs) have shown promise in identifying landforms from imagery, but its suitability for barchan research remains untested. Since CNNs are strongly influenced by the texture of the image, it can be questioned whether the classification is based on the image’s texture (which can vary due to solar angles and atmospheric disturbances) or the geometry of the landform. Additionally, deviations in shape and other morphometric properties can manifest as subtle alterations to the barchan’s geometry. This poses a challenge for CNNs which have difficulty in distinguishing between similarly shaped landforms. Using a small sample of dunes from the Kunene region in Namibia, it is shown that CNNs can: distinguish between different morphologic classes of barchans in the absence of image texture with accuracies exceeding 80%, and distinguish between similarly shaped landfroms. When used along with methods of barchan outline extraction, a suitably trained CNN can automatically extract barchan morphologic data from remotely sensed imagery. This can increase both the temporal and spatial resolution of barchan research.
Thesis (PhD (Geography))--University of Pretoria, 2022.