Application of a deep learning framework to explore transfer learning for dune mapping across regions

Dune mapping is a traditional task for aeolian geomorphologists and has made use of satellite imagery since the 1970s (Breed & Grow, 1979). Labour-intensive manual mapping approaches are increasingly substituted by (semi-)automated ones that apply progressive Machine Learning algorithms (Zheng et al., 2022). Advanced techniques such as neural networks enable the creation of powerful computational models to automatically map dune fields (Shumack et al., 2020; Rubanenko et al., 2021). Globally available satellite imagery datasets and the progression of computational infrastructure and power facilitate the operation of increasingly elaborate models and their application to spatially extensive regions. A lack of training and validation datasets for such dune mapping models and the subjective and time-consuming nature of their creation, however, remains a challenge.

We present a framework that uses Deep Learning and different types of satellite imagery to map dune crests. It comprises automated modules to (1) retrieve and pre-process training and prediction data, (2) train a neural network (U-Net; Ronneberger et al., 2015), and (3) identify dune crests in unlabelled target areas applying the trained model. The framework has shown good performance mapping linear dunefields in the Kalahari Desert using a small training and validation dataset (130 labelled 960mx960m tiles).

Addressing the lack of global training data, we use our model to explore the possibilities of transfer learning and the universality of regional training datasets. In our main case study, we assess whether a model trained on satellite data of linear dunes in the Kalahari can be applied to map linear dunes in regions containing morphologically similar dunes in the Australian deserts.

 

Breed, C. S., & Grow, T. (1979). Morphology and distribution of dunes in sand seas observed by remote sensing. A study of global sand seas, 1052, 253-302.

Ronneberger, O., Fischer, P., & Brox, T. (2015, October). U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention (pp. 234-241). Springer, Cham.

Rubanenko, L., Pérez-López, S., Schull, J., & Lapôtre, M. G. (2021). Automatic Detection and Segmentation of Barchan Dunes on Mars and Earth Using a Convolutional Neural Network. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 14, 9364-9371.

Shumack, S., Hesse, P., & Farebrother, W. (2020). Deep learning for dune pattern mapping with the AW3D30 global surface model. Earth Surface Processes and Landforms, 45(11), 2417-2431.

Zheng, Z., Du, S., Taubenböck, H., & Zhang, X. (2022). Remote sensing techniques in the investigation of aeolian sand dunes: A review of recent advances. Remote Sensing of Environment, 271, 112913.