Glacial Landform Classification with Vision Transformer and Digital Elevation Model

Classification of glacial landforms is a task in geomorphology that has not been widely explored with deep neural network methods. This study uses Vision Transformer (ViT) architecture to classify glacial landforms using Digital Elevation Model (DEM) in three study sites: Elise Glacier in Svalbard, Norway; Gardno-Leba Plain and Lubawa Upland in Poland. In datasets each of those sites has different DEM resolutions and terrain types which includes end moraines, hummocky moraines, kettle holes, outwash/glaciolacustrine plains, till plains and valleys. The results of the classification show that ViT architecture is a suitable method for this type of task and can achieve up to 97.5% of accuracy. The classification process described in this study can be reproducible and applied to other terrain types around the world.