C019-05
Supervised Classification of Slush and Open Water on Antarctic Ice Shelves using Landsat 8 Imagery
Wednesday, 9 December 2020: 04:16
Virtual
Rebecca L Dell1, Ian Willis1, Alison F Banwell2, Neil S Arnold1, Anna Ruth W Halberstadt3 and Hamish D Pritchard4, (1)University of Cambridge, Scott Polar Research Institute, Cambridge, United Kingdom, (2)University of Colorado Boulder, Cooperative Institute for Research in Environmental Sciences, Boulder, United States, (3)University of Massachusetts Amherst, Geosciences Department, Amherst, MA, United States, (4)NERC British Antarctic Survey, Cambridge, United Kingdom
Abstract:
The ponding of meltwater on an ice shelf surface has significant implications for ice shelf (in)stability and potential breakup. For example, the disintegration of the Larsen B Ice Shelf in 2002 followed the development of an extensive network of surface lakes, which resulted in rapid ice-shelf collapse via the drainage of > 2000 of the lakes by hydrofracture. While several studies have investigated the changing extent of surface lakes on ice shelves, few consider the changing extent of saturated firn (slush). It is important to do so, however, because melting and re-freezing within the firn pack over multiple seasons will reduce firn air content and porosity, eventually facilitating surface lake formation. Both firn air depletion and surface meltwater ponding may increase the susceptibility of ice shelves to hydrofracture, and therefore breakup.
We use Google Earth Engine to develop a classifier capable of accurately identifying all surface meltwater (comprising slush and open water) on a pan-Antarctic scale using the Landsat 8 satellite record. To do this, we test the classifier on six ice shelves characterised by extensive surface melt each austral summer: (1) Nivlisen, (2) Roi Baudouin, (3) Amery, (4) Shackleton, (5) Nansen, (6) George VI. Selected Landsat 8 scenes from these ice shelves are fed into a k-means clustering algorithm, and output clusters are manually interpreted, to generate suitable training classes (e.g. slush, open water, blue ice and dirty ice). These classes are subsequently used to train a Random Forest classifier to identify slush and open water. The performance of the classifier is assessed using expert elicitation, whereby a group of ‘experts’ is asked to manually interpret 200 pixels per validation scene. Early results show high accuracy levels across all training sites, with the Random Forest classifier successfully identifying both slush and open water on a pan-Antarctic scale, through the full austral summer.
