An investigation into pixel- and object-based classification approaches for the detection of dolines on the surface of Antarctic ice shelves

Ice dolines are distinctive bowl-shaped depressions observed on many Antarctic ice shelves. They are interpreted as the results of meltwater ponds at or near the surface draining into or through the shelf column. Dolines are characterised by a typical morphology of a raised rim surrounding the central doline basin caused by hydrostatic rebound following the drainage of the original lake. The first descriptions of ice dolines were made during the British Graham Land Expedition in 1936 and since the start of the satellite era, dolines have been observed on many Antarctic ice shelves. They act as markers of past lake drainage events, allowing focused evaluation of the evolution of supraglacial and sub-surface ponds and their drainage in specific places. Lake drainage can be facilitated by overflow and lateral channel incision or through the opening of fissures by hydrofracture. Vertical, meltwater-induced fracture propagation through the ice shelf column as well as ring fractures produced around dolines during hydrostatic rebound have been associated with catastrophic collapse events of ice shelves on the Antarctic Peninsula. The detection and investigation of dolines, therefore, holds the potential to give insight into the flexure-fracture processes that are integral to ice-shelf stability. Dolines have previously been systematically mapped on a few ice shelves; however, a continental quantification of dolines has eluded such manual mapping approaches. This study, therefore, sets out to investigate the applicability of automated image classification methods to the detection of dolines. On the basis of Sentinel-2 satellite imagery, pixel- and object-based approaches are compared. For both approaches, two random forest classifiers are trained, one based on manually sampled training data and one on a randomly generated dataset. The pixelbased approaches perform more reliably, typically detecting the inner rims of doline basins where they are illuminated by the sun. While manual evaluation was needed to establish whether the occurrence of the doline class in a classified image corresponded with the occurrence of dolines in the satellite image, the detection of two previously unreported dolines on Rio Baudouin Ice Shelf highlights the potential of employing similar classification approaches to detect doline on a continental scale. Improvements to the classification approaches in future are most likely to be made through the implementation of Digital Elevation Models in the classification workflow.

Level dc:type.qualificationlevel

Masters

Grantor dc:publisher.institution

University of Cambridge

Year dc:date.issued

2024

Author dc:creator

• Suchantke, Paula

Advisors dc:contributor.advisor

• Willis, Ian
• Dell, Rebecca

Licence

• http://purl.org/NET/rdflicense/allrightsreserved

Language dc:language

eng

DOI dc:identifier.doi

https://doi.org/10.17863/CAM.115388

OAI identifier oai:identifier

oai:www.repository.cam.ac.uk:1810/379225