Image Classification of Marine-Terminating Outlet Glaciers using Deep Learning Methods
A wealth of research has focused on elucidating the key controls on mass loss from the Greenland and Antarctic ice sheets in response to climate forcing, specifically in relation to the drivers of marine-terminating outlet glacier change. Despite the burgeoning availability of medium resolution sate...
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | unknown |
| Published: |
2020
|
| Subjects: | |
| Online Access: | http://etheses.dur.ac.uk/14003/ http://etheses.dur.ac.uk/14003/1/Marochov000853698_minor_corrections.pdf |
| Summary: | A wealth of research has focused on elucidating the key controls on mass loss from the Greenland and Antarctic ice sheets in response to climate forcing, specifically in relation to the drivers of marine-terminating outlet glacier change. Despite the burgeoning availability of medium resolution satellite data, the manual methods traditionally used to monitor change of marine-terminating outlet glaciers from satellite imagery are time-consuming and can be subjective, especially where a mélange of icebergs and sea-ice exists at the terminus. To address this, recent advances in deep learning applied to image processing have created a new frontier in the field of automated delineation of glacier termini. However, at this stage, there remains a paucity of research on the use of deep learning for pixel-level semantic image classification of outlet glacier environments. This project develops and tests a two-phase deep learning approach based on a well-established convolutional neural network (CNN) called VGG16 for automated classification of Sentinel-2 satellite images. The novel workflow, termed CNN-Supervised Classification (CSC), was originally developed for fluvial settings but is adapted here to produce multi-class outputs for test imagery of glacial environments containing marine-terminating outlet glaciers in eastern Greenland. Results show mean F1 scores up to 95% for in-sample test imagery and 93% for out-of-sample test imagery, with significant improvements over traditional pixel-based methods such as band ratio techniques. This demonstrates the robustness of the deep learning workflow for automated classification despite the complex characteristics of the imagery. Future research could focus on the integration of deep learning classification workflows with platforms such as Google Earth Engine (GEE), to classify imagery more efficiently and produce datasets for a range of glacial applications without the need for substantial prior experience in coding or deep learning. |
|---|