C004-0015
2020 Dickinson Sydkap Hackweek: Using a U-Net for Semantic Segmentation of Glaciers on Southern Ellesmere and Western Devon Islands, Canada
Monday, 7 December 2020
Poster
Will Bresnahan1, William Hardy Kochtitzky2, Benjamin R Edwards3, Hiba Aoid4, Luke Copland5, Isabel Ruff4, Celine Nour Ghattas Smith6, Patrick Kissell Noonan4 and Kaelan Felknor-Edwards6, (1)Hamilton College, Clinton, MA, United States, (2)University of Ottawa, Ottawa, Canada, (3)University of East Anglia, Carlisle, PA, United States, (4)Dickinson College, Carlisle, PA, United States, (5)University of Ottawa, Ottawa, ON, Canada, (6)Carleton College, Northfield, MN, United States
Abstract:
Quantifying glacier area change is a key metric that is crucial to understand changes in water supply and natural hazards around the world. Presently, few methods exist to accurately automatically extract glacier outlines. Here, we present a method to automatically delineate glaciers using a U-Net trained on manually delineated glacier outlines. Bands 1, 2, 3, 4, 5, and 7 from Landsat 7 and 8 are preprocessed and divided into 512x512 pixel overlapping tiles. The model predicts masks for where the glaciers are in the tiles, and the output is stitched back together. This process takes about 5 minutes running on a NVIDIA GeForce GTX 1080 Ti for a single 185 x 185 km Landsat image. We applied our algorithm to glaciers on southern Ellesmere and western Devon Islands using Landsat imagery from 1999 to 2020. We additionally manually digitized all these glaciers to compare the performance of our machine learning algorithm to manual delineation. The model is capable of identifying glaciers well enough that only a small amount of manual correction (less than 5 minutes per scene) is required.
Early results show the intersection over union on validation data is greater than 0.9 and binary cross entropy is around 0.15. We compared results from our machine learning model to manual digitization and found the areas of the glaciers to be within less than 1% of each other, well below the uncertainty of either method. Although this model struggles with identifying medial moraines and distinguishing sea ice and fog on lakes from glacial ice, it allows for rapid digitization of glacier outlines. The algorithm can therefore be applied to glaciers across Arctic Canada to rapidly and accurately quantify glacier change at higher temporal resolution and over longer periods of time than has previously been possible. With additional training data, this algorithm could be applied to other locations around the world.
