Temporal Deep Learning Approach to Bedfast and Floating Thermokarst Lake Ice Mapping using SAR imagery: Old Crow Flats, Yukon, Canada

• Main Author: Shaposhnikova, Maria
• Format: Master Thesis
• Language: English
• Published: University of Waterloo 2021
• Subjects: remote sensing; lake ice; SAR; deep learning; temporal convolutional neural network; bedfast ice; floating ice; Old Crow Flats; microwave remote sensing; TempCNN; thermokarst; permafrost; Arctic; Yukon; Canada; Monitor Lake; Climate change; Ice; Old Crow; Thermokarst
• Online Access: http://hdl.handle.net/10012/17414

Lake ice is a fundamental part of the freshwater processes in cold regions and a sensitive indicator of climate change. As such, in light of the recent climate warming, monitoring of lake ice in arctic and sub-arctic regions is becoming increasingly important. Many shallow arctic lakes and ponds of thermokarst origin freeze to bed in the winter months maintaining the underlying permafrost in its frozen state. However, as air temperatures rise and precipitation increases, less lakes are expected to develop bedfast ice. In fact, a consistent decrease in maximum ice thickness has been observed over the past decades. Synthetic aperture radar (SAR) offers a unique opportunity to monitor lake ice regimes remotely. Taking advantage of the growing temporal resolution of microwave remote sensing, we proposed applying a temporal deep learning approach to lake ice regime mapping. We employed a combination of Sentinel 1, ERS 1/2, and RADARSAT 1 SAR imagery for the Old Crow Flats (OCF), Yukon, Canada to create an extensive annotated dataset of SAR time-series labeled as either bedfast ice, foating ice, or land, used to train a temporal convolutional neural network (TempCNN). The trained TempCNN, in turn, allowed to automatically map lake ice regimes over a 29-year period (1993-2021). The classi ed maps aligned well with the available fi eld measurements and Canadian Lake Ice Model (CLIMo) simulated ice thickness. Reaching a mean overall classi cation accuracy of 95.05%, the temporal deep learning approach was found promising for automated lake ice regime classi cation. Change detection tools were utilized to determine lake ice regime changes in the OCF over the past 29 years. In the view of signi cant annual variability, no consistent transition towards more foating lake ice was observed. On the contrary, the overall change indicated an extensive transition to bedfast ice caused by a growing number of catastrophic drainages within the examined time period brought on by climate warming and thermokarst processes.