Graduate Thesis Or Dissertation


The Application of Statistical Learning Techniques to Studying Arctic Sea Ice Survivability Public Deposited

Downloadable Content

Download PDF
  • Arctic sea ice extent has continued to decline in recent years, and the fractional coverage of multi-year sea ice has decreased significantly during this period. These changes impact the hardiness of the ice pack against future climate forcing, and will affect the future evolution of the Arctic ice cover. These changes will also have substantial effects on Arctic inhabitants, military operations, commercial exploration, and global climate. There exist many sources of remotely sensed Arctic data that can be used to study these changes and determine which predictors account for much of the change in the Arctic ice cover. This thesis assesses the impact of several remotely sensed sea ice parameters on the survival of sea ice in the summer melt season. A Lagrangian track-based sea ice data product that combines sea ice parcel locations with coincident satellite-derived data is described herein. This database is used in conjunction with several statistical learning classifiers to determine the optimal technique for predicting sea ice extent at the end of the melt season. These statistical learning classifiers are then used to assess which remotely sensed sea ice parameters have the greatest impact on sea ice survival for the pan-Arctic domain. These methods are further combined with airborne data from NASA’s Operation IceBridge to investigate sea ice survival in the Beaufort Sea from 2009-2016. It is shown that sea ice parcel latitude and thickness prior to the onset of melt are the most important variables in estimating parcel survival. As the melt season progresses, broadband albedo becomes the greatest predictor of summer survival. Additionally, downwelling longwave radiation is observed to contribute to melt onset and the triggering of the sea ice albedo feedback in the Beaufort Sea. Coincident airborne ice thickness and snow depth offer less conclusive results, with some years exhibiting higher mean thicknesses and depths in the melted population. The statistical learning techniques described herein are relatively underutilized methods that will prove valuable in future studies of changing predictor importance in the Arctic.
Date Issued
  • 2018
Academic Affiliation
Committee Member
Degree Grantor
Commencement Year
Last Modified
  • 2019-11-14
Resource Type
Rights Statement