Abstract
Classification of open water fraction (OWF) from synthetic aperture radar (SAR) images in the marginal ice zone can be significantly difficult during the summer months, where melt-onset can alter the backscatter and melt ponds contaminate OWF estimates. In this dissertation, we explore five different machine learning algorithms including Neural Networks, Linear Support Vector Machines, Naïve Bayes, K-Nearest Neighbor and Discriminant Analysis to quantify OWF using TerraSAR-X Stripmap images during the boreal summer of 2014. To validate our methods, we use nearly-coincident high resolution panchromatic optical images. We find that overall, the classification algorithms attained comparable accuracies, however the Naïve Bayes achieved the fastest computation time. Faster computation can be very practical for users on vessels wishing to have accurate “on-the-fly” methods to calculate ice/water from SAR for navigational purposes and for modelers working with near real-time ice forecasting. We also present a prototype algorithm using linear support vector machines designed to quantify the evolution of melt pond fraction from the optical dataset in an area where several in-situ instruments were deployed by the British Antarctic Survey and the Marginal Ice Zone Program, during April-September 2014. We explore both the temporal evolution of melt ponds and spatial statistics such as pond fraction, pond area, and pond number density, to name a few. We also introduce a linear regression model that can potentially be used to estimate average pond area by ingesting several melt pond statistics and shape parameters.