Olivia Patterson Successfully Defended Her Master’s Thesis On Tuesday, November 17, 2026
Congratulations Olivia!
Olivia Patterson successfully defended on Tuesday, November 17, 2026
Olivia first joined iHARP as a Data Science & AI Scholar and continued her work as a Master’s student after she graduated from college. It’s been truly wonderful to see your growth over the years, and we cannot wait to see what you achieve in the future. Congratulations Olivia!
___________________________
Title: Data-Driven Representations of Antarctic Icebergs
Committee:
- Dr. Rebecca Williams, Chair/Advisor, UMBC CSEE
- Dr. Vandana Janeja, committee member, UMBC IS
- Dr. Don Engel, committee member, UMBC CSEE
Abstract:
Understanding iceberg dynamics supports climate science and our understanding of polar systems. This thesis explores data-driven representations of Antarctic icebergs through an educational game, graph-based analysis of Synthetic Aperture Radar (SAR) imagery, patch-level classification, and a comparison of fixed-square patch-based and superpixel-based graphs. These representations support both scientific analysis and public understanding of iceberg dynamics.
This research started with Glacial Guardians, an educational video game that combines tracking data and published observations of iceberg A68 with scientific storytelling and gameplay. Players explore its movement and fragmentation in an engaging virtual world grounded in real-life data. Available records do not capture every small fragment, motivating us to examine how these fragments are represented in synthetic aperture radar (SAR) imagery. We used a Vision Graph Neural Network (ViG) to connect square image patches through learned feature similarity. Using nine Sentinel-1 images of iceberg A-81, we examined graph neighborhoods across patch sizes and network depths. Early and mid-depth blocks better preserved spatial proximity, mutual connections, and consistency with ice and ocean labels than deeper blocks. Building on this work, we adapted ViG for binary and four-class patch-level classification. The binary task distinguishes ice from ocean, while the four-class task also considers boundaries and small ice fragments. Direct fine-tuning achieved 86.5% binary accuracy, outperforming reconstruction-based self-supervised pretraining followed by fine-tuning.
Both approaches struggle with boundary regions and small fragments, highlighting the challenge of how these regions are represented. Our superpixel study compares fixed square patches with regions that adapt to image content. We examine the differences in spatial relationships, ice/ocean label consistency between their graphs, and whether superpixels better capture small fragments. Together, these studies connect a data-driven game simulation of iceberg A68’s lifecycles with graph-based analysis and patch-level classification of SAR imagery. This thesis establishes a foundation for improving representation of small ice fragments and understanding their role in overall iceberg dynamics.