Graph-Based Modeling of Iceberg Dynamics from Synthetic Aperture Radar Imagery

Abstract

Understanding glacier and iceberg dynamics, such as calving, drifting, fragmentation, and melting, is critical in improving climate modeling and prediction. Synthetic Aperture Radar (SAR) has become one of the most important instruments for monitoring these dynamics, as it operates in all weather conditions, day or night, and offers a much higher revisit time compared to other optical satellites. In prior work using SAR for studying calving events, challenges include translating such large volumes of data into meaningful representations that capture both spatial and temporal information. In this work, we explore the use of isotropic graph-based representations of iceberg dynamics over time, extracted from SAR imagery. We use a Vision Graph Neural Network (ViG) architecture to transform the SAR image features into graph structures, enabling the modeling of relationships between small ice objects through dynamically updated neighbor connections. As a proof-of-concept, we use a temporal sequence of SAR images of A-81, a large iceberg that calved off the Brunt Ice Shelf in January 2023. By extracting graphs from multiple ViG blocks, we examine how spatial relationships change within the image. Our preliminary analysis focuses on qualitative visualization and limited quantitative investigation, including variations in patch size, neighborhood size, and simple neighborhood metrics. This work establishes a scalable pipeline that can be extended to include temporal graph connections and comprehensive quantitative analysis, enabling future investigation of fragment connectivity, clustering behavior, aggregation events, and neighborhood motion over time. By laying the groundwork for spatio-temporal graph-based modeling of iceberg dynamics from SAR imagery, this work supports the study of small untracked ice fragments and their contribution to overall iceberg dynamics.