Icebergs, Genealogy and Jigsaw Puzzles

Accurate estimates of iceberg populations, disintegration rates and iceberg movement are essential to understand ice sheet contributions to global sea level change, effects of freshwater inputs on ocean circulations and heat balances. Furthermore, there are operational imperatives to predict iceberg drift and fragmentation in order to ensure the safety and efficiency of polar shipping. The dynamics, persistence, fragmentation rates, melt rates and dispersal of icebergs are, however, poorly understood due to a lack of automated approaches for monitoring them.

We present an automated iceberg tracking approach that is capable of reconstructing iceberg paths, fragmentations and ultimately lineages through multiple generations based on satellite radar imagery. The method offers scope for the first time to relate iceberg fragments back to their original source computationally, which will allow scalable deployment and the development of improved predictive iceberg drift and disintegration models and a better understanding of contributions to nutrient and freshwater distributions. 

Tracking is developed using the Canadian Ice Island Drift, Deterioration and Detection (CI2D3) database. This contains manually-delineated observations of large tabular icebergs in the Canadian Arctic between 2008 and 2012 based on RADARSAT-1 and -2 imagery. Critically, CI2D3 documents the lineages of icebergs across fragmentation events and therefore provides a unique ground control dataset allowing evaluation of tracker performance.

Tracking of unchanging icebergs is achieved using a Bayesian tracking algorithm that makes linkages based upon a variety of geometric shape descriptors. Tracking across fragmentation events minimises Dynamic Time Warping distances between residual perimeter curves for candidate fragments and potential parents. This enables the matching of noisy, partial geometries and the automatic tessellation of fragments at one time step into the outline of their parent in a preceding observation irrespective of the intervening drift patterns. We evaluate tracker performance against bespoke metrics and those developed for cell tracking challenges that include mitotic division.

The system provides a generalisable geospatial tracking methodology based on object geometries that is applicable to other contexts and questions as well as a novel means of reconciling global invariances in geometries when conducting shape fingerprinting and matching.