Accuracy Assessment of Iceberg Drift Models Over Short Time Scales

Since the early 1980s, two approaches have been proposed to predict iceberg drift trajectories. The deterministic approach uses forecasts of relevant environmental drivers, including wind and ocean currents, and assumes the resultant forces acting on the iceberg can be predicted and used to calculate future drift. The statistical-plus approach holds that small-scale current eddies are inherently chaotic, and that the recent trajectory of icebergs contains information on these currents that can be used to improve drift prediction.

Although the statistical-plus theory was well-described in the 1980's, it has been largely neglected until very recently, in spite of its ability to assimilate observed iceberg trajectories in real-time. Most iceberg drift forecasts are still done using deterministic models, which must be parameterized using coarse assumptions about iceberg shape and drag.

This study uses iceberg drift and forecast data for Canadian East Coast icebergs that originate primarily in Greenland. Using the error in predicted position over time, we produce average positional error over time (APEOT) curves over a population of comparable icebergs. The APEOT calculated for a suitable number of comparable icebergs allows statistical analysis to determine whether different drift models produce significantly different accuracy at different prediction time scales. We show that a representative statistical-plus model is superior over short time scales (<16 to 30 h), while a representative deterministic model is superior over longer time scales (> 16 to 30 h). We further propose some absolute accuracy benchmarks and their interpretation.