Predicting Iceberg Severity off Newfoundland, Canada, Using a Control Systems Model and Machine Learning Tools

Abstract:

Icebergs pose a direct risk to stationary platforms, such as oil rigs, and to shipping. This is especially true in the region off Newfoundland, Canada, which is an area of popular commercial shipping routes. Icebergs also locally cool and freshen the surface ocean, which may enhance sea ice extent or thickness, and increase stratification, potentially reducing deep convection.

We have been working on improving a statistical model of iceberg numbers in the region off Newfoundland, developed in Bigg et al., 2019. Here, a Windowed Error Reduction Ratio (WERR) control systems method is used to predict the number of icebergs passing 48°N in the following iceberg year. This is because it is commonly held that annual variability in iceberg activity in the Labrador Sea can be represented by the number of icebergs passing 48°N. Findings for the 2019 and 2020 ice seasons were sent to the International Ice Patrol (IIP).

Due to an observable seasonal pattern in iceberg distribution, often showing a rapid increase in number over a few days, we decided that rate of change may be a useful indicator of the severity of the ice year, as well as overall iceberg numbers. The IIP also showed interest in this area, giving space in an Appendix of their 2018 Annual Report to the desirability of predicting this feature of the ice season. Therefore, we have been looking at ways to predict this, as well as the magnitude of the annual I48N, using machine learning tools. These are linear discriminant analysis, a linear Support Vector Machine algorithm (SVM) and a quadratic SVM algorithm. While this technique is still in the early stages, when combined with the control systems model it looks to provide a useful forecasting tool.

For the 2020 prediction, machine learning predicted a low ice year, so less than 230 bergs past 48°N in total. The control systems model predicted a low/medium ice year. Current observations support this prediction. Machine learning also predicted a high rate of change, which is consistent with a rapid onset to a low season.

Funding: AXA XL