Ice drift forecasts in Svalbard are now more accurate

by Panagiotis Kountouris on 2020-05-27

Co-Authors: Cyril Palerme (Norwegian Meteorological Institue), Lasse Rabenstein (Drift Noise GmbH)

Arctic weather forecasts are challenging

Everyone has some experience with forecasts either by watching weather news or the usage of weather Apps providing information about e.g wind and rain. However, we sometimes see that those predictions are not in line with the actual weather conditions. Especially severe is the situation for marine masters navigating cruise-, cargo- or fishing vessels in remote polar regions. They reported us that the difference between e.g. ice drift forecast and reality in these regions is often too large to consider the forecast for operational planning. SVALNAV will help those professionals by providing improved sea-ice drift forecasts.

Screenshot of the SVALNAV App. Colored background shows sea-ice concentration, grey images are Sentinel-1 radar satellite scenes and the red dots are improved sea-ice dift forecast trajectories.

SVALNAV displays improved sea-ice drift forecasts

The forecasts range 10 days into the future and cover the area around the Svalbard archipelago with a spatial resolution of 12.5 km. In the SVALNAV App the forecasts will be visualized with points anticipating the future location of ice patches. In the image above each subsequent red point symbolizes a new position after 24 hours for up to 10 days into the future. The location 10 days in the future is plotted with solid red and the location of today in most transparent red (like the tail of a comet started from today 10 days into the future)

Noticeable improvement in forecast accuracy using machine learning techniques

The SVALNAV sea-ice forecasts are improved by means of machine learning techniques developed at the Norwegian Meteorological Institute. In particular by comparisons between modelled forecasts and real observations of the ice drift using random forest algorithms. This technique allows us to estimate the uncertainties in magnitude and direction of the ice drift and to apply corrections to the forecast. Improvements by this method are noticeable: We did compare forecasts of the widely used TOPAZ4 modelling system and our optimized forecasts against real observations. On average, for a 24 hours forecast the error for the ice drift speed is reduced from roughly 3500 m/day to 3100 m/day (11.5 %), whilst the error for the drift direction is reduced from on average 45° to 41° (5%). Most considerable improvement has been achieved for longer range forecasts: The ice drift speed error for a 10 days forecasts is on average reduced from 5500 m/day to 4400 m/day (20%) (panel a). In total more than 50% of the forecasts in the Svalbard region were improved (see panels c,d).

This short news article is just a teaser into the topic. For more information and a complete picture we refer to the Copernicus Ocean State Report 5 which will be accessible on the copernicus site in 2021.