CIRFA Seminar 6. Mai: Geoffrey Dawson and Sindre Fritzner

Arctic sea ice thickness measurements from radar altimeters, are typically limited to the winter months. In the summer months, meltwater ponds at the surface of the sea ice produce strong specular reflections in a similar way leads do, making it challenging to distinguish between the two surface types. Without reliable lead detection, we cannot measure sea ice thickness and melt rates over the summer.

In this study, we use coincident images from Synthetic Aperture Radar (SAR) (Sentinel -1 and RADARSAT-2) and optical (Sentinel-2 and Landsat-8) satellites that are recorded within 15 minutes of CryoSat-2 tracks, as an independent measure of surface type. We use these coincident data, to create a new classification scheme using convolutional neural networks for summer sea ice lead detection. This allows us to create the first pan-Arctic summer freeboard map from CryoSat-2 radar altimetry. We validate these maps against winter radar freeboard maps, laser scanner freeboards from Operation IceBridge, Airborne EM and Beaufort Gyre Exploration Program moorings, and discuss biases that are present in the data.

Predicting the Arctic sea-ice cover is important for many reasons, e.g. shipping and weather forecast. Traditionally, the Arctic sea-ice cover has been predicted by using dynamical physics-based models. Over the years these models have grown into large complex beasts with a significant computational cost that requires supercomputers for reasonable prediction time. An alternative to these dynamical models is machine-learning methods which have received lots of attention in recent years. With Machine learning methods the prediction is based on simple mathematical relationships and there is no need to resolve physical mechanisms which greatly reduce the computational cost of the prediction. In this talk, the U-net, a machine-learning method based on convolutional operations, is shown to provide reasonably accurate sea-ice predictions for both monthly and daily predictions. The machine learning predictions are found to improve on both the climatological forecast and persistence.