We are thrilled to announce that the paper "Interrogating Sea Ice Predictability With Gradients", a collaborative paper between Visual Intelligence, The Alan Turing Institute (AT) and the British Antarctic Survey (BAS), was accepted in the journal IEEE Geoscience and Remote Sensing Letters on February 14th 2024.

The paper focuses on interrogating the effect of the IceNet's, a state-of-the-art ice prediction model, input feature with a gradient-based analysis, which takes advantage of the developments within the deep learning literature to open the so-called "black box".

The authors' analysis focuses on the unusually large sea ice extent event in September 2013 and indicates that IceNet places a strong emphasis on previous observations of SIC, linear trends, and seasonal components when making predictions. They further identify which input features are most influential for the prediction and also which spatial location these measurements are particularly influential.

The authors of this particular work from the Visual Intelligence side are post doc. Luigi Luppino (UiT), PhD students Harald Lykke Joakimsen and Iver Martinsen (UiT), and Robert Jenssen (UiT). From the UK side, the authors are Scott Hosking (BAS/AT) and Andrew McDonald (BAS).

You may read the paper via IEEE Explore.

‍

Predicting sea ice concentration (SIC) is an importanttask in climate analysis. The recently proposed deep learning system IceNet is the state-of-the-art sea ice prediction model. IceNet takes high-dimensional climate simulations and observational data as input features and forecasts SIC for the next6 months over a spatial grid over the northern hemisphere. The model has proven to be particularly good at predicting extreme sea ice events compared with previous dynamical models, but lacks interpretability. In the original IceNet paper, a permute and-predict approach was taken for assessing feature importance. However, this approach is not capable of revealing whether a feature contributes positively or negatively to the final prediction, nor can it reveal the importance of features over the spatial grid of predictions. In this letter, we take steps to instead interrogate the effect of the IceNet input feature with a gradientbased analysis, taking advantage of developments within the deep learning literature to open the so-called black box. Our analysis focuses on the unusually large sea ice extent event in September 2013 and indicates that IceNet places a strong emphasis on previous observations of SIC, linear trends, and seasonal components when making predictions. In our analysis, we identify which input features are most influential for the prediction and also which spatial location these measurements are particularly influential.

‍