Graph neural networks for antarctic sea ice concentration forecasting

Abstract

Antarctic sea ice is a crucial component of the climate system, yet in recent years it has undergone abrupt, poorly understood changes that motivate more accurate weeks-ahead forecasts. Existing statistical and dynamical approaches can struggle to represent the strongly nonlinear, heterogeneous, and rapidly evolving nature of Antarctic sea-ice dynamics, while conventional deep-learning models like convolutional neural networks (CNN) typically operate on regular Euclidean grids, making long-range spatial dependencies and irregular regional interactions difficult to represent explicitly. Here we introduce the first graph neural network (GNN) framework for forecasting weekly Antarctic sea ice concentration (SIC) and its anomalies (SICA) on subseasonal time scales of up to eight weeks ahead. Our framework couples an encoder–processor–decoder architecture, trained on satellite observations and ERA5 reanalysis, with an ablation of two contrasting graph-construction schemes: (1) a purely geometric icosahedral mesh, and (2) a data-driven graph built from super-pixel segmentation of the Antarctic domain. In controlled ablations, we found that how the graph is constructed matters more than the choice of message-passing operator (Interaction, GraphSAGE, DeepGAT or DConv). Benchmarked against five re-implemented CNN architectures (IceNet, SICNet-CBAM, SICNet-TSAM, Unicorn, ANTSIC-UNet), two statistical baselines, and the operational ECMWF S2S dynamical forecast, the GNN achieves the lowest error on three out of four metrics, cutting full-grid MAE by roughly 18% relative to the strongest CNN benchmark and by 48% relative to a persistence forecast, while using 4–24 times fewer trainable parameters. Only on root-mean-square error (RMSE) at the ice edge does it fall behind IceNet (by about 4%). To explain our model’s prediction, a feature attribution analysis identifies that it relies on sea-surface temperature as the dominant driver beyond the seasonal climatology.