Key points are not available for this paper at this time.
Observing and analyzing the changing polar ice patterns is crucial for understanding the climate crisis. Research works across the Circumpolar North use machine learning models to study and predict changes in sea ice. In this paper, we propose a deep learning model using satellite images of the Arctic, captured daily and monthly over a half-century period and curated at the National Snow and Ice Data Center (NSIDC), to forecast future ice extent. We perform a time-series analysis using a multimodal approach, combining a gated recurrent unit (GRU) with a transformer-based model to predict changes in Arctic ice. Our model explains approximately 92.02% of the variance in the true ice extent time series. The error metrics were low: We observed the Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) to be, respectively, 0.1362 and 0.1637. Preliminary assessments of our prototype show promising results in understanding past trends and making predictions.
Namir et al. (Sat,) studied this question.