Long-term prediction of Arctic sea ice concentrations using deep learning: Effects of surface temperature, radiation, and wind conditions

Abstract

Over the last five decades, Arctic sea ice has been shrinking in area and thickness. As a result, increased marine traffic has created a need for improved sea ice forecasting on seasonal to annual time-scales. In this study, we introduce a novel UNET-based deep learning model to forecast sea ice concentration up to 12 months. Based on yearly hindcast validation, the UNET 3-, 6-, 9-, and 12-month predictions provided more accurate and stable predictions than did the four baseline models: the Copernicus Climate Change Service (C3S), the damped anomaly persistence (DP) forecast, and two deep learning approach, the Convolutional Neural Network (CNN) models and Convolutional Long Short-Term Memory (ConvLSTM). During years with large departures from the long-term trend, the proposed UNET model exhibited promising SIC prediction results with root-mean-square errors (RMSEs), which were reduced from 17.35 to 7.07 % compared to the four baseline models. Our findings also confirmed the relative importance of each predictor variable (temperature, incoming solar radiation, wind speed and direction) in long-term prediction. Past SIC conditions, together with surface temperature emerged as the most important factors for SIC prediction, especially in the marginal ice zone. Incoming solar radiation and wind speed and direction showed greater sensitivity in predicting SICs in areas with thin ice. This model offers the potential to shape Arctic development and management plans and strategies, ensuring extended forecasting periods and enhanced prediction accuracy.