A deep learning-based bias correction model for Arctic sea ice concentration towards MITgcm

Abstract

Accurate prediction of Arctic sea ice is essential for ship navigation. The numerical forecast is an important method to predict sea ice. However, currently, it has significant bias from observation data. In this paper, we propose a deep learning-based bias correction model, Ice-BCNet, to post-process the weekly sea ice concentration (SIC) forecast data of MITgcm to improve its accuracy. Different from the existing bias correction models that only consider spatial features, Ice-BCNet embeds Convlstm into UNet, enabling it to extract spatiotemporal features from SIC forecast data. Ice-BCNet also corrects a monthly scale by iteration. Before the correction, we first assimilate the MASIE-AMSR2 (MASAM2) SIC observation into MITgcm to obtain a better numerical output, which can improve the accuracy of bias correction results. We evaluate the Ice-BCNet from the 2022 hindcasting and 2023 forecasting and use the binary accuracy classification coefficient (BACC) to measure the accuracy of the sea ice edge. We compare Ice-BCNet with statistical corrected methods (Simple Bias Correction, SimBC). The weekly corrected SIC’s average RMSE decreased by over 41%, and Ice-BCNet outperforms SimBC in correcting sea ice near the route. The monthly corrected SIC’s RMSE is below 0.1, with a BACC exceeding 94%. Ice-BCNet also shows a better performance in the extreme case of September 2020.