EGU2020-3485, updated on 30 Nov 2020
EGU General Assembly 2020
© Author(s) 2020. This work is distributed under
the Creative Commons Attribution 4.0 License.

Extended Range Arctic Sea Ice Forecast with Convolutional Long-Short Term Memory Networks

Yang Liu1,2, Laurens Bogaardt3, Jisk Attema1, and Wilco Hazeleger4
Yang Liu et al.
  • 1Netherlands eScience Center, Amsterdam, the Netherlands
  • 2Meteorology and Air Quality Group, Wageningen University, Wageningen, the Netherlands
  • 3Department for Statistics, Informatics and Modelling, Netherlands Institute for Public Health and the Environment, Utrecht, the Netherlands
  • 4Faculty of Geoscience, Utrecht University, Utrecht, the Netherlands

Operational Arctic sea ice forecasts are of crucial importance to commercial and scientific activities in the Arctic region. Currently, numerical climate models, including General Circulation Models (GCMs) and regional climate models, are widely used to generate the Arctic sea ice predictions at weather time-scales. However, these numerical climate models require near real-time input of weather conditions to assure the quality of the predictions and these are hard to obtain and the simulations are computationally expensive. In this study, we propose a deep learning approach to forecasts of sea ice in the Barents sea at weather time scales. To work with such spatial-temporal sequence problems, Convolutional Long Short Term Memory Networks (ConvLSTM) are useful.  ConvLSTM are LSTM (Long-Short Term Memory) networks with convolutional cells embedded in the LSTM cells. This approach is unsupervised learning and it can make use of enormous amounts of historical records of weather and climate. With input fields from atmospheric (ERA-Interim) and oceanic (ORAS4) reanalysis data sets, we demonstrate that the ConvLSTM is able to learn the variability of the Arctic sea ice within historical records and effectively predict regional sea ice concentration patterns at weekly to monthly time scales. Based on the known sources of predictability, sensitivity tests with different climate fields were also performed. The influences of different predictors on the quality of predictions are evaluated. This method outperforms predictions with climatology and persistence and is promising to act as a fast and cost-efficient operational sea ice forecast system in the future.

How to cite: Liu, Y., Bogaardt, L., Attema, J., and Hazeleger, W.: Extended Range Arctic Sea Ice Forecast with Convolutional Long-Short Term Memory Networks, EGU General Assembly 2020, Online, 4–8 May 2020, EGU2020-3485,, 2020

Comments on the presentation

AC: Author Comment | CC: Community Comment | Report abuse

Presentation version 1 – uploaded on 01 May 2020
  • CC1: 2 questions about the method setup and results. , Laurent Bertino, 06 May 2020

    The baseline statistical model seems to perform very well compared to other predictors. Is it used as a first guess for the ConvLSTM? 

    Intuitively, the atmospheric variables (winds) should be more important in winter and the oceanic variables (OHC) in summer. It is hard to see from the presentation if these actually give an advantage to the predictions. Can you say more if some parameters are more important that others in different seasons? 

    • AC1: Reply to CC1, Yang Liu, 06 May 2020

      Dear Laurent, thank you for your questions. The baseline statistical model is a generalized linear model with a logit link. It takes SIC, OHC and T2M as input fields and includes neighboring points and previous timesteps. Therefore it is indeed a nonlinear high-level statistical model that is pretty difficult to beat. It is not a first guess but a very "top-standard" baseline.

      The wind is important for the trans-basin drift. We only look at the Barents Sea (a cut-off region consisting of 24x56 points with a resolution of about 28km x 28km) and this region is strongly dominated by the energy transport in the ocean as well as the OHC (e.g. Onarheim et. al. 2015). Our results suggest that the energy budget components (OHC, net surface flux SFlux) play important roles in this region for sea ice forecast (see the difference between ConvLSTM SIC/OHC, SIC/SFlux and SIC/SLP in figure a and b on slide 10, basically SLP indicates the pressure gradient, which is the wind) and these are consistent with previous studies. But I do agree with you that the difference is pretty small when looking at different seasons (figure b on slide 10).

      • CC2: Reply to AC1, Laurent Bertino, 07 May 2020

        Thanks for your detailed answer, Yang. 

        This was promising results and I wish you good luck with the project. 

        • AC2: Reply to CC2, Yang Liu, 07 May 2020

          Dear Laurent, thank you very much for your question and your remark! Success!