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

View More View Less
  • 1 Netherlands eScience Center, Amsterdam, the Netherlands, Meoteorology and Air Quality Group, Wageningen University, Wageningen, the Netherlands
  • 2 Department for Statistics, Informatics and Modelling, National Institute for Public Health and the Environment, Utrecht, the Netherlands
  • 3 Netherlands eScience Center, Amsterdam, the Netherlands
  • 4 Faculty of Geosciences, Utrecht University, Utrecht, the Netherlands
© Get Permissions
Restricted access

Abstract

Operational Arctic sea ice forecasts are of crucial importance to science and to society in the Arctic region. Currently, statistical and numerical climate models are widely used to generate the Arctic sea ice forecasts at weather time-scales. Numerical models require near real-time input of relevant environmental conditions consistent with the model equations and they are computationally expensive. In this study, we propose a deep learning approach, namely Convolutional Long Short Term Memory Networks (ConvLSTM), to forecast sea ice in the Barents Sea at weather to sub-seasonal time scales. This is an unsupervised learning approach. It makes use of historical records and it exploits the covariances between different variables, including spatial and temporal relations. With input fields from reanalysis data, we demonstrate that ConvLSTM is able to learn the variability of the Arctic sea ice and can forecast regional sea ice concentration skillfully at weekly to monthly time scales. It preserves the physical consistency between predictors and predictands, and generally outperforms forecasts with climatology, persistence and a statistical model. Based on the known sources of predictability, sensitivity tests with different climate fields as input for learning were performed. The impact of different predictors on the quality of the forecasts are evaluated and we demonstrate that the surface energy budget components have a large impact on the predictability of sea ice at weather time scales. This method is promising to enhance operational Arctic sea ice forecasting in the near future.

Corresponding author address: Netherlands eScience Center, Science Park 140, Amsterdam, the Netherlands. E-mail: y.liu@esciencecenter.nl

Abstract

Operational Arctic sea ice forecasts are of crucial importance to science and to society in the Arctic region. Currently, statistical and numerical climate models are widely used to generate the Arctic sea ice forecasts at weather time-scales. Numerical models require near real-time input of relevant environmental conditions consistent with the model equations and they are computationally expensive. In this study, we propose a deep learning approach, namely Convolutional Long Short Term Memory Networks (ConvLSTM), to forecast sea ice in the Barents Sea at weather to sub-seasonal time scales. This is an unsupervised learning approach. It makes use of historical records and it exploits the covariances between different variables, including spatial and temporal relations. With input fields from reanalysis data, we demonstrate that ConvLSTM is able to learn the variability of the Arctic sea ice and can forecast regional sea ice concentration skillfully at weekly to monthly time scales. It preserves the physical consistency between predictors and predictands, and generally outperforms forecasts with climatology, persistence and a statistical model. Based on the known sources of predictability, sensitivity tests with different climate fields as input for learning were performed. The impact of different predictors on the quality of the forecasts are evaluated and we demonstrate that the surface energy budget components have a large impact on the predictability of sea ice at weather time scales. This method is promising to enhance operational Arctic sea ice forecasting in the near future.

Corresponding author address: Netherlands eScience Center, Science Park 140, Amsterdam, the Netherlands. E-mail: y.liu@esciencecenter.nl
Save