Developing a Deep Learning forecasting system for short-term and high-resolution prediction of sea ice concentration

Abstract

There has been a steady increase of marine activity throughout the Arctic Ocean during the last decades, and maritime end users are requesting skillful high-resolution sea ice forecasts to ensure operational safety. Different studies have demonstrated the effectiveness of utilizing computationally lightweight deep learning models to predict sea ice concentration in the Arctic, but few have explored the integration of real-time data to create an operational forecasting system. This thesis aims to develop a deep learning forecasting system which can predict sea ice concentration at one kilometer resolution for 1 to 3-day lead time. Deep learning models have been trained using sea-ice charts form the Norwegian Ice Service, and predictors from the AROME Arctic numerical weather prediction system hosted by the Norwegian Meteorological Institute and OSI SAF SSMIS passive microwave sea ice concentration observations to establish the deep learning forecasting system. The deep learning system has primarily been validated using the normalized integrated ice edge error, which is a sea ice edge aware skill-metric that ensures operational relevance. It is shown that the deep learning forecasting system achieves lower seasonal mean and median normalized integrated ice edge error for several sea ice concentration contours when compared against baseline-forecasts (persistence-forecasts and linear trend), as well as two state-of-the-art dynamical sea ice forecasting systems (neXtSIM and Barents-2.5) for all considered lead times and seasons. This result was repeated when changing the validational data to sea ice concentration from independent AMSR2 observations, demonstrating generalizability of the deep learning forecasts. The deep learning system was also investigated in terms of explainability. With different predictor-modifying experiments, it is shown that the contributions from AROME Arctic weather forecasts are essential for the deep learning forecasts to achieve performance beyond persistence-forecasting. However, through use of the novel segmentation gradient-weighted class activation mapping technique, it is suggested that 2-meter temperature from AROME Arctic may degrade deep learning performance by limiting the extent of important pixels to be contained within the sea ice extent.