Seasonal Arctic sea ice forecasting with probabilistic deep learning

The Arctic is experiencing warming at two to three times the global average ('Arctic amplification'), primarily due to positive feedbacks. This warming significantly reduces Arctic sea ice, with September sea ice extent now about half of what it was in 1979. This decline is projected to continue, even under optimistic greenhouse gas emission scenarios, potentially leading to an ice-free Arctic summer by 2050 or even earlier. This dramatic sea ice loss has severe consequences: it threatens polar bear populations, intensifies algal blooms, and poses significant challenges to Indigenous communities. Arctic sea ice also plays a crucial role in the global climate system, potentially influencing weather patterns beyond the Arctic. Current operational sea ice forecasting systems based on deterministic coupled atmosphere-ice-ocean models often perform no better than simple statistical forecasts at seasonal lead times of two months or more. This study addresses these challenges by introducing IceNet, a novel sea ice prediction system leveraging the power of deep learning to improve seasonal forecasts of Arctic sea ice.

Existing physics-based dynamical models for Arctic sea ice forecasting struggle to outperform simple statistical methods at longer lead times (two months or more), despite successfully predicting sea ice concentration over shorter periods. While inherent limitations in predictability exist due to chaotic atmospheric processes, studies suggest that potential predictability is higher, indicating room for improvement in forecasting accuracy. Deep learning, a powerful approach in various fields with abundant data, has shown promise in Earth sciences applications, particularly with satellite data. This study builds upon prior deep learning efforts for sea ice concentration (SIC) prediction, addressing the limitations of previous approaches (such as limited input receptive field) by utilizing a U-Net architecture that enables the modeling of long-range spatiotemporal interactions.

IceNet, an ensemble of U-Net convolutional neural networks, forecasts monthly averaged sea ice concentration (SIC) maps at 25 km resolution for the next six months. The input data comprises SIC, eleven climate variables (capturing atmospheric and oceanic couplings), statistical SIC forecasts, and metadata, stacked analogously to RGB image channels (totaling 50 channels). Each U-Net processes this data, outputting probability distributions over three SIC classes: open-water (SIC ≤ 15%), marginal ice (15% < SIC < 80%), and full ice (SIC ≥ 80%). The ensemble-mean forecast is obtained by averaging individual probability distributions across the 25 ensemble members, improving both performance and probability calibration. To address the relatively short observational data record (1979-2011), the models were pre-trained using 2220 years of climate simulation data from the Coupled Model Intercomparison Project phase 6 (CMIP6), covering 1850-2100. This transfer learning approach, followed by fine-tuning on observational data and temperature scaling for probability calibration, enhances the model's performance. The validation years (2012-2017) were used for hyperparameter optimization, early stopping, and calibration, while the test years (2018-2020) served for evaluating generalization ability. The model is evaluated using a binary accuracy metric assessing the percentage of correctly classified SIC (open water or ice) and compared to SEAS5 (a state-of-the-art dynamical model) and a linear trend model for various lead times and calendar months. A permute-and-predict method assesses variable importance. The calibrated probabilities are used in a framework to probabilistically bound the ice edge, improving the utility of forecasts beyond simple ice edge predictions. Details on data pre-processing, the U-Net architecture, training procedures, and model ensembling and calibration are also provided.

IceNet consistently outperforms SEAS5 and the linear trend model at lead times of two months or more, especially in August, September, and October. This improvement is particularly significant for extreme sea ice events (e.g., the record low sea ice extent in 2012 and the anomalously high extent in 2013). IceNet's predictive skill exhibits a seasonal dependence, reflecting the 'spring predictability barrier'. However, even during this barrier, IceNet's performance is comparable to or better than the other models. Comparisons with the Sea Ice Outlook (SIO) multi-model median September SIE predictions indicate that IceNet often matches or surpasses the SIO's accuracy, particularly in years with anomalously high or low sea ice extent. CMIP6 pre-training slightly improves overall accuracy, while model ensembling consistently enhances performance, particularly for long-range summer predictions. The combined effect of pre-training and ensembling leads to IceNet's high performance. IceNet's probabilistic sea ice probability (SIP) outputs demonstrate good calibration. The well-calibrated and sharp SIP forecasts allow for the reliable bounding of the ice edge between two SIP contours, which is a significant improvement in forecast utility, allowing for robust quantification of uncertainty. Variable importance analysis reveals that IceNet relies heavily on initial sea ice conditions and tropospheric conditions for short-range predictions, especially for September. The importance of initial conditions decreases with lead time, aligning with observed limits in sea ice predictability. IceNet shows remarkable speed, running over 2000 times faster than SEAS5 on a laptop.

IceNet's superior performance in seasonal Arctic sea ice forecasting compared to the state-of-the-art dynamical model SEAS5 highlights potential shortcomings in the fidelity or initialisation of dynamical models in certain regions and seasons. The information gained from IceNet's variable importance analysis can guide improvements in dynamical model parameterizations, data assimilation, and calibration techniques. The accurate and reliable probabilistic forecasts from IceNet, along with the ice edge bounding framework, have significant practical implications. They can improve safety in Arctic shipping by providing more accurate predictions of ice-covered areas, reducing the risk of accidents and environmental damage. Furthermore, as understanding of the links between sea ice extent and mid-latitude weather strengthens, IceNet's accurate seasonal forecasts could help predict mid-latitude weather conditions months in advance. Beyond shipping, IceNet's forecasts are also crucial for conservation efforts, offering early warnings for local communities, authorities, and environmental organizations. They can help anticipate events like walrus mega-haulouts, enabling better management of human access to prevent stampedes and mortality. Forecasts can also aid in the planning of dynamic marine protected areas (MPAs) to protect Arctic biodiversity.

IceNet demonstrates the potential of AI for improving seasonal Arctic sea ice forecasting, outperforming leading dynamical models. Its speed, calibrated probabilities, and ice edge bounding capabilities offer significant advantages for shipping safety and Arctic conservation. Future work will explore incorporating ice thickness into IceNet and implementing a daily resolution online version to enhance short-term forecast accuracy.

The study's reliance on monthly averaged data might limit its accuracy for very short-term forecasts (less than a month). The performance gains from CMIP6 pre-training were modest, highlighting the need for more accurate climate simulations, particularly regarding summer melt processes. While IceNet's variable importance analysis offers valuable insights, it cannot establish direct causal relationships.