Spotlight on seasonal ice change

IceNet, ice loss forecasts for the season ahead

IceNet is an AI tool that enables scientists to more accurately forecast Arctic sea ice conditions months into the future. Developed for understanding Arctic sea ice loss in collaboration with BAS researchers, IceNet has provided the first demonstration of an AI framework that can predict the concentration of sea ice more accurately and more efficiently than traditional physics-based approaches. The improved predictions could underpin new early-warning systems that protect Arctic wildlife and coastal communities from the impacts of sea ice loss.

IceNet provides a solution built on a deep learning model called a convolutional neural network (CNN), which is the technology behind facial recognition systems, medical imaging analysis, and self-driving cars. IceNet has been trained to automatically forecast the next six months of monthly average sea ice maps, based on climate simulations for 1850–2100 and satellite observations from 1979 to 2011. IceNet directly predicts probabilities of sea ice occurring, displaying a confidence level for each prediction. When evaluated on unseen data after training, IceNet was more accurate than similar physics-based prediction systems. IceNet offers a simple framework for probabilistically bounding the ice edge within a region of lower predictive confidence, which has added utility over deterministic ice edge forecasts. Finally, a variable importance method was used to identify the climate variables most important for IceNet’s forecasts. The next generation of IceNet will build on this exciting result, predicting at a more granular daily timescale and running in real-time across the polar regions (Andersson, Hosking et al, 2021).

IceNet began as a small research project at BAS but soon expanded at an early stage in collaboration with the Turing researchers, REG and RAM team members from ASG. To ensure rigour across all aspects of this highly interdisciplinary project, new international domain experts were further involved to help address any knowledge gaps in the team. Relying on complementary skills, the IceNet team was able to improve both the forecasting ability of the software and its speed of execution, as well as develop tools for forecast dissemination. Typically the numerical physics-based forecasting systems running on supercomputers require hours of run-time to produce forecasts. IceNet, on the other hand, can run on a laptop 2,000 times faster than equivalent numerical forecast models, taking less than ten seconds on a single graphics processing unit. By training AI models on larger more regionally diverse datasets, IceNet will be able to provide larger coverage and use cases for both the Northern and Southern Hemisphere. The goal for the next generation of IceNet is to go beyond monthly Arctic forecasts by making daily predictions of sea ice levels in both hemispheres.

Future extensions of IceNet will couple novel deep generative modelling and lifelong learning AI forecasting research with digital infrastructure developments to deliver data products and services, underpinning the development of polar digital twins. In the next stage, IceNet will deliver on three specific aims: weather-to-seasonal timescale sea ice forecasts to meet Arctic and Antarctic conservation challenges; research and operational tools providing forecasts to decision-focused software frameworks; and support for users to integrate these technologies into their day-to-day workflows. This ecosystem will draw data generated from deep learning pipelines to demonstrate real-world use cases and make them available through integrations and interfaces for programmatic access. By making it a fully interoperable module to work with other environmental systems and “digital twins”, IceNet will be easy to deploy, energy efficient and a powerful tool for instantaneous and real-time decision-making.