SEAL: Sea Ice Forecasting using Active Learning

1: So… what have we done so far? 3: Cool! And the AL part? 5: Sounds like SEAL is on its way to saving the ! 1: The ice I live on is disappearing faster every year! What’s happening? 3: That’s massive! Can you predict what’s next? 5: And that’s where I come in? 7: Sounds like teamwork. Science and seals saving the together! 2: Arctic sea ice has nearly halved since satellite records began. The Arctic is warming nearly 4X faster than the global average [1, 2]. 4: ML-based models such as IceNet [3, 4] show promise, but there’s still room for improvement. 6: Exactly! SEAL uses AL to forecast SIC efficiently with less data, improving predictions while reducing cost [5]. We also use DDPM for higher-resolution results. 2: We built a DDPM for sea ice forecasting, and it gives higher-resolution predictions! 4: That’s next! We’ll integrate AL to make training smarter and more efficient. References: [1] L. Polvani, M. England, J. Screen, A surprising, but not unexpected, multi-decadal hiatus in Arctic sea ice decline, EGU Gen. Assem. 2025, EGU25-7248. [2] WWF Arctic, https://www.arcticwwf.org/threats/climate-change/ (May 16, 2025). [3] T.R. Andersson, J.S. Hosking, M. Pérez-Ortiz, et al., Seasonal Arctic sea ice forecasting with probabilistic deep learning, Nat. Commun. 12, 5124 (2021). [4] GitHub, https://github.com/icenet-ai/icenet-notebooks/tree/main/pytorch/ (May 13, 2025). [5] M.C. Novitasari, J. Quaas, M. Rodrigues, ALAS: Active Learning for Autoconversion Rates Prediction from Satellite Data, Proc. 27th Int. Conf. Artif. Intell. Stat. (2024), Proc. Mach. Learn. Res. 238, 3358–3366. https://proceedings.mlr.press/v238/novitasari24a.html. U-Net DDPM U-Net DDPM ERA5 Temporal embedding Step sOthers Clean SIC (x0) Noisy SIC (xs)U-Net Pred v (vθ)Target v loss Random Noise xₛ ~ 𝒩(0, I) Denoising step x{s-1} from vθ Clean SIC (x0) Forecast ERA5 Temporal embedding Step s S → S-1 → … → 0 Others U-Net Pred v (vθ) Repeat: s=S-1, S-2, …, 0 DDPM Training DDPM Inference Oracle labels query 𝐿𝑡𝑟𝑎𝑖𝑛 𝑈𝑝𝑜𝑜𝑙 ML Model (DDPM) train add Select the best instances Active Learning evaluate DDPM-based stochastic query strategy 𝐿𝑣𝑎𝑙 General Framework SEAL: Sea Ice Forecasting using Active Learning Maria Carolina Novitasari1, N. Pedrazzini1, I. Fenton1, J. Wilkinson2, J.S. Hosking1,2, and L. van Zeeland1 1The Alan Turing Institute, UK; 2British Antarctic Survey, UK Craving for more? animation slides here! Want to connect? Here's my LinkedIn. Background Conclusions Methodology Results ` ` ` Want to check out the code? Here’s the GitHub.