Research map / Image generation / Fast diffusion sampling
IceDiff: High Resolution and High-Quality Arctic Sea Ice Forecasting with Generative Diffusion Prior
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes IceDiff, a two-module framework that pairs a Swin-Transformer U-Net forecaster at 25 km with a zero-shot, patch-based guided diffusion downscaler to produce Arctic sea ice concentration forecasts at 6.25 km.
Problem
Arctic sea ice concentration forecasting is limited to 25 km resolution from satellite products, and forecasting at finer scales is under-explored due to lack of high-resolution data.
Approach
Forecast at the native 25 km resolution with a Swin Transformer V2 U-Net (with TSAM), then downscale forecasts to 6.25 km by guiding an unconditional DDPM pre-trained on daily SIC maps via an optimizable up-scaling convolution kernel and a patch-based sampling strategy, with no paired high-resolution training.
Key results
- At 7-day lead on NT data, IceDiff forecasting reaches RMSE 0.0324 and R2 0.987 versus SICNet's RMSE 0.0430 and R2 0.978.
- For downscaling to 6.25 km, IceDiff achieves FID 29.34 at 7 days versus 40.15 for GDP and 54.71 for bicubic interpolation.
- At 6-month average lead on NT data, IceDiff reaches RMSE 0.0573 and R2 0.958, versus MT-IceNet RMSE 0.0777 and IceNet R2 0.567.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace SICNet with IceDiff | SICNet forecasting · 7 Days (Daily), NT | RMSE: 0.0430 -> 0.0324 | Helped |
| replace SICNet with IceDiff | SICNet forecasting · 7 Days (Daily), CDR | RMSE: 0.0490 -> 0.0396 | Helped |
| use IceDiff instead of MT-IceNet | seasonal SIC forecasting · 6 Months Average (Monthly), NT | RMSE: 0.0777 -> 0.0573 | Helped |
| use IceDiff downscaling instead of GDP | SIC downscaling · 7 Days | FID: 40.15 -> 29.34 | Helped |
| use IceDiff downscaling instead of GDP | SIC downscaling · 8 Week Avg. | FID: 52.40 -> 31.29 | Helped |
| use IceDiff downscaling instead of GDP | SIC downscaling · 6 Month Avg. | FID: 55.72 -> 39.79 | Helped |
| add TSAM (ablation) | IceDiff forecasting · 7 Days | RMSE: 0.0403 -> 0.0396 | Helped |
| add TSAM (ablation) | IceDiff forecasting · 8 Week Avg. | RMSE: 0.0535 -> 0.0553 | Hurt |
| add TSAM (ablation) | IceDiff forecasting · 6 Month Avg. | RMSE: 0.0763 -> 0.0648 | Helped |
| use equal input and forecast lengths instead of half-length input (ablation) | IceDiff forecasting · 7 Days | RMSE: 0.0416 -> 0.0396 | Helped |
| use equal input and forecast lengths instead of double-length input (ablation) | IceDiff forecasting · 7 Days | RMSE: 0.0401 -> 0.0396 | Helped |
| use equal input and forecast lengths instead of half-length input (ablation) | IceDiff forecasting · 8 Week Avg. | RMSE: 0.0580 -> 0.0553 | Helped |
Surprising findings
- The paper reports that bicubic interpolation and other baselines are not beaten on every metric; at 8-week average on CDR, IceDiff's RMSE (0.0553) is worse than its NT-trained variant, and the TSAM ablation does not help at 8-week lead on all metrics.
Limitations
- No evaluation against high-resolution in situ or finer-scale ground truth, since no 6.25 km observations exist.
- Downscaling quality is assessed with FID and Consistency rather than against true high-resolution data.
- SICNet was re-implemented by the authors because source code was unavailable, so baseline numbers may differ from originals.
Experimental setup
Datasets: NSIDC Climate Data Record of Passive Microwave Sea Ice Concentration Version 4 NASA Team sea ice concentration (NT)
Benchmarks: NSIDC CDR SIC 2016-2023 test split NASA Team SIC test split
Metrics: RMSE MAE R2 NSE IIEE SIEdif FID Consistency
Baselines: SICNet IceFormer SICNet90 IceNet MT-IceNet Nearest interpolation Bilinear interpolation Bicubic interpolation DGP GDP
Builds on: SICNet SwinV2 DDPM IceNet GDP
Closest papers by meaning
- MFogHub: Bridging Multi-Regional and Multi-Satellite Data for Global Marine Fog Detection and Forecasting (CVPR 2025 · similarity 0.83)
- FuXi-Ocean: A Global Ocean Forecasting System with Sub-Daily Resolution (NeurIPS 2025 · similarity 0.83)
- OmniCast: A Masked Latent Diffusion Model for Weather Forecasting Across Time Scales (NeurIPS 2025 · similarity 0.82)
- Common Task Framework For a Critical Evaluation of Scientific Machine Learning Algorithms (NeurIPS 2025 · similarity 0.82)
- DAWP: A framework for global observation forecasting via Data Assimilation and Weather Prediction in satellite observation space (NeurIPS 2025 · similarity 0.82)
- PostCast: Generalizable Postprocessing for Precipitation Nowcasting via Unsupervised Blurriness Modeling (ICLR 2025 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.