IceDiff: High Resolution and High-Quality Arctic Sea Ice Forecasting with Generative Diffusion Prior

Jingyi Xu, Siwei Tu, Weidong Yang, Ben Fei, Shuhao Li, Keyi Liu, Yeqi Luo, Lipeng Ma, Lei BaiCVPR 2025 · DOI · OpenReview

TL;DR

This paper introduces IceDiff, a novel two-module deep learning framework for high-resolution Arctic sea ice forecasting. By combining a forecasting module with a guided diffusion module, IceDiff achieves unprecedented 6.25 km resolution sea ice concentration predictions, significantly outperforming existing methods and paving the way for more practical applications.

What Problem Does This Paper Address?

Arctic sea ice plays a critical role in global climate, polar ecosystems, and human activities like shipping and coastal communities. Accurate forecasting of sea ice concentration (SIC) is crucial for these areas. While recent AI-driven methods have shown promise, they are often limited to a coarse resolution of 25 km, derived from passive microwave sensors. This resolution is insufficient for capturing fine-scale details vital for operational use and scientific research. The lack of high-resolution ground truth data further complicates training models at finer scales. This paper addresses the challenge of forecasting SIC at significantly higher resolutions (down to 6.25 km) by proposing a novel framework that leverages existing coarse-resolution data and advanced generative models.

Methods

IceDiff employs a two-module cooperative framework:

  1. Forecasting Module (FM): This module uses a vision transformer (specifically, a U-Net architecture built on Swin Transformer V2 with ResNetTSAM blocks) to generate skillful SIC forecasts at the standard 25 km resolution. It's trained to predict SIC for various temporal scales (7 days, 8 weeks average, 6 months average), aiming for superior accuracy compared to existing SIC forecasting models. This module provides the crucial "coarse" forecast.
  2. Guided Diffusion Module (GDM): This module takes the 25 km forecasts from the FM and uses a pre-trained unconditional diffusion model to downscale them to a much higher resolution (6.25 km). The key innovation here is a "zero-shot guided sampling strategy" combined with a "patch-based method."
    • Guided Diffusion: The diffusion model's reverse process is conditioned on the low-resolution forecast. A novel aspect is the introduction of an "optimizable convolutional kernel" that simulates the degradation from high to low resolution. The gradient of a distance function between the upscaled degraded map and the low-resolution forecast is used to guide the diffusion process and dynamically update the kernel's parameters.
  3. Patch-based Strategy: To handle arbitrary resolutions and the fixed-size output of pre-trained diffusion models, the input SIC maps are divided into patches. For each patch, an individual optimizable convolutional kernel is used, and the final output is an aggregation of these patch-wise results. This allows for more precise guidance and better capture of small-scale details.

Datasets

The primary dataset used is the Climate Data Record of Passive Microwave Sea Ice Concentration Version 4 (CDR) from NSIDC. This dataset provides daily SIC data at a 25 km grid resolution, covering the pan-Arctic region from October 1978 to 2023. The paper also mentions using additional data generated through the retrieval algorithm (NASA Team [9], NT in Table 1) for comparison.

Key Insights and Ideas

  • Bridging the Resolution Gap: The core insight is that a high-quality, albeit coarse, forecast can effectively guide a generative model to produce a high-resolution output, overcoming the lack of direct high-resolution ground truth.
  • Generative Priors for Downscaling: Diffusion models, known for their generative capabilities, are leveraged not just for image synthesis but for guided downscaling, incorporating information from a lower-resolution source.
  • Dynamic Kernel for Degradation Simulation: The introduction of an optimizable convolutional kernel that learns to simulate the upscaling degradation is a novel approach to better align the low-resolution guidance with the high-resolution generation process.
  • Patch-based for Flexibility: The patch-based strategy is crucial for making the diffusion model applicable to arbitrary resolutions and for aggregating information effectively from multiple sub-regions.

Key Results

  • State-of-the-Art Forecasting: The forecasting module (FM) of IceDiff demonstrates superior performance across various metrics (RMSE, MAE, R2, NSE) compared to existing deep learning models like SICNet and IceNet for subseasonal and seasonal forecasts.
  • Unprecedented Resolution: For the first time, IceDiff achieves sea ice forecasting at a 6.25 km resolution, a significant four-fold increase in spatial detail compared to the standard 25 km.
  • High-Quality Downscaling: Qualitative results (Figure 5) show that IceDiff captures fine-scale structures and details much better than traditional interpolation methods, producing more realistic and visually appealing SIC maps. Quantitative evaluation using FID and Consistency metrics also shows significant improvements over interpolation and other generative methods like GDP.
  • Consistent Performance: The FM's forecasting skills are consistent across different SIC datasets and temporal scales.

Limitations and Future Work

  • Lack of High-Resolution Ground Truth: The authors acknowledge that a significant limitation is the absence of finer-scale ground truth SIC data for direct evaluation. Current evaluations rely on downscaling existing 25 km data.
  • Future Work: The paper suggests focusing on evaluating IceDiff on higher-resolution SIC observations (when available) and further exploring its application in operational forecasting and sea ice research. The framework's ability to capture small-scale details is highlighted as vital for pragmatic usages.

Summary updated

Abstract

Variation of Arctic sea ice has significant impacts on polar ecosystems, transporting routes, coastal communities, and global climate. Tracing the change of sea ice at a finer scale is paramount for both operational applications and scientific studies. Recent pan-Arctic sea ice forecasting methods that leverage advances in artificial intelligence have made promising progress over numerical models. However, forecasting sea ice at higher resolutions is still under-explored. To bridge the gap, we propose a two-module cooperative deep learning framework, IceDiff, to forecast sea ice concentration at finer scales. IceDiff first leverages a vision transformer to generate coarse yet superior forecasting results over previous methods at a regular 25km grid. This high-quality sea ice forecasting can be utilized as reliable guidance for the next module. Subsequently, an unconditional diffusion model pre-trained on low-resolution sea ice concentration maps is utilized for sampling down-scaled sea ice forecasting via a zero-shot guided sampling strategy and a patch-based method. For the first time, IceDiff demonstrates sea ice forecasting with a 6.25km resolution. IceDiff extends the boundary of existing sea ice forecasting models and more importantly, its capability to generate high-resolution sea ice concentration data is vital for pragmatic usages and research.
BibTeX
@inproceedings{xu2025icediff,
  author = {Jingyi Xu and Siwei Tu and Weidong Yang and Ben Fei and Shuhao Li and Keyi Liu and Yeqi Luo and Lipeng Ma and Lei Bai},
  title = {{IceDiff: High Resolution and High-Quality Arctic Sea Ice Forecasting with Generative Diffusion Prior}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2025},
  doi = {10.1109/cvpr52734.2025.00988},
}
Figure