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

TL;DR

This work tackles the need for high-resolution Arctic SIC forecasts by introducing IceDiff, a two-stage framework that couples a Swin Transformer–based forecasting model with a diffusion-based super-resolution module. The first stage (IceDiff-FM) yields accurate 25 km SIC forecasts across subseasonal to seasonal leads, while the second stage (IceDiff-GDM) uses a guided diffusion prior to produce 6.25 km SIC maps via zero-shot sampling, guided by the low-resolution forecasts and aided by a patch-based downscaling strategy. The approach delivers state-of-the-art performance on three temporal scales and demonstrates the first practical 6.25 km SIC forecasts, offering a valuable tool for operational planning and scientific studies. By integrating a patch-based, optimizable upscaling mechanism with a pre-trained diffusion prior, IceDiff generalizes to arbitrary output sizes and enables high-resolution sea ice mapping with realistic texture and detail.

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 has 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-staged deep learning framework, IceDiff, to forecast sea ice concentration at finer scales. IceDiff first leverages an independently trained vision transformer to generate coarse yet superior forecasting over previous methods at a regular 25km x 25km grid. This high-quality sea ice forecasting can be utilized as reliable guidance for the next stage. Subsequently, an unconditional diffusion model pre-trained on 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 the 6.25km x 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.

Figures (4)

• Figure 1: Overview of IceDiff Framework. IceDiff first employs the forecasting model (FM) that takes history SIC data as input to generate subsequent predictions at 25km x 25km grid. N denotes the number of sequential time stamps, for instance, set N to 7 for seven days of forecasting leads, FM takes a tensor with shape of $\{7, 448, 304 \}$ as input and forecasts the SIC for the next 7 days. C represents the quantity of embedded channels within each layer of FM, the original SIC map data has a channel of 1 rather than 3 for regular images. The guided diffusion module (GDM) then leverages accurate forecasts as guidance to generate downscaled samples at 6.25km x 6.25km grid with high quality, see details in Figure \ref{['fig:enter-label']}.
• Figure 2: Overview of GDM for SIC map dowdnscaling.(a) Pre-trained diffusion model is used to estimate $\tilde{x}_0$ at reverse step $t$. The optimizable convolutional kernel of step $t$ reduces the resolution of $X_t$ so that guidance from low-resolution SIC maps can be incorporated to sample $X_{t-1}$. (b) Low-resolution SIC map $Y$ is integrated during the sampling process. A distance function is introduced to describe the distance between a low-resolution SIC map and the generated high-resolution SIC map $x$ after convolution. The gradient of distance function could be used to guide the sampling of $x_{t-1}$ and (c) dynamically update the parameter of the convolutional kernel.
• Figure 3: Illustration of the patch-based diffusion model pipeline. We set patch size $p=256$ and stride $r=128$ to fit the size of pre-trained models. This setting will divide the example map into $K=4$ sub maps boxed with different colors. For each sub map, individual optimizable convolution kernel with adaptive parameters is set to add guidance from corresponding low resolution SIC sub map. At each sampling step $t$, the average estimated noise on four overlapping blocks will be utilized to update the mean $\mu$ used to sample $x_{t-1}$.
• Figure 4: Qualitative analysis of down-scaled SIC map. We visualize the SIC map by coloring the land and the ocean in yellow and blue, respectively. Since SIC has a range of $[0,1]$, we map it to $[0,255]$ and replicate it for all 3 RGB channels. The darker the pixels appear, the SIC values are more closer to 0%. For better identification of the difference, we crop out a region and plot a grid size of 250km and further displays details on 4 x 4 grid.