Understanding the Limitations of Deep Transformer Models for Sea Ice Forecasting - ICCS 2025

Content

This repo contains code and results of long-term sea ice concentration forecasting of four models: 2D CNN, 3D CNN, TimeSformer and SwinLSTM.

Code and results for toy example with loop video also presented for all models.

Supplementary materials as PDF can be loaded as file.

Sea ice concentration data

We used OSI SAF Global Sea Ice Concentration product as source data. The spatial resolution of the images is reduced to 14 km. Five Arctic seas were selected as test areas, its spatial position presented on image:

Sea ice - Results

Table with main results of 52 weeks ahead forecast with implemented models and its comparison with SOTA numerical (SEAS5) and deep learning (IceNet) solutions for sea ice prediction problem:

Metric	MAE SEAS5	MAE 2D CNN	MAE 3D CNN	MAE TimeS former	SSIM SEAS5	SSIM 2D CNN	SSIM 3D CNN	SSIM TimeS former	Accuracy Ice Net	Accuracy 2D CNN	Accuracy 3D CNN	Accuracy TimeS former
Kara Sea	0.093	0.076	0.076	0.109	0.653	0.683	0.663	0.581	0.918	0.945	0.943	0.929
Barents Sea	0.073	0.063	0.060	0.129	0.634	0.684	0.672	0.489	0.906	0.922	0.944	0.916
Laptev Sea	0.101	0.068	0.072	0.146	0.703	0.722	0.706	0.608	0.967	0.982	0.980	0.966
East-Siberian Sea	0.098	0.074	0.069	0.177	0.723	0.718	0.714	0.685	0.980	0.990	0.990	0.988
Chukchi Sea	0.067	0.075	0.073	0.147	0.780	0.713	0.719	0.588	0.974	0.979	0.981	0.962

Visualization of one timestep of prediction for Kara sea presented on image:

Convergence plots for models, extended tables with metrics by quarters and SwinLSTM prediction results are presented in file.

Loop video prediction - Results

As a synthetic data gif-file with loop video was used. Images was set to gray scale and resized to 45x45 pixels.

To understand convergence process of 2D CNN and TimeSformer and get reason of TimeSformer's quality lack inference for each epoch during optimization was produced.

2D CNN convergence process:

TimeSformer convergence process:

Convergence plots and metrics estimation are in supplementary and in the paper.