Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Seasonal Forecasting of Pan-Arctic Sea Ice with State Space Model

Wei Wang, Weidong Yang, Lei Wang, Guihua Wang, Ruibo Lei

TL;DR

The study tackles the challenge of seasonal Pan-Arctic SIC forecasting by introducing IceMamba, a deep-learning model that embeds sophisticated attention within a state-space framework to capture multiscale sea-ice dynamics. Leveraging ERA5 and ORAS5 reanalysis alongside SIC data at 25 km resolution, IceMamba combines Residual Efficient State Space Blocks with Vision State Space Blocks and Efficient Channel Attention to forecast monthly SIC over multiple lead times. In a rigorous benchmark against 25 dynamical, statistical, and DL models, IceMamba achieves leading performance in $RMSE$ and $ACC$, with strong edge-forecast skill ($IIEE$), and demonstrates enhanced performance with subsurface ocean data, particularly for longer horizons. The work underscores the potential of integrating state-space dynamics with neural attention for physically consistent, interpretable climate forecasts, offering a scalable tool for climate adaptation planning and advancing DL-based sea-ice forecasting.

Abstract

The rapid decline of Arctic sea ice resulting from anthropogenic climate change poses significant risks to indigenous communities, ecosystems, and the global climate system. This situation emphasizes the immediate necessity for precise seasonal sea ice forecasts. While dynamical models perform well for short-term forecasts, they encounter limitations in long-term forecasts and are computationally intensive. Deep learning models, while more computationally efficient, often have difficulty managing seasonal variations and uncertainties when dealing with complex sea ice dynamics. In this research, we introduce IceMamba, a deep learning architecture that integrates sophisticated attention mechanisms within the state space model. Through comparative analysis of 25 renowned forecast models, including dynamical, statistical, and deep learning approaches, our experimental results indicate that IceMamba delivers excellent seasonal forecasting capabilities for Pan-Arctic sea ice concentration. Specifically, IceMamba outperforms all tested models regarding average RMSE and anomaly correlation coefficient (ACC) and ranks second in Integrated Ice Edge Error (IIEE). This innovative approach enhances our ability to foresee and alleviate the effects of sea ice variability, offering essential insights for strategies aimed at climate adaptation.

Seasonal Forecasting of Pan-Arctic Sea Ice with State Space Model

TL;DR

The study tackles the challenge of seasonal Pan-Arctic SIC forecasting by introducing IceMamba, a deep-learning model that embeds sophisticated attention within a state-space framework to capture multiscale sea-ice dynamics. Leveraging ERA5 and ORAS5 reanalysis alongside SIC data at 25 km resolution, IceMamba combines Residual Efficient State Space Blocks with Vision State Space Blocks and Efficient Channel Attention to forecast monthly SIC over multiple lead times. In a rigorous benchmark against 25 dynamical, statistical, and DL models, IceMamba achieves leading performance in and , with strong edge-forecast skill (), and demonstrates enhanced performance with subsurface ocean data, particularly for longer horizons. The work underscores the potential of integrating state-space dynamics with neural attention for physically consistent, interpretable climate forecasts, offering a scalable tool for climate adaptation planning and advancing DL-based sea-ice forecasting.

Abstract

The rapid decline of Arctic sea ice resulting from anthropogenic climate change poses significant risks to indigenous communities, ecosystems, and the global climate system. This situation emphasizes the immediate necessity for precise seasonal sea ice forecasts. While dynamical models perform well for short-term forecasts, they encounter limitations in long-term forecasts and are computationally intensive. Deep learning models, while more computationally efficient, often have difficulty managing seasonal variations and uncertainties when dealing with complex sea ice dynamics. In this research, we introduce IceMamba, a deep learning architecture that integrates sophisticated attention mechanisms within the state space model. Through comparative analysis of 25 renowned forecast models, including dynamical, statistical, and deep learning approaches, our experimental results indicate that IceMamba delivers excellent seasonal forecasting capabilities for Pan-Arctic sea ice concentration. Specifically, IceMamba outperforms all tested models regarding average RMSE and anomaly correlation coefficient (ACC) and ranks second in Integrated Ice Edge Error (IIEE). This innovative approach enhances our ability to foresee and alleviate the effects of sea ice variability, offering essential insights for strategies aimed at climate adaptation.
Paper Structure (4 sections, 5 equations, 8 figures, 1 table)

This paper contains 4 sections, 5 equations, 8 figures, 1 table.

Table of Contents

  1. Introduction
  2. Result
  3. Discussion
  4. Method

Figures (8)

  • Figure 1: Comparison of IceMamba-6 and Anomaly Persistence. (a), (c), and (e): MAE, RMSE, and ACC of IceMamba-6, averaged over the test years (2016-2022), presented for each target month (month of prediction) and lead time, with heatmap values in each grid cell. (b), (d), and (f): Heatmaps showing the differences in MAE, RMSE, and ACC between IceMamba-6 and Anomaly Persistence. All metrics are calculated over non-land regions in the Pan-Arctic.
  • Figure 2: Seasonal cycle of IIEE and its two components, overestimated error (OE) and underestimated error (UE), during the test period from sea ice forecast of IceMamba-6, IceMamba-6-ERA5, and Anomaly Persistence, averaged over six lead times.
  • Figure 3: Comparative visualization of predicted sea ice edge boundaries (colored contours) and Integrated Ice-Edge Error (IIEE) for August-September 2016 forecasts generated by IceMamba-6 (red) and the Anomaly Persistence benchmark (black), evaluated across 1 to 6 month lead times.
  • Figure 4: RMSE, ACC, and IIEE for September SIC forecast (2001–2020). (a), (c), (e) show results for dynamical models, and (b), (d), (f) for statistical models. RMSE and ACC are averaged over regions where SIC standard deviation $> 10\%$. Models are color-coded, with reference forecasts in grey. Skill metrics are shown for each initialization from June 1 to September 1. Bracketed numbers in the legend indicate the years of data each model contributed over 20 years.
  • Figure 5: Box plot illustration: RMSE, ACC, and IIEE for September SIC forecast (2001–2020) from models that contribute a full 20-year forecast. Panels (a), (b), and (c) show RMSE, ACC, and IIEE across models, averaged over regions with SIC standard deviation $> 10\%$. Models are color-coded: grey (reference), orange (statistical), and purple (dynamical). IceMamba-4 and IceMamba-1-only-SIC are highlighted in red and blue, respectively.
  • ...and 3 more figures