[Paper Review] Sea Ice Forecasting using Attention-based Ensemble LSTM

Sahara Ali, Yiyi Huang+2 more|arXiv (Cornell University)|Jul 27, 2021

Arctic and Antarctic ice dynamics10 references18 citations

TL;DR

This paper proposes an attention-based ensemble Long Short-Term Memory (LSTM) model to forecast monthly Arctic sea ice extent up to one month in advance, leveraging daily and monthly atmospheric, oceanic, and sea ice data from ERA5 and NSIDC over 1979–2018. The method combines multi-temporal models using an attention mechanism that improves prediction accuracy, achieving a 4.11% RMSE—significantly outperforming baseline and recent deep learning models.

ABSTRACT

Accurately forecasting Arctic sea ice from subseasonal to seasonal scales has been a major scientific effort with fundamental challenges at play. In addition to physics-based earth system models, researchers have been applying multiple statistical and machine learning models for sea ice forecasting. Looking at the potential of data-driven sea ice forecasting, we propose an attention-based Long Short Term Memory (LSTM) ensemble method to predict monthly sea ice extent up to 1 month ahead. Using daily and monthly satellite retrieved sea ice data from NSIDC and atmospheric and oceanic variables from ERA5 reanalysis product for 39 years, we show that our multi-temporal ensemble method outperforms several baseline and recently proposed deep learning models. This will substantially improve our ability in predicting future Arctic sea ice changes, which is fundamental for forecasting transporting routes, resource development, coastal erosion, threats to Arctic coastal communities and wildlife.

Motivation & Objective

• To improve sub-seasonal to seasonal Arctic sea ice forecasting using data-driven deep learning models.
• To address the challenge of low predictability during the Arctic melting season (July–September).
• To develop an ensemble method that integrates models trained on different temporal resolutions (daily and monthly) for enhanced forecasting accuracy.
• To incorporate an attention mechanism that dynamically weights contributions from constituent models to reduce prediction error.
• To provide a scalable, interpretable, and accurate forecasting framework for climate resilience and Arctic planning.

Proposed method

• The model uses two primary branches: a daily-to-monthly LSTM (d-LSTM) and a monthly-to-monthly LSTM (m-LSTM), trained on daily and monthly time-series data, respectively.
• An ensemble model (E-LSTM) combines predictions from d-LSTM and m-LSTM using learned attention weights to prioritize more informative model outputs.
• The attention mechanism computes dynamic weights for each constituent model based on their predictive performance, minimizing overall loss during training.
• The final model, EA-LSTM, integrates attention into the ensemble framework, allowing the network to focus more on high-performing models across different temporal and seasonal conditions.
• The method is trained on 34 years (1979–2012) of data and tested on the final 5 years (2013–2018), using area-weighted averaging over the Arctic north of 25°N.
• The architecture is extensible to include additional model branches trained on different temporal or spatial resolutions.

Experimental results

Research questions

• RQ1 Can an ensemble of multi-temporal deep learning models improve monthly Arctic sea ice extent forecasting accuracy compared to single-model baselines?
• RQ2 How does incorporating an attention mechanism enhance the performance and robustness of the ensemble model, especially during the low-predictability melting season?
• RQ3 To what extent does the attention-based ensemble outperform recent state-of-the-art deep learning models in sea ice forecasting?
• RQ4 Can the attention mechanism effectively learn to weight model contributions based on temporal resolution and seasonal variability?
• RQ5 How does the model perform in capturing inter-annual variability and seasonal transitions in sea ice extent?

Key findings

• The proposed EA-LSTM model achieves a test RMSE of 4.11% for monthly sea ice extent prediction, significantly outperforming baseline models.
• The attention-based ensemble reduces model loss and improves accuracy compared to both the E-LSTM (without attention) and individual d-LSTM/m-LSTM models.
• The EA-LSTM model shows superior performance during the challenging melting season (July–September), where predictability is typically low.
• Compared to Chi & Kim (2017), the model achieves a 1.65% reduction in RMSE, and compared to Kim et al. (2020), a 8% reduction, demonstrating strong performance gains.
• The attention mechanism effectively learns to weight model contributions, with higher attention assigned to models that perform better under specific conditions.
• Visualizations show that EA-LSTM predictions closely follow observed sea ice trends from 2014 to 2018, particularly in capturing seasonal decline and recovery patterns.