Frequency-Compensated Network for Daily Arctic Sea Ice Concentration Prediction

Jialiang Zhang, Feng Gao, Yanhai Gan, Junyu Dong, Qian Du

Introduction

Accurately predicting Arctic Sea Ice Concentration (SIC) is a critical task for climate modeling, navigation safety in emerging shipping lanes, and the preservation of Arctic ecosystems. While deep learning models have shown promise, they often struggle with the "fine print" of sea ice dynamics—specifically the jagged edges of the marginal ice zone and sudden fractures in the ice pack. A recent study introduces the Frequency-Compensated Network (FCNet), a dual-branch architecture designed to capture both broad-scale patterns and high-frequency details.

The Challenge: Spectral Bias in Sea Ice Prediction

Predicting SIC is inherently difficult due to the highly dynamic and non-linear nature of the Arctic environment. Existing deep learning models, such as ConvLSTMs or U-Nets, typically face two major hurdles:

  1. Long-term Dependencies: Standard convolutional layers often fail to capture long-range spatial and temporal relationships that are more apparent in the frequency domain.
  2. Spectral Bias: Neural networks naturally gravitate toward learning low-frequency patterns (the "big picture") while smoothing over high-frequency details (the "edges"). In the Arctic, these high-frequency details represent the marginal ice zone—the most critical area for navigation and the most sensitive to climate change.

Methodology: The FCNet Architecture

To address these challenges, the authors proposed FCNet, which splits feature extraction into two parallel branches. This design explicitly counters spectral bias by dedicating resources to high-frequency components.

Comparison between standard CNN frameworks and the proposed dual-branch FCNet.

Dual-Branch Feature Extraction

The FCNet architecture consists of two specialized branches:

  • Frequency Feature Extraction Branch: This branch utilizes an Adaptive Fourier Filter Block (AFFB). By transforming input data into the frequency domain via a Fast Fourier Transform (FFT), the model can apply learnable filters to capture global dependencies and periodic climate patterns that standard spatial convolutions might miss.
  • Convolutional Feature Extraction Branch: This branch features a High-Frequency Enhancement Block (HFEB). It separates data into high and low-frequency components. High-frequency features are boosted through channel-wise attention, while low-frequency features are processed via a temporal attention unit to capture long-term trends.
Detailed architecture of the Frequency-Compensated Network (FCNet).

Frequency-Aware Loss

A key innovation is the introduction of a Frequency Loss ($L_{freq}$) function. This function employs a spectral weight matrix to adaptively penalize the model more heavily for errors in difficult-to-learn high-frequency components. This forces the network to prioritize the fine details, such as "polynyas" (openings in the ice) and "leads" (cracks), which traditional models often blur.

Data and Experimental Setup

The study utilized the OSI-450-a dataset, providing satellite-derived SIC at a 25 km resolution from 1978 to 2020. The model was trained on data from 1991–2010 and tested on the 2016–2020 period, focusing on the AMAP area north of the Arctic Circle.

Key Results and Insights

FCNet was benchmarked against state-of-the-art models like IceNet and PredRNN, as well as the SEAS5 numerical model.

  • Quantitative Performance: FCNet achieved an average Mean Absolute Error (MAE) of 2.128% and a Nash-Sutcliffe Efficiency (NSE) of 0.967, consistently outperforming competitors.
  • Edge Accuracy: Using Binary Accuracy (BACC) to evaluate the ice edge, FCNet proved superior to the SEAS5 numerical model, particularly during the volatile summer melting season.
  • Robustness in Extremes: The model remained accurate during record-low ice years (e.g., 2012 and 2020), where traditional statistical models often fail due to a lack of historical analogs.

Conclusion and Future Directions

FCNet demonstrates that explicitly addressing spectral bias can significantly improve the prediction of complex geophysical phenomena. Future work may involve integrating multi-source data—such as sea surface temperature and ice thickness—and exploring "Physics-Informed Machine Learning" to bake thermodynamic laws directly into the neural architecture.

Reference: Zhang, J., Gao, F., Gan, Y., Dong, J., & Du, Q. (2025). Frequency-Compensated Network for Daily Arctic Sea Ice Concentration Prediction. IEEE Transactions on Geoscience and Remote Sensing.

Summary updated

BibTeX
@article{zhang2025frequency,
  author = {Jialiang Zhang and Feng Gao and Yanhai Gan and Junyu Dong and Qian Du},
  title = {{Frequency-Compensated Network for Daily Arctic Sea Ice Concentration Prediction}},
  journal = {IEEE Trans. Geosci. Remote. Sens.},
  year = {2025},
  doi = {10.1109/tgrs.2025.3564457},
}
Figure