⚡ electrical engineering

Frequency-Enhanced Hilbert Scanning Mamba for Short-Term Arctic Sea Ice Concentration Prediction

This paper proposes the Frequency-enhanced Hilbert scanning Mamba (FH-Mamba) framework, which integrates a 3D Hilbert scan mechanism, wavelet transforms, and a Hybrid Shuffle Attention module to effectively address temporal correlation and boundary detail challenges, thereby achieving superior short-term Arctic sea ice concentration prediction compared to state-of-the-art baselines.

Original authors: Feng Gao, Zheng Gong, Wenli Liu, Yanhai Gan, Zhuoran Zheng, Junyu Dong, Qian Du

Published 2026-02-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Feng Gao, Zheng Gong, Wenli Liu, Yanhai Gan, Zhuoran Zheng, Junyu Dong, Qian Du

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

🧊 The Big Picture: Predicting the Melting Ice

Imagine the Arctic Ocean as a giant, shifting puzzle made of ice and water. Scientists need to predict how this puzzle will look a few weeks from now. This is crucial because melting ice changes global weather, and knowing where the ice is helps ships navigate safely.

For a long time, computers have tried to solve this puzzle. Some use complex physics equations (like trying to calculate every drop of water), while others use Deep Learning (AI that learns by looking at pictures). The newest AI stars are called Mamba models. They are incredibly fast and good at remembering long sequences of events, like reading a book from start to finish without forgetting the beginning.

But there's a problem: When standard Mamba models look at the Arctic ice, they get a little confused. They tend to miss the "fuzzy edges" where ice meets water, and they sometimes lose track of how the ice moves from one day to the next.

🚀 The Solution: FH-Mamba

The authors of this paper built a new, super-charged AI called FH-Mamba. Think of it as giving the AI a pair of high-tech glasses and a better map. Here is how it works, broken down into three simple tricks:

1. The "Hilbert Scanning" Trick (The Golden Thread)

The Problem: Imagine you have a 3D block of ice (width, height, and time). A standard AI reads this block like a typewriter: it reads line by line, left to right, top to bottom.

The Flaw: If you read a page this way, the pixel at the bottom-right of "Tuesday" is physically next to the top-left of "Wednesday" in the computer's memory. But in reality, they are far apart in time and space! The AI gets confused about what happened next.

The Fix: The authors use a 3D Hilbert Scan.

The Analogy: Imagine a golden thread (the Hilbert curve) that winds through the entire 3D block of ice. It doesn't just go in straight lines; it snakes back and forth, ensuring that every point it touches is physically close to the next point it touches.
The Result: The AI reads the data in a way that keeps "neighbors" together. It sees that the ice at a specific spot on Tuesday is right next to the ice at that same spot on Wednesday. This helps the AI understand the flow of time and space much better.

2. The "Wavelet" Trick (The High-Definition Lens)

The Problem: The most dangerous and important parts of the Arctic are the margins—the edges where thick ice turns into slush and then open water. Standard AI models are great at seeing the big, solid ice blocks but often blur these messy, jagged edges. It's like looking at a photo that's slightly out of focus.

The Fix: They added a Wavelet Transform.

The Analogy: Think of the ice map as a song. The big ice sheets are the low, bass notes. The jagged, messy edges are the high-pitched, sharp notes. Standard AI often ignores the high notes.
The Result: The Wavelet module acts like a specialized ear that listens only to the high-pitched sounds. It amplifies the details of the edges, ensuring the AI doesn't blur the boundary between ice and water.

3. The "Hybrid Shuffle" Trick (The Team Huddle)

The Problem: Now the AI has two pieces of information: the "flow" of time (from the Hilbert scan) and the "sharp details" of the edges (from the Wavelet). How do you combine them without them getting in each other's way?

The Fix: They built a Hybrid Shuffle Attention (HSA) module.

The Analogy: Imagine two experts in a meeting. One is a Time Traveler (knows the history), and the other is a Microscope User (knows the tiny details). Instead of them talking over each other, the HSA module acts like a smart moderator. It shuffles their notes, mixes them up, and says, "Okay, Time Traveler, look at this specific edge detail. Microscope User, look at this time sequence."
The Result: The two experts work together perfectly, creating a prediction that is both historically accurate and visually sharp.

🏆 The Results: Why It Matters

The team tested their new AI on real satellite data from the Arctic.

Better Accuracy: It predicted the ice concentration more accurately than any other method currently available.
Sharper Edges: When they looked at the maps, the FH-Mamba predictions showed crisp, clear ice edges, whereas other models looked blurry.
Efficiency: Despite being smarter, it wasn't slower or heavier than the other models. It's like upgrading a sports car's engine without making it heavier.

💡 The Takeaway

The Arctic is changing fast, and we need tools that can keep up. This paper shows that by changing how an AI reads data (the Hilbert scan) and by giving it a special tool to see the tiny details (Wavelets), we can create a much better crystal ball for predicting the future of our planet's ice. It's a small tweak in the math that leads to a huge leap in understanding our climate.

1. Problem Statement

The paper addresses the challenge of short-term Arctic Sea Ice Concentration (SIC) prediction, a critical task for climate modeling and safe navigation. While deep learning models have improved SIC forecasting, existing approaches face two primary limitations when applied to this specific domain:

Ineffective Temporal Correlation Modeling: Standard "Vanilla" Mamba models (State Space Models) typically flatten 3D spatiotemporal data (Time $\times$ Height $\times$ Width) into a 1D sequence using fixed scanning orders (e.g., raster scan). This disrupts the inherent spatiotemporal locality, causing adjacent pixels in time and space to become distant in the sequence, making it difficult to capture local evolution patterns of sea ice.
Loss of Boundary Details: Mamba models tend to erode fine spatial details, particularly in Arctic marginal ice zones where boundaries are subtle and scattered. Standard attention mechanisms often fail to preserve these critical high-frequency edge details.

2. Methodology: FH-Mamba Framework

The authors propose FH-Mamba (Frequency-enhanced Hilbert Scanning Mamba), a novel framework designed to overcome the above limitations. The architecture consists of a feature encoder, a series of Frequency-enhanced State Space Modules (FSSM), and a feature decoder.

Key Components:

3D Hilbert Scanning Mechanism:
- Instead of standard raster scanning, the model employs a 3D Hilbert curve to traverse the spatiotemporal grid.
- Mechanism: The Hilbert curve is a space-filling curve that maps a 1D interval to a multi-dimensional space while preserving locality. In FH-Mamba, it ensures that adjacent indices in the flattened sequence correspond to neighboring voxels in both spatial and temporal dimensions.
- Benefit: This preserves the topological structure of sea ice evolution, allowing the State Space Model (SSM) to effectively capture local temporal correlations and spatial continuity.
Wavelet Transform for High-Frequency Enhancement:
- To address the loss of boundary details, the model incorporates a Wavelet Transform branch.
- Mechanism: This branch decomposes the input to amplify high-frequency details, which correspond to the sharp edges and fine structures of the sea ice margin.
Hybrid Shuffle Attention (HSA) Module:
- Function: A novel attention mechanism designed to fuse the sequence features (from the 3D Hilbert Mamba branch) and frequency features (from the Wavelet branch).
- Process: It applies spatial average pooling, concatenates features, performs a "shuffle" operation to interleave channels, and then uses group convolution to generate channel-wise attention weights. These weights are unshuffled and used to perform a weighted summation of the sequence and frequency features.
- Benefit: This allows the model to adaptively aggregate complementary information from both domains, ensuring that high-frequency boundary details are preserved while maintaining global spatiotemporal consistency.
Loss Function:
- The model is trained using a combination of L1 Loss (for pixel-level reconstruction) and Gradient Loss (to explicitly preserve edges and fine details).

3. Key Contributions

Novel Scanning Strategy: Introduction of a 3D Hilbert scanning mechanism for Vision Mamba, which simultaneously traverses temporal and spatial dimensions to preserve locality, solving the temporal correlation gap in standard Mamba models.
Frequency-Aware Architecture: Development of a dual-branch framework combining Wavelet transforms (for edge enhancement) and Hybrid Shuffle Attention (HSA) to effectively fuse sequence and frequency features.
State-of-the-Art Performance: Demonstration of superior prediction accuracy and edge reconstruction compared to existing deep learning baselines (including ConvLSTM, Transformers, and other Mamba variants) on two distinct datasets.
Open Source: Public release of the code to support the climate research community.

4. Experimental Results

The model was evaluated on two benchmark datasets: OSI-450a1 (coarse resolution, 25km) and AMSR2 (fine resolution, 3.125km).

Quantitative Performance:
- On OSI-450a1, FH-Mamba achieved the lowest RMSE (6.60%–7.07%) and MAE (1.99%–2.21%), outperforming SOTA methods like IceMamba, SwinLSTM, and FCNet.
- On AMSR2, it achieved an RMSE of 10.05% and MAE of 2.54%, demonstrating strong generalization across different sensor resolutions.
- Statistical significance tests (paired t-tests) confirmed that improvements over baselines like VMRNN are significant ( $p < 0.001$ ).
Qualitative Analysis:
- Edge Preservation: Visualizations of extreme low-ice events (e.g., 2007, 2012, 2020) show FH-Mamba produces sharper boundaries and fewer biases in marginal ice zones compared to competitors.
- Sea Ice Extent (SIE): The model achieved the highest Intersection over Union (IoU) score (0.8944 average) for delineating the sea ice edge, crucial for navigation safety.
Efficiency:
- FH-Mamba maintains linear computational complexity. It has 24.87M parameters and 236.92G FLOPs, making it more computationally efficient than many Transformer-based models while offering higher accuracy.
Ablation Studies:
- Removing any of the three core components (3DHSS, Wavelet, HSA) resulted in performance degradation, confirming their complementary roles.
- Scanning Direction: Temporal-first Hilbert scanning yielded slightly better results than spatial-first, highlighting the importance of modeling temporal continuity.
- Wavelet Basis: The Haar wavelet was selected for its computational simplicity and effectiveness in capturing high-frequency edges.

5. Significance

This paper makes a significant contribution to geoscience and remote sensing by bridging the gap between efficient sequence modeling (Mamba) and the specific physical requirements of sea ice prediction.

Scientific Impact: It provides a more accurate tool for understanding Arctic climate dynamics, particularly the rapid changes in marginal ice zones which are sensitive indicators of global warming.
Practical Application: The improved accuracy in boundary reconstruction directly benefits Arctic navigation safety and geopolitical planning by providing more reliable short-term forecasts.
Methodological Advancement: It establishes a new paradigm for applying State Space Models to 3D spatiotemporal data, demonstrating that locality-preserving scanning and frequency-domain enhancement are critical for high-fidelity environmental forecasting.