Abstract
Hyperspectral image (HSI) denoising requires capturing complex dependencies across both spatial and spectral dimensions. While Convolutional Neural Networks (CNNs) excel at local feature extraction and Transformers offer global context, they respectively struggle with long-range modeling and quadratic computational complexity. This paper introduces SSUMamba, a novel U-shaped architecture that adapts the Mamba state space model for 3D HSI data. By employing a Spatial-Spectral Continuous Scan (SSCS) strategy, SSUMamba achieves global receptive fields with linear complexity. Experimental results demonstrate that SSUMamba outperforms state-of-the-art Transformer models in both restoration quality and memory efficiency.
Introduction
Hyperspectral images (HSIs) provide rich information by capturing hundreds of narrow spectral bands for every pixel. This 3D "data cube" is invaluable for applications such as mineral mapping, precision agriculture, and object detection. However, the high-dimensional nature of HSIs makes them highly susceptible to noise from sensor mechanics and environmental factors, which can degrade the performance of downstream analysis.
Effective HSI denoising must balance local texture preservation with global spectral correlation. Traditional CNNs are effective at local modeling but often fail to capture long-range dependencies. Conversely, Transformers have set new benchmarks in global modeling but suffer from $O(n^2)$ computational complexity, making them prohibitively memory-intensive for high-resolution 3D data. SSUMamba addresses these challenges by leveraging the linear-time complexity of State Space Models (SSMs).
The SSUMamba Architecture
The core innovation of SSUMamba is the adaptation of the Mamba architecture—originally designed for 1D sequences—to the 3D HSI domain. The model utilizes a U-shaped encoder-decoder framework integrated with the Spatial-Spectral Continuous Scan (SSCS) Mamba block.
Spatial-Spectral Continuous Scan (SSCS)
Standard scanning methods often "jump" from the end of one row to the beginning of the next, which breaks the spatial continuity of the data. SSCS introduces a Continuous Zigzag Scanning pattern. This ensures that pixels adjacent in the flattened 1D sequence remain neighbors in the original 3D cube, preserving the "locality" essential for image processing and preventing the model from "forgetting" local pixel relationships.
Six-Directional Modeling
To fully exploit the volumetric nature of HSIs, SSUMamba employs six different scan orders (e.g., Row-Column-Band, Band-Row-Column, etc.). These scans are processed using bidirectional SSMs, resulting in a total of 12 directional modeling paths. This comprehensive coverage allows the model to capture dependencies across all spatial and spectral axes simultaneously.
Hybrid Local-Global Design
While the Mamba blocks handle long-range correlations, the authors embed 3D residual convolutional blocks within the architecture. This hybrid approach ensures that the model maintains high-frequency local details (via CNNs) while benefiting from the global receptive field provided by the SSMs.
Experimental Results
Quantitative Performance
The model was benchmarked against leading CNN and Transformer-based denoisers (such as TRQ3D, SST, and SERT) using both synthetic and real-world datasets. On the ICVL dataset under mixture noise conditions, SSUMamba achieved a PSNR of 43.07 dB, a significant improvement over the 40.54 dB achieved by the next-best model (NSSNN).
Efficiency and Scalability
One of the most significant advantages of SSUMamba is its efficiency. Because Mamba scales linearly with sequence length, the model avoids the "memory wall" typical of Transformers. SSUMamba utilizes approximately 10M parameters, compared to the ~15M parameters required by models like SERT, while maintaining lower memory consumption per batch.
Qualitative Analysis
Visual evaluations on real-world data, such as the Gaofen-5 Wuhan HSI, indicate that SSUMamba is particularly effective at removing complex noise patterns, including "striping" and impulse noise. Unlike competing models that often produce blurred edges or spectral distortions, SSUMamba preserves sharp spatial boundaries and maintains high spectral fidelity—the reflectance curve across bands remains consistent with the physical properties of the materials imaged.
Conclusion
SSUMamba represents a significant step forward in HSI denoising by successfully adapting State Space Models to 3D data. By combining the efficiency of linear-complexity modeling with a continuous scanning strategy that respects the physical structure of images, it offers a superior alternative to Transformer-based architectures. Future work will explore the application of the SSCS Mamba block to other HSI tasks, such as super-resolution and classification, to further validate its versatility in high-dimensional signal processing.
Summary updated