SMDS-Net: Model Guided Spectral-Spatial Network for Hyperspectral Image Denoising.

Area of Science:

• Remote Sensing
• Computer Vision
• Signal Processing

Background:

• Deep learning (DL) models for hyperspectral image (HSI) denoising often lack interpretability.
• These models typically ignore the inherent physical characteristics of HSIs, limiting their understanding and performance.
• Existing methods struggle to effectively leverage spatial redundancy, spectral low-rankness, and spectral-spatial correlations.

Purpose of the Study:

• To develop a novel model-guided interpretable network for hyperspectral image denoising.
• To address the lack of interpretability in current DL-based HSI denoising methods.
• To improve denoising performance by considering the physical properties of HSIs.

Main Methods:

• A subspace-based multidimensional sparse (SMDS) model was established using tensor notation, incorporating HSI characteristics.
• The SMDS model was unfolded into an end-to-end network, SMDS-Net, integrating denoising and model optimization.
• Discriminative training was employed to obtain key variables, enabling clear physical interpretations.

Main Results:

• SMDS-Net demonstrated strong denoising capabilities on both synthetic and real-world HSIs.
• The network exhibited robust learning and generalization abilities compared to state-of-the-art methods.
• Experiments confirmed the high interpretability of SMDS-Net, aligning with its physical model basis.

Conclusions:

• SMDS-Net offers a significant advancement in interpretable HSI denoising by integrating physical models into a DL framework.
• The proposed method effectively balances denoising performance with model interpretability.
• The publicly available code and data facilitate reproducible research in HSI denoising.