HSDSSM: a hybrid spectral denoising state-space model for hyperspectral images

Abstract

Hyperspectral image (HSI) denoising is a critical yet challenging task due to the complexity of noise patterns and the need to preserve spectral and spatial details. Addressing these challenges requires models that can effectively capture subtle spectral variations and broader spectral relationships. To tackle this, we introduce HSDSSM: A Hybrid Spectral Denoising StateSpace Model for Hyperspectral Images, a novel approach that combines the strengths of Spectral State-Space Models (S-SSM) and Spectral Self-Attention (SSA) within a unified framework. At the core of HSDSSM is the Sequential Mamba and SelfAttention Block (SMSAB), where Spectral Self-Attention (SSA) first captures fine-grained dependencies, followed by the Spectral State-Space Model (S-SSM) to enhance long-range correlations. Finally, a Simple Gate (SGate) enhances the processed features, ensuring effective noise suppression and spectral coherence. We evaluate HSDSSM on the ICVL dataset across various noise scenarios, including non-iid Gaussian, stripe, impulse, deadline, and mixed noise. Our model consistently outperforms state-ofthe-art methods, achieving higher Mean Peak Signal-to-Noise Ratio (MPSNR) and Mean Structural Similarity Index Measure (MSSIM) scores while maintaining the lowest Spectral Angle Mapper (SAM) values.