SVD-Mamba: A Spatial–Spectral Model for Denoising Hyperspectral Images
Abstract
Convolutional neural networks (CNNs) are effective for denoising hyperspectral images (HSIs), but are limited to local feature modeling. Transformers capture global spatial–spectral dependencies but incur high computational and memory costs. State space models (SSMs), such as Mamba, provide an efficient alternative by modeling long-range dependencies with linear time complexity. Despite this advantage, a limitation of Mamba is its difficulty in adequately capturing the fine-grained local structure and the strong spectral correlations inherent in HSIs. To address this limitation, we leverage the effective guidance provided by low-rank (LR) representations derived from singular value decomposition (SVD). We propose a hybrid framework that integrates an LR prior with Mamba. Experiments on real and simulated datasets demonstrate the superiority of the proposed method.