A Self-Supervised Conditional Diffusion Network With Multihead Wavelet Attention for Remote Sensing Image Denoising

Abstract

Hyperspectral remote sensing images (HSIs) inevitably suffer from multiple sources of noise during acquisition, including Gaussian noise, impulse noise, and stripe noise, which severely degrade the reliability of subsequent interpretation tasks. To address the limitations of existing denoising methods, including their reliance on paired clean reference data and the difficulty in simultaneously preserving global structures and recovering local details, we propose a self-supervised conditional diffusion-based denoising network termed DM-WAN. Specifically, DM-WAN incorporates a detail feature prompting strategy into the diffusion denoising process, in which a time-dependent feature modulation mechanism is introduced to dynamically enhance edge and texture information. Moreover, a multihead wavelet attention mechanism is designed to exploit the band-separation capability of wavelet decomposition, enabling the model to capture global structures and local details across different frequency bands, thereby achieving collaborative multiscale feature fusion and structure-preserving reconstruction. During training, a self-supervised learning paradigm is adopted by constructing pseudo-supervised pairs from degraded images, which eliminates the dependence on paired ground-truth supervision. Experimental results demonstrate that the proposed method exhibits favorable performance in terms of detail preservation, structural consistency, and denoising robustness under various synthetic and real-noise scenarios, validating the effectiveness of DM-WAN for HSI denoising.