A hyperspectral image denoising method based on land cover spectral autocorrelation

Developing denoising algorithms for hyperspectral remote sensing images (HSIs) can alleviate noise problem, improve data utilization as well as the accuracy of subsequent applications. However, existing denoising techniques are usually unstable due to the variations of landscapes, resulting in local distortion of HSIs, especially in heterogeneous areas. To tackle this issue, we propose a spatial-spectral interactive restoration (SSIR) framework by exploiting the complementarity of model-based and data-driven methods. Specifically, a deep learning-based denoising module that incorporates both convolutional neural networks (CNN) and Swin Transformer (TF) blocks is designed. This denoiser can achieve local-global dependencies modeling and content-based interactions to better capture global heterogeneity differences in HSIs. Moreover, we introduce an unsupervised unmixing module that utilizes spectral autocorrelation as prior information to effectively capture the differences in reflectance characteristics among different land cover components. This parameter-free module further improves the generalization ability of SSIR and enables stable denoising performance across different scenarios. Both modules are iteratively updated and fuel each other in SSIR. The proposed SSIR is shown to outperform others in preserving spatial details, maintaining spectral fidelity, and adapting to different landscapes based on simulated and real experiments conducted on various HSIs under diverse noise conditions.