WFD-GAN: A Filter-Based GAN for Denoising in Strip Noise Remote Sensing Images for Removal

Abstract As remote sensing technology rapidly advances, stripe noise has become a significant problem that impacts data quality in satellite remote sensing images. To enhance the quality of remote sensing images and thereby improve the accuracy of remote sensing products, such as aerosol products, this study introduces a strip noise removal method for remote sensing images based on Wave Filter Denoise Generation Adversarial Network (WFD-GAN). This method innovatively integrates waveform filters as prior knowledge to guide the generative adversarial network (GAN) in more precisely separating noise from image structures during end-to-end training. It employs a combination of frequency domain loss (FFT loss), perceptual loss, and L1 reconstruction loss to jointly optimize the generator, while introducing the SE (Squeeze-and-Excitation) channel attention mechanism to strengthen the network’s ability to suppress stripe noise. Through these designs, the method achieves effective denoising without altering the essential information of the original image.Experimental results demonstrate that when WFD-GAN is applied to a dataset of 9,760 noisy sub-images generated from 130 AGRI scenes, it outperforms traditional methods in removing stripe noise while better preserving image details and structures. Ablation studies further confirm that the wave filter and SE channel attention module play a critical role in enhancing denoising performance. Moreover, WFD-GAN surpasses existing state-of-the-art methods across multiple evaluation metrics, offering not only new insights for high-quality remote sensing image denoising, but also practical potential for improving the accuracy of downstream applications such as aerosol product retrieval, land surface monitoring, and climate observation.