Research on strip noise removal in remote sensing images based on frequency-spatial collaborative networks

Abstract

Stripe noise in remote sensing images, caused by inconsistent detector responses in push-broom sensors, severely degrades radiometric quality and downstream applications. Existing deep learning methods predominantly rely on wavelet transforms or generic convolutional networks, suffering from large parameter counts, slow training, and insufficient exploitation of the frequency-domain structure of stripe noise. This paper proposes a Frequency-Spatial Collaborative Network (FSCNet) with the following key designs: (1) Learnable Fourier-domain mask—a lightweight CNN predicts an attention mask on the FFT magnitude spectrum to explicitly model stripe frequency patterns; (2) Dual-domain collaborative architecture—a frequency branch localizes and attenuates stripe spectral components while a spatial branch captures non-periodic local details, with complementary fusion; (3) Residual stripe learning—directly estimating the stripe noise residual rather than the full clean image, reducing task complexity; (4) Frequency-domain consistency loss—constraining the FFT magnitude spectrum of the destriped output to match the clean image. FSCNet contains fewer than 5K parameters (grayscale mode) and processes a 256×256 image in approximately 1.3ms. Experiments across five stripe noise types on synthetic remote sensing images demonstrate that FSCNet outperforms moment matching, Wavelet-FFT, and total variation denoising baselines on all evaluation metrics, with PSNR improvements of up to 20 dB and SSIM consistently above 0.995, while achieving the fastest inference speed. Results confirm that explicit frequency-domain modeling of stripe structure significantly improves destriping performance.