On the impact of despeckling for supervised SAR super-resolution

Abstract:
Enhancement of SAR resolution is essential for various applications in earth observation. Since SAR images are highly corrupted by speckle noise, we propose to help super-resolution neural network learning with a despeckling preprocessing step. Unlike optical images, low-resolution SAR images are extracted from the sub-apertures of the original SAR image. To evaluate the impact of the despeckling, SwinIR, SRCNN, and ESPCN neural networks are trained in three ways: Noisy2Noisy, Noisy2Denoised, and Denoised2Denoised. The ONERA SAR database experiments show the despeckling improvement gap and the slight enhancement of SwinIR over SRCNN and ESPCN according to the visual reconstruction and to L1, L2, PSNR, and SSIM metrics.