Despeckling Sentinel-1 GRD Images by Deep-Learning and Application to Narrow River Segmentation

Abstract

This paper presents a despeckling method for Sentinel-1 GRD images based on the recently proposed framework "SAR2SAR": a self-supervised training strategy. Training the deep neural network on collections of Sentinel 1 GRD images leads to a despeckling algorithm that is robust to space-variant spatial correlations of speckle. Despeckled images improve the detection of structures like narrow rivers. We apply a detector based on exogenous information and a linear features detector and show that rivers are better segmented when the processing chain is applied to images pre-processed by our despeckling neural network.

Water-Body Segmentation for SAR Images: Past, Current, and Future

“…After 2020, scholars gradually tried to modify U-Net to achieve better performance, or to explore the applicability of more advanced semantic segmentation models in the field of computer vision to segment water in SAR images [114]. In 2021, Bayesian was introduced in U-Net for water-body segmentation [115], and the Bayesian neural network was created, which considers dropout as a random sampling layer in U-Net architecture.…”

Fast strategies for multi-temporal speckle reduction of Sentinel-1 GRD images

“…This framework has first been developed on single-look Sentinel-1 images. An adaptation to GRD data, presenting a different number of looks and a spatially varying speckle correlation, is presented in [14]. This network, available at https://gitlab.telecom-paris.fr/RING/ SAR2SAR, will be used in the following multi-temporal strategies.…”

“…As already presented in [16], we propose to use SAR2SAR for the denoising of the ratio image τ t of date t. In this paper we are moslty interested in GRD data and therefore we have to take into account the multi-look processing applied on these data. As mentioned in section 2, this is done by using the network presented in [14] which was trained on GRD images.…”

“…Please note that a temporal stack is constituted by finely registered images acquired on the same orbit with the same configuration (either ascending or descending). As a despeckling method, we consider the adap-tation of SAR2SAR for GRD images proposed in [14], and presented in section 1. The two multi-temporal frameworks considered, Quegan [2] and RABASAR [15], are described in section 2.…”