NBDNet: A Deep Learning Algorithm for Despeckling of SAR Data

Abstract

In the recent years, remote sensing has gained high momentum in varied applications. Satellite imaging and processing is one of the most sorted techniques followed by researchers. Synthetic aperture radar (SAR) images are popular among the remote sensing community due to its capability of imaging in all weather conditions. The practical applications of SAR data is limited due to presence of speckle noise. In the past, deep learning methods are developed to despeckle the SAR images. This paper proposes a convolutional neural network based non-blind denoising network (NBDNet) for the despeckling of SAR images. In the proposed NBDNet model, attention blocks are introduced to preserve the structural and texture details of the captured scene globally. Further, squeeze and excitation module and convolutional block attention module have been used in the proposed NBDNet to capture the minute structural information of the artefacts. The experimental results of proposed NBDNet and benchmark algorithms are evaluated on synthetic UC merced land-use images and real SAR images. Quantitative and visual results of of proposed NBDNet yield better texture and structural detail preservation as compared to benchmark algorithms on both datasets. © The Author(s), under exclusive licence to Springer Nature Singapore Pte Ltd. 2024.