Applying Sublook Diversity: A New Self-Supervised Deep Learning Method for SAR Image Despeckling

Abstract

Synthetic Aperture Radar (SAR) imaging is a pivotal technology in remote sensing, offering the ability to capture high-resolution data of the Earth's surface under any weather conditions and at all times. However, the quality of SAR images is often compromised by speckle noise, a granular interference that severely impacts the usability of the data for analysis and interpretation. Traditional despeckling methods often struggle to reduce noise without sacrificing image detail and resolution. This thesis introduces a new self-supervised deep learning approach, which is intended to improve the quality of SAR images by effectively reducing speckle noise while preserving resolution and finer image details. The proposed model for the reduction of speckle noise leverages the inherent sublook diversity in SAR images. By adopting the Noise2Noise training paradigm, the method trains directly on the noisy images without requiring clean ground truth data, thus circumventing one of the major limitations faced by previous despeckling approaches. The proposed model utilizes a custom U-Net architecture tailored to the specific characteristics of SAR data, optimizing it for both noise reduction and detail preservation. The performance of the proposed approach is quantitatively evaluated using metrics that allow us to compare the results of our method with other contemporary state-of-the-art methods, qualitative visual evaluations are also performed on a variety of SAR scenes, which demonstrate the effectiveness of our approach in reducing the speckle and preserve the resolution of the images.