A Sure-Based Unsupervised Deep Learning Method for Sar Despeckling

A Sure-Based Unsupervised Deep Learning Method for Sar Despeckling

연구 분야: Artificial Intelligence

학회: 2021 IEEE International Geoscience and Remote Sensing Symposium IGARSS

Synthetic Aperture Radar (SAR) images most often get corrupted by multiplicative (granular) signal variations known as ‘speckle’ during the remote acquisition. In this paper, we propose an unsupervised deep learning approach for SAR despeckling without the need for clean SAR references. To this end, we employ Stein's unbiased risk estimator (SURE) that facilitates estimation of the true mean-squared error between the output of the de-speckling network and the ground truth, however, using only available SAR images. We utilize thus unsupervised trained U-Net architecture, coupled with the proposed SURE-based loss function. Experiments are conducted on synthetic as well as real Sentinel-l SAR images. The results are compared both qualitatively and quantitatively with popular methods in SAR despeckling.