Deep learning based Filtering of Polarimetric SAR Images

Conference: EUSAR 2022 - 14th European Conference on Synthetic Aperture Radar
07/25/2022 - 07/27/2022 at Leipzig, Germany

Proceedings: EUSAR 2022

Pages: 4Language: englishTyp: PDF

Authors:
Zaman, Roghayeh; Vitale, Sergio; Ferraioli, Giampaolo (Universita degli Studi di Napoli “Parthenope”, Napoli, Italy)
Aghababaei, Hossein (Faculty of Geo-Information Science and Earth Observation (ITC), Department of Earth Observation Science (EOS), University of Twente, Enschede, the Netherlands)

Abstract:
This paper presents a new data-driven approach for despeckling of polarimetric SAR (PolSAR) images. The problem of denoising covariance matrix is addressed in the context of deep convolutional neural networks and based on the definition of a new multi-objective loss function that takes into account not only the divergence of the covariance matrices of the data, but also the statistical properties of the multi-channel polarimetric data. The proposed multi-objective function balances the spatial details and the statistical properties of speckle in the denoising process. The proposed methodology is experimentally validated using synthetic and real datasets and compared with existing despeckling approaches.