Multi-Objective CNN-Based Algorithm for SAR Despeckling

Deep learning (DL) in remote sensing has nowadays become an effective operative tool: it is largely used applications, such as change detection, image restoration, segmentation, and classification. With reference to the synthetic aperture radar (SAR) domain, application of DL techniques not straightforward due nontrivial interpretation SAR images, especially caused by presence speckle. Several solutions for despeckling have been proposed last few years. Most these focus on definition different network architectures with similar cost functions, involving properties. In this article, a convolutional neural (CNN) multi-objective function taking care spatial statistical properties proposed. This achieved peculiar loss obtained weighted combination three terms. Each terms dedicated mainly one following characteristics: details, speckle properties, strong scatterers identification. Their allows balancing effects. Moreover, specifically designed architecture effectively extract distinctive features within considered framework. Experiments simulated real images show accuracy method compared state-of-art algorithms, both from quantitative qualitative point view. The importance considering crucial correct noise rejection details preservation underlined scenarios, homogeneous, heterogeneous, extremely heterogeneous.