A multi-objective deep learning based approach for SAR image reconstruction in urban environment

SAR sensors have become a crucial instrument for Earth observation. Interferometric SAR (InSAR) systems provide complex images whose phases contain information on the height profile of the scene, but at the same time are affected by noise. Therefore, InSAR phase denoising is a fundamental pre-processing step for easing and improving the performance of further applications. Urban areas are very challenging to be characterized and usually require a specific filtering design. In this work the result in urban areas of a CNN solution trained on natural scenarios with a realistic dataset, are exploited. The results show a good generalization ability thanks to the wide and realistic dataset and to the multi-objective nature of the designed cost function.