Persistent Scatterers Detection Supported by Deep Learning: A Solution Based on U-Net

Multitemporal differential interferometric synthetic aperture radar (MT-DInSAR) allows accurate and long-term monitoring of displacements of ground persistent scatterers (PS). PS are typically detected using suitable statistical tests, built to strictly control the probability of false alarm (PFA). At full resolution, this detection strategy can lead to the rejection of PS characterized by spatial consistency of the estimated parameters. Reducing the density of PS measurements can impact the interpretation of the results. In this work, we investigate the integration of a deep-learning (DL) solution, specifically U-Net, at the stage of PS detection. A three stream U-Net is proposed to replace the typical thresholding of the classical statistical indicators. Results on simulated data and on data acquired by the sensors of the COSMO-SkyMed (CSK) and COSMO-SkyMed second generation (CSG) constellation, demonstrate the superior performances of the proposed DL- PS detection scheme over the classical one.