Physically Informed Neural Networks for Synthetic Aperture Imaging

We propose two Physically Informed Neural Networks (PINNs) for locating multiple point scatterers using synthetic aperture radar (SAR). The first network takes as input the data collected on the synthetic aperture and generates an image of the scatterers. The second network uses the traditional SAR beamforming image as input and generates a sharper image of the scatterers by performing a deconvolution. Both networks are resilient to measurement noise modeled by additive Gaussian noise. When tested on datasets where the number of scatterers is larger than those present in the training dataset, both networks are shown to be able to generalize well.