A Physics-Based Neural Estimation of the Direction of Arrival Over Sea Surfaces.

This article presents a physics-based machine learning (ML) approach to estimate the direction of arrival (DOA) over sea surfaces. We designed a recurrent neural network (RNN) to accept two receiving channel radar data to ensure angular coherence between the receiving signals from different directions, given the scattered signal by rough surfaces having angular memory. Specifically, the radar scattering coefficients of sea surfaces were simulated at the C-band and for both co- and cross-polarizations; investigations show that the sea speckles significantly and negatively impact the dual receiver's DOA estimation performance; to mitigate the speckle interferences, we further designed an optimal configuration of dual-channel observation for DOA estimations over the sea surface, i.e., the co-polar (CP) observation mode performed best compared to that of the co-azimuth (CA) and the full-bistatic (FB) observation mode, and at moderate sea speckle impacts, the root-mean-square error of CP observation mode was about 1 & DEG; for incident angle and 5 & DEG; for incident azimuth angle estimations; results demonstrate the superior performances of the proposed physics-based neural DOA estimator.