Radar Sensing via Geometric Machine Learning Over Riemannian Manifolds

The imperative for autonomously detecting radar signals is paramount in the context of emerging shared-spectrum wireless networks, such as the Citizens Broadband Radio Service (CBRS) band. The dynamic allocation of this spectrum hinges upon a specialized sensor network tasked with identifying the presence of federal incumbent radar signals. In this paper, we propose a radar sensing strategy using received signals at base stations. More specifically, the sample covariance matrices of received signals lie over Riemannian manifolds (i.e., curved surfaces) thanks to their symmetric positive definite (SPD) properties. Consequently, we propose to use support vector machine (SVM) learning models over Riemannian manifolds for classification of radar existence. Our findings reveal that the model attains more than 90% radar detection accuracy considering Signal-to-noise ratio (SNR) values up to 14 dB.