A Hybrid Raytracing and Deep Learning-Based Approach for Single-Site Localization of Radio Emitters in Urban Environment

Abstract

Efficient passive radio monitoring in dense urban areas is obstructed by multipath propagation and NLOS conditions, causing conventional single-site localization methods to suffer from compounding geometric errors and sparse data availability. To address these limitations, this paper introduces a novel hybrid framework that fuses physics-based wave propagation geometry with data-driven machine learning. The proposed methodology utilizes an advanced, Circulant Modified MUSIC (CM-MUSIC) algorithm to reliably resolve highly correlated wave components on a 5-element Uniform Circular Arrays (UCA) antenna. The extracted directional inputs are processed via a backward raytracing engine to generate a continuous spatial likelihood map, which is then filtered by a Convolutional Neural Network (CNN) trained as a sector semantic mask. Empirical field validation using a low-cost Software Defined Radio (SDR) platform capturing analog FM transmissions demonstrated high tracking precision and robust noise immunity under severe spatial constraints, while an ideal sector-classification scenario analysis confirmed that the proposed framework closely converges to the theoretical accuracy limits of the geometric model. This framework provides a scalable, highly cost-effective, and hardware-preserving single-site solution for modern civilian radio spectrum regulatory enforcement.