2025/11/29 by Bicer, H. Nazim
#FOS: Electrical engineering #Signal Processing (eess.SP) #electronic engineering #information engineering
paper · doi:10.48550/arxiv.2512.00510
With more devices competing for limited spectrum, dynamic spectrum sharing is increasingly vulnerable to interference from unauthorized emitters. This motivates fast detection and localization of these emitters using low-cost, distributed sensors that do not require precise time synchronization. This paper presents two convolutional neural network (CNN) approaches for multi-emitter detection and localization from sparsely sampled power maps. The first method performs single-stage prediction of existence probabilities and positions. The alternative two-stage method first estimates an occupancy map as an interpretable intermediate representation and then localizes emitters. A unified training objective combines binary cross entropy with coordinate regression loss and can handle an unknown emitter count. Small footprint networks, on the order of 70 k parameters, are trained and evaluated on simulated free-space and urban scenes. Experiments demonstrate that both approaches localize multiple emitters from sparse measurements across diverse environments, with the logits based two-stage variant remaining competitive, and in some cases superior, under extreme sensor sparsity. The findings indicate that small CNNs with a unified objective can be deployed for spectrum monitoring and localization.