Abstract
The proliferation of drones, or unmanned aerial vehicles (UAVs), has raisedsignificant safety concerns due to their potential misuse in activities such asespionage, smuggling, and infrastructure disruption. This paper addresses thecritical need for effective drone detection and classification systems thatoperate independently of UAV cooperation. We evaluate various convolutionalneural networks (CNNs) for their ability to detect and classify drones usingspectrogram data derived from consecutive Fourier transforms of signalcomponents. The focus is on model robustness in low signal-to-noise ratio (SNR)environments, which is critical for real-world applications. A comprehensivedataset is provided to support future model development. In addition, wedemonstrate a low-cost drone detection system using a standard computer,software-defined radio (SDR) and antenna, validated through real-world fieldtesting. On our development dataset, all models consistently achieved anaverage balanced classification accuracy of >= 85% at SNR > -12dB. In the fieldtest, these models achieved an average balance accuracy of > 80%, depending ontransmitter distance and antenna direction. Our contributions include: apublicly available dataset for model development, a comparative analysis of CNNfor drone detection under low SNR conditions, and the deployment and fieldevaluation of a practical, low-cost detection system.