From Lab to Field Trials: Real-Time Multi-Model Drone Detection in Low SNR Environments

Unmanned Aerial Vehicles (UAVs) have numerous applications across various industries, including surveillance and communication services. However, concerns about their potential misuse have led to the development of counter-drone technologies. In this paper, we proposed LowSNR_DroneRF, a novel drone radio frequency (RF) dataset gathered in low signal-to-noise ratio (SNR) environments. This dataset involves capturing and analyzing RF signals in both indoor (lab) and outdoor (field) environments across various drone modes (armed, off, on, connected, and flying). We propose Fine-Tuned and Light Weight Convolutional Neural Networks (FTLW-CNN), artificial neural networks (FTLW-ANN), and random forest (FTLW-RF) models that significantly reduce the number of model parameters on the custom RF dataset. The extracted features are further classified using the proposed multi-model binary classification (MMBC) approach. Our results show that the FTLW-RF classifier outper forms other models, achieving a 27.41% improvement in accuracy over state-of-the-art models for real-time drone detection and classification while maintaining a small number of parameters. The proposed models are also tested in the field under significant interference from other signals at low SNR levels, where the proposed model outperforms the conventional models.