Tackling the Effect of Noise in the Context of a Neural Network-Based Unmanned Aerial Vehicles (UAVs) Classification from Radar Data

This work evaluates and enhances a deep learning method for classifying Unmanned Aerial Vehicles (UAVs) using micro-Doppler signatures (mDs) extracted from raw radar signals, with a focus on addressing performance degradation across varying signal-to-noise ratio (SNR) levels. By modeling the relationship between SNR and distance, we simulate realistic noise conditions to assess the method robustness over extended ranges. Leveraging Gated Recurrent Unit (GRU) layers to capture temporal dependencies in raw radar signals, we investigate the impact of radar signal length, feature representation, and dropout regularization on classification robustness using simulated noisy signals. Our experiments, conducted on the DIAT- μSAT dataset, show that processing 100 samples per time step significantly improves noise resilience, while moderate dropout rates (4−8%) enhance generalization without compromising performance. The refined method achieves consistent accuracy (F1-score >0.9) across a broad SNR range, effectively simulating real-world distance variations. These findings advance radar-based UAVs classification and offer a scalable framework for deployment in operational environments with dynamic SNR conditions.