Acoustic UAV Detection Framework Improves F1 Score to 78.6% in Battlefield Conditions

A recent study published on arXiv (2608.14287) introduces a comprehensive framework for the passive acoustic detection of small unmanned aerial vehicles (UAVs) in actual battlefield environments. It tackles two primary issues: high levels of environmental noise and domain shifts due to varying hardware. The method combines Per-Channel Energy Normalization (PCEN) with attention-based pooling to improve feature extraction in low signal-to-noise scenarios. Furthermore, a domain-aware training approach utilizes auxiliary classes and multi-microphone data to reduce performance drops across domains. Tested on a distinctive dataset of combat recordings from the Ukrainian frontlines, the method notably enhances the F1 score from 55.4% to 78.6%. This research emphasizes passive acoustic sensing as a viable and cost-effective solution for UAV detection in military contexts.