Robust Acoustic Drone Detection Using Noise-Resilient Feature Extraction

Abstract

The increasing accessibility of small Unmanned Aerial Vehicles, commonly referred to as drones, has created significant challenges for surveillance and airspace security, necessitating reliable and robust detection techniques. Conventional drone detection methods, including radio frequency, radar, and vision-based approaches, are often limited by signal dependency, high operational cost, and reduced effectiveness under non-line-of-sight and adverse environmental conditions. In contrast, acoustic-based detection provides a passive and cost-effective alternative by exploiting the distinct sound signatures generated by drone propellers. However, the performance of acoustic detection is highly affected by environmental noise. To address this challenge, a robust acoustic-based drone de-tection framework is developed for reliable operation under severe noise conditions. Synthetic Gaussian noise at a signal-to-noise ratio of −10 dB is incorporated to simulate real-world acoustic environments. The noisy acoustic signals are transformed into time–frequency representations to capture distinctive drone acoustic signatures. A denoising autoencoder is employed to learn compact and noise-resilient latent feature representations that preserve discriminative information while suppressing noise. The extracted features are subsequently used for binary classification to distinguish drone and non-drone acoustic signals. A Gradient-Regularized Convolutional Neural Network (GRCNN) and a CatBoost classifier (CBC) are implemented for comparative evaluation. Experimental results demonstrate that the CBC achieves superior detection performance with an accuracy of 99%, compared to 94% achieved by the GRCNN. The proposed framework demonstrates strong robustness and effectiveness, making it suitable for real-world acoustic drone surveillance applications.