Ensemble Learning Models for Micro-Drone Detection Using Integrated Acoustic Signatures

The rapid evolution of unmanned aerial vehicles (UAVs) has transformed both civilian and defense sectors. However, unauthorized UAV use poses significant challenges to public safety and airspace regulation, highlighting the need for robust anti-drone defense systems. Such systems aim to detect, track, and mitigate/neutralize aerial threats. Detecting drones in complex environments remains challenging due to diverse operational conditions and ambient noise. Numerous acoustic signature-based deep learning methods have been proposed for reliable micro-drone detection. While these methods outperform traditional statistical signal processing techniques in detection accuracy, they are computationally intensive and require large amounts of training data. Hence, developing lightweight techniques that require less training data while maintaining comparable detection performance on real-time edge devices is essential. To achieve this, Autocorrelation Coefficients (temporal features) and Mel-Frequency Cepstral Coefficients (spectral features) are extracted from captured acoustic data. The integrated features are applied to multiple ensemble learning models, and their detection performance is evaluated, including the Random Forest algorithm, AdaBoost model, Extreme Gradient Boosting model, and Stacking Ensemble model. Experimental analysis shows that the Stacking Ensemble model achieves detection performance comparable to deep learning models while requiring lower computational complexity and reduced training data, attaining an overall detection accuracy of 98\%. These findings demonstrate that integrating acoustic features with ensemble algorithms provides a reliable and scalable framework for real-time acoustic drone detection, making it a promising solution for next-generation aerial surveillance systems.