Acoustic-Based Detection of UAVs using Machine Learning: Analysis of Distance and Environmental Effects

This paper presents a study of the distance dependence of a detection system based on acoustic signals from unmanned aerial vehicles (UAVs). The detection system uses machine learning algorithms fed with relevant frequency domain features extracted from the acoustic signals emitted by UAVs. The feature set includes Mel frequency cepstral coefficients (MFCC), delta MFCC, delta-delta MFCC, pitch, centroid, harmonic ratio, spectral flux, and spectral roll-off point. Five machine learning methods are used to evaluate the detection performance: linear discriminant analysis, multilayer perceptron, radial basis function network, support vector machine and random forest. Evaluation is carried out for different distances to assess the effect of noise and attenuation on the detection performance. The results show that UAVs can be detected effectively, but the performance degrades with increasing distance. Our study provides an overview of how increasing the distance between the UAV to be detected and the sensor affects simple detection methods due to the decrease in the signal-to-noise ratio, as the signal of interest is attenuated due to distance. The results of this study reveal that, considering the difficult evaluation environment, it can be concluded that acoustic detection is feasible at distances less than 200 meters, and could be feasible at longer distances in scenarios where the interfering signal power is more realistic, and the interferers are not as close to the acoustic signal produced by UAVs.