Classifying Acoustic Signals for Wildlife Monitoring and Poacher Detection on UAVs

Acoustic detection and classification methods for wildlife monitoring have recently been used as an alternative to video-based methods. With the advent of reliable and cheaper mini-UAVs, using acoustic information for detection and classification of targets of interests can provide a cheaper solution than employing video-based or distributed wireless acoustic sensors for wildlife monitoring. In reality, we would like to identify or classify individual targets from a mixture of interest signals (polyphonic). Also, these signals when recorded by UAVs have a low signal to noise ratio due to noise coming from the UAV propeller, motor, and wind. In this paper, we explore the performance of a SVM classifier when used to classify a mixture of acoustic signals of interests (polyphonic) under noisy environments. The performance of the classifier is compared to the performance in the absence of noise and the performance of individual signals of interest (monophonic). The classifier performance is evaluated for precision, recall, and F1-score on an extensively built database that contains acoustic sources from mammals, birds, vehicles, and firearms.