Smartphone-powered citizen science for bioacoustic monitoring

Citizen science is the involvement of amateur scientists in research for the purpose of data collection and analysis. This practice, well known to different research domains, has recently received renewed attention through the introduction of new and easy means of communication, namely the internet and the advent of powerful “smart” mobile phones, which facilitate the interaction between scientists and citizens. This is appealing to the field of biodiversity monitoring, where traditional manual surveying methods are slow and time consuming and rely on the expertise of the surveyor. This thesis investigates a participatory bioacoustic approach that engages citizens and their smartphones to map the presence of animal species. In particular, the focus is placed on the detection of the New Forest cicada, a critically endangered insect that emits a high pitched call, difficult to hear for humans but easily detected by their mobile phones. To this end, a novel real time acoustic cicada detector algorithm is proposed, which efficiently extracts three frequency bands through a Goertzel filter, and uses them as features for a hidden Markov model-based classifier. This algorithm has permitted the development of a cross-platform mobile app that enables citizen scientists to submit reports of the presence of the cicada. The effectiveness of this approach was confirmed for both the detection algorithm, which achieves an F1 score of 0.82 for the recognition of three acoustically similar insects in the New Forest; and for the mobile system, which was used to submit over 11,000 reports in the first two seasons of deployment, making it one of the largest citizen science projects of its kind. However the algorithm, though very efficient and easily tuned to different microphones, does not scale effectively to many-species classification. Therefore, an alternative method is also proposed for broader insect recognition, which exploits the strong frequency features and the repeating phrases that often occur in insects songs. To express these, it extracts a set of modulation coefficients from the power spectrum of the call, and represents them compactly by sampling them in the log-frequency space, avoiding any bias towards the scale of the phrase. The algorithm reaches an F1 score of 0.72 for 28 species of UK Orthoptera over a small training set, and an F1 score of 0.92 for the three insects recorded in the New Forest, though with higher computational cost compared to the algorithm tailored to cicada detection. The mobile app, downloaded by over 3,000 users, together with the two algorithms, demonstrate the feasibility of real-time insect recognition on mobile devices and the potential of engaging a large crowd for the monitoring of the natural environment.