Automated detection of bat species in Belgium

Details

Supervisors

Bonaventure, Olivier ; Nijssen, Siegfried

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en informatique, à finalité spécialisée

Abstract

Many bat species are endangered, mostly by human activity. To address this, environmental organisations census bats, among others, by manually classifying bat call recordings which is a tedious and repetitive task requiring lots of experience. The objective of this thesis is to design an automated, robust and open-source bat call detection and classification tool for the twenty-three bat species in Belgium. Not only do we explore multi-class classification but also multi-label classification. We investigate the performance obtained when combining different machine learning algorithms, namely convolutional neural networks, support vector machines and extreme gradient boosting. The best multi-class classification performance is obtained by combining a CNN for the detection with another CNN followed by an XGBoost for the classification. This gives a precision of 78% and a recall of 75%. For the multi-label classification, our best performing model is composed of a CNN followed by an XGBoost and performs both detection and classification at once. This architecture has a global precision of 73% and a recall of 65%. During this project, we designed a robust tool that runs in real-time, making it possible to use it instead of performing manual classification of bat calls.