Adaptation of BirdNET for bats detection

Details

Supervisors

Bonaventure, Olivier

Faculty

Ecole polytechnique de Louvain

Degree label

Master [60] en sciences informatiques

Abstract

With the development of artificial intelligence, tools as well as projects in the service of biodiversity could make their appearance. In this thesis, a project based on Raspberry Pi, BirdNET-Pi, initially dedicated to the detection and identification of birds has been adapted to bats and their constraints. Three artificial intelligence models were tested on the new device, dubbed BatNET-Pi. Each model had its CPU, RAM and storage performance measured to assess the limits of the new system. The best performing model turned out to be BatDetect2, a recent bat detection and identification model, being able to process 10 hours of recording per day on a Raspberry pi 4. Particular attention has been paid to the problem of local storage by compressing high acoustic frequency wav files. Different lossy and lossless algorithms have been tested, such as WavePAck, FLAC and OptimFROG. It emerged that only lossless algorithms manage to obtain a good compression ratio while perfectly preserving the data. OptimFROG was identified as the best algorithm among the 3 in this specific context. The compression and the slight modifications made have made it possible to multiply the initial storage capacity by 8. However, the resulting device can be greatly improved to allow better deployment in the wild, such as the power source or Internet connectivity for instance.