Automated monitoring of bat species in Belgium

Anthropogenic change has become a recurrent subject in today's society. The population's growth, pollution and the over-consumption of resources is heavily damaging the biodiversity. A way to identify these changes is by monitoring bioindicators. Bat species can fit this role as they can easily show qualitative information on the environment they are in. The monitoring of bats can be achieved through bioacoustics. In fact, bats produce ultrasounds for most of their activities. The objective of this work is to create a reliable Belgian bat species classification tool using machine learning techniques trained on bat calls audio files. Data provided by the association Natagora was processed to obtain the most accurate data available to train our model. The project Batmen was created to answer this need with the help of a convolutional neural network. Audio files were transformed into spectrograms before being fed to the neural network. Our results show that the classification through convolutional neural networks was accurate provided that the dataset was big and balanced enough. Batmen showed average results for the classification of bat species using groups to gather species containing similarities. Improvements can be done with more accurate labeling, more balanced data and different multiclass classification strategies through convolutional neural networks improvements.