Automatic fish sounds classification
Résumé
The context of this work is environmental monitoring. Specifically, we focus on acoustic systems for monitoring fish populations. By investigating recorded fishes sounds it is possible to monitor the spatial and temporal evolution of fish populations. The aim of this work is to present an automatic fish sounds classification system. The solution we propose is new and based upon supervised machine learning (in particular, classification is performed by Random Forest). The features used in input of the learning algorithm come from an extensive state of the art in various domains of classification such as speech, music, animal calls, environmental acoustic landscape and human induced noises. Our system is trained and tested on nighttime recordings from a 20 meters depth seagrass habitat (Calvi, Corsica, France). From this study, we propose to consider 66 different features (shape and/or statistical description in time and frequency). Fish sounds are automatically classified into four different classes (drums, grunt, impulse, FM) and our system reaches 94% of correct classification rate compared to 77% when considering MFFC features.In order to deal with large datasets and to study the evolution of fish populations, we are currently developing an approach based on dynamic classification with rejection (negative class).