Automated cataloguing of American silver perch (Bairdiella chrysoura) calls using machine learning

The American silver perch (Bairdiella chrysoura) is a numerically dominant and ecologically important species found throughout coastal habitats along the eastern United States and Gulf of Mexico. During spawning in the spring and summer, male silver perch produce distinctive knocking sounds to attract females. These sounds are readily identifiable through aural and visual analysis of underwater acoustic recordings, providing a means to track the distribution and spawning activity of these fish. However, as the volume of passive acoustic datasets grows, there is an essential need to automate the process of cataloguing silver perch vocalisations. The approach presented here utilises a (1) detection stage, where candidate calls are identified based on the properties of signal kurtosis and signal-to-noise ratio, (2) a feature extraction stage where layer activations are returned from the pre-trained ResNet-50 convolutional neural network operating on a wavelet scalogram of these signals, and (3) a one-vs-all support-vector-machine classifier. The labelled data used to build the classifier consists of 6000 perch calls and 6000 other signals that sample diverse acoustic conditions within the Pamlico Sound estuary, USA. The model accuracy is 98.9%, and the accompanying software provides an efficient tool to investigate silver perch calling patterns within passive acoustic data.