Automated acoustic monitoring methods are frequently used to survey bat activity around wind turbines. The algorithms are often based on spectral features or threshold values of the recordings. Due to the generality of these features, a lot of recordings are noise, making manual analysis and labelling of the recordings time consuming. In this paper, we present an approach based on convolutional neural networks to detect and classify bat calls respectively. Recordings are converted to Mel-frequency cepstral coefficients (MFCCs), which are then fed as an image into the convolutional neural networks (CNNs) for classification. A dataset consisting of 43585 recordings gathered at 5 m height was used to train and test this method. An accuracy of 99.7% was achieved on a test set for the binary classification of noise and bat calls. For the species classification, this approach achieved an accuracy of 96%. Both networks, trained on data gathered at 5 m, were also tested on recordings gathered at heights of 33 m, 65 m and 95 m. In case of the binary classification task, the results showed an increased rate of misclassifications among noise recordings. For species classification, there was a higher amount of misclassifcations among all species.
Bat echolocation calls, classification, convolutional neural networks, field test