Over het archief
Het OWA, het open archief van het Waterbouwkundig Laboratorium heeft tot doel alle vrij toegankelijke onderzoeksresultaten van dit instituut in digitale vorm aan te bieden. Op die manier wil het de zichtbaarheid, verspreiding en gebruik van deze onderzoeksresultaten, alsook de wetenschappelijke communicatie maximaal bevorderen.
Dit archief wordt uitgebouwd en beheerd volgens de principes van de Open Access Movement, en het daaruit ontstane Open Archives Initiative.
Basisinformatie over ‘Open Access to scholarly information'.
| one publication added to basket [321391] |
Knock knock who's there? Bat detection along the Belgian coast
Fojtek, C. (2020). Knock knock who's there? Bat detection along the Belgian coast. MSc Thesis. Faculty of Engineering and Architecture, Ghent University: Gent. xiii, 78 pp.
|
| Abstract |
To date, the majority of research focuses on the extraction of certain temporal and frequency characteristics from the spectrogram of the bat echolocation call to classify bat species. The downside of this method is the loss of information about the shape of the bat call and the inability to classify the more complex social calls of bats. Tied to this is the process of identifying the bat calls in the audio recordings and extracting the features; two important steps, but often neglected in current research. Therefore, this article presents a method to fully automate the classification of bat species at the Belgian coast from audio recordings. The accomplish this, the calls were detected in the audio files with the recent BatDetective convolutional neural network (CNN) [1]. Different image segmentation techniques were then tried to denoise the spectrogram of the call. Three sets of features, including the call characteristics, a fitted curve through the call and the image of the spectrogram, were extracted from the denoised spectrogram. Finally, a variety of models, both supervised and unsupervised were tested on these features and compared. The final method denoises the spectrogram using a combination of background removal, global thresholding and region growing. Only when high intensity echos are present, the method does not fully denoise the spectrogram. The best classification performance was achieved with the pretrained VGG16 [2] model, trained on the images of the spectrogram. In particular when adding reference lines at fixed frequencies in the image the CNN outperformed all other methods by a large margin. This indicates that other methods, apart form extracting the call characteristics, should be considered in classifying bat species. |
IMIS is ontwikkeld en wordt gehost door het VLIZ.
