Automatic annotation of spectrograms – Comparison of traditional energy detectors and ResNet deep learning to analyse 20-Hz fin whale calls

Passive underwater acoustic measurement systems produce very large amounts of data, which need to be analysed to detect sources of noise (e.g. ships, marine life, natural physical processes). Supervised/semi-supervised machine learning applications rely on annotated datasets for training. In this study, the annotated dataset comes from manual picking and the aim is that machine learning will produce automated detections fast and repeatably which are in agreement with the analyst’s annotations. We consider data from two different ocean observatories (namely, Lofoten-Vesterålen (LoVe) in Norway and the Ascension Island station of the Comprehensive Nuclear-Test-Ban Treaty network), and three sampling rates (32 or 64 kHz at LoVe, 250 Hz at Ascension Island). We look at how the annotation of data, spectrogram parameters (such as window length and frequency resolution), and signal-to-noise in the training data affect performance. As well as examining whether or not the signals of interest are detected, accuracy in determining the start and end times of the signals is also considered.