Semantic segmentation for feature detection in ocean bottom seismometer data

We curate a small, manually annotated dataset of 500 spectrograms containing a range of feature classes in the 1 to 50 Hz frequency band from a single ocean bottom seismometer (OBS) from the UPFLOW array in the mid-Atlantic region. We explore several machine learning (ML) training techniques that are specialised for low-data training regimes, and compare their performances for two feature classes (instrument resonances and blue whale calls). We find that a synthetic pre-training step significantly improves performance relative to semi-supervised approaches and finetuning an off-the-shelf model, with a ~5% improvement in performance for well-represented features, and an enhancement of over 90% for rare features. Despite the small dataset, our method can be utilised to accurately and efficiently segment spectrogram data across 43 OBSs with high-quality data of the large-scale UPFLOW array, as well as for earlier OBS deployments. We next investigate a range of applications for the trained segmentation models. We demonstrate that our ML algorithm identifies current-induced instrument resonances accurately enough to extract a tidal signal. In addition, it reliably detects blue whale calls across the entire UPFLOW array, and it even enables automated tracking of individual whales detected simultaneously at multiple OBSs.