Deep neural networks to predict foraging behaviour: saltwater immersion data can accurately predict diving in seabirds

The accurate identification of foraging locations is critical for wildlife conservation. While remote sensing and biologging devices provide much of the necessary data, their deployment is often complicated by factors such as weight, battery life and sensor capacity, limiting their effectiveness for long-term tracking of wide-ranging species. In this study, we evaluate the effectiveness of saltwater immersion data from light-level geolocation loggers (global location sensor; GLS) as a predictor of foraging behaviour in a pursuit-diving seabird, the red-footed booby (Sula sula rubripes). Using co-deployed tri-axial acceleration data as a high-resolution benchmark, we compare the performance of deep learning models for classifying dive and non-dive states. Predictions are cross-validated on withheld individuals for generalizability. Using a small pilot dataset, we find that models trained solely on GLS data only slightly underperform those trained on acceleration data despite the resolution discrepancy, classifying the diving behaviours of unseen birds with 93.65% accuracy. These findings suggest that GLS data alone may be sufficient to reliably infer dive events and, by extension, foraging locations, for pursuit-diving seabirds, providing a minimally invasive, scalable method to enrich year-round GLS migratory tracking studies using models derived from co-deployment of GLS and global positioning system devices.