Open wide: Identifying North Atlantic right whale feeding behaviour using camera-validated kinematic data and machine learning.

Whales are at risk from human activities and behavioural information can be important for informing risk policy. Underwater behaviour is increasingly being characterised with camera tags. Here, we characterise the kinematic signature of endangered North Atlantic right whales during feeding in high-risk areas. Customized Animal Tracking Solutions (CATS) tags were deployed (n = 23) to record the body position and movement of while simultaneously recording underwater video. Manual validation of tag camera footage was used to train classification learners, with a Support Vector Machine (SVM, 95% test accuracy) being the best performing. The SVM confirmed previously identified kinematic signatures important to identifying ram filtration. where decreased swim speed and increased fluke stroke occur due to increased drag from the opened mouth. Individual whales were predicted to spend 11-76% of their time actively filtering prey, averaging 11.4 hours day-1 feeding. Through video audit, we observed right whales feeding close to the seafloor during the day; our model also predicted previously unreported feeding close to the surface at night. This night feeding was underestimated by traditional feeding classification using dive shape relative to both audit (43.1%) and SVM-predicted feeding (54.9%). Based on pseudotracks, tagged whales collectively spent 74.6% of their time within a vessel exclusion zone. This research provides standardised techniques for classifying behaviour and also highlights potential gaps in policy-relevant information, such as elevated vessel strike risk at night.