Table 1.
Individual information for tagged North Atlantic right whales (Eubalaena glacialis), presented with predicted individual filtration rate (volume of water filtered/ total deployment time) and proportion of time spent feeding predicted by a support vector machine model.
Fig 1.
Pseudotracks of n = 21 tagged North Atlantic right whales (Eubalaena glacialis) in the Shediac Valley, Gulf of St. Lawrence, Canada in July 2023 and 2024. Overlaid is the vessel exclusion zone (pink-shaded rectangle) enforced from June through September by Transport Canada. Coastline data obtained via Natural Earth (https://www.naturalearthdata.com) and bathymetric data obtained from GEBCO and interpolated [52].
Fig 2.
An annotated time-depth recorder for training data.
Time-depth recorder data for four CATs tag deployments, manually annotated using corresponding video data to identify when the mouth was open (black) or closed (gray).
Fig 3.
Example kinematic signals of North Atlantic right whales.
Example fluking signal (radians), speed (m s-1) and depth over time during one U-shaped feeding dive of a North Atlantic right whale (Eubalaena glacialis) tagged with CATS (www.cats.is) inertial sensing tags in the Gulf of St Lawrence, Canada, from manually validated feeding event (EG4903) on 7th July 2024, on EG4903. The figures show kinematics of swimming at the surface for two minutes before dive descent at 12:50:30, a ~ 4-minute dive, and after dive ascent at 12:56.
Fig 4.
Validated and Estimated Feeding Kinematics of North Atlantic right whales- Normalised fluking signal, speed (m s-1) and depth (m) over time of North Atlantic right whale (Eubalaena glacialis) tagged with CATS (www.cats.is) inertial sensing tags in the Gulf of St Lawrence, Canada.
Feeding detection (darker colour) manually annotated using corresponding video data (Panel a) or predicted using a trained Support Vector Machine (SVM). Panel (a) shows kinematic data from a manually validated feeding event on 7th July 2024, on EG4903. Overlaid picture of mouth confirmed open from corresponding timestamp during bottom phase of dive, most easily identifiable by high contrast inner mouth. Panel (b) shows SVM-classified deep dive (EG2605, Smoke, 8th July 2024) and panel (c) shows SVM-classified shallow dive (EG1419, 6th July). Coloured markers on photos indicate approximate time of the screen grab from manual validation plot (panel a).
Fig 5.
Key classification variables of feeding predicted by Supported Vector Machine (SVM).
Relative density of speed (m s-1; red), fluking rate (flukes min-1; blue) and depth (m; grey) as they relate to classification of feeding by trained Supported Vector Machine model, built using audited video.
Table 2.
Predicted feeding time proportions from manually-validated and Supported Vector Machine (SVM) model output, compared to classification using dive shape [58]. We derived false positive and negative rates of feeding detection for dive shape relative to SVM and manually audited video data.
Table 3.
NARW feeding times predicted by Support Vector Machine (SVM) or manually audited using onboard camera. Data classified by whale depth strata and daylight classification and averaged across all available data for the given depth/daylight classification.
Fig 6.
Time-depth heatmap of feeding and closed predicted by supported vector machine.
Log density heatmap of North Atlantic right whale (Eubalaena glacialis) presence by hour of the day by 2m depth bins, normalised by total time per hour. Day-night cycles marked with white lines; twilight considered surrounding hour to sunrise/sunset. Feeding was determined through a trained Supported Vector Machine (SVM) and Long-Short-term Memory (LSTM) neural network models based on processed 10 Hz accelerometer and video audit CATs tag data.
Fig 7.
Estimated North Atlantic right whale feeding rates.
Volume water filtered per hour by tagged North Atlantic right whales (Eubalaena glacialis) based on Support Vector Machine (SVM) trained to identify feeding, using training data from manual behaviour audit and 10 Hz kinematic data. Mouth gape was calculated per individual [2] with updated body length at age estimates from Fortune et al., [69].