Direct observation of foraging behavior is not always possible, especially for marine species that hunt below the surface. However, biologging and tracking devices in particular have provided very detailed information about how various species use their habitat. From these indirect observations, researchers have tried to infer foraging and prey catching events for a more accurate definition of these species’ ecological niches. In this study, we deployed video cameras in addition to GPS and time-depth recorders on chinstrap penguins during the brood phase of the 2018-19 breeding season at various colonies on the Gourlay peninsula (South Orkney Islands). More than 57 hours of footage from 16 birds covering 770 dives were scrutinized by two independent observers. The outcome of each dive was classified as unsuccessful, individual krill encounter or krill swarm encounter. In addition, the number of prey items caught was recorded for successful dives. We then used various predicting variables derived from the other logging devices or from the environment to train a machine-learning algorithm to predict the outcome of each dive. Our results show that despite some limitations, the data collected from the footage was reliable as there was a high agreement from both annotators. We also demonstrate that it was possible to accurately predict the outcome of each dive from basic dive patterns and horizontal movement characteristics that have not been used for penguins previously. Finally, we discuss how video footage can help build more accurate habitat models and gain wider knowledge about predator behavior or prey distribution.

During the breeding season 2018-19, 22 chinstrap penguins (Pygoscelis antarcticus) from Signey in the South Orkney Islands were equipped with a GPS tracker (Pathtrack nanoFix® GEO) set to record one location every minute, a time-depth recorder (TDR, Lotek LAT 1800 series) recording one pressure measure every 2 seconds and a video camera (Little Leonardo DVL400M) to record approximatively 5 hours of footage starting at a preset time.

The data from the TDR was processed using the R diveMove package to detect each dive and calculate some dive metrics. The dives were linked to the GPS data in order to add the water depth at the location of the dive. The camera footage was viewed by two independent researchers who classified the outcome of each dive as: no krill encounter, single krill encounter or krill swarm encounter. The number of prey items caught during each dive was also estimated.

Two random forest models were used to predict the outcome and the number of prey cauht for each dive from an arrey of environmental variables. These were derived from the dive metrics, the surface tracking or the environment. The dataset presents the response and the predicting variables for 770 dives.