Identifying prey capture events in chinstrap penguins using accelerometer data and deep learning

Modern biologging technology (Global Positioning System [GPS], accelerometer, time-depth recorders [TDR], animal-borne cameras, etc.) coupled with advances in analytical techniques (e.g., convolutional neural networks) and increased computing power provide an opportunity to improve estimates of functional responses of marine predators to changes in their prey field. Studies linking the prey capture rates of marine predators to their environmental context (e.g., measures of environmental variability, prey abundance, or fishing pressure) have the potential to inform management decisions. This working paper introduces our efforts to develop predictive models that can identify prey capture events of chinstrap penguins in acceleration data. We used animal-born video cameras and accelerometer-dive loggers to obtain visual and acceleration-based measures of prey captures of chinstrap penguins, and to train deep neural network algorithms to predict prey capture events in new acceleration data. This work is a step towards monitoring indices such as foraging efficiency (prey capture over overall energy expenditure) that can show how chinstrap penguin functional responses vary over time and in areas of differing environmental conditions or fishing pressure.