Data and code from: To stay close or go far? When streaked shearwaters decide the duration of their trips at sea

Revealing the timing of decision-making by animals is crucial to understand the behavioral strategy of animals and their relationships with the surrounding environment. Although GPS tracking data of animals in the wild enable computational analysis of animal movement, it is difficult to identify the timing of decision-making from multi-dimensional time-series data consisting of dozens of parameters such as speed and orientation. In this study, we developed a machine learning method based on a group of classifiers to determine the timing when streaked shearwaters (Calonectris leucomelas) decide to conduct short (an average of 15 h) or long (>1 day) foraging trips by analyzing GPS trajectories. In our method, each classifier is responsible for different time periods of data and is designed to predict the type of the foraging trip by processing the multi-dimensional time-series data extracted from trajectory segments within the period. We found that the types of trips were predicted with high accuracy (approximately 70%) only from the initial 5 h of trajectory data, indicating that the shearwaters decided the types of trips at a very early phase after starting their trips. Following the machine learning method, an in-depth analysis of the initial 5 h period of the trips suggested that the shearwaters conducted long-term trips when they could not catch fish during this time. Our framework opens new directions for the analysis of tracking data and provides evidence of in-situ decision-making by wild animals.