Abstract
The study of seabird movement is fundamental to understanding how seabirds interact with their environment, and to developing conservation policies for marine ecosystems. In response to the recent accumulation of data describing their movements and habitats, this thesis proposes to characterize and simulate the movements of tropical seabirds using artificial intelligence approaches, more specifically deep learning. On one hand, convolutional neural networks are used to identify seabird habitats from satellite data and to describe their behavior from GPS trajectories. These tools are able to exploit heterogeneous data by taking into account several spatio-temporal scales. They automatically define relevant data mining metrics that can be generalized to the analysis of the movement of other bird species. On the second hand, generative networks are used to simulate foraging trajectories of seabirds. Most of the existing movement simulation tools, such as random walks, focus mainly on small-scale statistics of trajectory data and fail to reproduce realistic large-scale movement patterns. Here we show that generative adversarial networks are able to reproduce geometric features at both small and large scales. These results suggest that generative models can become a pragmatic solution for simulating and predicting complex stochastic processes, such as seabird trajectories, for which the underlying mechanistic rules are unclear or too difficult to be anatically formulated.