Deep inference of seabird dives from GPS-only records: performance and generalization properties

Abstract

Seabirds are considered as suitable indicators for the study of marine ecosystems, since their foraging strategies provide a real-time response to complex ecosystem dynamics. By deploying GPS sensors on seabirds it is possible to obtain their trajectories, and deep learning have recently shown promising results for the classification of animal behaviour from trajectory data. Yet there is still lot of investigation needed in terms of network architectures, data representation but also to demonstrate the generalization properties of these approaches. From a database of about 250 foraging trajectories derived from GPS data deployed simultaneously with pressure sensors for the identification of dives, this work has consisted in training deep networks in a supervised manner for the prediction of dives from trajectory data. In this study, we confirm that deep learning allows better dive prediction than usual methods such as Hidden Markov Models for two distinct seabirds species. We propose a novel deep learning model for trajectory data. It combines the computational efficiency of convolutional neural networks to distance-matrix-based representations of trajectory data. Our model considerably increases the ability of deep networks to infer behaviour, as well as their stability to different data inputs. The considered trajectory data representation might enable deep networks to better capture spatial information than from longitude and latitude time-series considered in previous works.