Movement is a fundamental characteristic of life that enables organisms to find food and mates, avoid predators, and adapt to environmental changes. In marine systems, studying animal movement is key to designing marine protected areas and assessing how species interact with human pressures, such as fisheries. Decades of tagging and advances in tracking technologies have produced large datasets, positioning movement ecology as a data-rich discipline and opening new opportunities for data-driven research. Machine learning, in particular, offers a suite of techniques tailored to extracting complex patterns from this high-dimensional information. In this thesis, we explore its potential through two distinct tasks: the automated identification of marine species from their movement trajectories and the probabilistic forecasting of their locations.
First, we developed a model to identify species from satellite tracking data. We trained a deep learning classifier on thousands of trajectories from more than 70 species and combined it with a confidence estimator that allows the model to abstain when faced with an unknown species. In regions where the model was trained, it achieved high accuracy and correctly refrained from classifying most unknown species. Performance varied among taxa, with fishes and turtles requiring more data to reach high accuracy than birds or seals. Our analysis showed that trajectories reflect species-specific movement patterns, with geographic location being the main factor influencing predictions.
We then focused on predicting marine animal movements, recognizing that their inherent randomness—driven by external factors such as prey availability and internal ones such as physiological state—demands probabilistic models that quantify uncertainty. We adapted a transformer-based deep learning model to predict the future locations of southern elephant seals and to fill gaps in their tracking records, providing both point estimates and uncertainty regions. When tested in familiar geographical areas, our model outperformed standard state-space models, yielding more accurate point predictions and better-calibrated uncertainty estimates that contained the true positions within smaller areas. Additionally, we determined the model’s optimal operational regime, which occurred for animals moving at low-to-intermediate speeds along the continental shelf.
This thesis demonstrates that advanced machine learning models can effectively analyze marine animal movement and outperform traditional methods. Our work provides practical tools for confidence-aware predictions, uncertainty quantification and performance diagnostics, as well as procedures for assessing data requirements and preprocessing satellite-tracking data. Altogether, it serves as a foundational step toward automated metadata extraction and probabilistic movement forecasting.