Transfer Learning for Conservation

Cross-species transfer learning to overcome data scarcity in endangered species monitoring, demonstrated on hawksbill sea turtles.

Data scarcity is a major challenge when applying deep learning to ecology — particularly for endangered species where collecting large labeled datasets is logistically difficult or ethically constrained. Transfer learning offers a solution by reusing models trained on data-rich species.

We demonstrated that a deep learning model trained on green turtle acceleration data can be successfully transferred to classify behaviors of critically endangered hawksbill sea turtles, achieving an 8% F1-score improvement over training from scratch.

Key contributions

  • First application of cross-species transfer learning for biologging data
  • Successful transfer from green turtles to hawksbill turtles
  • Additional validation using human activity recognition datasets
  • Practical framework for extending deep learning to data-scarce species

References