Sea Turtle Behavior Classification

Deep learning approaches for automatic identification of underwater sea turtle behaviors using biologging sensors (accelerometers, gyroscopes, depth recorders).

Sea turtles spend most of their lives underwater, making direct observation of their behavior extremely challenging. By deploying miniaturized multi-sensor loggers (accelerometers, gyroscopes, and time-depth recorders) on free-ranging turtles, we can record continuous movement data that encodes their behavioral repertoire.

Our work progressed from classical machine learning (Random Forest, CART) to fully convolutional neural networks (V-Net), achieving increasingly accurate automatic identification of behaviors including swimming, resting, feeding, and scratching — activities of crucial ecological importance that were previously impossible to monitor remotely.

Key contributions

  • First validation of accelerometer-based behavior identification in sea turtles
  • Combined supervised learning approach (Random Forest + CART) for behavior discrimination
  • Development of V-Net architecture adapted from biomedical image segmentation to 1D sensor data
  • Automatic estimation of reproductive outputs (egg counts) from accelerometer data

References