Automated Detection and Recognition of Wild Dolphin Behaviors Using Deep Learning

We presented a deep convolutional neural network approach that was able to automatically detect and recognize the behavior of wild dolphins. This study used a deep learning approach to automatically identify and analyze three visu-ally distinct behaviors (exiting behavior, wandering behavior, and entering be-havior) in dolphins. Using data directly from video recordings by both drones and handheld cameras, an action recognition model trained on this data achieved high levels of accuracy (0.99 for exiting, 0.92 for wandering, and 0.88 for entering). The method is the first to achieve visual action recognition of dol-phins behavior, opening new possibilities for using large datasets in dolphins research.