Developing deep machine learning of vocal biomarkers for non-invasive monitoring of the health and welfare status of the Navys mine-hunting dolphins
Abstract
The Navy mine hunting dolphins (Tursiops truncatus) are an integral part of Navy operations, as their ability to detect sea mines provides safe lanes for passage of our ships. These dolphins are extremely valuable because of their utility to U. S. defenses and their extensive training. While it is well established that marine mammals rely on their acoustic capabilities for seeing, hearing, and communicating underwater, acoustic behaviors are currently absent from marine mammal welfare evaluations. The overarching goal is to develop tools to identify vocal biomarkers of rare instances of illness or injury and, just as importantly, acoustic indicators of healthy and thriving dolphins. A thorough exploration of dolphin vocal behavior and acoustic characteristics during times of health and during rare instances of illness or injury has the potential to identify vocal biomarkers of health and welfare for bottlenosedolphins.Access to this healthy and long-lived group of animals provides anunparalleled opportunity for studying the foundation of what comprises a healthy bottlenose dolphin vocal repertoire. As is the case for any animal, there are still rare instances of the occurrence of an injury or infection. Unlike their wild counterparts, these animals receive the gold standard of veterinary care in response to these experiences. Therefore, we have the opportunity to record individual animals throughout times of optimal health, sub-optimal health, and throughout their recovery. We propose to do this both at the group level by monitoring dolphin group acoustic activity 24/7, and at the individual level, by recording dolphins when isolated from groupmates during local training procedures, transport, and medical assessments. Based on the mammalian trend for non-linear phenomenon in vocalizations to be a common cue for respiratory and/or vocal tract disease, we will also provide the first thorough review of the presence and persistence of non-linear phenomena in dolphin vocalizations for both healthy and unhealthy animals. Finally, we will continue to be at the forefront of technological advancement and to employ cost-saving techniques for health monitoring of Navy bottlenose dolphins. In order to do this, we will build deep machine learning artificial neural networks to identify, classify and learn the vocal biomarkers present in the Navy dolphin whistle emissions. Frequently, complex patterns exist in animal calls that may be salient to the communicating species but human bias for our own language makes them difficult to discern. Transfer learning of deep neural networks for the detection of marine mammal call features paired with the success of artificial intelligence for identifying vocal biomarkers of human health, has the potential to culminate in a novel tool for identifying health status from marine mammal vocalizations. The current project is aimed at reducing the number of dolphin sick days by using non-invasive tools to early predict rare instances of illness in the Navy bottlenose dolphins. Early diagnosis facilitates a preventative medicine approach, effective treatment plans, fewer lost work days, and better overall health and welfare. This should improve readiness and overall robustness of dolphins in the MK-7 mine hunting system.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 05, 2021
- Source ID
- N000142112414
Entities
People
- Sam H Ridgway
Organizations
- National Marine Mammal Foundation
- Office of Naval Research
- United States Navy