Passive Acoustic Monitoring for Fish Sounds in the Delaware Bay

• Ethan Pang, Electrical Engineering, Swarthmore College

• Mohsen Badiey, Electrical and Computer Engineering, University of Delaware

Passive acoustic monitoring (PAM) is a non-invasive and cost-effective way to monitor our oceans and the organisms inside of them. PAM techniques have been applied to sea animals like cetaceans, but applications to fish have been relatively less explored. This research specifically deals with soniferous fish in the Delaware Bay. Using databases such as Sounds of Western Atlantic Fishes by Marie P. Fish and William H. Mowbray, audio recordings of relevant fish calls can be visualized as spectrograms, which can then be matched to calls in spectrograms from collected Delaware Bay data. Given the capacity to record a single location for long periods of time, automation of this identification process through machine learning algorithms is necessary. An extensive literature review of the state of the art in fish sound detection with machine learning was performed, and viable methods were tested with collected data from the Delaware Bay. A popular choice is YOLO (You Only Look Once), an object detection model that not only produces bounding boxes around areas of interest but also classifies them into categories. By training this model on spectrograms from the aforementioned existing databases, fish sounds in novel spectrograms generated from the Delaware Bay can be automatically detected and classified, increasing efficiency when compared with human data annotators. The results of these classifications can provide insight not only about the species themselves and how their sounds may change based on different factors, but also about the overall health of the ecosystem.