Applying Machine Learning to Marine Animal Detection - UW Marine Renewable Energy Seminar

Abstract: The Adaptable Monitoring Package (AMP) is an integrated instrumentation system co-developed by the University of Washington and the Applied Physics Laboratory (APL) that permits unprecedented in-situ observation of marine environments. The AMP’s sensor suite includes two stereo-vision cameras, an imaging sonar, a multi-beam sonar, an echo sounder, an ADCP, a water clarity sensor, and a hydrophone array. Because these instruments can produce as much as 60 TB of data per day, automated data review tools are necessary to economically and expediently analyze AMP data.

In this talk, Mitchell will present an approach to the detection and segmentation of marine animals from AMP observations that utilizes the the popular You Only Look Once (YOLO) instance segmentation algorithm, as well as traditional computer vision processing approaches. He will compare the performance of models built from data collected during AMP deployments in Sequim, WA and the Wave Energy Test Site (WETS) in Kaneohe Bay, HI, and from additional data collected at traditional hydroelectric dams. Preliminary results suggest that the current approach is adept at detecting fish while maintaining a low false positive rate.

Speaker Bio: Mitchell Scott is a software engineer at the University of Washington Applied Physics Laboratory (APL). His primary focuses include computer vision for underwater perception and marine robotic mapping, localization, and reconstruction. Prior to joining APL in 2018, he attended Washington State University where he received his B.S. (2016) and M.S. (2018) degrees in mechanical engineering, where his graduate study focused on multi-robot systems.

For more videos from the AMP, visit

For more information about PMEC, visit

Pacific Marine Energy Center,PMEC,marine energy,Adaptable Monitoring Package,marine monitoring,fish detection,machine learning,Applied Physics Lab,University of Washington,