Performance of machine learning algorithms for classifying benthic habitats and species

Session: Seeing Below the Surface: Quantifying the Underwater Environment with Image Analysis

Peter Esselman, U.S. Geological Survey, pesselman@usgs.gov
Christopher Roussi, Michigan Tech Research Institute, croussi@mtu.edu
Meryl Spencer, Michigan Tech Research Institute, mspencer@mtu.edu
Samuel Pecoraro, US Geological Survey, specoraro@usgs.gov
Jennifer Wardell, US Geological Survey, jwardell@usgs.gov
Scott Dwyer, US Geological Survey, sdwyer@contractor.usgs.gov
Anthony Arnold, US Geological Survey, ajarnold@contractor.usgs.gov

Abstract

The benthification of Great Lakes foodwebs facilitated by dreissenid mussels places new emphasis on the need for quantifying benthic ecosystem attributes including habitats, mussels, round goby, and Cladophora.  US Geological Survey and Michigan Tech Research Institute are working collaboratively to develop a robot-assisted computer vision system that includes an autonomous underwater vehicle carrying camera systems and lights, as well as algorithms to condition acquired imagery and quantify features of interest.  A subset of images collected during the 2018 field campaign were manually annotated for substrate types, round goby individuals, Cladophora presence, and dreissenid mussel presence.  The annotated images were used as training data with several machine learning approaches (support-vector machines, convoluational neural networks) to conduct pixel-level classification of the features of interest in novel images.  Initial findings from the algorithm development process will be shared, along with plans for subsequent improvements to the classifiers in 2019 and beyond.