Characterization of fine scale variation in substrates using robots and machine learning

Session: Mud, Macrofauna and Microbes: Benthic Organism-Abiotic Interactions at Varying Scales (2)

Peter Esselman, U.S. Geological Survey, [email protected]
Christopher Roussi, Michigan Tech Research Institute, [email protected]
Meryl Spencer, Michigan Tech Research Institute, [email protected]
Samuel Pecoraro, US Geological Survey, [email protected]

Abstract

Despite being a major control on pattern and process in benthic ecosystems, spatial variation in lake surface geology remains poorly mapped and understood for most areas of the Great Lakes.  Widespread application of new and existing technologies is needed to map variation in surface geology to inform our understanding of ecological constraints on benthic ecosystems from local to regional scales.  Here we present our initial findings from application of a robotic computer vision system to the problem of local-scale physical substrate classification in lakes Ontario, Huron, and Michigan.  Color and grayscale stereo images of the lake bottom were collected by SCUBA divers and an autonomous underwater vehicle during 2018 and 2019.  A subset of images was annotated manually for different substrate classes, and these 'labeled' images used as training data for supervised classification using support-vector machines and deep convolutional neural networks.  Initial findings and advantages of the different approaches are discussed.  Priorities for future work are identified relative to regional initiatives striving to support benthic management of the Great Lakes ecosystem.