MACHINE LEARNING APPROACHES FOR BROAD-SCALE CHARACTERIZATION OF SEAFLOOR GEOLOGY ON THE NORTHWEST ATLANTIC SHELF

Abstract

Mapping seafloor surficial geology provides information necessary for effective marine spatial planning, assessing natural and anthropogenic disturbances to the environment, and science-based fisheries and natural resource management. Traditional methods for seafloor geological mapping depend on by-eye interpretation and delineation by experts in the field. However, these methods are generally subjective, non-repeatable, and lacking in statistical validation. Recent advances in computing technology and modelling techniques, such as machine learning, have allowed scientists to use spatial predictive modelling to efficiently produce statistically accurate and spatially continuous map products that characterize various aspects of seafloor geology. This research applied common machine learning algorithms to map seabed sediment properties and classify seafloor morphology across the Northwest Atlantic Shelf using coarse resolution (> 400 m) open-source datasets. Sediment properties (hard substrate occurrence, modified Folk class, mean grain size, and % mud/sand/gravel) were modelled using random forest, trained with observations from NRCan Seabed grainsize analysis and seafloor photograph databases for offshore Canada. A semi-automated approach was developed to classify seafloor morphology from the GEBCO 2020 bathymetric grid, using a k-means clustering algorithm and manual assignment of feature names based on standardized feature definitions. The standardized workflow from this study enables the integration of datasets from a variety of sources and provides output maps that are comparable over large regions and across a variety of ocean governance boundaries.