Using machine learning to achieve simultaneous, georeferenced surveys of fish and benthic communities on shallow coral reefs

Abstract

Surveying coastal systems to estimate distribution and abundance of fish and benthic organisms is labor-intensive, often resulting in spatially limited data that are difficult to scale up to an entire reef or island. We developed a method that leverages the automation of a machine learning platform, CoralNet, to efficiently and cost-effectively allow a single observer to simultaneously generate georeferenced data on abundances of fish and benthic taxa over large areas in shallow coastal environments. Briefly, a researcher conducts a fish survey while snorkeling on the surface and towing a float equipped with a handheld GPS and a downward-facing GoPro, passively taking ~ 10 photographs per meter of benthos. Photographs and surveys are later georeferenced and photographs are automatically annotated by CoralNet. We found that this method provides similar biomass and density values for common fishes as traditional scuba-based fish counts on fixed transects, with the advantage of covering a larger area. Our CoralNet validation determined that while photographs automatically annotated by CoralNet are less accurate than photographs annotated by humans at the level of a single image, the automated approach provides comparable or better estimations of the percent cover of the benthic substrates at the level of a minute of survey (~ 50 m² of reef) due to the volume of photographs that can be automatically annotated, providing greater spatial coverage of the site. This method can be used in a variety of shallow systems and is particularly advantageous when spatially explicit data or surveys of large spatial extents are necessary.