Automated Observation and Prediction of Fine-Scale Spatial Distributions of Benthic Fauna

Nader Boutros1, Jacquomo Monk2, Oscar Pizarro1,3, Stefan B Williams4 and Neville Barrett2, (1)University of Sydney, Australian Centre for Field Robotics, Sydney, NSW, Australia, (2)University of Tasmania, Institute for Marine and Antarctic Studies, Hobart, TAS, Australia, (3)ACFR, University Of Sydney, Australia, (4)The University of Sydney, Australian Centre for Field Robotics, Sydney, NSW, Australia
Abstract:
Benthic imaging AUVs can be used to map the visual appearance and 3D structure of relatively large areas (~10-10000m2) of the seafloor while preserving the spatial distributions of observed habitat features and associated biology. However, these surveys result in large volumes of image data which can be difficult and time-consuming to manually annotate. This study used a region proposal and classification convolutional neural network to automatically detect sea urchins and rock lobsters in survey imagery, with each detected individual’s position is estimated using the vehicle’s navigation data at St Helen’s Reef, south-eastern Australia. This results in a geo-referenced location of each individual urchin and lobster with minimal manual annotation.

Habitat structure and the presence or absence of kelp are known drivers of the distributions of both urchins and rock lobsters. The 3D seafloor reconstruction was used to calculate structural complexity across the survey site, and we estimated the distribution of kelp from the reef photomosaic. Complexity, quantified as rugosity and slope, was calculated over virtual quadrats draped on the surface reconstruction in the form of a triangular mesh. We varied quadrat size to capture the spatial scales relevant to the species’ distributions.

Species distribution models were trained from the complexity and kelp distribution maps, using the location of each individual generated by the neural network detector as presence points for each species. These models accurately predicted the likely distributions of the urchins and lobsters in new areas, as well as suggesting that a combination of kelp presence and complexity at larger spatial scales (> ~1.5m) were most important in defining suitable habitat for these benthic fauna. This approach facilitates a semi-automated workflow to examine the structural drivers of the distributions benthic marine organisms at spatial resolutions previously not explored.

Back to: Artificial Intelligence Systems for Advancing the Study of Aquatic Ecosystems I