Assessing Deep-Sea Fauna at Scale: An Automated Approach

Lead Research Organisation: University of Oxford
Department Name: Environmental Research DTP

Abstract

The deep sea is the largest ecosystem on Earth. Nevertheless its marine fauna and communities are at best sparsely explored, surveying considered a major challenge due to accessibility and scale. However such barriers are now beginning to fall as advances in submarine drones widen our research capabilities and bring deep-sea exploration and large-scale marine-life appraisal much closer.Submersible drones, for example autonomous underwater vehicles (AUVs), when fitted with video capture technology are increasingly making possible new studies of both fish and benthic (seabed) communities. They can record life in great depths, in extreme environments and move through vast areas, and offer potential to generate extensive fresh data on marine-dwelling species and their habitats. Such surveying is important to help map biodiversity, communities and establish new ecological baselines. It can also evaluate community change. Such analyses can support oceanic modelling and endorse environmental frameworks, as appropriate. However, although we can now envision extensive image capture, a major "choke" is the current process used to analyse image-data. To date, major delays are common in marine image analysis. This is mainly because contemporary methods are reliant on manual Video Annotation Systems (VASs). To alleviate this bottleneck research is proposed herein to ultimately develop automated approaches to analysis. It is considered that automated approaches to image classification based on machine learning could revolutionise the use of deep-submergence platforms for surveying biodiversity, offering an increase in speed of video / photo analysis of orders of magnitude, massively accelerating deep-sea science. Such approaches would also dramatically improve the prospects of monitoring impacts of human activities such as deep-sea mining over extended time periods.
Initially the project intends to focus on seafloor communities with deployment of remotely operated vehicles (ROVs) and AUVs in benthic environs of the "pristine" waters of Antarctica. Study will form part of the Weddell Sea Expedition 2019 investigating the Larsen C ice shelf and iceberg A-68. Megafaunal imagery captured will be used in machine learning exercises harnessing Citizen Science (CS) platform Zooniverse. Over the research period, under project name "Poseidon's Eye", a global forum will be encouraged to identify and characterise marine assemblages according to taxonomic category, anomalous, and additionally (visible) anthropogenic debris. Respective CS is expected to provide crowd-sourcing of image-characterised data-sets for leverage in algorithm generation and synthesis of artificial intelligence (AI) pathways for animal classification. Intention is machine learning will encompass characterised imagery from diverse geographies and research would ultimately provide new "tools of choice" for large-scale and expedient seabed imagery appraisals, with product extensions being envisaged. Please note that sensitivities lie within the project, each step of the proposal being demanding in respect of delivery and potential hurdles arising.

Publications

10 25 50