Combining acoustic and visual survey data through multimodal machine learning for wide area seafloor habitat classification

Autonomous underwater vehicles (AUVs) are increasingly used for a wide range of marine surveys, for example to support marine spatial planning, conservation or monitoring. Recently, robotic camera surveys have begun to complement wide-area acoustic mapping efforts, as visual images can provide more detailed information that can be used for more reliable classification of seafloor habitats than acoustic data alone. However, the strong attenuation of light in water requires visual data to be gathered at close range, limiting the areas that can be observed. This tradeoff between survey extent and detail remains an open challenge in marine survey.

The aim of this project is to develop novel objective approaches to build probabilistic habitat models that learn relationships between habitat classes derived from visual imagery and infer these onto wide-area environmental (primarily acoustic) data. The new techniques will be applied to extensive datasets of visual seabed images, typically collected by AUVs in benthic environments of conservation interest (e.g. cold-water coral reefs, submarine canyons,..). This work will establish best practice to estimate seafloor habitat distribution at the regional scales needed for marine conservation and monitoring.