Probabilistic Habitat Modeling for Benthic Surveys

Autonomous underwater vehicles (AUVs) are utilised for benthic habitat surveying, collecting imagery of the seafloor to facilitate scientific research and monitoring. The areas that need to be characterized are vast, however, the AUV can only visually sample a small portion of these areas. As the cost of deployment is high, the information gathered should be maximised for each deployment. The relationship between remotely-sensed acoustic data and the sampled imagery can be learned, creating predictive habitat models. These habitat models can be used to plan more efficient AUV surveys. This research applies Bayesian neural networks to predict visually-derived habitat classes when given broad-scale bathymetry and backscatter data. These networks are effective at learning the many-to-many relationship between remotely-sensed data and habitat classes. Furthermore, Bayesian neural networks estimate the uncertainty associated with a prediction, which can be deconstructed into its aleatoric (data) and epistemic (model) components. We demonstrate how these structured uncertainty estimates can be utilised to decide where further samples should be collected to most improve the model. Such adaptive approaches to benthic surveys also have the potential to reduce costs prioritizing sampling efforts. This approach is applied to AUV-collected imagery and corresponding bathymetry and backscatter in the Waikaloa region of Hawaii.