Predictive habitat models from AUV-based multibeam and optical imagery

In AUV habitat mapping and exploration missions, a prior habitat map with associated uncertainty has the potential to guide the design of AUV deployments more effectively than a bathymetric map alone. We present and characterize an approach for learning predictive models of benthic habitats as a function of seabed terrain features. The models were learned by correlating limited-coverage high resolution imagery with full-coverage multibeam bathymetry data, both collected by an AUV at a site off the Tasman Peninsula in Tasmania, Australia. Correlations observed where these data overlapped were extrapolated to the much larger area covered by the multibeam survey. Accuracies of 0.69 - 0.78 were attained using a 10-fold cross-validation. A feature ranking analysis using bootstrap aggregation was also carried out revealing features were more informative at the larger scales of 5 × 5m
²
. Using bootstrap aggregation we learn probabilistic habitat maps for the site along with map of entropy that indicates areas of uncertainty. We discuss the implications for the planning of AUV missions and for the generation of adaptive trajectories aimed at improving map quality.