Predictive habitat mapping and iterative planning of image surveys given existing bathymetry and imagery - a machine learning perspective.

Machine learning research offers flexible and powerful approaches to handling observations at multiple scales and of different modalities to construct predictive models with meaningful representations of uncertainty. Beyond providing a sense of the quality of the models, these representations can guide further collection of observations to improve predictive capabilities. The traditionally-resource constrained problem of generating habitat maps from full coverage acoustic multibeam data and targeted optical surveys can be viewed through the lens of machine learning and adaptive sampling. This paper investigates the use of state-of-the-art techniques in machine learning including deep learning methods, Gaussian Processes and Dirichlet-Multinomial regressors to generate habitat maps and to suggest where further sampling would be most useful. We discuss the interpretations of the different measures of uncertainty associated with these models and their implications for the choice of observations and survey design. We also discuss novel ways of viewing the relationships between image data and bathymetry that are possible with these models. We present results based on surveys performed in tropical and temperate reefs in Australia using ship-borne multibeam sonar and precisely georeferenced imagery collected with AUVs as part of the benthic monitoring program run by Australia’s Integrated Marine Observing System.