Benthic Mapping of Geological, Biogeochemical and Biodiversity Parameters through an Autonomous Vehicle and Deep Learning Software Workflow

The time has come to integrate the capabilities we have developed for real-time habitat processing on shipboard with the HabCam towed vehicle, into an autonomous vehicle with 3D reconstruction of seafloor topology, substrate classification, single target identification, hyperspectral imaging for physiological information, and plankton classification as an index of ecosystem health. Integrated together, these data streams represent a full description of habitat that supports a variety of organisms, communities, and a defined biodiversity. Designed and built at WHOI by Gallager’s team, HARIM (Habitat Aware Reconnaissance and Imaging Module), is a complete package of sensors and processing capability to survey habitat at depths from 0 to 6000m, depending on which REMUS vehicle is used. Our goal for this Phase I project is to fully integrate information being collected by HARIM with the vehicles’ navigational system to create a dynamic sampling capability depending on habitat information. Deep learning CDNN models of habitat will be built from stereo images in a variety of habitats, Habitat Suitability Modeling will be used to project habitats using statistical inference, and topic modeling will be used to label habitat components and specific targets to ascertain the degree of information content. A dynamic sampling scheme will understand when habitat information is changing or when it is stable and guide the vehicle to maximize information content. The market for habitat characterization is large- from wind farm siting and monitoring to oil and gas prospecting and environmental monitoring to research and exploration of novel environments and assessment of damage to both shallow and deep coral reef systems. This Phase I will complete the software workflow for HARIM and test it under rigorous field conditions. Phase II will commercialize the product by rugedizing the hardware and hardening the software. The product will be a habitat characterization module that fits on a variety of autonomous and human controlled platforms from AUVs to ROVs.