Improved ground-truthing and benthic habitat characterization with machine learning models and 3D photogrammetry

Benthic mapping relies on a range of methods to ground truth remotely sensed data to map physical and biotic features. Traditional methods are labor-intensive and time-consuming with limited scalability. Autonomous underwater vehicles (AUVs) and remote operated vehicles (ROVs) can collect high volumes of quality data but require automated processing routines and pose other logistic challenges. This paper explores the automated use of multiple machine learning (ML) models to detect and delineate benthic habitat components from structure-from-motion (SfM) orthomosaics. Overlapping AUV photos were processed into high-resolution, seamless orthomosaics for a portion of the Lake Michigan lakebed. ML models were trained to delineate rocks and live mussels and subsequently applied to the SfM orthomosaics. Results demonstrate that a robust pipeline with SfM and ML models is an effective approach to efficiently, accurately, and comprehensively map benthic habitats.