Seagrass coverage estimation and depth limit analysis from unlabeled underwater videos

Visual coverage estimation of seagrass for ground truth verification is one of the most critical aspects of marine ecosystem monitoring programs worldwide. It has traditionally been an arduous and tedious task. Commonly used tools like a scuba diver and underwater video transects require manual investigation by domain experts to assess seagrass status. Supervised machine learning methods have had a limited role in automating this process due to the lack of labeled seagrass images. This paper proposes two robust algorithms for seagrass coverage estimation from unlabeled underwater videos obtained from scuba divers and investigates their different potentials. Two seagrass-specific features are extracted and modeled for coverage estimation (0%–100%), matching the domain expert’s prediction. We also show that these algorithms detect and rectify rare labeling mistakes from the domain expert. Coverage estimates from one of the methods are then used to estimate the depth limit and its associated uncertainty.