A Closer Look at Seagrass Meadows: Semantic Segmentation for Visual Coverage Estimation

Underwater imaging enables marine researchers to collect large datasets of seagrass images. These images can be used to monitor the health state of underwater meadows by estimating the area that is covered by seagrass and how this area changes over time. Since the manual analysis of such images is slow and error-prone, we follow the path of deep learning for automatic image analysis.Our contribution is the investigation of deep semantic segmentation for the specific task of seagrass coverage estimation. We evaluated multiple Deep Neural Network Architectures including the DeepLabv3Plus Network which performs best, with a mean intersection over union of 87.78%. The qualitative results in our experiments indicate that the Deep Learning approach is not only more accurate than a human but also multiple times faster in annotating underwater meadows. Our code is available on GitHub: https://enviewfulda.github.io/LookingForSeagrassSemanticSegmentation/.