A comparative study on CNN-based semantic segmentation of intertidal mussel beds

Abstract

Mussels are globally distributed and abundant species on intertidal habitats such as rocky shores where they form dense beds and provide a variety of ecosystem services. The ability to measure mussel abundance on shores with high efficiency and accuracy will enhance assessment of their facilitative services to intertidal ecosystems. Traditional sampling methods, however, are labour intensive and consequently limited in spatial and temporal coverage. To advance such monitoring and quantification processes, four state-of-the-art convolutional neural networks (CNN); namely U-Net, FastSCNN, DeepLab v3+, and HRNet, were compared in their performance to segment mussels at the pixel level either with or without Cityscapes pre-trained weights. All proposed CNN were trained on a photographic dataset of 1455 images generated from field images of mussel beds. CNN trained with the Cityscapes pre-trained weights showed an overall increased model performance as illustrated by mean intersection-over-union (mIoU), loss, overall accuracy (oAcc), and Cohen's kappa coefficient. DeepLab v3+ and HRNet with the combination of Cityscapes pre-trained weights achieved over 95% mIoU and 97% oAcc for assessing distribution and abundance of these important intertidal species. Such combination of approaches is, therefore, recommended for ecologists who wish to conduct image-driven surveys of intertidal ecosystems and the code for this study is available at: https://github.com/Vicellken/CNN-SS-mussel.