SeagrassFinder: Deep Learning for Eelgrass Detection and Coverage Estimation in the Wild

TL;DR

This study demonstrates that deep neural networks, particularly Vision Transformers, can effectively automate eelgrass presence detection in underwater videos and, when combined with underwater image enhancement, achieve high AUROC (>0.95). It introduces a data-efficient workflow with a custom annotation platform, transfer learning from ImageNet, and a novel data-driven approach to estimate eelgrass coverage from image streams using a temporal mean of binary predictions. The work also discusses practical deployment considerations, including the potential integration with autonomous underwater vehicles for automated transect surveys and scalable environmental monitoring. Together, these contributions offer a scalable, objective, and repeatable framework for marine ecological monitoring and EIAs, with implications for coastal management and blue-carbon assessments.

Abstract

Seagrass meadows play a crucial role in marine ecosystems, providing benefits such as carbon sequestration, water quality improvement, and habitat provision. Monitoring the distribution and abundance of seagrass is essential for environmental impact assessments and conservation efforts. However, the current manual methods of analyzing underwater video data to assess seagrass coverage are time-consuming and subjective. This work explores the use of deep learning models to automate the process of seagrass detection and coverage estimation from underwater video data. We create a new dataset of over 8,300 annotated underwater images, and subsequently evaluate several deep learning architectures, including ResNet, InceptionNetV3, DenseNet, and Vision Transformer for the task of binary classification on the presence and absence of seagrass by transfer learning. The results demonstrate that deep learning models, particularly Vision Transformers, can achieve high performance in predicting eelgrass presence, with AUROC scores exceeding 0.95 on the final test dataset. The application of underwater image enhancement further improved the models' prediction capabilities. Furthermore, we introduce a novel approach for estimating seagrass coverage from video data, showing promising preliminary results that align with expert manual labels, and indicating potential for consistent and scalable monitoring. The proposed methodology allows for the efficient processing of large volumes of video data, enabling the acquisition of much more detailed information on seagrass distributions in comparison to current manual methods. This information is crucial for environmental impact assessments and monitoring programs, as seagrasses are important indicators of coastal ecosystem health. This project demonstrates the value that deep learning can bring to the field of marine ecology and environmental monitoring.

Figures (13)

• Figure 1: The SeagrassFinder Project Pipeline: (1) a vessel performs transects surveys by towing a sled with a camera along the seabed and (2) records videos of the underwater environment. The extracted images (3) are then labeled by human annotators (4) and used to train deep neural networks (DNNs) on the task of detecting eelgrass (5). A trained DNN can now be used to replace the time-consuming manual annotation process.
• Figure 2: (a) - (c) Sample images of areas with no, sparse and dense submerged aquatic vegetation (SAV). (d) A map of the Copenhagen harbor with transect survey lines (blue) and SAV data overlaid. All SAV maps are based on © Copernicus Sentinel-2 data from 2023-06-13. All SAV maps are based on analysis of Sentinel-2 imagery from 2018 (map created with QGIS v.3.30.2).
• Figure 3: A screenshot of the SeagrassFinder Annotation Platform accessed through a browser window. The labeling interface is simple and intuitive.
• Figure 4: Two images before (left) and after (right) applying underwater image enhancement.
• Figure 5: Transfer learning: (1) we use DNNs pre-trained on a generic data set (blue), here ImageNet russakovsky2015imagenet. (2) The trained layers of each DNN model are frozen and infused into our custom training process. (3) We fine-tune a small number of additional layers on our eelgrass data set (green) to perform the target task of eelgrass detection.