Multi-dataset-integrated Coral-Lab segmentation with enhanced towed camera array for rapid large-scale coral reef monitoring and mapping

Abstract

Highly efficient and reliable monitoring of coral reef ecosystems is imperative for effective conservation and management under increasing anthropogenic and climatic pressures. However, current survey techniques either offer limited coverage and low efficiency or incur substantial manual costs for data processing. In this study, we propose a highly efficient towed optical camera array system, Speedy Sea Scanner version 2.0 (SSSv2), with an advanced electrical system supporting a stable power supply, reliable communications, and underwater illumination, which enables continuous video data collection and real-time monitoring. We also develop a semantic segmentation model, Coral-Lab, with high accuracy and robustness in coral identification task, which enables fully automated coral reef identification and coral coverage calculation. Coral-Lab model achieved an F-score of 0.802 and an mIoU of 0.665 on our test set. Leveraging SSSv2, we conducted field surveys off the northern coast of Kumejima Island, Okinawa, Japan, on July 14, 2024 and July 15, 2024 across seven sampling areas comprising 29 transect lines. Over two days survey, we collected video data covering a total seafloor area of 47,950 m², which was converted into a georeferenced orthomosaic at an average spatial resolution of 2.5 mm via Structure-from-Motion (SfM) and Multi-View Stereo (MVS) techniques. This approach achieved an effective survey efficiency of approximately 7200 m² per hour. Applying Coral-Lab model to 25,658 orthomosaic tiles at a 0.25 m grid resolution, we generated detailed coral-cover distribution maps in under 75 min of inference time, processing each 512 × 512-pixel tile in $\sim$0.17 s. These results demonstrate the synergistic potential of integrating advanced imaging hardware with deep learning algorithms, enabling rapid, large-scale coral reef monitoring and assessments.