Automated Detection and Counting of 12 Macrobenthic Taxa on the Basque Coast using Computer Vision on ROV Videos

Abstract

Developing robust and cost-effective methods for benthic biodiversity assessments is crucial for mapping vulnerable habitats and designing effective management strategies. Achieving this objective requires increasing the spatial and temporal coverage of observations and optimizing the cost-effectiveness of sampling efforts. Remotely operated vehicles (ROVs) together with computer vision-based tools present great potential for cost-effectively mapping benthic taxa. However, the processing of hours of footage manually is time-consuming, making the process inefficient and costly. To address this challenge, a semi-automated system for the detection and abundance estimation of 12 key benthic taxa was developed. The proposed approach analyses videos obtained with a ROV and provides abundance counts to support habitat monitoring. It was developed and tested using a total of 42 sampling transects, resulting in approximately 20 hours of high-quality video footage. Three approaches were evaluated: YOLOv8x, Faster R-CNN, and weighted boxes fusion (WBF) ensemble combining the predictions of both detectors. Their performance was evaluated using mAP50, reaching values of 0.7 ± 0.028 for YOLOv8x,0.792 ± 0.029 for Faster R CNN, and 0.832±0.028 for WBF, indicating strong detection capability. Since WBF showed a slightly higher performance, it was selected and used for abundance estimation. Comparison of automated abundance estimates with manual counts across 10 transects gave a transect-level mean absolute percentage error (MAPE) of 20.2%. However, taxon-level error was larger (MAPE = 28.6%), indicating that abundance estimates are currently more reliable at transect level than for individual taxa. Consequently, this semi-automatic approach reduces the effort required for macrobenthic biodiversity monitoring by reducing the image-analysis time, thereby enabling an efficient detection of ecological changes and offering valuable insights for benthic habitat preservation