Using AI computer vision algorithm (YOLOv11) for automatic video annotation of coral taxa on the Moroccan Atlantic Coast

Abstract. Marine biodiversity is essential for maintaining healthy and resilient ocean ecosystems, supporting fisheries, regulating climate, and providing vital resources for human well-being. It underpins ecosystem services such as carbon sequestration and oxygen production, making it important for both environmental and economic sustainability. Consequently, the conservation and monitoring of highly species rich and vulnerable marine ecosystems, such as shallow-water gorgonian coral populations, are important for biodiversity preservation. This study tests the ability of the AI computer vision algorithms YOLOv11, to detect and count coral colonies belonging to six common taxa on video records from coral gardens on the Moroccan Atlantic Coast. These videos were recorded using an Remotely Operated Vehicle (ROV) with the objective to map coral habitats with the research vessel Dr. Fridjof Nansen as part of the FAO Nansen program. Focusing on three gorgonian species: Eunicella verrucosa (Pallas, 1766), Ellisella paraplexauroides Stiasny, 1936, and Leptogorgia viminalis (Pallas, 1766), two sea pen species: Veretillum cynomorium (Pallas, 1766), and Pennatula rubra Ellis, 1764, and the hard coral Dendrophyllia ramea (Linnaeus, 1758). The research aims to develop an efficient solution to help improving video annotation by making it faster and easier. A dataset of 658 coral images was collected from Google Image and the DORIS database (Données d'Observations pour la Reconnaissance et l'Identification de la faune et la flore Subaquatiques). The images were divided into training, validation and test sets. To enhance model performance, we applied data augmentation. The YOLOv11 includes five different variants (n, s, m, l, x) for which detection precision was compared. Based on precision, recall, F1-score and mAP metrics, YOLOv11n proved to be the best model for coral detection regarding balance of accuracy and efficiency and with a mAP of 88% and a F1-score of 81%. This model was used for all subsequent ROV video analyses. The prediction results were applied to ROV video recordings from shallow water areas, demonstrating the potential of YOLOv11 as a powerful tool for the automated detection and monitoring of coral gardens. This approach offers significant contributions to marine biodiversity assessment along the Moroccan Atlantic coast.