Coral Intel: a YOLOv8 deep learning framework for monitoring Caribbean corals

Abstract This study evaluated the performance of YOLOv8 object detection framework for identifying Caribbean scleractinian corals in underwater imagery. A dataset of 10373 coral images, expanded to 32423 images via augmentation, was used to train genus- and species-level models spanning 59 taxa. In the held-out test imagery, the genus-level model achieved a mean average Precision ( mAP 50 ) of 0.954, whereas the species-level model reached 0.938, with the performance varying by taxon and generally the highest for morphologically distinctive genera such as Isophyllia and Scolymia . However, independent field validation using transect-based photoquadrat imagery revealed a substantially lower observation-level performance (species-level Precision = 0.389; Recall = 0.336), reflecting the influence of domain shift, morphological similarity among coral taxa, and sparse point-based annotations. Restricting evaluation to the nine species present in the transects improved separability, but did not fully eliminate false positives or missed detections, indicating that label-set reduction does not recover the test set performance under transect survey conditions. The top-down perspective of the field validation imagery lacked key morphological details from oblique-angled training imagery, suggesting model performance is best for near real-time classification from oblique, rather than using top-down, imagery. To facilitate applied deployment for general use, the trained models were integrated into Coral Intel, a web-based framework for automated coral detection currently available in prototype form. Although this study focused on the Caribbean coral taxa, this framework established a transferable foundation for regionally tuned deep learning models that may support scalable reef monitoring.