‘Find-a-HALO’ – An AI-assisted detection and segmentation framework for ecological monitoring of Halophila ovalis using YOLOv8 and the Segment Anything Model (SAM)

Fragmentation of seagrass habitats impact the ecosystem services of sensitive coastal lagoons, necessitating periodic, high-resolution, non-intrusive monitoring of the spatio-temporal dynamics to adequately support conservation efforts. The present investigation details the development of a “coupled” deep learning framework by combining You-Only-Look-Once (YOLO) v8 model with the Segment Anything model (SAM) to detect the spatial extent of seagrass patches from underwater images. Using publicly available global database images, the supervised detection of Halophila ovalis was automated. H. ovalis was selected due to its global distribution and for its unique leaf morphology which make it an ideal candidate to apply segmentation models. Also, patches of H. ovalis habitats were found and observed during the field work in Chilika lagoon. The YOLOv8 model was trained and tested to detect H. ovalis signatures from underwater images using the publicly available datasets which were modified and annotated for the study. These datasets were filtered for clear H. ovalis images, which comprised of 1501 images in total out of which 1050, 300 and 151 images are from the training, validation and testing sets respectively. The YOLOv8 model achieved 75.7% Precision, 76.7% Recall, 82.7% mAP@0.5 and 76.2% F1-score. Differentiating H. ovalis from the matrix of the substratum, provided a useful methodology to study the ambient habitat of H. ovalis within a single sampling framework. Later, SAM was employed to segment the seagrass area which provided a useful method to study the patchiness. The combined framework (YOLOv8 + SAM) achieved a mean IoU of 0.8069 while processing images at 1.8 ms postprocess per image. The area coverage estimated by the coupled model, was compared with the global standards for visual interpretation of the coverage to validate the real-time usage of our proposed methodology. The current methodology provides a proof of concept for monitoring of seagrass meadows repeatedly during a single over pass, using boat-bound or a drone-attached camera for rapid, scalable and non-invasive identification. With accurate field validation programme, it can be used to study the habitat fragmentation of H. ovalis successfully.