Automatic Classification of Seagrass using Deep Learning Methods from Underwater Images

Seagrass meadows are among the most important coastal ecosystems, providing essential ecological functions such as carbon sequestration, biodiversity support, sediment stabilization, and water quality improvement. Effective monitoring of seagrass habitats is therefore critical for understanding ecosystem health and supporting sustainable coastal management. However, the analysis of underwater imagery collected during field surveys is often labor-intensive and time-consuming, creating a need for automated classification approaches. This study investigates the potential of deep learning techniques for multi-class seagrass classification using the publicly available DeepSeagrass dataset. The dataset contains underwater images representing different seagrass morphologies and environmental backgrounds acquired from multiple survey locations. Two classification scenarios are considered: a four-class configuration consisting of Strappy, Rounded, Ferny, and Background classes, and a five-class configuration in which the Background category is further separated into Substrate and Water Column classes. The proposed framework employs state-of-the-art YOLO-based image classification models within a transfer learning strategy. Image preprocessing and data augmentation techniques are applied to improve model robustness against common underwater imaging challenges, including variations in illumination, water turbidity, viewing geometry, and habitat composition. Model performance is evaluated using standard classification metrics, including overall accuracy, precision, recall, and F1-score. In addition, comparative analysis is conducted to examine the influence of class configuration on classification performance and generalization capability. The outcomes will provide insights into the suitability of modern YOLO architectures for underwater ecological applications and contribute to the development of efficient computer vision tools for large-scale seagrass monitoring and marine ecosystem assessment.