Towards Multi-Class Segmentation of Seafloor Images with Posidonia Oceanica

Posidonia oceanica meadows are key components of the Mediterranean marine ecosystem, offering crucial ecological services such as carbon sequestration, sediment stabilization, and habitat for diverse marine species. Increasing anthropogenic pressures and climate change threaten their stability, highlighting the need for efficient and scalable monitoring solutions. While traditional diver-based surveys offer limited spacial coverage and consistency, advancements in autonomous robotics and image analysis provide promising alternatives. This study investigates the application of Deep Learning (DL) for multi-class segmentation of underwater images in P. oceanica-dominated environments. Building on prior work focused on binary segmentation, this approach extends the task to four ecologically relevant seabed classes: alive P. oceanica, dead P. oceanica, sand, and rock. The DeepLabv3+ architecture was adopted and trained on a heterogeneous dataset of 806 real underwater images collected by an underwater robot across five distinct marine sites. To address challenges related related to data scarcity and overfitting, the training process incorporated transfer learning, data augmentation, and early stopping techniques. The model's performance was quantitatively evaluated using Intersection over Union as metric, yielding a mean value of 96.21% on the test set. Results suggest the effectiveness of the proposed DL-based approach for accurate and detailed segmentation, supporting its potential for integration into scalable tools for marine habitat monitoring and conservation.