Underwater Inspection Platform for Vision-Based Biodiversity Identification

Marine biodiversity monitoring is crucial for understanding and mitigating the impacts of environmental change, overfishing, and habitat degradation. In this work, we present an autonomous, fish-inspired robotic platform designed for in situ identification of marine species. The system allows for realtime perception and classification without reliance on external computation. A lightweight convolutional neural network (CNN) was trained on a custom dataset comprising over $\mathbf{2, 0 0 0}$ annotated images of marine organisms, including dolphins, belugas, jellyfish, sharks, and rays, collected from public datasets, fieldwork, and major aquariums worldwide. The combination of agile underwater navigation and embedded neural processing offers a scalable and non-invasive solution for marine biodiversity inspection. The synergy between robotic platforms and deep learning represents a fundamental advancement toward autonomous, in situ ecological monitoring in dynamic and unstructured aquatic habitats.