Monitoring shallow-water marine ecosystems requires precise, low-cost and reproducible methods, yet traditional diver-based surveys are labor-intensive, subject to observer bias and limited in spatial coverage. In this study, we present an autonomous framework that integrates deep learning techniques with open-source data acquisition platforms to overcome these challenges. Our approach uses a transformer-based model to classify different coral morphotypes, seagrass species and benthic habitats. This model is trained on the publicly available Seatizen Atlas image dataset . Data were collected using low-cost autonomous surface vehicles (ASVs) equipped with action cameras, differential GNSS receivers, single-beam echosounders and inertial measurement units. The ASVs provide precise geolocation, depth and attitude measurements, synchronized with image timestamps and embedded as standardized metadata, ensuring interoperability and reproducibility. Beyond image classification, the ASVs enable bathymetric surveys and photogrammetric reconstruction for three-dimensional mapping of surveyed areas. Field validation at Réunion Island and Aldabra Atoll demonstrates that this framework delivers cost-effective, non-invasive and standardized monitoring, while automating laborious annotation tasks and increasing survey scalability. By combining open-source datasets, AI models and versatile ASV functionalities, this methodology supports reproducible ecological assessments and informs conservation decisions.
Contini et al. (2026) studied this question.