Experimenting with learning-based image orientation approaches for photogrammetric mapping of Posidonia oceanica meadows

Abstract. Posidonia oceanica (L.) Delile (PO) is an endemic seagrass of the Mediterranean Sea, where it grows in the form of dense meadows extending from the surface up to 40 m depth. PO plays a key role in the underwater realm, providing numerous ecosystem services, but it is nowadays endangered by climate change and anthropogenic pressure. Its evolution is therefore monitored following protocols recommended by national environmental agencies. In the literature, optical imaging technologies have been tested for mapping and monitoring PO, although no studies have systematically investigated how the non-static, threadlike, characteristics of PO negatively impact the underwater photogrammetry workflow. To optimize and complement current monitoring practices, the POSEIDON project is currently investigating a multi-resolution, multi-technique geomatic approach. Within POSEIDON, this study focuses on the use of beyond ultra-high resolution (BUHR) underwater photogrammetry and highlights the critical aspects involved in surveying a complex, moving environment such as extended continuous PO meadows. A comparative analysis of traditional algorithms and AI-driven approaches for image orientation is presented on datasets that differ by acquisition protocols, depth, season, platform type, and imaging system. Although some learning-based methods seemed to perform better than hand-crafted ones, we could not identify a winning method. Moreover we verified that, in such a complex scenario, it is crucial to adjust processing thresholds at the different stages of SfM (from matching to bundle adjustment) and take manual intervention measures to improve image orientation.