Abstract
In a context of climate change and the erosion of marine biodiversity, ecological monitoring of the most sensitive marine habitats is of paramount importance. In particular, there is a need for operational methods that enable decision-makers and managers to establish relevant conservation measures and to evaluate their effectiveness. TEMPO and RECOR are two monitoring networks focusing on Posidonia meadows and coralligenous reefs, the two richest and most sensitive habitats in the Mediterranean. The objective of this thesis is to meet the needs of effective monitoring of marine habitats by developing methods for assessing their health, based on two key image analysis methods: convolutional neural networks and photogrammetry. The results show that convolutional neural networks are capable of recognizing the main species of coralligenous assemblages in underwater photographs from RECOR, with a precision similar to that of an expert taxonomist. Furthermore, we have shown that photogrammetry can reproduce a marine habitat in three dimensions with a high degree of accuracy, sufficient for monitoring habitat structure and species distribution at a fine scale. Based on these reconstructions, we have developed a method for automatic mapping of Posidonia meadows, enabling temporal monitoring of the ecological quality of this sensitive habitat. Finally, we characterized the three-dimensional structure of coralligenous reefs based on their photogrammetric reconstructions and studied the links with the structuring of the assemblages that make them up. This PhD work has led to the development of operational methods that are now integrated into the TEMPO and RECOR monitoring networks. Results of this work paves the way for future research, in particular concerning characterization of the biological activity of coralligenous reefs thanks to the coupling of photogrammetry, neural networks and underwater acoustics.