Résumé
Tropical coral reefs are the most biodiverse ecosystems on the planet. Unfortunately, increasing direct (e.g., overfishing, change in land use) and indirect (e.g., ocean acidification, sea-level rise, global warming, storms) human pressures are causing rapid declines in marine biodiversity globally, calling for more careful monitoring of sensitive ecosystems. If environmental baseline data like marine habitat maps are increasingly required to manage coastal environments, they are often lacking or are only available at a resolution that do not meet the needs of management. Such deficiencies hamper the implementation of effective management strategies and monitoring efforts.Here, we developed a method for conducting very high-resolution (sub-meter) mapping of benthic habitats semi-automatically across the 195km of Mayotte’s reef flat, a French island of the southwestern Indian Ocean. The mapping was performed using 50-cm resolution Pleiades satellite imagery from 2021-2023 and LiDAR-derived bathymetry. Satellite images were supplemented by underwater images acquired along 185 transects using a GoPro camera fixed to a diver propulsion vehicle. To achieve a very high spatial accuracy, underwater images were positioned with a few centimeters accuracy using a Global Navigation Satellite System (GNSS) system and post-processed kinematic (PPK) corrections. Geo-referenced image frames were extracted from the videos and integrated for annotations into CoralNet, a free online benthic image analysis software using semi-automatic deep learning analysis. About 10% of the video frames (3045 images) were annotated, allowing the analysis of over 27,000 image and calculate habitats’ covers on the reef. Satellite images were segmented using an Object-Based Image Analysis (OBIA) approach (i.e., the "Large-Scale Mean-Shift" (LSMS) segmentation method) and segments properties were used with other variables in a Random Forest pixel-based classification. To ensure replicability, data processing and analyses were conducted using free and open source software Orfeo ToolBox, QGIS, and R Stats.Thirteen benthic habitats, extending from coastal mangroves to the outer reef slope, were mapped at the pixel level. The LSMS segmentation method applied to Pléiades images created54,706 segments. CoralNet classifiers achieved 65% accuracy, while the Random Forest models reached an accuracy of 76%. The composition of benthic assemblages varied significantly aroundthe island, with changes in live corals accompanied by a rise in muddy habitats, turf, and dead corals as anthropogenic pressures intensify.The detailed habitat maps produced through this project will serve as an essential management tool to inform the management of Mayotte’s coastal waters. In December 2024, Mayotte experienced the very destructive Chido cyclone that strongly impacted the island, its population, and its terrestrial and marine ecosystems. By using satellite images acquired just before the cyclonic event, maps can serve as reference data, and help assess the cyclone's impact on the island's marine ecosystems.