Multi-Temporal Drone Mapping of Coastal Ecosystems in Restoration: Seagrass, Salt Marsh, and Dune

James, D.; Collin, A.; Bouet, A.; Perette, M.; Dimeglio, T.; Hervouet, G.; Durozier, T.; Duthion, G., and Lebas, J-F., 2024. Multi-temporal drone mapping of coastal ecosystems in restoration: Seagrass, salt marsh, and dune. In: Phillips, M.R.; Al-Naemi, S., and Duarte, C.M. (eds.), Coastlines under Global Change: Proceedings from the International Coastal Symposium (ICS) 2024 (Doha, Qatar). Journal of Coastal Research, Special Issue No. 113, pp. 524-528. Charlotte (North Carolina), ISSN 0749-0208. Ecotones at the land-sea interface provide valuable geo-ecological benefits, providing services such as biodiversity support, coastal protection, and environmental amenities. In temperate sandy-muddy regions, blue carbon ecosystems (seagrass meadows, salt marshes) store more carbon than temperate or tropical forests. Temperate sandy areas, with dunes and vegetation, trap sediment and significantly limit coastal erosion. Recognizing their importance in the face of global change, these ecosystems are being protected and restored. In Europe, coastal development has fragmented these ecosystems into smaller patches, making them difficult to map by satellite or aircraft. Drone imagery is now revealing their spatial patterns and processes, with the potential for ultra-high resolution (UHR) pixel size, although it's high temporal resolution remains a niche in research. The study addresses this issue by generating time series of UHR classifications for three ecosystems under restoration: seagrass, salt marsh, and dune. Three coastal sites in northern Brittany were selected: a subtidal seagrass meadow (for anchoring and fishing), a restricted salt marsh (an old polder reconnected to the sea), and a dune dealing with coastal infrastructure issues (road). Aerial drones with red-green-blue sensors flew over these sites annually: since 2021 for the seagrass (8 times), since 2020 for the salt marsh (8 times), and since 2022 for the dune (5 times). Training and validation data were located on identical areas on overlaid orthomosaics, and two pixel-oriented supervised classification algorithms were tested: probabilistic and fast Maximum Likelihood (ML), and decision and efficient Random Forest (RF). Overall accuracies from confusion matrices were compared for each algorithm. For the seagrass meadow, results ranged from 78.33% to 87.18% (ML) and from 86.06% to 93.10% (RF); for the salt marsh, variations were between 71.61% to 95.32% (ML) and 88.69% to 98.59% (RF); for the sand dune, results ranged from 90.04% to 95.5% (ML) and from 93.8% to 96.06% (RF). Regardless of the geo-ecosystem, the RF algorithm consistently outperformed the ML algorithm in mapping.