Mapping and monitoring seagrass meadows in Seychelles using Sentinel-2 imagery: Model training at Anse-aux-Pins and transferability to nearby lagoons

Details

Supervisors

Hanert, Emmanuel ; Culot, Alexis

Faculty

Faculté des bioingénieurs

Degree label

Master [120] : bioingénieur en chimie et bioindustries, à finalité spécialisée

Abstract

Seagrass meadows provide essential ecological functions, including habitat provision, sediment stabilization, carbon storage and coastal protection. In Seychelles, recent national mapping has improved knowledge of seagrass distribution, but temporal dynamics remain less documented. This thesis aimed to develop a reproducible Sentinel-2 workflow to map and monitor seagrass extent at Anse-aux-Pins, Mahé and assess its application to other reef-lagoon systems. The workflow combined Sentinel-2 preprocessing, ACOLITE correction, masking, feature extraction, classification and temporal aggregation. Several atmospheric correction, classifier and thresholding options were compared. The selected pipeline combined DSF correction, XGBoost classification and a Macro-F1 optimized threshold. It gave the best compromise between classification performance, temporal robustness and interpretability, reaching a mean Macro-F1 of 0.808 ± 0.048. Applied to the 2016-2026 archive, the workflow showed that mapped seagrass extent at Anse-aux-Pins was dynamic in space and time. The meadow did not follow a linear trend, but showed repeated phases of expansion and contraction. A clear seasonal cycle was detected, with mapped extent highest around February and lowest around September-October. Spatial outputs distinguished a persistent core, a seasonal envelope and an apparent recovery zone in the central-southern lagoon. The workflow was also applied to Sainte Anne, Côte d’Or, Grande Anse and Baie Sainte Anne. The method was technically transferable, but classifier generalization varied between sites, reflecting differences in lagoon morphology, substrate brightness, water-column conditions, tide levels and seagrass species. Overall, Sentinel-2 can give coherent indicators of mapped seagrass extent in Seychelles shallow lagoons. Broader use would mainly require multi-site and multi-season field data to improve calibration, validation and ecological meaning.