Benchmarking Segmentation Models for Scalable Mangrove Mapping with Sentinel-1/2 Imagery

Mangroves are dynamic coastal ecosystems that contribute to carbon storage, coastal protection and biodiversity conservation, but are vulnerable to climate change. Very-high-resolution imagery such as Pléiades enables detailed mangrove mapping but it is costly and have a low revisit rate. Large-scale monitoring requires free and regularly available data such as Sentinel-1 and Sentinel-2 but their spatial resolution poses challenges for detailed mangrove mapping. In this work, we benchmark deep learning approaches for mangrove mapping from Sentinel-1 radar and Sentinel-2 optical imagery using reference maps derived from very-high-resolution Pléiades imagery. We compare four segmentation configurations based on CROMA and BigEarthNet-pretrained ResNet50 representations with different segmentation decoders. We construct a mangrove mapping dataset covering French Guiana, Suriname, and Amapá. CROMA + UPerNet achieves the best performance, with an F1-score of 0.9218 and an IoU of 0.8549. These results highlight the potential of pretrained multimodal Earth observation representations and open Sentinel data to support mangrove resilience through scalable monitoring.