Towards Transformer-Based Semantic Segmentation of Seagrass in the Baltic Sea with High-Resolution Satellite Images

Seagrass meadows are blue carbon hotpots. Mapping is an integral part to manage and understand meadows and their dynamics and thereby assist conservation and restoration efforts. Since traditional monitoring methods often become unviable at large spatial scales, satellite remote sensing has emerged as a supplementary tool. However, this approach is often constraint by the sensors' spatial resolution and required training data. Here, we test a transformer-based approach to segment seagrass and compare it against a ResNet50 and MobileNetV3 on very high-resolution (VHR) Pléiades data (0.5 m spatial resolution). Our cross-validation approach demonstrates high performances of all methods, with transformers and CNN approaches being almost equal in performance (~ 95 %), whereby our small dataset might have have promoted overfitting. Large seagrass areas are well recognised, while very small patches of just a few pixels size are detected less accurately. However, the smallest patches detected by our models are only few meters in size, demonstrating that VHR data allows to resolve significantly more spatial details compared to data of publicly available medium-resolution sensors like Sentinel-2.