Earth Foundation Models outperform traditional deep learning architectures for coastal ecosystem semantic segmentation and large-scale habitat mapping

Coastal ecosystems face escalating threats from climate-driven sea-level rise and anthropogenic modification, necessitating precise, scalable monitoring for effective conservation and blue-carbon accounting. Traditional deep learning models (e.g., UNet and DeepLabV3) struggle with boundary delineation and spectral confusion in heterogeneous coastal transition zones. Earth Foundation Models (EFMs) pre-trained on massive Earth observation archives encode transferable spectral–spatial–temporal representations that address these challenges through dynamic channel adaptation, multi-scale feature extraction, and long-range dependency capture. This study establishes a benchmark of fine-tuned EFMs against traditional architectures for coastal semantic segmentation, using annual Sentinel-2 composites across 85 patches from 17 nations and evaluating six metrics including boundary IoU and edge F1-score across five spectrally ambiguous classes. DOFA achieved the highest performance (0.98 mean accuracy, 0.95 mIoU, 0.46 Bd-IoU, 0.62 Edge-F1), outperforming the best traditional model by 4.4% absolute mIoU. Ablation studies show encoder-only fine-tuning preserves 98.9% of end-to-end performance (0.94 vs. 0.95 mIoU), whereas decoder-only fine-tuning reduces trainable parameters by 95.5% but catastrophically degrades accuracy to 0.83 mIoU. Computational analysis confirms operational feasibility at regional scales. These findings establish EFMs as transformative, actionable tools for high-fidelity coastal conservation, ecological restoration, and blue-carbon inventorying.