Flood mapping using EO foundation model with limited data

Flood maps derived from remote sensing data, especially synthetic aperture radar (SAR), are crucial for situational awareness and risk assessment during flood events. In recent years, deep learning models, such as U-Net have been applied successfully to flood mapping based on SAR data. However, these models typically require large amounts of labeled training data. Earth observation (EO) foundation models offer a promising alternative. By pretraining a neural network encoder on large, diverse remote sensing datasets, using self-supervised learning, they enable efficient fine‑tuning of small decoders for specific downstream tasks, potentially requiring only limited amounts of annotated data.

In this study, we evaluate THOR, a pretrained EO foundation model, for flood mapping and compare its performance against a U‑Net baseline with a pretrained ResNet backbone. To assess the dependence on training dataset size, we prepare multiple datasets of varying scales using Sentinel‑1 SAR data and water body masks from Norway. These datasets are used both to train the U‑Net model and to fine‑tune a decoder on top of THOR. The resulting models are tested on an independent dataset of flood events and systematically compared.

We analyze how model performance changes with decreasing dataset size and identify conditions under which the foundation model outperforms the U‑Net baseline. In particular, we investigate the threshold at which THOR becomes advantageous for limited-data scenarios. Finally, we assess whether the performance achieved by the foundation-model-based approach is sufficient for operational flood mapping when only small, labeled datasets are available.