Do Geospatial Foundation Models Improve SAR-Based Flood Mapping?

Accurate and timely flood delineation is a cornerstone of disaster response and hydrological risk management. Synthetic Aperture Radar (SAR) is uniquely suited to this task because it operates independently of cloud cover and illumination, yet its interpretation remains challenging due to speckle, terrain effects, vegetation scattering, and ambiguities between flooded and permanent water as well as shadows and smooth surfaces such as tarmac. While deep learning has substantially advanced SAR-based flood segmentation, most existing models are trained from scratch and often struggle to generalize across regions and flood regimes. Recently, geospatial foundation models (GFMs) pretrained on massive satellite archives have shown promise, but their benefits for SAR-based flood mapping remain insufficiently quantified. This paper presents a controlled, large-scale global scale evaluation and benchmarking of a vision-transformer based GFM (NASA IBM Prithvi) against two task-specific segmentation architectures, the SegFormer (hierarchical transformer) and the commonly used U-Net (convolutional neural network), including lightweight variants, for post-event SAR-based flood mapping. All models were trained and evaluated under a standardized pipeline that explicitly addresses extreme class imbalance via stratified negative sampling and weighted loss functions. Training and validation used the expert-annotated Kuro Siwo dataset (43 flood events, 67,490 Sentinel-1 VV/VH tiles), while generalization is assessed on both the in-distribution Kuro Siwo test set and the out-of-distribution Sen1Floods11 hand labelled benchmark dataset. Results show that stratified negative sampling (controlling how many background-only tiles are shown to the model in each training epoch) increases precision by approximately 6% and mean Intersection-over-Union (mIoU) by about 7% relative to no sampling, while stabilizing training loss dynamics. On the in-distribution data, all architectures reach similar performance (mIoU ≈ 0.82), indicating that well-designed task-specific models remain competitive with GFMs. However, under out-of-distribution conditions, the foundation model Prithvi (mIoU 0.768) closely matches the performance of the SegFormer (mIoU 0.772) and clearly outperforms the U-Net (mIoU 0.712), highlighting the robustness of transformer-based representations when transferring across datasets. Pretraining on optical imagery yields only modest gains for SAR (+3.4% mIoU), suggesting that architectural inductive biases and data handling matter more than cross-modal pretraining. Notably, lightweight GFM variants achieve comparable accuracy with up to 94% fewer parameters, demonstrating strong potential for operational deployment. Scene-level analysis reveals that CNNs suppress scattered false alarms due to the neighborhood contextualization but miss large, continuous floods, while transformers preserve spatial coherence yet overpredict along complex boundaries and scattered surface water ponding, especially near permanent water bodies. Findings demonstrate that while SAR-based flood mapping accuracy requires a combination of appropriate model architectures and class imbalance-aware training, rather than foundation-scale pretraining alone. However, for spatial and statistical transfer to out of distribution datasets, GFMs offer substantial advantages and provide above-average performance for unseen cases, even without localized fine-tuning.