Evaluating Cross-Country Generalization of Deep Learning-Based Flood Mapping Using Sentinel-1 SAR and Sentinel-2 Imagery

Nitya D Dhagat 

Benchmark performance in flood mapping does not reliably predict how a model will behave in a new country. This paper tests that directly. Using the Sen1Floods11 dataset, we evaluate twelve model configurations—two architectures (Vanilla U-Net and ResNet34-encoder U-Net) across three sensing modalities (Sentinel-1 SAR, Sentinel-2 augmented with NDWI and MNDWI spectral indices, and multimodal Fusion)—under a standard random train-test split and a geographically disjoint country-level split where test countries are held out entirely from training. The gap between these two conditions is large and consistent. The worst case is Vanilla U-Net with Sentinel-2: IoU drops from 0.727 under random splitting to 0.464 under country splitting, a 36.2% relative loss. SAR degrades least. ResNet34+Fusion is the strongest country-level configuration, reaching IoU 0.555 and average precision 0.850. NDWI and MNDWI provide the clearest class separation of all input features (mean flood-to-background difference of 2.21 and 2.27 standard deviations respectively), and their inclusion improves country level IoU consistently across architectures. Somalia is harder than Paraguay in every configuration, with SAR-only models showing a 2.4× country ratio that narrows to 1.3× for ResNet34+Fusion. The key finding: the evaluation protocol matters as much as the model design.