A transferable deep learning framework for flood mapping: Spatial generalization across hydro-climatic regimes using satellite imagery

• Two U-Net models trained on Sentinel-1 and Sentinel-2 imagery were developed for transferable water segmentation. • Hydro-climatic similarity analyses were introduced to evaluate representativeness and guide model transferability. • Structural biases were systematically profiled to inform model design. • A Leave-One-Location-Out cross-validation was implemented to measure true spatial generalization across 18 flood events. • Sensor-specific patterns revealed complementary strengths of radar and optical data for operational flood mapping. Flood hazard assessment has traditionally relied on hydraulic models that require extensive boundary conditions, detailed topographic data, and substantial computational resources. Their reliability, however, is often undermined by the scarcity of in-situ observations during extreme events, which increases uncertainty in flood extent delineation. Remote sensing has therefore emerged as a valuable alternative for flood mapping. Yet, most recent deep learning studies have focused on maximizing in-sample accuracy while providing limited evaluation of spatial transferability and operational readiness when applied outside the training domain. To address these constraints and move beyond the usual focus on in-sample accuracy in flood mapping, this study introduces an operationally oriented deep learning framework designed for spatial transferability. The approach combines two modality-specific U-Net models, one for Sentinel-1 (SAR) and one for Sentinel-2 (optical), trained on a globally sourced dataset (C2S-MS) of flood events. Beyond conventional model training, we incorporate key steps to support generalization: (i) a hydro-climatic and geomorphological characterization of the training dataset, combined with a multivariate similarity analysis to assess the environmental representativeness of the target region, (ii) an assessment of structural biases to orient model design, and (iii) a deployment-driven validation strategy to rigorously evaluate spatial transferability under operational constraints. The S1 model achieved an explanatory performance (within the training dataset) of 85 % IoU and 92 % F1-score, while the S2 model reached 89 % and 94 %, respectively. Predictive performance, evaluated on an independent flood event in the target region, remained consistent with these results, with only a limited degradation observed (−2 % in IoU and − 1 % in F1-score). These findings provide empirical support for the role of environmental representativeness in supporting spatial transferability. By unifying environmental representativeness assessment, modality-specific deep learning architectures, and rigorous spatial validation, this work advances a transferable and operationally oriented framework for satellite-based flood mapping. More broadly, the results highlight the importance of aligning training data coverage, validation strategy, and deployment context to achieve reliable generalization under domain shift. Beyond this specific application, the proposed methodology provides insights into model design and evaluation for Earth observation studies confronted with spatial extrapolation.