Graph Transformer Transferability for Flood Scusceptibility mapping in contrasting regions

Graph Transformers have demonstrated considerable promise for spatial prediction tasks, yet their capacity to generalize across contrasting flood regimes remains poorly understood. This study examined whether a GT model trained on the French Riviera (mountainous, ~2,650 km²) could be transferred to two regions with fundamentally different hydrological characteristics: Île-de-France (IDF, flat urban terrain, ~12,000 km²) and Rhône (intermediate topography, ~14,000 km²). Marked distributional divergence was observed across key conditioning features: mean elevation in the Riviera training domain was 321.8 m, compared to 117.1 m in IDF and 440.3 m in Rhône; stream power index differed from 5,703 in the Riviera to 17,058 in IDF and 7,193 in Rhône; and mean channel distance increased from 0.8 km to 3.2 km in IDF. Transfer performance was poor in both target regions, with near-random discrimination in IDF (AUC-ROC: 0.56, sensitivity: 20%) and only marginal improvement in Rhône (AUC-ROC: 0.68, sensitivity: 22%). These results indicate that the model failed to generalize beyond its training domain, primarily because the physical drivers of flooding differ substantially between mountainous Mediterranean and urban pluvial environments. The findings suggest that direct model transfer is not viable across regions with such contrasting hydrological regimes, and that domain-specific training data remain a prerequisite for reliable flood susceptibility assessment.