Graph transformer-based flood susceptibility mapping: Application to the French Riviera and railway infrastructure under climate change

Study region The French Riviera watershed in southeastern France (≈2650 km²) extends from steep alpine headwaters to a densely urbanised Mediterranean coast traversed by the SNCF rail network, and has suffered repeated destructive floods, including the 1994 and 2010 Var floods, the 2015 Côte d’Azur flood, and the 2020 Storm Alex. Study focus Traditional machine-learning (ML) approaches achieve high point-wise metrics (AUC-ROC, accuracy) but fail to produce the spatially coherent predictions needed for reliable infrastructure-exposure estimation. We propose a graph transformer (GT) architecture incorporating a feature-similarity graph regularised by Laplacian positional encodings and spatial attention, and benchmark it against six traditional ML models, and show that these point-wise metrics alone are insufficient and should be complemented by spatial autocorrelation metrics (Moran's I, Geary's C) when evaluating multi-class flood predictions. Using topographic, hydrologic, and environmental features, we further project flood susceptibility and railway-track exposure for 2050 under RCP 4.5 and RCP 8.5 scenarios. New hydrological insights for the region The GT achieved AUC-ROC ≈ 0.974 (slightly below XGBoost ≈ 0.990) but substantially improved spatial coherence (Moran’s I ≈ 0.612, Geary’s C ≈ 0.473 versus XGBoost 0.409 and 0.589, and RF 0.474 and 0.523). Its smooth boundary delineation produced markedly different distributions: the GT classified 6.2% of the area as very-high with 35.6% railway exposure, versus 55–57% very-high area and 46–52% railway exposure for ML models; the GT very-high extent closely matches the 172.6 km² historical inundation inventory. Accumulated Local Effects analysis confirmed elevation, stream power, and precipitation as dominant monotonic drivers consistent with flood hydrology. Scenario-conditional projections for 2050 under RCP 4.5 (Q95) and RCP 8.5 (Q95) indicate very-high areas could expand to ≈ 28% and ≈ 36.3%, with railway exposure reaching ≈ 64% and ≈ 67.4%. The GT framework provides a replicable blueprint for regions seeking spatially accurate flood susceptibility mapping.