Graph Transformer-Based Flood Susceptibility Assessment of the French Riviera: Implications for Railway Infrastructure Resilience

Floods are one of the deadliest natural hazards worldwide, owing to their frequency and widespread impact. Moreover, flood hazard is escalating continuously due to climate change, exacerbating the risks to human life, infrastructure, and ecosystems globally. In the context of railway systems, flooding causes track submersion, bridge scouring, embankment failures, and damage to electrical equipment, resulting in an estimated annual damage of €581 million in the European Union alone under current climate conditions. The French Riviera region has been particularly susceptible to severe flooding in recent decades. The 2020 Storm Alex, in particular, affected railway infrastructure in the Roya and Vésubie valleys north of Nice, resulting in enormous damage. It took two years of work to restore the infrastructure to operating condition. Despite recurring floods (e.g., the 1994 and 2010 Var floods, the 2015 Côte d'Azur floods, and the 2020 Storm Alex) that significantly affected railway infrastructure in this region, to the best of our knowledge, there are no reported flood susceptibility analyses for this area. Furthermore, machine learning (ML) techniques have been shown to significantly outperform traditional modeling approaches in this domain. This study, therefore, aims to address this critical research gap by considering multiple flooding mechanisms and developing a comprehensive flood susceptibility analysis for the region using ML, and assessing its influence on the railway infrastructure. We leverage a state-of-the-art Graph Transformer to capture complex spatial interactions among topographic, hydrological, environmental, geological, and anthropogenic factors influencing flood susceptibility. In this architecture, nodes represent discrete points within the watershed, encoding multi-dimensional feature vectors such as elevation, slope, precipitation, land use classification, and distance from roads. Edges depict connectivity between the nodes. This model’s key feature self-attention mechanism, prioritizes key node-edge connections, and its capacity to process heterogeneous data types and learn complex patterns enhances prediction accuracy. Apart from floods, this region is also affected by landslides and earthquakes. For instance, a landslide in August 2023 blocked the Fréjus rail tunnel with 3,000 cubic meters of rock, disrupting railway services between France and Italy. Consequently, the developed network for flood susceptibility has been extended to include analyses of landslide and earthquake susceptibility to achieve a holistic understanding. This comprehensive multi-hazard analysis approach provides valuable insights for enhancing infrastructure resilience and informing future planning.