Natural hazards such as earthquakes, landslides, and floods are increasingly recognized as interconnected processes, with cascading impacts that challenge traditional risk assessment frameworks. The devastating 2023 Emilia Romagna floods in Italy highlighted the urgent need for advanced tools capable of capturing the complex spatial and temporal interactions between multiple hazards. This work is part of the AMHEI project, which aims to revolutionize multi-hazard assessment using artificial intelligence (AI) and Earth Observation (EO) data.

Our approach leverages recent advances in deep learning, particularly graph neural networks and transformer architectures, to model multi-hazard processes as dynamic graphs. Each node represents a spatial unit (e.g., slope, catchment), with edges encoding physical or functional relationships (e.g., hydrological connectivity, proximity to faults). The graph transformer is trained on multi-source EO data, including:

– Sentinel-1 InSAR time series for ground deformation and landslide detection.
– Sentinel-2 optical imagery for land cover and flood mapping.
– Hydrological and meteorological datasets (rainfall, river discharge).
– Ancillary data (topography, lithology, infrastructure, population density).

The model captures both short-term triggers (e.g., extreme rainfall) and long-term predisposing factors (e.g., ground shaking, land use change), enabling the prediction of hazard cascades and their evolution over time.

We present the first results of applying the graph transformer to the Emilia Romagna region, which experienced catastrophic flooding in 2023. The model successfully identified areas of increased landslide susceptibility following the floods, capturing both direct and indirect hazard interactions.

These preliminary results demonstrate the potential of graph transformer architectures for advancing multi-hazard risk assessment. The approach enables a more holistic understanding of hazard dynamics, supporting early warning, emergency response, and long-term resilience planning. Ongoing work includes expanding the model to the whole country of Italy and incorporating additional hazard types (e.g., earthquakes, wildfires).