Machine Learning Flood Risk Assessment at Street Level: A Graph Embedding Approach in the Tamanduateí River Watershed

Urban flooding poses significant risks to urban infrastructure and public safety. Although conventional hydrological models are well-established, they frequently demand extensive computational resources and face challenges in providing real-time, street-level detail. This research presents a model for assessing street-level flood susceptibility within the Tamanduateí River Watershed, São Paulo, employing a Graph Machine Learning methodology. We utilize diverse geospatial data -- Digital Terrain Models and hydrography -- into a road network graph. Structural properties of the street network are captured via Node2Vec embeddings, combined with engineered features (Topographic Wetness Index, Slope, Distance to River). A CatBoost model was trained using spatial cross-validation to prevent data leakage, achieving a ROC AUC of 0.69 on an unseen microbasin.