Probabilistic urban flood prediction with multivariate assimilation of sewer and surface observations

Aging drainage infrastructure and intensifying rainfall are increasing urban flood risk. However, few data assimilation studies in urban flooding jointly consider rainfall uncertainty, drainage-system degradation, and observation-network configuration within a unified framework. We develop a probabilistic urban flood prediction framework that couples a 1D-2D hydrodynamic model with particle-filter data assimilation (DA), jointly integrating surface inundation depths and sewer water levels. Synthetic experiments conducted for an urban catchment testbed in Osaka, Japan evaluate how observation type, sensor placement, and update frequency influence assimilation performance. Multivariate DA consistently surpassed the open-loop baseline, improving sewer-level prediction and flood-extent mapping and raising spatial skill (CSI) from 0.64 to 0.86 (+34.4%). Among univariate settings, assimilating sewer levels outperformed surface-only DA, while combining both data types delivered the most balanced gains. Observation configuration strongly influenced performance: frequent updates and a greater number of observation points reduced errors, and under sparse updates, downstream sewer sensors provided more stable constraints than upstream ones, while dense networks reduced sensitivity to placement. These results demonstrate the suitability of particle filter-based multivariate DA for capturing nonlinear sewer-surface interactions and guiding sensor placement and update strategies in urban flood prediction. These advances highlight the potential of multivariate DA to support infrastructure resilience and risk management in data-scarce environments.