An integrated statistical – physical modeling approach for multivariate flood risk assessment

Abstract:

Hydrologic models require spatio-temporal weather data as forcing, and maintaining spatio-temporal dependence across these variables is important in stochastic simulations. The dependence across the weather variables can be nonlinear, and each variable may follow a different probability model. The dimension of the problem can be relatively large to be represented through a parametric modeling framework. We introduce a new space-time simulator for multiple variables and spatial locations and demonstrate its application for modeling floods in two different river basins with different climatologies. Given multiple variables, each of which can be simulated from its marginal or conditional distribution, and a historical data set for these variables, the method appropriately preserves multivariate and the temporal dependence across the variables. We use the VIC hydrologic model, which is physically based, to show the necessity of preserving both the spatial and temporal patterns in flood modeling. We demonstrate that the spatial structure is not needed for small basin areas, but as one moves to larger river basins, it becomes increasingly important. This research bridges the divide between stochastic weather simulators and flood models, demonstrating that merging the two approaches can produce more robust estimates of flood distributions and return period estimation.