Linking statistical, hydrodynamic and machine learning models for assessment of compound floods

Many large population centres are located along estuaries where freshwater flows merge with tidally-driven sea water. In these intertidal zones the river water levels are directly affected by the upstream flow and the downstream coastal conditions. Naturally, such coastal zones can be vulnerable to flood events both from a single driver or several drivers acting in a combination. The compound coastal floods levels may generate extreme impacts even if hazards from individual drivers in isolation would be unlikely. Moreover, the complexity of compound flooding is exacerbated by the presence of interactions (e.g. tide and surge) or dependencies between drivers (e.g. river discharge and surge). To fully understand the multi-driver flood dynamics, the multiple drivers and their impacts need to be assessed in an integrated manner.

In this study the statistical and hydrodynamic models are linked to determine probabilities of multiple-driver flood events and associated risks. Cork City on the south coast of Ireland, frequently subject to complex coastal-fluvial flooding is used as a study case.  The research shows that in Cork Harbour and estuary, the tide-surge interactions have a damping effect on the total water level while dependencies between the surge residual and river flow amplify the risk of flooding. The study also shows that for the most accurate assessment of flood hazard, these phenomena need to be accounted for in the joint probability analysis. From a range of uni- and multivariate scenarios, the multivariate joint exceedance probability AND scenario that includes dependence between multiple drivers represents the most realistic representation of flood probabilities. The outputs from the statistical analysis were used to force the hydrodynamic model of Cork City floodplains. The MNS_Flood model was found to be a robust tool for mapping coastal flood hazards in tidally active river channels. Ultimately, the model results were used to build a machine-learning-based flood forecasting tool. A range of machine learning algorithms were tested to explore relationships between the flood drivers and the resulting spatially variable inundation patterns.

The information derived from the integrated statistical, hydrodynamic and machine learning tools can provide a significant support for short-term early-warning applications as well as for the long-term flood management.