Stochastic Generation and Hybrid Downscaling for Improving the Estimation of Return Periods of Flooding

Flooding (coastal, pluvial and riverine) is one of the most dangerous natural disasters worldwide. In the case of riverine flooding, impacts mostly derive from the velocity and the depth of water induced by extreme flows, which tend to be forced by extreme precipitation.

Extreme flows are normally characterized by their return periods which are computed through the analysis of river flow time series or by propagating precipitation time series through a hydrological model to generate a synthetic river flow time series that is then analyzed. Flooding maps are computed for every return level of the river flow time series and are assigned that same return period.

These methodologies assume that flooding is a univariate variable and that the variation of one magnitude (aggregated rainfall or extreme river discharge) explains the variation of flooding. However, flooding is controlled by many factors (antecedent moisture condition, spatial and temporal distribution of rainfall, etc.) and thus a more robust estimation is obtained considering flooding as a multivariate problem.

In the present work, flooding maps are constructed comparing three methodologies. The first one analyzing a river flow discharge time series. The second one propagating the recorded rainfall time series through a hydrological model. And the third one making use of stochastic generation to augment the range of observations to include situations that had not been observed in the historic time series but that were likely to have occurred for being similar to other situations that were observed. This synthetic generation allows to account for a larger natural variability, which in turn improves the estimation of flooding maps and flooding return periods.

Synthetic generation is used to create additional situation with the potential to induce flooding. However, this increment in the number of situations to be considered also increases the modeling requirements, and therefore the computational cost of computing flooding maps and return periods. In order to reduce the computational cost of the third method, hybrid downscaling techniques, mostly based on machine learning methods, are applied to reduce the amount of model simulations to be carried, while maintaining the accuracy of the technique.

The third techniques proves to provide more robust and complete solution for the flooding maps, explaining the apparent lack of relation between the computed and observed flooding extremes in some areas.