Spatiotemporal Uncertainties in Flood Predictability in the Missouri River Basin

The Generalized Extreme Value theory (GEV), as well as Peak Over the Threshold (POT) are used to estimate the probability of occurrence of historical hydrometeorologic and Climate Extreme Events (HCEE). Spatial distribution and temporal variability of precipitation evidence the dynamix and complexity of the interdependence between the climate and Water Systems, which influences predictability of HCEE and their associated uncertainties. We hypothesize that uncertainty in such forecasts can be reduced by adopting techniques such as neural networks (NN) and boostraping (BS) on GEV and POT. Our objectives are (1) Develop comprehensive guidelines to obtain the likelihood of occurrence of flooding events on a monthly bases; (b) implement and test the application of GEV and POT's NN and BS approaches; (3) Characterize the uncertainty associated in the forecast of a historical flood event. We will use the sub-continental 1/16th degree and daily precipitation (minimum and maximum temperatures, and wind speed) dataset that spans from 1950 to 2013. Our forecast window is up to 5 days and will consist of a mathematical model that targets the predictability of flooding events based on a 64-year dataset (1950-2013) and an enhanced dataset iteratively generated through the use of NN and BS applied into GEV and POT. Both datasets' precipitation will be defined by a GEV distribution and defined by pareto principles, respectively. NN and BS sample size will be obtained iteratively by matching the values obtained from the empirical and theoretical models. We will test our approach by forecasting historical events such as the July 1993 flooding event in the Northern High Plains. Results from the GEV model are encouraging and reflect the role of precipitation surplus or deficit on the spatial distribution of returning periods and the number of interactions needed that make empirical and theoretical models converge.