Probabilistic Flash Flood Forecasting using Stormscale Ensembles
Rainfall error characteristics of the individual members are first diagnosed and quantified in terms of structure, amplitude, and location (SAL; Wernli et al., 2008). Amplitude and structure errors are readily correctable due to their diurnal nature, and the fine scales represented by the CAPS QPF members are consistent with radar-observed rainfall, mainly showing larger errors with afternoon convection. To account for the spatial uncertainty of the QPFs, we use an elliptic smoother, as in Marsh et al. (2012), to produce probabilistic QPFs (PQPFs). The elliptic smoother takes into consideration underdispersion, which is notoriously associated with stormscale ensembles, and thus, is good for targeting the approximate regions that may receive heavy rainfall. However, stormscale details contained in individual members are still needed to yield reasonable flash flood simulations. Therefore, on a case study basis, QPFs from individual members are then run through the hydrological model with their predicted structure and corrected amplitudes, but the locations of individual rainfall elements are perturbed within the PQPF elliptical regions using Monte Carlo sampling. This yields an ensemble of flash flood simulations. These simulated flows are compared to historically-based flow thresholds at each grid point to identify basin scales most susceptible to flash flooding, therefore, deriving PFFF products. This new approach is shown to: 1) identify the specific basin scales within the broader regions that are forecast to be impacted by flash flooding based on cell movement, rainfall intensity, duration, and the basin's susceptibility factors such as initial soil moisture conditions; 2) yield probabilistic information about on the forecast hydrologic response; and 3) improve lead time by using stormscale NWP ensemble forecasts.