On Latin Hypercube sampling for efficient uncertainty estimation of satellite rainfall observations in flood prediction

With the advent of the Global Precipitation Measurement (GPM) in 2009, satellite rainfall measurements are expected to become globally available at space–time scales relevant for flood prediction of un-gauged watersheds. For uncertainty assessment of such retrievals in flood prediction, error models need to be developed that can characterize the satellite's retrieval error structure. A full-scale assessment would require a large number of Monte Carlo (MC) runs of the satellite error model realizations, each passed through a hydrologic model, in order to derive the probability distribution in runoff. However, for slow running hydrologic models this can be computationally expensive and sometimes prohibitive. In this study, Latin Hypercube Sampling (LHS) was implemented in a satellite rainfall error model to explore the degree of computational efficiency that could be achieved with a complex hydrologic model. It was found that the LHS method is particularly suited for storms with moderate rainfall. For assessment of errors in time to peak, peak runoff, and runoff volume no significant computational advantage of LHS over the MC method was observed. However, the LHS was able to produce the 80% and higher confidence limits in runoff simulation with the same degree of reliability as MC, but with almost two orders of magnitude fewer simulations. Results from this study indicate that a LHS constrained sampling scheme has the potential to achieve computational efficiency for hydrologic assessment of satellite rainfall retrievals involving: (1) slow running models (such as distributed hydrologic models and land surface models); (2) large study regions; and (3) long study periods; provided the assessment is confined to analysis of the large error bounds of the runoff distribution.