Investigating error propagation in flood prediction based on remotely-sensed rainfall

Date of Completion

January 2004


Engineering, Environmental|Remote Sensing




Quantifying hydrologic prediction uncertainty based on remotely-sensed rainfall requires understanding the error interaction between hydrologic modeling and rainfall estimation. This study attempts to address this issue. A Remote Sensing Rainfall Error Model (RSREM) is developed to characterize the error structure in rainfall. The Generalized Likelihood Uncertainty Estimation (GLUE) framework (Beven and Binley, 1992) is implemented for quantification of hydrologic modeling uncertainty. Application of the GLUE framework on radar observations from a mountainous basin demonstrated that estimates adjusted for mean-field bias and vertical reflectivity profile effects yield runoff error propagation similar to that by a dense rain gauge network. A probabilistic discharge prediction scheme developed on the basis of the GLUE framework yielded 50% less variability in prediction error of time to peak than the conventional optimum parameter set approach. In terms of satellite Passive Microwave (PM) rain estimation, the 3-hourly sampling, as anticipated during the Global Precipitation Measurement ( GPM) mission after 2009, is found to be adequate for flood prediction of weeklong events. It is further shown that the current PM sampling is associated with almost twice the flood prediction uncertainty expected from the planned 3-hourly GPM sampling. ^ The GLUE and RSREM error frameworks are subsequently applied on a land surface model to study the dependency of soil moisture prediction accuracy on satellite rainfall estimation, modeling uncertainties and site characteristics. Satellite rainfall estimation uncertainty under-represented to a considerable degree the total prediction uncertainty. However this under-representation is observed to be influenced by the level of model accuracy. The study finally investigated techniques for accelerating random sampling used in error propagation studies. The Latin Hypercube Sampling technique applied to sample rainfall error structure yielded upper confidence limits (>80%) in runoff simulation consistent with the unconstrained Monte Carlo method, but with two orders fewer computational burden. A stochastic response surface method developed for accelerating sampling of model parameters within the GLUE framework is shown to reduce the computational burden (up to 70%) of the uniform Monte Carlo sampling. ^