Traditionally, rainfall estimates by a remote-sensing platform such as radar or satellite are evaluated by the use of independent, in-situ instruments such as rain gauges. Several studies have pointed out the error characteristics of these gauges as well as the vastly differing spatial and temporal sampling characteristics as compared to radar. Recent studies have integrated the use of hydrologic models to evaluate the performance of competing rainfall algorithms or quantitative precipitation forecasts by analysis of the hydrologic predictions as compared to streamflow observations. The downfall of such an approach is the introduction of uncertainty caused by the hydrologic model structure such as the governing equations, the model parameters, and the observations of streamflow. The study reported herein proposes a model ensemble approach to evaluating quantitative precipitation algorithms at the scale of application, a watershed.

Five different precipitation algorithms, some of which incorporate rain gauge estimates, are evaluated for nine cases of significant runoff in the Blue River near Blue, OK in the U.S. Hourly rainfall estimates from each of these algorithms are input independently to a distributed-parameter, physics-based hydrologic model. The hydrologic model employs three sensitive parameters, each having defined ranges that have physical relevance. Hundreds of simulations are performed using a given rainfall input with all combinations of model parameterizations, thus remedying the need to recalibrate the model for each input. A probability distribution is then derived that describes the probability that a given input is capable of producing the correct hydrologic prediction. A ranked probability score is then used to summarize the overall performance of each input. It is believed that this ensemble approach can be extended to other hydrologic models and rainfall estimates under development, such as those from polarization diverse radars.