Evaluation of a physics-based distributed hydrologic model for flood forecasting

A fully distributed, physics-based rainfall–runoff model called r.water.fea is applied within the Distributed Model Inter-comparison Project (DMIP) organized by the US National Weather Service. Simulations are performed for two basins, the Illinois River and Blue River in Oklahoma. The r.water.fea model is an event-based model that derives parameters and is calibrated using geospatial data. Longstanding research on the Blue and Illinois River basins resulted in a calibrated model using eight events. In order to draw statistical comparisons, the number of events was augmented for the purposes of DMIP. Model performance is evaluated for the Blue and Illinois for the initial and augmented set of storm events. An important finding related to the stability of calibrated parameters from the original 8 to 18-event storm series was observed. As more events were added to expand the number of storms, parameter values changed only slightly. Beyond the calibration phase, a verification period was also used to test the validity of the calibrated parameters. Consistent results were found between the calibration and verification period. In fact prediction accuracy was better in some cases during the verification period, which adds to the confidence in the calibrated model predictions and the methodology. Interior points are used to identify internal model consistency and achievable accuracy. At the interior points located at Watts and Savoy, predictions were biased at Savoy but had better R 2 values than obtained at Tahlequah in terms of volume and peak. Watts had comparable bias and nearly identical prediction accuracy compared to Tahlequah. During the verification period for the Blue and Illinois, volume predictions had an accuracy of RMSE=17 and 19 mm. Peak discharge in the two basins was predicted with an accuracy of RMSE=105 and 292 m 3/s, respectively. Closer agreement in volume than peak or timing was found in both watersheds, which may indicate the need for improved channel characteristics and routing. The peak discharge predictions achieved by this model in the Illinois during the verification period are biased towards over-prediction by 16% with an R 2 of 0.716. Peak discharge prediction accuracy in the Blue River during the verification period is biased towards under-prediction by 13% with an R 2 of 0.438. The performance demonstrates that geospatial data may be used to parameterize and calibrate a fully distributed physics-based model, and is capable of makin