Institute of Hydraulic Engineering Department of Hydrology and Geohydrology Prof. Dr. rer. nat. Dr.-Ing. András Bárdossy Pfaffenwaldring 61, 70569 Stuttgart, Germany www.iws.uni-stuttgart.de Addressing several issues like sedimentation, water quality, conservation measures, environmental and geomorphologic studies etc, needs the prediction of erosion patterns which, in turn, needs runoff source areas within the catchment. Several modeling alternatives exist, all with certain potential and limitations. The use of a distributed rainfall- runoff model is basis for identification of such areas. Such model, even in case of physically- based, needs prior calibration of some or many parameters. The optimization and prediction capability of those distributed models is being assessed based on their ability to correctly predict lumped hydrograph at watershed outlet. The presented work aims to show the unreasonable consequences that we have encountered while calibrating and applying a distributed rainfall runoff model. The model used was WaSiM-ETH, a physically based spatially distributed rainfall-runoff model. At first to apply for events in a small agricultural catchment in central Belgium, its 11 parameters were calibrated using Gauss-Marcquardt-Levenberg algorithm. As is the trend, the calibration was done with objective function of minimizing prediction errors in the catchment outlet. Very nice results were obtained with closely matching hydrographs and Nash-Sutcliffe efficiency as high as 0.97 in calibration and 0.81 in validation. But when the modeled runoff source areas within the catchment were investigated, a very much unrealistic patterns were observed with almost all the runoff are coming from a small isolated patch in the catchment. Further we calibrated the model using more accepted Schuffle Complex Evolution (SCE-UA) algorithm and, in addition, sets of equally well performing parameter vector are estimated based on Tukey’s half space depth function. They are applied to a bigger Rems catchment in southern Germany where also we found that very good model performance were not accompanied by the reasonable runoff patterns within the catchment. Very good prediction of a distributed rainfall runoff model but for all wrong reasons Thapa, P.K.; Bárdossy, A. [email protected] Conclusions Well performing parameter sets may lead to good results with high model efficiency but these can be for all the wrong reasons Better hydrographs prediction by models do not guarantee better hydrology representation by them. References Bardossy A. & Singh S. (2008): Robust estimation of hydrological parameters. HESSD (in press) Doherty J. PEST. (2002): Model-Independent Parameter Estimation,Watermark Numerical Computing, Australia, 2002. Duan, Q., Sorooshian, S. and Gupta, V.K. (1994): Optimal use of the SCE-UA global optimization method for calibrating catchment models, J. Hydrol., 158, 265-284. Oost KV, Govers G, Cerdan O, Thaure D, Rompaey AV, Steegena A, Nachtergaele J, Takken I & Poesen J. (2005): Spatially distributed data for erosion model calibration and validation: The Ganspoel and Kinderveld datasets. Catena 61, 105– 121. Schulla J. & Jasper K. (1999, 2006, 2007): Model Description WASIM-ETH. Institut für Atmosphäre und Klima. ETH Zürich. Some Results : Court of Miracles, A scientific workshop, Paris, June 2008 Catchment Area: 111 ha Mean annual precipitation: 740 mm LU: farmland; scarce built-up areas Soil: loess (Haplic Luvisols) Ganspoel catchment (central Belgium) 3.39 o Slope +3.17 o 21.62 o Slope 41.8 o +6.4 o -6.4 o 8.7 o Rems catchment (southern Germany) Catchment Area: 580 sq. km Mean annual precipitation: 900 mm LU: agriculture; forest; built-up areas Soil: light sandy on highs; loamy clay on lows Study Area Model used : Water balance & flow Simulation Model [WaSiM-ETH] Parameters estimation : . Gauss-Marcquardt-Levenberg method [PEST package] . Schuffle complex evolution algorithm [SCE-UA] . Use of statistical depth function [Tukey‘s Half space depth] The model is calibrated for an event in Ganspoel catchment and very well matching of hydrograph is obtained. But the spatial runoff within the catchment producing this hydrograph is completly unrealistic Gauge 1 [Schwäbisch Gmünd] 0 5 10 15 20 25 30 1/1/1993 2/20/1993 4/11/1993 5/31/1993 7/20/1993 9/8/1993 10/28/1993 12/17/1993 time [d] Discharge [m 3 /s] Obs. Sim lin. NS: 0.90 log NS: 0.70 Gauge 2 [Haubersbronn] 0 2 4 6 8 10 12 14 1/1/1993 2/20/1993 4/11/1993 5/31/1993 7/20/1993 9/8/1993 10/28/1993 12/17/1993 time [d] Discharge [m 3 /s] Obs. Sim lin. NS: 0.77 log NS: 0.66 Gauge 3 [Schorndorf] 0 20 40 60 80 100 120 1/1/1993 2/20/1993 4/11/1993 5/31/1993 7/20/1993 9/8/1993 10/28/1993 12/17/1993 time [d] Discharge [m 3 /s] Obs. Sim lin. NS: 0.85 log NS: 0.60 Gauge 4 [Neustadt] 0 20 40 60 80 100 120 140 1/1/1993 2/20/1993 4/11/1993 5/31/1993 7/20/1993 9/8/1993 10/28/1993 12/17/1993 time [d] Discharge [m 3 /s] Obs. Sim lin. NS: 0.85 log NS: 0.77 The model is calibrated for each subcatchment individually in Rems catchment and reasonably matching hydrographs are obtained. But the runoff patterns within the catchment is again unacceptable as they vary highly among subcatchments. SCE UA parameters The model is calibrated for identical parameter set for all subcatchments in Rems catchment using SCE-UA algorithm. 11 different parameter sets performing equally good are also obtained using depth function. All of them produce resonable model performance and at least, uniform runoff patterns within catchment. But the total surface runoff from the catchment is highly varying with different parameter sets. Introduction : Pest parameters SCEUA parameters Parameter set 2