• Abdulla, F. A., and Lettenmaier D. P. , 1997: Development of regional parameter estimation equations for a macroscale hydrologic model. J. Hydrol., 197, 237257.

    • Search Google Scholar
    • Export Citation
  • Ajami, N. K., Gupta H. V. , Wagener T. , and Sorooshian S. , 2004: Calibration of a semi-distributed hydrological model for streamflow estimation along a river system. J. Hydrol., 298, 112135.

    • Search Google Scholar
    • Export Citation
  • Asante, K. O., Dezanove R. M. , Artan G. , Lietzow R. , and Verdin J. , 2007: Developing a flood monitoring system from remotely sensed data for the Limpopo Basin. IEEE Trans. Geosci. Remote Sens., 45, 17091714.

    • Search Google Scholar
    • Export Citation
  • Balsamo, G., Viterbo P. , Beljaars A. , den Hurk B. V. , Hirchi M. , Betts A. , and Scipal K. , 2009: A revised hydrology for the ECMWF model: Verification from field site to terrestrial water storage and impact in the Integrated Forecast System. J. Hydrometeor., 10, 623643.

    • Search Google Scholar
    • Export Citation
  • Balsamo, G., Pappenberger F. , Dutra E. , Viterbo P. , and van den Hurk B. , 2011: A revised land hydrology in the ECMWF model: A step towards daily water flux prediction in a fully-closed water cycle. Hydrol. Processes, 25, 10461054, doi:10.1002/hyp.7808.

    • Search Google Scholar
    • Export Citation
  • Bartholmes, J. C., Thielen J. , Ramos M.-H. , and Gentilini S. , 2009: The European Flood Alert System EFAS – Part 2: Statistical skill assessment of probabilistic and deterministic operational forecasts. Hydrol. Earth Syst. Sci., 13, 141153.

    • Search Google Scholar
    • Export Citation
  • Bogner, K., Pappenberger F. , Thielen J. , and De Roo A. , 2010: Wavelet based error correction and predictive uncertainty of a hydrological forecasting system. Geophysical Research Abstracts, Vol. 12, Abstract EGU2010-10245. [Available online at http://meetingorganizer.copernicus.org/EGU2010/EGU2010-10245.pdf.]

    • Search Google Scholar
    • Export Citation
  • Buizza, R., 2010: Horizontal resolution impact on short- and long-range forecast error. Quart. J. Roy. Meteor. Soc., 136, 10201035.

  • Buizza, R., Richardson D. S. , and Palmer T. N. , 2003: Benefits of increased resolution in the ECMWF ensemble system and comparison with poor-man’s ensembles. Quart. J. Roy. Meteor. Soc., 129, 12691288.

    • Search Google Scholar
    • Export Citation
  • Buizza, R., Bidlot J.-R. , Wedi N. , Fuentes M. , Hamrud M. , Holt G. , and Vitart F. , 2007: The new ECMWF VAREPS (Variable Resolution Ensemble Prediction System). Quart. J. Roy. Meteor. Soc., 133, 681695.

    • Search Google Scholar
    • Export Citation
  • Buizza, R., Leutbecher M. , and Isaksen L. , 2008: Potential use of an ensemble of analyses in the ECMWF Ensemble Prediction System. Quart. J. Roy. Meteor. Soc., 134, 20512066.

    • Search Google Scholar
    • Export Citation
  • Clark, M. P., and Hay L. E. , 2004: Use of medium-range numerical weather prediction model output to produce forecasts of streamflow. J. Hydrometeor., 5, 1532.

    • Search Google Scholar
    • Export Citation
  • Clark, M. P., Gangopadhyay S. , Hay L. , Rajagopalan B. , and Wilby R. , 2004: The Schaake shuffle: A method for reconstructing space–time variability in forecasted precipitation and temperature fields. J. Hydrometeor., 5, 243262.

    • Search Google Scholar
    • Export Citation
  • Cloke, H. L., and Pappenberger F. , 2009: Operational flood forecasting: Review of ensemble techniques. J. Hydrol., 375, 613626.

  • Demargne, J., Brown J. , Liu Y. , Seo D.-J. , Wu L. , Toth Z. , and Zhu Y. , 2010: Diagnostic verification of hydrometeorological and hydrologic ensembles. Atmos. Sci. Lett., 11, 114122, doi:10.1002/asl.261.

    • Search Google Scholar
    • Export Citation
  • De Roo, A. P. J., and Coauthors, 2003: Development of a European flood forecasting system. Int. J. River Basin Manage., 1, 4959.

  • Feyen, L., Kalas M. , and Vrugt J. , 2008: Semi-distributed parameter optimization and uncertainty assessment for large-scale streamflow simulation using global optimization. Hydrol. Sci. J., 53, 293308, doi:10.1623/hysj.53.2.293.

    • Search Google Scholar
    • Export Citation
  • Funk, C., Michaelsen J. , Verdin J. , Artan G. , Husak G. , Senay G. , Gadain H. , and Magadzire T. , 2003: The collaborative historical African rainfall model: Description and evaluation. Int. J. Climatol., 23, 4766, doi:10.1002/joc.866.

    • Search Google Scholar
    • Export Citation
  • Ghelli, A., and Primo C. , 2009: On the use of the extreme dependency score to investigate the performance of an NWP model for rare events. Meteor. Appl., 16, 537544, doi:10.1002/met.153.

    • Search Google Scholar
    • Export Citation
  • Gupta, H. V., Kling H. , Yilmaz K. K. , and Martinez G. F. , 2009: Decomposition of the mean squared error and NSE performance criteria for improving hydrological modeling. J. Hydrol., 377, 8091, doi:10.1016/j.jhydrol.2009.08.003.

    • Search Google Scholar
    • Export Citation
  • Hamill, T. M., and Colucci S. J. , 1997: Verification of Eta–RSM short-range ensemble forecasts. Mon. Wea. Rev., 125, 13121327.

  • Hamill, T. M., and Juras J. , 2006: Measuring forecast skill: Is it real skill or is it the varying climatology? Quart. J. Roy. Meteor. Soc., 132, 29052923.

    • Search Google Scholar
    • Export Citation
  • Hamill, T. M., and Whitaker J. S. , 2006: Probabilistic quantitative precipitation forecasts based on reforecast analogs: Theory and application. Mon. Wea. Rev., 134, 32093229.

    • Search Google Scholar
    • Export Citation
  • Hamill, T. M., Whitaker J. S. , and Mullen S. L. , 2006: Reforecasts: An important dataset for improving weather predictions. Bull. Amer. Meteor. Soc., 87, 3346.

    • Search Google Scholar
    • Export Citation
  • Hersbach, H., 2000: Decomposition of the continuous ranked probability score for ensemble prediction systems. Wea. Forecasting, 15, 559570.

    • Search Google Scholar
    • Export Citation
  • Hong, Y., Adler R. F. , Hossain F. , Curtis S. , and Huffman G. J. , 2007: A first approach to global runoff simulation using satellite rainfall estimation. Water Resour. Res., 43, W08502, doi:10.1029/2006WR005739.

    • Search Google Scholar
    • Export Citation
  • Hopson, T. M., and Webster P. J. , 2010: A 1–10-day ensemble forecasting scheme for the major river basins of Bangladesh: Forecasting severe floods of 2003–07. J. Hydrometeor., 11, 618641.

    • Search Google Scholar
    • Export Citation
  • Hossain, F., and Katiyar N. , 2006: Improving flood forecasting in international river basins. Eos, Trans. Amer. Geophys. Union, 87, doi:10.1029/2006EO050001.

    • Search Google Scholar
    • Export Citation
  • Huffman, G. J., and Coauthors, 2007: The TRMM Multisatellite Precipitation Analysis (TMPA): Quasi-global, multiyear, combined-sensor precipitation estimates at fine scales. J. Hydrometeor., 8, 3855.

    • Search Google Scholar
    • Export Citation
  • Joyce, R. J., Janowiak J. E. , Arkin P. A. , and Xie P. , 2004: CMORPH: A method that produces global precipitation estimates from passive microwave and infrared data at high spatial and temporal resolution. J. Hydrometeor., 5, 487503.

    • Search Google Scholar
    • Export Citation
  • Liang, X., Lettenmaier D. P. , Wood E. F. , and Burges S. J. , 1994: A simple hydrologically based model of land surface water and energy fluxes for general circulation models. J. Geophys. Res., 99, 14 41514 428.

    • Search Google Scholar
    • Export Citation
  • Lohmann, D., Nolte-Holube R. , and Raschke E. , 1996: A large-scale horizontal routing model to be coupled to land surface parametrization schemes. Tellus, 48A, 708721.

    • Search Google Scholar
    • Export Citation
  • Lohmann, D., Raschke E. , Nijssen B. , and Lettenmaier D. P. , 1998: Regional scale hydrology: I. Formulation of the VIC-2L model coupled to a routing model. Hydrol. Sci. J., 43, 131141.

    • Search Google Scholar
    • Export Citation
  • Maurer, E. P., Wood A. W. , Adam J. C. , Lettenmaier D. P. , and Nijssen B. , 2002: A long-term hydrologically based dataset of land surface fluxes and states for the conterminous United States. J. Climate, 15, 32373251.

    • Search Google Scholar
    • Export Citation
  • Nash, J. E., and Sutcliffe J. V. , 1970: River flow forecasting through conceptual models. Part I—A discussion of principles. J. Hydrol., 10, 282290.

    • Search Google Scholar
    • Export Citation
  • Neil, J., Schumann G. , Bates P. D. , Buytaert W. , Matgen P. , and Pappenberger F. , 2009: A data assimilation approach to discharge estimation from space. Hydrol. Processes, 23, 36413649.

    • Search Google Scholar
    • Export Citation
  • O’Donnell, G. M., Czajkowski K. P. , Dubayah R. O. , and Lettenmaier D. P. , 2000: Macroscale hydrological modeling using remotely sensed inputs: Application to the Ohio River basin. J. Geophys. Res., 105 (D10), 12 49912 516.

    • Search Google Scholar
    • Export Citation
  • Palmer, T. N., and Coauthors, 2007: The ECMWF Ensemble Prediction System: Recent and on-going developments. ECMWF Research Dept. Tech. Memo. 540, 53 pp.

    • Search Google Scholar
    • Export Citation
  • Palmer, T. N., Buizza R. , Doblas-Reyes F. , Jung T. , Leutbecher M. , Shutts G. J. , Steinheimer M. , and Weisheimer A. , 2009: Stochastic parameterization and model uncertainty. ECMWF Research Dept. Tech. Memo. 598, 42 pp.

    • Search Google Scholar
    • Export Citation
  • Pappenberger, F., and Beven K. J. , 2006: Ignorance is bliss: Or seven reasons not to use uncertainty analysis. Water Resour. Res., 42, W05302, doi:10.1029/2005WR004820.

    • Search Google Scholar
    • Export Citation
  • Pappenberger, F., and Buizza R. , 2009: The skill of ECMWF predictions for hydrological modelling. Wea. Forecasting, 24, 749766.

  • Pappenberger, F., Beven K. J. , Hunter N. , Gouweleeuw B. , Bates P. , de Roo A. , and Thielen J. , 2005: Cascading model uncertainty from medium range weather forecasts (10 days) through a rainfall-runoff model to flood inundation predictions within the European Flood Forecasting System (EFFS). Hydrol. Earth Syst. Sci., 9, 381393.

    • Search Google Scholar
    • Export Citation
  • Pappenberger, F., Harvey H. , Beven K. , Hall J. , and Meadowcroft I. , 2006: Decision tree for choosing an uncertainty analysis methodology: A wiki experiment. Hydrol. Processes, 20, 37933798.

    • Search Google Scholar
    • Export Citation
  • Pappenberger, F., Buizza R. , Bodis K. , and Ghelli A. , 2009a: The skill of probabilistic forecasts under observational uncertainties within the generalized likelihood uncertainty estimation framework for hydrological applications. J. Hydrometeor., 10, 807819.

    • Search Google Scholar
    • Export Citation
  • Pappenberger, F., Cloke H. L. , Balsamo G. , Ngo-Duc T. , and Oki T. , 2009b: Global runoff routing with the hydrological component of the ECMWF NWP system. Int. J. Climatol., 30, 21552174, doi:10.1002/joc.2028.

    • Search Google Scholar
    • Export Citation
  • Schaake, J., and Coauthors, 2007: Precipitation and temperature ensemble forecasts from single-value forecasts. Hydrol. Earth Syst. Sci. Discuss., 4, 655717.

    • Search Google Scholar
    • Export Citation
  • Schumann, G., Bates P. D. , Horritt M. S. , Matgen P. , and Pappenberger F. , 2009: Progress in integration of remote sensing–derived flood extent and stage data and hydraulic models. Rev. Geophys., 47, RG4001, doi:10.1029/2008RG00027.

    • Search Google Scholar
    • Export Citation
  • Seo, D.-J., Herr H. , and Schaake J. C. , 2006: A statistical post-processor for accounting of hydrologic uncertainty in short-range ensemble streamflow prediction. Hydrol. Earth Syst. Sci. Discuss., 3, 19872035.

    • Search Google Scholar
    • Export Citation
  • Shepard, D. S., 1984: Computer mapping: The SYMAP interpolation algorithm. Spatial Statistics and Models, G. L. Gaile and C. J. Willmott, Eds., D. Reidel, 133–145.

    • Search Google Scholar
    • Export Citation
  • Shi, X., Wood A. W. , and Lettenmaier D. P. , 2008: How essential is hydrologic model calibration to seasonal streamflow forecasting? J. Hydrometeor., 9, 13501363.

    • Search Google Scholar
    • Export Citation
  • Su, F., Hong Y. , and Lettenmaier D. P. , 2008: Evaluation of TRMM Multi-satellite Precipitation Analysis (TMPA) and its utility in hydrologic prediction in La Plata basin. J. Hydrometeor., 9, 622640.

    • Search Google Scholar
    • Export Citation
  • Talagrand, O., and Vautard R. , 1997: Evaluation of probabilistic prediction systems. Proc. ECMWF Workshop on Predictability, Reading, United Kingdom, ECMWF, 1–25.

    • Search Google Scholar
    • Export Citation
  • Thielen, J., Bartholmes J. , Ramos M.-H. , and de Roo A. , 2009: The European Flood Alert System – Part 1: Concept and development. Hydrol. Earth Syst. Sci., 13, 125140.

    • Search Google Scholar
    • Export Citation
  • Thiemig, V., Pappenberger F. , Thielen J. , Gadain H. , de Roo A. , Bodis K. , Del Medico M. , and Muthusi F. , 2010: Ensemble flood forecasting in Africa: A feasibility study in the Juba-Shabelle River basin. Atmos. Sci. Lett., 11, 123131, doi:10.1002/asl.266.

    • Search Google Scholar
    • Export Citation
  • Uppala, S. M., and Coauthors, 2005: The ERA-40 Re-Analysis. Quart. J. Roy. Meteor. Soc., 131, 29613012.

  • Van der Knijff, J. M., Younis J. , and de Roo A. P. J. , 2008: LISFLOOD: A GIS-based distributed model for river basin scale water balance and flood simulation. Int. J. Geogr. Inf. Sci., 24, 189212, doi:10.1080/13658810802549154.

    • Search Google Scholar
    • Export Citation
  • Vitart, F., and Coauthors, 2008: The new VAREPS-monthly forecasting system: A first step towards seamless prediction. Quart. J. Roy. Meteor. Soc., 134, 17891799.

    • Search Google Scholar
    • Export Citation
  • Voisin, N., Schaake J. C. , and Lettenmaier D. P. , 2010: Adaptation of calibration and downscaling methods for quantitative ensemble precipitation forecasts. Wea. Forecasting, 25, 16031627.

    • Search Google Scholar
    • Export Citation
  • Westrick, K. J., Storck P. , and Mass C. F. , 2002: Description and evaluation of a hydrometeorological forecast system for mountainous watersheds. Wea. Forecasting, 17, 250262.

    • Search Google Scholar
    • Export Citation
  • White, D., Johnston K. , and Miller M. , 2005: The Ohio River basin. Rivers of North America, A. C. Benke and C. E. Cushing, Eds., Elsevier, 375–412.

    • Search Google Scholar
    • Export Citation
  • Wilks, D. S., 2006: Statistical Methods in the Atmospheric Sciences. 2nd ed. Academic Press, 627 pp.

  • Yadav, M., Wagener T. , and Gupta H. , 2007: Regionalization of constraints on expected watershed response behavior for improved predictions in ungauged basins. Adv. Water Resour., 30, 17561774, doi:10.1016/j.advwatres.2007.01.005.

    • Search Google Scholar
    • Export Citation
  • Yapo, P. O., Gupta H. V. , and Sorooshian S. , 1998: Multi-objective global optimization for hydrologic models. J. Hydrol., 204, 8397.

  • Yilmaz, K. K., Adler R. F. , Tian Y. , Hong Y. , and Pierce H. , 2010: Evaluation of a satellite-based global flood monitoring system. Int. J. Remote Sens., 31, 37633782, doi:10.1080/01431161.2010.483489.

    • Search Google Scholar
    • Export Citation
  • View in gallery

    Schematic of the probabilistic quantitative hydrological forecast system for global application. The calibration and downscaling technique was developed in V2010. In this paper, the calibrated and downscaled meteorological forecasts derived and verified in V2010 force the hydrology model. Derived hydrologic forecasts (runoff and discharge) are then verified.

  • View in gallery

    The verification of the system includes four hydrologic experiments: (middle) the reference (1), (left) the interpolation method (2), (right) the RMSDmean method (3), and (far right) the climatology forecast with zero precipitation (4).

  • View in gallery

    TMPA precipitation climatology over (left) the Ohio River basin and (right) the drainage areas of the four streamflow forecast verification locations.

  • View in gallery

    (top to bottom) The 5-day moving average of the TMPA and the gauge-station-based (Maurer et al. 2002) daily basin average precipitation. The 5-day moving average of the difference between ECMWF 24-h precipitation forecast and the gauge–station precipitation is shown in green. Daily calibration of the VIC-routing model at Metropolis (October 2005–September 2007), and verification in Metropolis (October 2002–September 2005) and at three other stations (October 2002–September 2007).

  • View in gallery

    The 2003–07 15-day daily forecast verification for the spatially distributed (solid lines) and basin-average (dashed lines) runoff, for two categories of forecasts: all and upper terciles for the two experiments (interpolation, gray; RMSDmean, black). The reference is the 2003–07 runoff simulated by forcing VIC with ECMWF analysis fields and TMPA precipitation.

  • View in gallery

    Talagrand histograms for daily (a) distributed and (b) basin-average runoff for two forecast categories (all, upper tercile) at days 1, 5, and 10.

  • View in gallery

    Day 1–15 daily flow forecast verification at Metropolis (outlet) for two forecast categories (all and upper terciles) for two downscaling methods (interpolation, gray; RMSDmean, black) and climatology with zero precipitation with respect to the reference (dashed).

  • View in gallery

    As in Fig. 7, but for the Monongahela River at Elizabeth.

  • View in gallery

    Talagrand histograms for daily flow forecasts for two forecast categories (all and upper terciles) at two different stations (Metropolis and Elizabeth, with decreasing drainage areas) at days 2, 5, and 8 (rows 1 and 2, 3 and 4, and 5 and 6, respectively).

All Time Past Year Past 30 Days
Abstract Views 0 0 0
Full Text Views 196 99 11
PDF Downloads 147 64 7

Application of a Medium-Range Global Hydrologic Probabilistic Forecast Scheme to the Ohio River Basin

View More View Less
  • 1 Department of Civil and Environmental Engineering, University of Washington, Seattle, Washington
  • | 2 European Centre for Medium-Range Weather Forecasts, Reading, United Kingdom
  • | 3 Department of Civil and Environmental Engineering, University of Washington, Seattle, Washington
  • | 4 European Centre for Medium-Range Weather Forecasts, Reading, United Kingdom
  • | 5 NOAA/NWS/Office of Hydrologic Development, Silver Spring, Maryland
Full access

Abstract

A 10-day globally applicable flood prediction scheme was evaluated using the Ohio River basin as a test site for the period 2003–07. The Variable Infiltration Capacity (VIC) hydrology model was initialized with the European Centre for Medium-Range Weather Forecasts (ECMWF) analysis temperatures and winds, and Tropical Rainfall Measuring Mission (TRMM) Multisatellite Precipitation Analysis (TMPA) precipitation up to the day of forecast. In forecast mode, the VIC model was then forced with a calibrated and statistically downscaled ECMWF Ensemble Prediction System (EPS) 10-day ensemble forecast. A parallel setup was used where ECMWF EPS forecasts were interpolated to the spatial scale of the hydrology model. Each set of forecasts was extended by 5 days using monthly mean climatological variables and zero precipitation in order to account for the effects of the initial conditions. The 15-day spatially distributed ensemble runoff forecasts were then routed to four locations in the basin, each with different drainage areas. Surrogates for observed daily runoff and flow were provided by the reference run, specifically VIC simulation forced with ECMWF analysis fields and TMPA precipitation fields. The hydrologic prediction scheme using the calibrated and downscaled ECMWF EPS forecasts was shown to be more accurate and reliable than interpolated forecasts for both daily distributed runoff forecasts and daily flow forecasts. The initial and antecedent conditions dominated the flow forecasts for lead times shorter than the time of concentration depending on the flow forecast amounts and the drainage area sizes. The flood prediction scheme had useful skill for the 10 following days at all sites.

Current affiliation: Pacific Northwest National Laboratory, Richland, Washington.

Consultant.

Corresponding author address: Dennis Lettenmaier, Dept. of Civil and Environmental Engineering, University of Washington, Seattle, WA 98195. E-mail: dennisl@u.washington.edu

Abstract

A 10-day globally applicable flood prediction scheme was evaluated using the Ohio River basin as a test site for the period 2003–07. The Variable Infiltration Capacity (VIC) hydrology model was initialized with the European Centre for Medium-Range Weather Forecasts (ECMWF) analysis temperatures and winds, and Tropical Rainfall Measuring Mission (TRMM) Multisatellite Precipitation Analysis (TMPA) precipitation up to the day of forecast. In forecast mode, the VIC model was then forced with a calibrated and statistically downscaled ECMWF Ensemble Prediction System (EPS) 10-day ensemble forecast. A parallel setup was used where ECMWF EPS forecasts were interpolated to the spatial scale of the hydrology model. Each set of forecasts was extended by 5 days using monthly mean climatological variables and zero precipitation in order to account for the effects of the initial conditions. The 15-day spatially distributed ensemble runoff forecasts were then routed to four locations in the basin, each with different drainage areas. Surrogates for observed daily runoff and flow were provided by the reference run, specifically VIC simulation forced with ECMWF analysis fields and TMPA precipitation fields. The hydrologic prediction scheme using the calibrated and downscaled ECMWF EPS forecasts was shown to be more accurate and reliable than interpolated forecasts for both daily distributed runoff forecasts and daily flow forecasts. The initial and antecedent conditions dominated the flow forecasts for lead times shorter than the time of concentration depending on the flow forecast amounts and the drainage area sizes. The flood prediction scheme had useful skill for the 10 following days at all sites.

Current affiliation: Pacific Northwest National Laboratory, Richland, Washington.

Consultant.

Corresponding author address: Dennis Lettenmaier, Dept. of Civil and Environmental Engineering, University of Washington, Seattle, WA 98195. E-mail: dennisl@u.washington.edu

1. Introduction

In situ precipitation observations derived from gauges and precipitation radars are often used for real-time or near-real-time flood forecasting (Hopson and Webster 2010; Thielen et al. 2009), or are used to downscale precipitation forecasts issued by global weather forecast models for short- (up to 48 h, according to the American Meteorological Society’s Glossary of Meteorology) and medium-range (up to 15 days) flood forecasts (Schaake et al. 2007; Clark and Hay 2004). When available, regional-scale atmospheric models can provide medium-range forecasts at a finer level of spatial resolution than is available from global weather models, and the forecasts so derived can be used for flood forecasting (Westrick et al. 2002; De Roo et al. 2003; Pappenberger et al. 2005). Pappenberger et al. (2009b) implemented a global routing model in connection with the hydrological component of the Tiled European Centre for Medium-Range Weather Forecasts (ECMWF) Scheme for Surface Exchanges over Land (H-TESSEL; Balsamo et al. 2009) land surface model to evaluate global hydrological forecast capabilities. They concluded that a preprocessing of the weather forcing was required for accurate daily discharge simulation. Flood forecasting capabilities are especially limited in areas where in situ observations are sparse and/or where there are no regional-scale atmospheric models, a situation that includes much of the underdeveloped world (Hossain and Katiyar 2006).

Improvements in global weather prediction, and in precipitation observations and “nowcasts,” both from satellite and numerical weather prediction systems, are beginning to be implemented in parts of the world that are undergauged. Asante et al. (2007) describe a flood monitoring system that uses satellite precipitation data and a semidistributed hydrologic model that they apply over the poorly gauged Limpopo River basin in Africa. Hopson and Webster (2010) describe a method of producing flood forecasts at two locations in the Ganges and Brahmaputra River basins in Bangladesh that the Flood Forecasting and Warning System of Bangladesh now integrates in their automated flood forecast system. Their streamflow forecasts are derived using ECMWF Ensemble Prediction System (EPS) precipitation forecasts, Tropical Rainfall Measuring Mission (TRMM) 3B42 (Huffman et al. 2007) and Climate Prediction Center (CPC) morphing technique (CMORPH; Joyce et al. 2004) precipitation nowcasts, and National Oceanic and Atmospheric Administration (NOAA) Global Telecommunication System (GTS) precipitation gauge data. Thiemig et al. (2010) describe an African flood alert system with application to the Juba–Shabelle Rivers basin. It draws heavily from the European Flood Alert System (EFAS; Thielen et al. 2009; Bartholmes et al. 2009). The system uses ECMWF EPS weather forecasts, which drive the fully distributed rainfall-runoff model LISFLOOD (Van der Knijff et al. 2008) at a spatial resolution of 0.1°. The model is initialized using either 40-yr ECMWF Re-Analysis [ERA-40; Uppala et al. (2005); at 1° resolution] or Collaborative Historical African Rainfall Model [CHARM; Funk et al. (2003); at 0.1° resolution] atmospheric forcings. An alternative global flood forecast approach that uses TRMM data in conjunction with a rainfall–runoff model has been developed for near-real-time global flood monitoring (Hong et al. 2007; Yilmaz et al. 2010) and shows promise for filling the void in ungauged basins.

Buizza et al. (2008) and Cloke and Pappenberger (2009) summarize the benefits of ensemble-based probabilistic forecasting; not only is the “most-likely scenario” provided by the deterministic forecast (or ensemble mean forecast) but, in addition, information on the uncertainty of this scenario (i.e., the probabilistic forecasts) is also given, with longer lead times than the basin concentration time. In view of those benefits, global ensemble forecasts will be used instead of the finer spatial resolution deterministic forecasts. Similarly, several operational flood forecast systems have been upgraded to incorporate those probabilistic forecasts; see Cloke and Pappenberger (2009) for a summary.

Our objective is to test a global approach to producing hydrological ensemble forecasts in river basins where in situ data are sparse. The strategy is to incorporate remotely sensed precipitation observations and to use precipitation forecasts produced by global weather prediction models in conjunction with a semidistributed macroscale land surface hydrology model; this strategy is similar to the designing of EFAS (Fig. 1). The analysis fields along with satellite precipitation data warm up the hydrology model and create the initial conditions for the hydrology forecasts. We expand on previous approaches by combining satellite precipitation data with the downscaling of global model ensemble weather forecasts to force the hydrology model. In a companion paper, Voisin et al. (2010, hereafter V2010) evaluated several methods for calibrating and spatially downscaling ensemble precipitation forecasts, produced by global weather prediction models using globally available remote sensing observations, to the scale used by macroscale hydrology models. Specifically, we considered a 0.25° latitude–longitude spatial resolution, which is consistent with the capabilities of current-generation global land models (see, e.g., Balsamo et al. 2011). In V2010, three methods were evaluated for downscaling ECMWF EPS 10-day daily 51-ensemble precipitation forecasts from 1° to 0.25° spatial resolution over the Ohio River basin with respect to remotely sensed precipitation: simple interpolation, a bias correction with spatial disaggregation, and an analog method. The goal was to select which approach was most appropriate for calibrating and downscaling global ensemble weather forecasts for daily hydrologic forecasting with realistic precipitation patterns that maintain or improve the original ensemble forecast skill. They found that the analog method had the smallest absolute biases, competitive RMSEs, and the best reliability of the downscaled precipitation ensemble forecasts, but it also had a lower level of predictability than the interpolation method. The analog method had the most realistic high-resolution precipitation patterns, due at least in part to the fact that the analogs were patterns that had occurred in the historical record. V2010 evaluated three variations of the analog method: two variations calibrated each ensemble member individually while the third used the ensemble mean forecast in order to regenerate a calibrated and downscaled weather forecast. V2010 argued that this latter analog approach (RMSDmean) was most appropriate for the further development of a hydrological forecast scheme for global application.

Fig. 1.
Fig. 1.

Schematic of the probabilistic quantitative hydrological forecast system for global application. The calibration and downscaling technique was developed in V2010. In this paper, the calibrated and downscaled meteorological forecasts derived and verified in V2010 force the hydrology model. Derived hydrologic forecasts (runoff and discharge) are then verified.

Citation: Weather and Forecasting 26, 4; 10.1175/WAF-D-10-05032.1

In this paper, the V2010 0.25° calibrated ECMWF EPS 10-day daily ensemble forecasts using the RMSDmean variation of the analog method force the Variable Infiltration Capacity (VIC) semidistributed hydrology model (Liang et al. 1994) to produce ensemble daily runoff and streamflow forecasts in the Ohio River basin (Fig. 1). The paper has three objectives. First, we assess the skill of the system. Second, we evaluate the improvements in the skill of the spatially distributed runoff forecasts and of flow forecasts when the spatially distributed meteorological forecasts are calibrated (preprocessor). Third, we quantify the differences in skills for basins of different sizes. Because the approach is using a semidistributed model and distributed weather forecasts, skill may be different for basins of different sizes, in particular, when the concentration times are different. To address those objectives, we compare two system setups: one that uses interpolated forecasts and one using the calibrated weather forecast as input into the hydrological model. Spatially distributed runoff forecast verifications similar to the precipitation forecast verifications in V2010 are performed. Streamflow forecast verifications are then made at several locations with drainage areas ranging from 17 000 to 525 000 km2. This range of drainage areas allows us to assess the effects of different concentration times on the forecast skill.

The analog method used to calibrate and downscale the weather forecasts is based on resampling remotely sensed precipitation data [from the TRMM Multisatellite Precipitation Analysis (TMPA) 3B42 V6 research product; Huffman et al. (2007)]. In V2010, gridded in situ gauge station data were substituted for the remote sensing precipitation data to evaluate potential differences in skill when using different observed precipitation datasets. The Ohio River basin, which has an extensive in situ gauge network, was used for this evaluation. This paper follows from V2010, which motivates our choice of the Ohio River basin for evaluation of the hydrologic forecasts.

The remainder of the paper is organized as follows. Section 2 summarizes the observations, forecasts, and analysis fields, and the application domain. Section 3 describes the experiments that were designed to evaluate the flood forecast system. Results are presented in section 4, with discussion in section 5, and conclusions in section 6.

2. Datasets and domain

TMPA precipitation and ECMWF wind and temperature analysis fields were used to force the hydrology model up to the time of forecast. ECMWF EPS forecasts (precipitation, temperature, and wind) were calibrated and downscaled as in V2010 (summarized below) and then used to force the hydrology model during the forecast period (Fig. 1). A VIC simulation, forced with TMPA and ECMWF analysis fields, was used as a surrogate for observations for streamflow forecast verification purposes.

a. ECMWF EPS forecasts

The ECMWF EPS has been operational since 1992; it simulates initial condition uncertainties using singular vectors, which are the perturbations with the fastest growth over a finite time interval, and model uncertainties using a stochastic scheme (Palmer et al. 2009). The ECMWF EPS products available for this study are 10-day forecasts from 2003 to 2007 at 6-h time increments, which we aggregated or averaged to a daily time step (0000–2400 UTC). The horizontal spectral resolution of EPS started at ~80-km (T399) grid spacing at midlatitudes in November 2000 (T255), and increased to ~50-km in February 2006 and ~30-km (T639) in January 2010. During September 2006, the ECMWF EPS was extended to 15 days with a variable resolution approach (VAREPS; Buizza et al. 2007) and to 32 days once a week in March 2008 (Vitart et al. 2008), that is, a 63-km spatial resolution version was run for lead times beyond day 10. For consistency, we aggregated the forecasts to 1° spatial resolution and considered only the 10-day forecast period over the entire 2003–07 period. The following surface variables, which are required to force the VIC hydrologic model, were taken from the EPS forecasts: daily precipitation (convective and stratiform fields), maximum and minimum temperatures (MN2T and MX2T), and wind speed (U10 and V10). Those variables had to be calibrated and spatially downscaled from the 1° spatial resolution to 0.25°, the spatial scale at which the VIC model was implemented (as explained below). We reduced the 51 ensemble members to 15 as described in V2010 in order to reduce computation time. Because the random selection of the 15 ensemble members follows a discrete uniform distribution, the impacts on the ensemble mean and spread skills are minimized. With constant development of the forecast models [see section 2a(4)], the calibration and downscaling of the forecasts (analog method) are expected to work best when using a resampling of a retroforecast dataset, as in Hamill and Whitaker (2006), which most often is a reduced-size ensemble with coarser spatial resolution. The number 15 was chosen to be consistent with the National Centers for Environmental Prediction (NCEP) Global Forecast System (GFS) reforecasts dataset (Hamill et al. 2006) used in Hamill and Whitaker (2006).

The following subsections describe how the calibration and downscaling of the different forecast variables were performed. Note that the calibration and downscaling of the ensemble precipitation forecasts were performed in V2010 and readers are referred to that paper for details. The downscaling of the other forecast variables described below (specifically use of a Schaake shuffle) was performed here following recommendations from V2010.

1) Temperature and wind forecasts

Daily minimum (maximum) temperatures were first derived by extracting the minimum (maximum) temperature of the four 6-hourly minimum (maximum) temperature forecasts (MN2T and MX2T fields) of each day of the 10-day forecast. They were then interpolated to 0.25° spatial resolution using the SYMAP algorithm (Shepard 1984): a simple inverse squared distance-from-nearest-neighbors algorithm, where interpolated temperatures are also lapsed using the elevation difference between the 1° and the 0.25° cells using a −0.65°C (100 m)−1 rate (approximately pseudo-adiabatic). Different interpolation algorithms can (but must not) lead to significantly different results [Pappenberger et al. (2009a) compared interpolation methods]. The SYMAP algorithm is used here for its computational efficiency and quality for this area in comparison to other approaches because the forecasts are already gridded and have uniform spatial and temporal densities.

Six-hourly wind forecasts were computed as the square root of the sum of the squares of U10 and V10 fields. These values were averaged to a daily time scale and then interpolated from 1° to 0.25° using the SYMAP algorithm of Shepard (1984).

2) Precipitation forecasts

We calibrated and spatially downscaled the EPS 10-day 1° ensemble precipitation forecasts from 1° to 0.25° on a daily basis using what V2010 define as the analog RMSDmean method. The reader is referred to V2010 for a step-by-step description and evaluation of the method. In brief, the analog method is based on resampling from remotely sensed precipitation (TMPA) fields. A moving (5° × 5°, or 25 forecast points) spatial window (rather than the entire Ohio River basin) was used as the spatial domain for purposes of choosing the analogs; hence, the method is applicable to any domain (see V2010 for details). Over the 5° × 5° window, the daily ensemble mean forecast at day n, year X, for lead time Y, was compared with all daily retrospective ensemble mean forecasts in the same spatial window, for the same lead time Y, in a ±45 days temporal window around day n over the 2002–07 period (excluding day n of year X). The 15 retrospective ensemble mean forecasts that corresponded most closely to the ensemble mean forecast for the given lead time Y in the root-mean-square sense (15 smallest sums of the root-mean-square differences over the 25 grid points in the spatial window) became the 15 analogs for the center point of the spatial moving window for lead time Y. The corresponding finer spatial resolution remote sensing observations (four 0.25° grid cells at the center point of the moving spatial window) for those days (analog dates) became the downscaled ensemble forecasts, following V2010, Hamill and Whitaker (2006), and Hamill et al. (2006).

A Schaake shuffle (Clark et al. 2004) was then performed in order to construct a spatiotemporal rank structure over the 10-day period and over the Ohio River basin domain that was necessary for subsequent hydrologic simulation. This Schaake shuffle (V2010, appendix A.3) was also necessary to create the correlation between all downscaled forecast fields (precipitation, temperature, and wind). In this last step, 15 dates (and the subsequent 9 days) were selected in the 2002–07 period in the ±45-day window around day n. This resample was done for each daily forecast for the entire Ohio River basin domain and the same resample was used for precipitation, temperature, and wind forecast fields. For each forecast lead time (day in the 10-day resampled period) and 0.25° grid cell, each downscaled forecast member (resampled observed date) was ranked from 1 to 15 (number of ensemble members). This ranking was done for each forecast field (precipitation, temperature, and wind) where 0.25° TMPA results were used as the observed precipitation and 0.25° downscaled ECMWF analysis fields were used as the observed temperature and wind). Precipitation, temperature, and wind forecast ensemble members were then reordered so that the forecast rank matrix (space, time, variable, member) was similar to the 15 resampled 10-day observed periods. The calibrated and downscaled ensemble forecast values at each grid cell were not modified by the Schaake shuffle; only the ordering of the members was changed.

V2010 showed that the downscaled ensemble precipitation forecasts were more accurate and largely more reliable than those downscaled using a simple interpolation technique. V2010 also showed that the observed–predicted Spearman rank correlation was maintained, although it was slightly lower.

3) Consistency of the raw EPS forecasts

The quality of probabilistic weather predictions in general has significantly improved over the last decade (Palmer et al. 2007). Precipitation forecasts, which are of particular significance for this study, have gained about 1 day of lead time every 10 yr over Europe for the ECMWF forecasts (Ghelli and Primo 2009). It is important to point out that improvements over the years have varied with parameter as well as verification region (Palmer et al. 2009; Pappenberger and Buizza 2009).

Over the period considered here, the ECMWF forecast model has undergone several changes (spatial resolution, convection scheme), of which it is likely that the change in spatial resolution has had the largest impact on meteorological forecast skill (Buizza 2010; Buizza et al. 2003). Further changes in the meteorological forcing usually occur in two to three updates per year and are documented online at the ECMWF Web site. In this paper, the impacts of the changes in spatial resolution (the major change over the 2002–07 period) are minimized because we have aggregated the EPS forecasts to a 1° spatial resolution over the entire period. The current spatial resolution of the EPS forecasts (~30 km since 2010) would allow us to bypass the spatial disaggregation step of the V2010 calibration and downscaling approaches, but a long retrospective period is necessary for the calibration step. The need for the longest possible forecast period motivated us to choose the 2003–07 period since it has reasonably consistent model physics.

b. Observed precipitation

The 3-hourly TMPA 3B42 V6 research product (Huffman et al. 2007) is a near-global dataset at 0.25° latitude–longitude spatial resolution. It was used as the observation dataset from which the analogs were derived during the calibration and downscaling of the EPS precipitation forecasts (V2010). Daily TMPA estimates are accumulations of the 3-hourly values from 0000 to 2100 UTC.

TMPA rescales the monthly sums of the 3-hourly fields to a monthly gauge analysis. One evolution in the TMPA product during the 2002–07 period is that the Climate Assessment and Monitoring System (CAMS) 5° × 5° monthly gauge analysis is used to adjust the TMPA estimates from January 1998 to March 2005 and the Global Precipitation Climatology Center (GPCC) 5° × 5° monthly monitoring product is used to adjust the TMPA estimates after March 2005, resulting in some inconsistencies (Su et al. 2008).

c. ECMWF analysis

Daily 2002–07 ECMWF analysis wind and temperature fields were downscaled from 1° to 0.25° using the same method as was used to downscale the forecasts. Similarly to the EPS forecasts, the spatial resolution of the analysis fields increased from T511 (~40 km) to T799 (~25 km) between 2002 and 2007, and underwent the same model changes (convective scheme upgrades in 2003 and 2007). These analysis data were used with TMPA precipitation for the spinup of the VIC hydrological model and for the reference hydrological simulation, which produces the VIC initial conditions (soil moisture and snow accumulation) for the hydrologic forecasts, as explained in the methodology section.

d. The Ohio River basin

The Ohio River basin is defined here as being the area upstream of the downstream-most U.S. Geological Survey (USGS) gauging station at Metropolis, Illinois (Fig. 3). The river basin boundary upstream of Metropolis was identified from the HYDRO1k basin delineation [U.S. Geological Survey Center for Earth Resources Observation and Science (EROS); information online at http://eros.usgs.gov/#/Find_Data/Products_and_Data_Available/gtopo30/hydro] and the area was gridded at 0.25° spatial resolution into 848 grid cells. The elevation in the basin ranges from 100 to over 1100 m in the Appalachians Mountain headwaters (mean grid cell elevation). The climate is humid and temperate. The precipitation is relatively uniform throughout the year and throughout the basin. Precipitation types are variable, with more than 40 days a year of thunderstorms (convective) and with significant snow accumulation in wintertime in the northern Appalachians (White et al. 2005). The Ohio River basin has a number of small impoundments along its main stem for flood control and navigation. The Tennessee River, one of the Ohio’s main tributaries, has many larger impoundments for hydropower generation. With a relatively small gradient and therefore small reservoir capacities, the daily flow regulation does not have much effect on the monthly flows, at least upstream of Louisville, Kentucky (see O’Donnell et al. 2000). Overall, the Ohio River basin is a large river basin whose daily flows are closer to natural than many of its large U.S. river basin counterparts (e.g., the Columbia, Colorado, Missouri, and Sacramento). It offers substantial variability in precipitation types and has a high-density precipitation gauge network that was necessary for the evaluation of the calibration and downscaling of the weather forecasts (V2010). As mentioned later, the daily regulation nonetheless provides some challenges in terms of flow observation uncertainties.

3. Methodology

a. Experimental design

The study was based on four sets of runs (Fig. 2).

Fig. 2.
Fig. 2.

The verification of the system includes four hydrologic experiments: (middle) the reference (1), (left) the interpolation method (2), (right) the RMSDmean method (3), and (far right) the climatology forecast with zero precipitation (4).

Citation: Weather and Forecasting 26, 4; 10.1175/WAF-D-10-05032.1

Fig. 3.
Fig. 3.

TMPA precipitation climatology over (left) the Ohio River basin and (right) the drainage areas of the four streamflow forecast verification locations.

Citation: Weather and Forecasting 26, 4; 10.1175/WAF-D-10-05032.1

1) Reference

The first set was the reference, in which the semidistributed VIC hydrology model was forced with the downscaled ECMWF analysis temperatures and wind fields, and TMPA precipitation for the 2002–07 period (one simulation). The gridded runoff was then directed using the Lohmann et al. (1996, 1998) routing model to four stream gauge locations (forecast points) for the 2003–07 simulation period; 2002 was used as a spinup of the hydrology and routing models. The four forecast points are the Monongahela River at Elizabeth, Pennsylvania (USGS 03075070, 13 830 km2), the Wabash River at Mount Carmel, Illinois (USGS 03377500, 74 165 km2), the Ohio River at Louisville, Kentucky (USGS 03294500, 236 130 km2), and the Ohio River at Metropolis (USGS 03611500, 525 767 km2), which have drainage areas of varying size. Having four forecast points allows the assessment of the role of drainage area in the resulting streamflow forecast errors. The reference run serves as a surrogate for the observed spatial distribution of daily runoff over the Ohio River basin and for observed daily streamflow at the four locations for the forecast verifications. Use of the reference run (essentially the hydrology model forced with observations) avoids confounding forecast errors with the effects of hydrological modeling error, and as such allows for a focused assessment of individual uncertainty contributions. Nonetheless, a final system will need to integrate all error sources (Pappenberger and Beven 2006; Pappenberger et al. 2006).

Regulation of the Ohio River basin for water management purposes is mostly concerned with navigation and flood control, the effects of which are minimal on the monthly time scale. However, these effects, in addition to several small hydropower plants, can affect daily flows significantly. The VIC model does not simulate these effects; however, Fig. 4 shows that the combination of the VIC hydrology and routing model did capture the observed daily streamflow reasonably well at all four stations following calibration (performance described in detail in the next section). Given this agreement, the reference simulation is considered to be “truth,” rather than the daily streamflow observations, to focus attention on the hydrologic effects of the errors in the downscaled forecasts. One could also apply postprocessing and error corrections to the observed and forecasted flows (Seo et al. 2006, Bogner et al. 2010); however, our eventual intent is to apply the system described herein to ungauged catchments, for which observed flow may not be readily available (Schumann et al. 2009; Neil et al. 2009).

Fig. 4.
Fig. 4.

(top to bottom) The 5-day moving average of the TMPA and the gauge-station-based (Maurer et al. 2002) daily basin average precipitation. The 5-day moving average of the difference between ECMWF 24-h precipitation forecast and the gauge–station precipitation is shown in green. Daily calibration of the VIC-routing model at Metropolis (October 2005–September 2007), and verification in Metropolis (October 2002–September 2005) and at three other stations (October 2002–September 2007).

Citation: Weather and Forecasting 26, 4; 10.1175/WAF-D-10-05032.1

2) Interpolation method (interpolated ensemble weather forecasts)

The second set of runs used the reference forcing to spin up the hydrologic model until the day of the forecast. The runs subsequently used the downscaled ECMWF EPS temperatures and wind speeds, and ECMWF EPS interpolated precipitation forecasts to force the hydrology model. Each of the fifteen 10-day ensemble members was run through the hydrology model for all days during the 2003–07 period (1826 forecasts). The routing model was also spun up using the gridded runoff from the reference run. Next, the forecasted gridded runoff for each ensemble member was routed to derive daily ensemble streamflow forecasts at the four forecast points. This simulation is referred to as the interpolation method.

3) Analog RMSDmean method (calibrated and downscaled weather forecasts)

The third set of runs used the same spinups and setup as the interpolation method, except that the ECMWF EPS precipitation forecasts were calibrated and downscaled using the RMSDmean analog technique from V2010 and described above. This method is termed RMSDmean hereafter.

4) Zero precipitation and climatology forecast

The fourth simulation isolates the impacts of the initial conditions (and concentration time) from the skill of the forecasts in the early lead times. The VIC hydrology model was run as for the reference run until the day of the forecast, and then was forced with a 15-day deterministic forecast with null precipitation and temperature and wind climatology monthly means (Fig. 2).

Depending on the size of the drainage basin (hence concentration time), the forecasted daily streamflow for the first few days of the forecasts responds mostly to the hydrologic initial and antecedent conditions. The initial conditions include snow water equivalent (SWE) and soil moisture. Initial and antecedent conditions also include streamflow throughout the basin at the time of forecast. Table 1 shows the daily fraction of grid cells that contributes to the daily flow at each of four stations. These fractions are derived using unit hydrographs created by the calibrated routing model (see the next section for discussion of the calibration of the routing model). On day 1 (day of the precipitation event), only a small fraction of the grid cells in each drainage area contributes to the flow at the forecast points (i.e., the initial conditions prevail). As expected, the fraction of contributing grid cells increases faster with time for smaller drainage areas. Table 1 also shows the daily fraction of grid cells for which 100% of the generated runoff has been routed to the station. Because of the assumed grid cell unit hydrograph shape (Table 2), no grid cell (and hence no basin) can be fully drained in less than 4 days. Table 1 shows that the initial conditions will impact the flow at Metropolis (outlet) for up to 14 days, and up to 6 days at Elizabeth, the smallest drainage area considered in this paper. This also implies that there could be skill in the forecasted streamflow beyond the 10 days of the weather forecast. We therefore extended by 5 days each of the 10-day ensemble forecast members of the interpolation and the RMSDmean methods using the downscaled ECMWF analysis monthly mean temperatures and wind (climatology), and null precipitation, which resulted in an ensemble streamflow forecast period of 15 days (Fig. 2).

Table 1.

Daily fraction of grid cells that contribute to the flow at the station, i.e., part of the generated runoff at this grid cell has been routed to the station (contribute), and for which all (≥99%) of the generated daily runoff has been routed to the station (fully routed).

Table 1.
Table 2.

Estimated daily unit hydrograph for the outlet of one grid cell (0.25° × 0.25°) assuming a land-phase runoff velocity of 0.2 m s−1 and a rainfall event around 1600 local time (LT).

Table 2.

b. Calibration of the hydrology and routing models

The semidistributed VIC hydrology and routing models are two independent entities. For monthly applications, a default calibration is usually used for the routing model (routing makes little difference for monthly streamflow aggregations) and VIC is calibrated independently. For daily application, VIC and the routing component were considered to be one model and were calibrated jointly. This one-step approach avoided overfitting because the calibration of the two entities cannot be fully independent due to the absence of observed gridded runoff. The calibration of the VIC-routing model was performed at a daily time scale, from October 2005 through September 2007 at the downstream-most station (Metropolis). The period October 2002–September 2005 was used for evaluation. VIC was forced with ECMWF analysis daily minimum and maximum temperatures and wind, as well as TMPA daily precipitation.

The calibration was performed using the Multiobjective Complex Evolution of the University of Arizona (MOCOM-UA) method (Yapo et al. 1998), as applied in Shi et al. (2008) but with different objective functions. Following Gupta et al. (2009), the Nash–Sutcliffe efficiency (NSE) factor (Nash and Sutcliffe 1970) is inappropriate for calibration because bias, variability, and correlation all impact NSE and the optimization of the latter does not imply an optimization of each of those independent components. Four objective functions were used instead: the variance explained, the relative bias, the relative standard deviation difference, and the absolute value of annual mean volume error. Multiobjective automatic calibration is time saving and allows for fitting the simulated hydrograph to different characteristics of the observed hydrograph.

A lumped calibration approach was chosen here for simplicity. Ajami et al. (2004) showed that representing spatial variability in the calibrated parameters (distributed calibration approach) did not improve the flow calibration at their basin outlet. Feyen et al. (2008), however, showed that a spatially distributed approach for calibration parameters (in particular for the most uncertain soil parameters like soil depth) was necessary to represent subbasin hydrological processes. They argued that the distributed approach reduces the overall model uncertainty and therefore increases the accuracy of the flow predictions, especially at the subbasin level. In ungauged basins, regionalization of the calibration parameters can be performed by linking the calibration parameters to long-term climatic and hydrologic properties, as well as physical basin characteristics (e.g., Abdulla and Lettenmaier 1997; Yadav et al. 2007). The implementation of regional calibration approaches for ungauged basins will be the topic of future work. Here, we use the lumped approach, which is appropriate given our emphasis on forecast errors at the basin level.

Table 3 shows the summary statistics for the daily flow at Metropolis for the calibration period (October 2005–September 2007) and for a verification period (October 2002–September 2005). The top panel in Fig. 4 shows the 5-day moving average of the daily basin average difference between TMPA and a gridded station precipitation dataset (extension of Maurer et al. 2002). Figure 4 also shows reasonable agreement between the simulated and observed daily flows at Metropolis, and at the three other locations (Louisville, Mount Carmel, and Elizabeth).

Table 3.

Calibration and verification statistics for the one-step VIC and routing models’ daily calibration of the Ohio River at Metropolis, using either ECMWF analysis fields and TMPA precipitation, or ECMWF analysis fields and ECMWF 24-h precipitation as forcing.

Table 3.

c. Runoff and streamflow forecast verification

1) Runoff forecast verification

Runoff forecast verification at the grid cell level allows evaluation of the value of a spatially distributed preprocessor and is not dependent on the location of the streamflow gauge or its drainage area. It therefore avoids the problem of convolution of skills that is inherent in the use of streamflow for verification. Runoff forecasts were verified by comparing the daily gridded runoff forecast (interpolation and RMSDmean methods) with the reference runoff (derived from the model forced with observations). Similar to V2010, commonly used skill measures like biases (mean errors), RMSEs (accuracy), and predicted–observed Pearson correlation (predictability) were used to verify the daily forecasted runoff. Other skill measures specific to probabilistic quantitative forecasts were used: ensemble spread (range) reliability (Talagrand histograms; Hamill and Colucci 1997; Talagrand and Vautard 1997) and continuous rank probability skill score [CRPSS; see Hersbach (2000), Wilks (2006), and appendix C in V2010 for details]. Those measures were assessed as an average over the 2003–07 period over the Ohio River basin for the 15-day lead times for three forecast categories: all forecasts in the 2003–07 period, and the lower and upper terciles of the forecasts. It is important to note that the categories used in the evaluations are conditioned on the forecasts and not the observations. Conditioning on observations is common when comparing, for instance, forecasts from different sources (Demargne et al. 2010). Here, however, we want to assess the value of hydrologic forecasts for potential real-time decisions, “What can I expect if the forecasts falls into a certain forecast category? What should I do with this forecast?,” rather than how good are those forecasts for those particular events. The analysis was made on both spatially distributed and basin-average runoffs. The distributed runoff analysis provides a parallel with V2010, where skill scores are computed at each grid cell then averaged over the basin as recommended by Hamill and Juras (2006). As noted in V2010, the spatiotemporal structure in the five-dimensional forecast matrix (variable, space, lead time, ensemble member) is lost when forecasted variables for each ensemble member are calibrated and downscaled either (i) separately, (ii) in space, and/or (iii) in time. This results in an ensemble of basin-average precipitation–runoff, which is very narrow and centered more or less around climatology with no discrimination, reliability, or skill in general (not shown). The basin-average runoff analysis (scores computed for basin-averaged values) confirms the conservation of the spatial rank structure in the calibrated ensemble weather forecasts via the Schaake shuffle.

2) Streamflow forecast verification

Daily streamflow (routed runoff) at the four forecast points for each of the 15 lead times was verified using the reference simulation as truth. Not only are the ensemble flow forecasts the end product of the flood prediction scheme, but their accuracy also confirms the spatiotemporal rank structure of the downscaled weather forecasts; that is, if there was unsatisfactory spatiotemporal rank structure in the ensemble weather forecasts forcing the hydrology model, the ensemble routed runoff (flow) forecasts would have similar biases and RMSEs as using a deterministic climatology mean forecast with a narrow (and unreliable) ensemble spread that would not grow in size with increasing lead times. Furthermore, we assess the skill due to the initial conditions rather than the weather forecast skill in the first few lead times of the flow forecasts by comparing the scores of the RMSDmean and interpolation methods with those of the climatology and zero precipitation 15-day deterministic forecast experiment.

4. Results

The forecast verifications for the interpolation and the RMSDmean methods were made with respect to the reference simulation for the 2003–07 period over the entire Ohio River basin. Unless indicated otherwise, an improvement indicates that the RMSDmean method performed better than the interpolation method. Three forecast categories were evaluated: all forecast (no conditioning), and lower and upper terciles of the runoff/flow forecasts. For simplicity, only the all and high runoff forecast categories and two flow forecast points are shown. The means and relative biases (i.e., the ratio of bias to the mean) in each category are shown. The relative RMSEs, CRPSSs, and Talagrand histograms are indicative of the accuracy. CRPSS in particular is indicative of how the cumulative distribution function (CDF) of the forecast ensemble (daily forecast CDF based on 15 values, i.e., members, for one lead time) matches the observation for each individual forecast; it is indicative of bias, but also of the ensemble spread reliability, resolution, and predictability. The Talagrand histograms or rank histograms (Hamill and Colucci 1997; Talagrand and Vautard 1997) show the normalized rank of the observation within the ensemble forecast. Perfect ensemble reliability corresponds to a uniform histogram, a U-shaped histogram indicates an ensemble spread that is too small (forecast is overconfident), and an asymmetric histogram indicates a systematic bias in the forecast. We define here the ensemble spread as the ensemble range, that is, the difference between the maximum and the minimum member values. The predicted–observed Pearson correlation is indicative of the predictability of the ensemble mean forecast. The reader is referred to V2010’s appendix C for a detailed description of the measures.

a. Forecast verification for daily runoff over the 2003–07 period, Ohio River basin

The forecast verification for the forecasted spatially distributed runoff was performed by computing the various performance measures at each grid cell independently and then averaging over the basin (V2010, appendix C). For the basin-average runoff forecast verification, the scores were computed using the basin-average runoff values. Figure 5 shows the forecast verification for spatially distributed (solid lines) and for basin-average (dashed lines) runoff forecasts for the interpolation method (gray, setup 2) and the RMSDmean method (black, setup 3) with respect to the simulated truth (setup 1); the mean, relative bias, RMSE, CRPSS, and Pearson correlation are all shown. Results for the spatially distributed and basin-average runoff forecasts are consistent. The basin-average runoff scores are usually higher, as expected. On day 11 of the extended forecasts, the mean runoff values drop sharply, and consequently, relative biases and RMSEs increase sharply. This implies that for runoff there is no extension in skill beyond 10 days (through the initial conditions) when forecasts are extended using the assumed zero precipitation and temperature climatology beyond forecast day 10. Note that with the EPS forecast extension to 15 days in 2006 and 32 days in 2008, runoff forecasts for more recent periods should have a useful skill for longer than 10 days, but no extension in skill through the initial conditions. In agreement with the forecast verifications for spatially distributed precipitation in V2010, the biases of the runoff forecasts produced by the analog method improved relative to those of the interpolation method.

Fig. 5.
Fig. 5.

The 2003–07 15-day daily forecast verification for the spatially distributed (solid lines) and basin-average (dashed lines) runoff, for two categories of forecasts: all and upper terciles for the two experiments (interpolation, gray; RMSDmean, black). The reference is the 2003–07 runoff simulated by forcing VIC with ECMWF analysis fields and TMPA precipitation.

Citation: Weather and Forecasting 26, 4; 10.1175/WAF-D-10-05032.1

The RMSDmean CRPSS was lower than the interpolation CRPSS, due to an increase in ensemble spread, as shown by the Talagrand histograms (Fig. 6). The predictability (Pearson correlation) was similar between the interpolation and the analog RMSDmean methods. In V2010, the RMSDmean Spearman rank correlation for the daily precipitation forecasts was slightly lower than for the interpolation method. Figure 5 shows an improvement in RMSDmean reliability relative to the interpolation method for both the distributed and basin-average runoff. At day 1, there was a large improvement in reliability for the RMSDmean method for all forecast categories for both spatially distributed and basin-average runoff. The improvement for basin-average runoff for the RMSDmean method was important at small lead times. However, the interpolation and RMSDmean reliabilities were similar after day 5 (averaging effect). The RMSDmean method improved the reliability for all forecast categories and all lead times for spatially distributed runoff, which is consistent with precipitation as reported in V2010. The RMSDmean method showed high forecast ensemble spread reliability. The interpolation method lacked reliability (U-shape histograms), and also showed a consistent overestimation of distributed runoff (asymmetric histogram), which is explained by the fact that the VIC model is calibrated to the reference forcings (TMPA precipitation) and the interpolated precipitation forecasts are wetter than TMPA in general (Fig. 3 in V2010, all forecast categories, all lead times).

Fig. 6.
Fig. 6.

Talagrand histograms for daily (a) distributed and (b) basin-average runoff for two forecast categories (all, upper tercile) at days 1, 5, and 10.

Citation: Weather and Forecasting 26, 4; 10.1175/WAF-D-10-05032.1

b. Forecast verification for daily streamflow at four locations over the 2003–07 period

Figures 7 and 8 show the daily ensemble streamflow forecast verifications for lead times 1–15 days for the Ohio River in Metropolis and the Monongahela River in Elizabeth, with respect to the reference. The number of lead times (days) when the initial conditions control the streamflow forecasts decreases with decreasing drainage area as shown by the lead time at which the interpolation and RMSDmean method performance scores diverge from the climatology and zero precipitation deterministic forecast scores (dashed line, fourth simulation). As shown in the first and second rows of Fig. 7, the initial conditions control the forecast mean flow values in each forecast category for the first 2 days at Metropolis (and Louisville; not shown) forecast point. For the Mount Carmel forecast point, the climatology and zero precipitation deterministic mean flow and relative bias values dropped slightly on day 2; that is, the initial conditions control the forecast flow values on day 1 and partially on day 2 (not shown). For the Elizabeth forecast point (Fig. 8), which has the smallest drainage area, the initial conditions control the forecast value only on day 1. On the basis of the correlation between the forecast and observations (Figs. 7 and 8), the initial conditions had an influence on the forecasts’ performance for up to 3 days at Elizabeth and Mount Carmel, and 4 days at Louisville and Metropolis. The influence of the initial conditions on the correlation was shorter (longer) for low (high) flow forecasts (i.e., for lower- and upper-tercile forecast categories). Similarly, the control through the initial conditions of the flow forecasts for short lead times results in an extension in skill beyond 10 days, when interpolated and analog RMSDmean forecasts are extended to 15 days using the assumed zero precipitation and temperature climatology (forecast discontinuity). This extension in skill spans from day 11 to when the performance skill scores drop sharply: day 11 (1 day beyond the duration of the forecasts) at Elizabeth; day 12 at Mount Carmel, Louisville, and Metropolis, in terms of relative biases; and day 12 at Elizabeth and Mount Carmel and day 14 at Louisville and Metropolis on the basis of the forecast–observed correlation. These are average maximum lead times and it could be possible that longer lead times would be achieved for particular events (see .g. Thielen et al. 2009). Following the extension of the EPS forecasts beyond day 10 in 2006, it should be possible to generate flood forecasts with longer lead times when using more recent forecasts.

Fig. 7.
Fig. 7.

Day 1–15 daily flow forecast verification at Metropolis (outlet) for two forecast categories (all and upper terciles) for two downscaling methods (interpolation, gray; RMSDmean, black) and climatology with zero precipitation with respect to the reference (dashed).

Citation: Weather and Forecasting 26, 4; 10.1175/WAF-D-10-05032.1

Fig. 8.
Fig. 8.

As in Fig. 7, but for the Monongahela River at Elizabeth.

Citation: Weather and Forecasting 26, 4; 10.1175/WAF-D-10-05032.1

The relative biases were reduced by the RMSDmean method relative to the interpolation method at all forecast points. Relative RMSEs were usually improved, or remained equivalent, with the exception of Elizabeth. The analog RMSDmean method focused on improving bias and reliability, in contrast to RMSEs (V2010). Consequently, the RMSE for flow was not improved for short concentration time forecast points, but was reduced or maintained for longer concentration times due to an improved bias. The RMSDmean method CRPSSs were either equivalent (Elizabeth) or lower than those for the interpolation method. As with forecast verification for precipitation (V2010) and runoff (section 4a), improvement in CRPSSs is due to increased reliability, discussed below. The observed–predicted correlations were similar for the interpolation and the RMSDmean methods.

Figure 9 shows Talagrand histograms at two stations (columns) for two forecast categories and lead times of 1, 5, and 8 days. As expected, on day 1 (not shown), the ensemble spreads were too narrow and therefore the reliability was poor for both methods at all stations because the initial conditions controlled the streamflow forecasts. On day 2, the RMSDmean method improved the reliability with respect to the interpolation method, especially at Elizabeth and Mount Carmel (not shown) as the effects of the initial conditions decayed more rapidly for the smaller drainage areas. By day 5, the RMSDmean method improved the reliability of the forecasts considerably at all four stations for all forecast categories. At day 8, the RMSDmean method still improved the reliability for the larger basin. Reliabilities between the interpolation and the RMSDmean methods were equivalent at day 8 for smaller basins and by day 10 for larger basins due to a very large ensemble spread.

Fig. 9.
Fig. 9.

Talagrand histograms for daily flow forecasts for two forecast categories (all and upper terciles) at two different stations (Metropolis and Elizabeth, with decreasing drainage areas) at days 2, 5, and 8 (rows 1 and 2, 3 and 4, and 5 and 6, respectively).

Citation: Weather and Forecasting 26, 4; 10.1175/WAF-D-10-05032.1

5. Discussion

We have described a hydrological forecast scheme that is intended for global application, particularly in regions where in situ data are sparse. The potential improvement in ensemble streamflow forecast skill that is realizable from calibrating the ensemble weather forecasts (upstream of the hydrology model) in comparison to a simple interpolation of the forecasts was investigated.

a. Preprocessor improvement and time of concentration

The system using the calibrated weather forecasts (analog RMSDmean method) reduced the relative biases and improved the reliability of the spatially distributed and basin-average runoff and daily streamflow forecasts relative to no calibration (interpolated forecasts). Relative RMSEs and the observed–predicted correlations, on the other hand, were similar for the interpolation and RMSDmean methods for all forecast categories. The RMSDmean CRPSS values were lower than those for the interpolation method due mostly to the improved reliability of the ensemble spread. As a result of the routing time in the runoff to streamflow transformation, the initial conditions controlled the forecast performance for the first 1–4 days (depending on drainage basin size). Useful streamflow forecast skill was, therefore, derived from the 10-day forecasts for up to four additional days depending on the drainage area and the forecast category (flow magnitude). This is valid even for locations in which no site-specific calibration of the hydrologic-routing model was performed (surrogate for ungauged basins). Our results suggest that spatially distributed runoff and flow can be forecasted in regions with sparse precipitation observation networks in a more reliable and accurate way when the hydrology model is forced with calibrated and downscaled precipitation forecasts using the RMSDmean method as compared with the interpolation method. Forecast reliability is an important consideration in real-time forecast applications: “What can I expect if the forecast falls into a certain forecast category? What should I do with this forecast?” The improvement in streamflow forecast accuracy achieved through the use of the method we evaluated is similar for basins with varying drainage areas, but the timing of the improvement varies with the time of concentration at the forecast point, which in turn depends on the drainage area and the streamflow magnitude. For lead times shorter than the time of concentration, the calibration of the weather forecasts has little impact on the streamflow forecasts. The ensemble mean streamflow forecasts are of expected accuracy (as good as the streamflow simulation forced with the observations) but the probabilistic streamflow forecasts are not reliable (the ensemble spread is too narrow). For lead times longer than the time of concentration, the analog RMSDmean ensemble flow forecasts are more accurate and reliable than the flow forecasts without calibration upstream of the hydrology model.

b. Uncertainties in model parameterization

Because the VIC reference simulation was calibrated using TMPA precipitation rather than ECMWF 24-h precipitation forecasts, the forecast verification was more favorable for the RMSDmean method than for the interpolation method. We therefore performed an additional experiment (not shown) in which VIC was calibrated using ECMWF analysis fields and 24-h precipitation forecasts. Calibration performances were similar to the calibration using TMPA and ECMWF analysis fields (Table 3). The set of runs for the interpolation forecast experiment was updated, including a spinup using this new reference. The forecast verification was renewed with respect to the new reference. Note that this setup is not optimal because the fine spatial resolution of TMPA is not exploited and an adjustment between the spinup forcing and the forecast is still lacking (precipitation forecasts on day 10 tend to be too wet). The comparison of the forecast verification scores of the two experiments (RMSDmean and interpolation) with respect to their respective optimal reference showed generally similar results as those presented earlier in terms of reliability and bias improvement (albeit somewhat larger improvements). RMSE, CRPSS, and correlation were relatively unchanged. Using the updated VIC calibration and forecast verification reference for evaluation of the interpolation method, the bias in runoff increased with lead time because no adjustment was made for the precipitation forecasts, whose magnitudes tend to increase with lead time (see Fig. 3 in V2010). Use of the revised reference confirmed that an adjustment is needed to avoid biases resulting from precipitation forecasts becoming wetter with lead times. The fact that the overall relative performance of the two methods was not changed by the use of a reference more favorable to the interpolation method suggests that the improvement resulting from use of the preprocessor is larger than the model parameter uncertainties. This is likely a result in part from the fact that the ECMWF 24-h precipitation forecasts were close to TMPA (Fig. 4) and the resulting hydrological model parameterization was also close to the reference used in the earlier analysis.

c. Accumulation of uncertainties

We evaluated the use of observed streamflow rather than the model as the reference. Forecast skill levels when observations rather than the model were used as the reference were generally comparable in character (not shown). However, when the observations were used as the reference, the overall bias was low; also in addition, low flows tended to be underestimated and high flows overestimated. The key improvement resulting from the preprocessor, improved reliability, could not be assessed as the Talagrand histograms became asymmetric by construct due to the systematic biases relative to the observed flows.

It is important to note that the forecast preprocessor has the potential to reduce only one of several sources of forecast errors. Other error sources include (i) errors in the hydrologic initial conditions resulting from errors in the hydrological model forcings (as shown in V2010), (ii) hydrological model parameter and structural errors, (iii) errors in the observed streamflows, and (iv) effects of streamflow regulation not represented in the hydrological model. In particular, the minimum daily low flow maintained throughout the main stem for navigation is likely to artificially increase the observed low flows, and flood control will reduce the high flows, relative to what would be observed in the absence of regulation. We argue nonetheless that these remaining sources of predictive uncertainty are essentially independent of the forecast error sources we evaluate here and are, therefore, subject to other approaches to error reduction.

d. EPS underdispersion

The EPS short-range under dispersion [U-shape Talagrand diagrams at forecast days 1 and 2 for precipitation (Fig. 3 in V2010), and runoff forecasts; see Fig. 6] was due to the strategy used to define the EPS initial perturbations. Since June 2010, the perturbations have been defined by adding to the singular vectors a new set of perturbations defined by ECMWF’s new Ensembles of Data Assimilation System (Palmer et al. 2009). This change has lead to improved spread–skill reliability in the extratropics during the first two days (Buizza 2010). Work is progressing at ECMWF to further improve the EPS spread–skill reliability by further improving the simulation of the initial uncertainties by introducing land surface perturbations, and by further improving the simulation of model uncertainties by upgrading the stochastic scheme. These changes are expected to have a positive impact on EPS-based probabilistic flood applications. Future work could include the calibration and evaluation of the accuracy and reliability of routed flow forecasts in ungauged basins.

6. Conclusions

The hydrologic forecasting scheme we have outlined consists of the semidistributed VIC macroscale hydrology model forced by EPS precipitation forecasts calibrated and downscaled using the RMSDmean method (developed in V2010). It is of most practical interest for ungauged basins where calibration of the ensemble flow forecasts (postprocessor, downstream of the hydrology model) is complicated by the absence of or large uncertainties in observed streamflow. It should also have useful applications for related purposes such as landslide prediction and forecasting of flood inundation extent.

Our main conclusions include the following points:

  • The system has useful skill and reliability for spatially distributed runoff and streamflow forecasts for about 10 days beyond the time of concentration, more or less independent of the basin drainage area. Following the extensions of the EPS forecasts beyond day 10 in 2006 and day 15 in 2008, the system should have useful skill for longer lead times beyond the time of concentration of the basin. For forecast lead times shorter than the time of concentration, the hydrologic initial conditions control the streamflow forecast skill.

  • Reduction in the number of ensemble members (which helps computational feasibility) was successfully compensated by calibrating the reduced ensemble of forecasts. Furthermore, the improvements in reliability and bias that result from calibration of the weather forecast ensemble are realized in the runoff and streamflow forecasts.

  • Finally, the Schaake shuffle was successful for imposing a lost rank-correlation structure in the spatial and temporal ensemble forecast matrix (space, time, variable, ensemble member). This is the key element for transforming a spatially distributed calibrated ensemble weather forecast (with focus on reliability) into a reliable flow forecast. It also allows the system to be appropriate for river basins large enough that the average time of travel exceeds the time step of the hydrological model (1 day in the application we have illustrated here).

Acknowledgments

This work was supported in part by NASA Grant NNX10AG87G to the University of Washington.

REFERENCES

  • Abdulla, F. A., and Lettenmaier D. P. , 1997: Development of regional parameter estimation equations for a macroscale hydrologic model. J. Hydrol., 197, 237257.

    • Search Google Scholar
    • Export Citation
  • Ajami, N. K., Gupta H. V. , Wagener T. , and Sorooshian S. , 2004: Calibration of a semi-distributed hydrological model for streamflow estimation along a river system. J. Hydrol., 298, 112135.

    • Search Google Scholar
    • Export Citation
  • Asante, K. O., Dezanove R. M. , Artan G. , Lietzow R. , and Verdin J. , 2007: Developing a flood monitoring system from remotely sensed data for the Limpopo Basin. IEEE Trans. Geosci. Remote Sens., 45, 17091714.

    • Search Google Scholar
    • Export Citation
  • Balsamo, G., Viterbo P. , Beljaars A. , den Hurk B. V. , Hirchi M. , Betts A. , and Scipal K. , 2009: A revised hydrology for the ECMWF model: Verification from field site to terrestrial water storage and impact in the Integrated Forecast System. J. Hydrometeor., 10, 623643.

    • Search Google Scholar
    • Export Citation
  • Balsamo, G., Pappenberger F. , Dutra E. , Viterbo P. , and van den Hurk B. , 2011: A revised land hydrology in the ECMWF model: A step towards daily water flux prediction in a fully-closed water cycle. Hydrol. Processes, 25, 10461054, doi:10.1002/hyp.7808.

    • Search Google Scholar
    • Export Citation
  • Bartholmes, J. C., Thielen J. , Ramos M.-H. , and Gentilini S. , 2009: The European Flood Alert System EFAS – Part 2: Statistical skill assessment of probabilistic and deterministic operational forecasts. Hydrol. Earth Syst. Sci., 13, 141153.

    • Search Google Scholar
    • Export Citation
  • Bogner, K., Pappenberger F. , Thielen J. , and De Roo A. , 2010: Wavelet based error correction and predictive uncertainty of a hydrological forecasting system. Geophysical Research Abstracts, Vol. 12, Abstract EGU2010-10245. [Available online at http://meetingorganizer.copernicus.org/EGU2010/EGU2010-10245.pdf.]

    • Search Google Scholar
    • Export Citation
  • Buizza, R., 2010: Horizontal resolution impact on short- and long-range forecast error. Quart. J. Roy. Meteor. Soc., 136, 10201035.

  • Buizza, R., Richardson D. S. , and Palmer T. N. , 2003: Benefits of increased resolution in the ECMWF ensemble system and comparison with poor-man’s ensembles. Quart. J. Roy. Meteor. Soc., 129, 12691288.

    • Search Google Scholar
    • Export Citation
  • Buizza, R., Bidlot J.-R. , Wedi N. , Fuentes M. , Hamrud M. , Holt G. , and Vitart F. , 2007: The new ECMWF VAREPS (Variable Resolution Ensemble Prediction System). Quart. J. Roy. Meteor. Soc., 133, 681695.

    • Search Google Scholar
    • Export Citation
  • Buizza, R., Leutbecher M. , and Isaksen L. , 2008: Potential use of an ensemble of analyses in the ECMWF Ensemble Prediction System. Quart. J. Roy. Meteor. Soc., 134, 20512066.

    • Search Google Scholar
    • Export Citation
  • Clark, M. P., and Hay L. E. , 2004: Use of medium-range numerical weather prediction model output to produce forecasts of streamflow. J. Hydrometeor., 5, 1532.

    • Search Google Scholar
    • Export Citation
  • Clark, M. P., Gangopadhyay S. , Hay L. , Rajagopalan B. , and Wilby R. , 2004: The Schaake shuffle: A method for reconstructing space–time variability in forecasted precipitation and temperature fields. J. Hydrometeor., 5, 243262.

    • Search Google Scholar
    • Export Citation
  • Cloke, H. L., and Pappenberger F. , 2009: Operational flood forecasting: Review of ensemble techniques. J. Hydrol., 375, 613626.

  • Demargne, J., Brown J. , Liu Y. , Seo D.-J. , Wu L. , Toth Z. , and Zhu Y. , 2010: Diagnostic verification of hydrometeorological and hydrologic ensembles. Atmos. Sci. Lett., 11, 114122, doi:10.1002/asl.261.

    • Search Google Scholar
    • Export Citation
  • De Roo, A. P. J., and Coauthors, 2003: Development of a European flood forecasting system. Int. J. River Basin Manage., 1, 4959.

  • Feyen, L., Kalas M. , and Vrugt J. , 2008: Semi-distributed parameter optimization and uncertainty assessment for large-scale streamflow simulation using global optimization. Hydrol. Sci. J., 53, 293308, doi:10.1623/hysj.53.2.293.

    • Search Google Scholar
    • Export Citation
  • Funk, C., Michaelsen J. , Verdin J. , Artan G. , Husak G. , Senay G. , Gadain H. , and Magadzire T. , 2003: The collaborative historical African rainfall model: Description and evaluation. Int. J. Climatol., 23, 4766, doi:10.1002/joc.866.

    • Search Google Scholar
    • Export Citation
  • Ghelli, A., and Primo C. , 2009: On the use of the extreme dependency score to investigate the performance of an NWP model for rare events. Meteor. Appl., 16, 537544, doi:10.1002/met.153.

    • Search Google Scholar
    • Export Citation
  • Gupta, H. V., Kling H. , Yilmaz K. K. , and Martinez G. F. , 2009: Decomposition of the mean squared error and NSE performance criteria for improving hydrological modeling. J. Hydrol., 377, 8091, doi:10.1016/j.jhydrol.2009.08.003.

    • Search Google Scholar
    • Export Citation
  • Hamill, T. M., and Colucci S. J. , 1997: Verification of Eta–RSM short-range ensemble forecasts. Mon. Wea. Rev., 125, 13121327.

  • Hamill, T. M., and Juras J. , 2006: Measuring forecast skill: Is it real skill or is it the varying climatology? Quart. J. Roy. Meteor. Soc., 132, 29052923.

    • Search Google Scholar
    • Export Citation
  • Hamill, T. M., and Whitaker J. S. , 2006: Probabilistic quantitative precipitation forecasts based on reforecast analogs: Theory and application. Mon. Wea. Rev., 134, 32093229.

    • Search Google Scholar
    • Export Citation
  • Hamill, T. M., Whitaker J. S. , and Mullen S. L. , 2006: Reforecasts: An important dataset for improving weather predictions. Bull. Amer. Meteor. Soc., 87, 3346.

    • Search Google Scholar
    • Export Citation
  • Hersbach, H., 2000: Decomposition of the continuous ranked probability score for ensemble prediction systems. Wea. Forecasting, 15, 559570.

    • Search Google Scholar
    • Export Citation
  • Hong, Y., Adler R. F. , Hossain F. , Curtis S. , and Huffman G. J. , 2007: A first approach to global runoff simulation using satellite rainfall estimation. Water Resour. Res., 43, W08502, doi:10.1029/2006WR005739.

    • Search Google Scholar
    • Export Citation
  • Hopson, T. M., and Webster P. J. , 2010: A 1–10-day ensemble forecasting scheme for the major river basins of Bangladesh: Forecasting severe floods of 2003–07. J. Hydrometeor., 11, 618641.

    • Search Google Scholar
    • Export Citation
  • Hossain, F., and Katiyar N. , 2006: Improving flood forecasting in international river basins. Eos, Trans. Amer. Geophys. Union, 87, doi:10.1029/2006EO050001.

    • Search Google Scholar
    • Export Citation
  • Huffman, G. J., and Coauthors, 2007: The TRMM Multisatellite Precipitation Analysis (TMPA): Quasi-global, multiyear, combined-sensor precipitation estimates at fine scales. J. Hydrometeor., 8, 3855.

    • Search Google Scholar
    • Export Citation
  • Joyce, R. J., Janowiak J. E. , Arkin P. A. , and Xie P. , 2004: CMORPH: A method that produces global precipitation estimates from passive microwave and infrared data at high spatial and temporal resolution. J. Hydrometeor., 5, 487503.

    • Search Google Scholar
    • Export Citation
  • Liang, X., Lettenmaier D. P. , Wood E. F. , and Burges S. J. , 1994: A simple hydrologically based model of land surface water and energy fluxes for general circulation models. J. Geophys. Res., 99, 14 41514 428.

    • Search Google Scholar
    • Export Citation
  • Lohmann, D., Nolte-Holube R. , and Raschke E. , 1996: A large-scale horizontal routing model to be coupled to land surface parametrization schemes. Tellus, 48A, 708721.

    • Search Google Scholar
    • Export Citation
  • Lohmann, D., Raschke E. , Nijssen B. , and Lettenmaier D. P. , 1998: Regional scale hydrology: I. Formulation of the VIC-2L model coupled to a routing model. Hydrol. Sci. J., 43, 131141.

    • Search Google Scholar
    • Export Citation
  • Maurer, E. P., Wood A. W. , Adam J. C. , Lettenmaier D. P. , and Nijssen B. , 2002: A long-term hydrologically based dataset of land surface fluxes and states for the conterminous United States. J. Climate, 15, 32373251.

    • Search Google Scholar
    • Export Citation
  • Nash, J. E., and Sutcliffe J. V. , 1970: River flow forecasting through conceptual models. Part I—A discussion of principles. J. Hydrol., 10, 282290.

    • Search Google Scholar
    • Export Citation
  • Neil, J., Schumann G. , Bates P. D. , Buytaert W. , Matgen P. , and Pappenberger F. , 2009: A data assimilation approach to discharge estimation from space. Hydrol. Processes, 23, 36413649.

    • Search Google Scholar
    • Export Citation
  • O’Donnell, G. M., Czajkowski K. P. , Dubayah R. O. , and Lettenmaier D. P. , 2000: Macroscale hydrological modeling using remotely sensed inputs: Application to the Ohio River basin. J. Geophys. Res., 105 (D10), 12 49912 516.

    • Search Google Scholar
    • Export Citation
  • Palmer, T. N., and Coauthors, 2007: The ECMWF Ensemble Prediction System: Recent and on-going developments. ECMWF Research Dept. Tech. Memo. 540, 53 pp.

    • Search Google Scholar
    • Export Citation
  • Palmer, T. N., Buizza R. , Doblas-Reyes F. , Jung T. , Leutbecher M. , Shutts G. J. , Steinheimer M. , and Weisheimer A. , 2009: Stochastic parameterization and model uncertainty. ECMWF Research Dept. Tech. Memo. 598, 42 pp.

    • Search Google Scholar
    • Export Citation
  • Pappenberger, F., and Beven K. J. , 2006: Ignorance is bliss: Or seven reasons not to use uncertainty analysis. Water Resour. Res., 42, W05302, doi:10.1029/2005WR004820.

    • Search Google Scholar
    • Export Citation
  • Pappenberger, F., and Buizza R. , 2009: The skill of ECMWF predictions for hydrological modelling. Wea. Forecasting, 24, 749766.

  • Pappenberger, F., Beven K. J. , Hunter N. , Gouweleeuw B. , Bates P. , de Roo A. , and Thielen J. , 2005: Cascading model uncertainty from medium range weather forecasts (10 days) through a rainfall-runoff model to flood inundation predictions within the European Flood Forecasting System (EFFS). Hydrol. Earth Syst. Sci., 9, 381393.

    • Search Google Scholar
    • Export Citation
  • Pappenberger, F., Harvey H. , Beven K. , Hall J. , and Meadowcroft I. , 2006: Decision tree for choosing an uncertainty analysis methodology: A wiki experiment. Hydrol. Processes, 20, 37933798.

    • Search Google Scholar
    • Export Citation
  • Pappenberger, F., Buizza R. , Bodis K. , and Ghelli A. , 2009a: The skill of probabilistic forecasts under observational uncertainties within the generalized likelihood uncertainty estimation framework for hydrological applications. J. Hydrometeor., 10, 807819.

    • Search Google Scholar
    • Export Citation
  • Pappenberger, F., Cloke H. L. , Balsamo G. , Ngo-Duc T. , and Oki T. , 2009b: Global runoff routing with the hydrological component of the ECMWF NWP system. Int. J. Climatol., 30, 21552174, doi:10.1002/joc.2028.

    • Search Google Scholar
    • Export Citation
  • Schaake, J., and Coauthors, 2007: Precipitation and temperature ensemble forecasts from single-value forecasts. Hydrol. Earth Syst. Sci. Discuss., 4, 655717.

    • Search Google Scholar
    • Export Citation
  • Schumann, G., Bates P. D. , Horritt M. S. , Matgen P. , and Pappenberger F. , 2009: Progress in integration of remote sensing–derived flood extent and stage data and hydraulic models. Rev. Geophys., 47, RG4001, doi:10.1029/2008RG00027.

    • Search Google Scholar
    • Export Citation
  • Seo, D.-J., Herr H. , and Schaake J. C. , 2006: A statistical post-processor for accounting of hydrologic uncertainty in short-range ensemble streamflow prediction. Hydrol. Earth Syst. Sci. Discuss., 3, 19872035.

    • Search Google Scholar
    • Export Citation
  • Shepard, D. S., 1984: Computer mapping: The SYMAP interpolation algorithm. Spatial Statistics and Models, G. L. Gaile and C. J. Willmott, Eds., D. Reidel, 133–145.

    • Search Google Scholar
    • Export Citation
  • Shi, X., Wood A. W. , and Lettenmaier D. P. , 2008: How essential is hydrologic model calibration to seasonal streamflow forecasting? J. Hydrometeor., 9, 13501363.

    • Search Google Scholar
    • Export Citation
  • Su, F., Hong Y. , and Lettenmaier D. P. , 2008: Evaluation of TRMM Multi-satellite Precipitation Analysis (TMPA) and its utility in hydrologic prediction in La Plata basin. J. Hydrometeor., 9, 622640.

    • Search Google Scholar
    • Export Citation
  • Talagrand, O., and Vautard R. , 1997: Evaluation of probabilistic prediction systems. Proc. ECMWF Workshop on Predictability, Reading, United Kingdom, ECMWF, 1–25.

    • Search Google Scholar
    • Export Citation
  • Thielen, J., Bartholmes J. , Ramos M.-H. , and de Roo A. , 2009: The European Flood Alert System – Part 1: Concept and development. Hydrol. Earth Syst. Sci., 13, 125140.

    • Search Google Scholar
    • Export Citation
  • Thiemig, V., Pappenberger F. , Thielen J. , Gadain H. , de Roo A. , Bodis K. , Del Medico M. , and Muthusi F. , 2010: Ensemble flood forecasting in Africa: A feasibility study in the Juba-Shabelle River basin. Atmos. Sci. Lett., 11, 123131, doi:10.1002/asl.266.

    • Search Google Scholar
    • Export Citation
  • Uppala, S. M., and Coauthors, 2005: The ERA-40 Re-Analysis. Quart. J. Roy. Meteor. Soc., 131, 29613012.

  • Van der Knijff, J. M., Younis J. , and de Roo A. P. J. , 2008: LISFLOOD: A GIS-based distributed model for river basin scale water balance and flood simulation. Int. J. Geogr. Inf. Sci., 24, 189212, doi:10.1080/13658810802549154.

    • Search Google Scholar
    • Export Citation
  • Vitart, F., and Coauthors, 2008: The new VAREPS-monthly forecasting system: A first step towards seamless prediction. Quart. J. Roy. Meteor. Soc., 134, 17891799.

    • Search Google Scholar
    • Export Citation
  • Voisin, N., Schaake J. C. , and Lettenmaier D. P. , 2010: Adaptation of calibration and downscaling methods for quantitative ensemble precipitation forecasts. Wea. Forecasting, 25, 16031627.

    • Search Google Scholar
    • Export Citation
  • Westrick, K. J., Storck P. , and Mass C. F. , 2002: Description and evaluation of a hydrometeorological forecast system for mountainous watersheds. Wea. Forecasting, 17, 250262.

    • Search Google Scholar
    • Export Citation
  • White, D., Johnston K. , and Miller M. , 2005: The Ohio River basin. Rivers of North America, A. C. Benke and C. E. Cushing, Eds., Elsevier, 375–412.

    • Search Google Scholar
    • Export Citation
  • Wilks, D. S., 2006: Statistical Methods in the Atmospheric Sciences. 2nd ed. Academic Press, 627 pp.

  • Yadav, M., Wagener T. , and Gupta H. , 2007: Regionalization of constraints on expected watershed response behavior for improved predictions in ungauged basins. Adv. Water Resour., 30, 17561774, doi:10.1016/j.advwatres.2007.01.005.

    • Search Google Scholar
    • Export Citation
  • Yapo, P. O., Gupta H. V. , and Sorooshian S. , 1998: Multi-objective global optimization for hydrologic models. J. Hydrol., 204, 8397.

  • Yilmaz, K. K., Adler R. F. , Tian Y. , Hong Y. , and Pierce H. , 2010: Evaluation of a satellite-based global flood monitoring system. Int. J. Remote Sens., 31, 37633782, doi:10.1080/01431161.2010.483489.

    • Search Google Scholar
    • Export Citation
Save