• Alfieri, L., , Thielen J. , , and Pappenberger F. , 2012: Ensemble hydro-meteorological simulation for flash flood early detection in southern Switzerland. J. Hydrol., 424–425, 143153, doi:10.1016/j.jhydrol.2011.12.038.

    • Search Google Scholar
    • Export Citation
  • Alfieri, L., , Burek P. , , Dutra E. , , Krezeminski B. , , Murato D. , , Thielen J. , , and Pappenberger F. , 2013: GloFAS—Global ensemble streamflow forecasting and flood early warning. Hydrol. Earth Syst. Sci., 17, 11611175, doi:10.5194/hess-17-1161-2013.

    • Search Google Scholar
    • Export Citation
  • Cloke, H. L., , and Pappenberger F. , 2009: Ensemble flood forecasting: A review. J. Hydrol., 375, 613626, doi:10.1016/j.jhydrol.2009.06.005.

    • Search Google Scholar
    • Export Citation
  • Coe, M. T., , Costa M. H. , , and Howard E. A. , 2008: Simulating the surface waters of the Amazon River basin: Impacts of new river geomorphic and flow parameterizations. Hydrol. Processes, 22, 25422553, doi:10.1002/hyp.6850.

    • Search Google Scholar
    • Export Citation
  • Cuo, L., , Pagano T. C. , , and Wang Q. J. , 2011: A review of quantitative precipitation forecasts and their use in short- to medium-range streamflow forecasting. J. Hydrometeor., 12, 713728, doi:10.1175/2011JHM1347.1.

    • Search Google Scholar
    • Export Citation
  • Dietrich, J., , Trepte S. , , Wang Y. , , Shumann A. H. , , Voß F. , , Hesser F. B. , , and Denhard M. , 2008: Combination of different types of ensembles for the adaptive simulation of probabilistic flood forecasts: Hindcasts for the Mulde 2002 extreme event. Nonlinear Processes Geophys., 15, 275286, doi:10.5194/npg-15-275-2008.

    • Search Google Scholar
    • Export Citation
  • Dole, R., and Coauthors, 2011: Was there a basis for anticipating the 2010 Russian heat wave? Geophys. Res. Lett., 38, L06702, doi:10.1029/2010GL046582.

    • Search Google Scholar
    • Export Citation
  • Hoffman, R. N., , and Kalnay E. , 1983: Lagged average forecasting, and alternative to Monte Carlo forecasting. Tellus, 35A, 100118, doi:10.1111/j.1600-0870.1983.tb00189.x.

    • Search Google Scholar
    • Export Citation
  • Hong, C.-C., , Hsu H.-H. , , Lin N.-H. , , and Chiu H. , 2011: Roles of European blocking and tropical-extratropical interaction in the 2010 Pakistan flooding. Geophys. Res. Lett., 38, L13806, doi:10.1029/2011GL047583.

    • Search Google Scholar
    • Export Citation
  • Kain, J. S., , and Fritsch J. M. , 1993: Convective parameterization for mesoscale models: The Kain–Fritsch scheme. The representation of cumulus convection in numerical models. The Representation of Cumulus Convection in Numerical Models, Meteor. Monogr., No. 46, Amer. Meteor. Soc., 165170.

    • Search Google Scholar
    • Export Citation
  • Kalnay, E., 2003: Atmospheric Modeling, Data Assimilation and Predictability. Cambridge University Press, 341 pp.

  • Kubota, T., , Ushio T. , , Shige S. , , Kida S. , , Kachi M. , , and Okamoto K. , 2009: Verification of high-resolution satellite-based rainfall estimates around Japan using a gauge-calibrated ground-radar dataset. J. Meteor. Soc. Japan, 87A, 203222, doi:10.2151/jmsj.87A.203.

    • Search Google Scholar
    • Export Citation
  • Lau, W. K. M., , and Kim K.-M. , 2012: The 2010 Pakistan flood and Russian heat wave: Teleconnection of hydrometeorologic extremes. J. Hydrometeor., 13, 392403, doi:10.1175/JHM-D-11-016.1.

    • Search Google Scholar
    • Export Citation
  • Lehner, B., , Verdin K. , , and Jarvis A. , 2008: New global hydrography derived from spaceborne elevation data. Eos, Trans. Amer. Geophys. Union, 89, 9394, doi:10.1029/2008EO100001.

    • Search Google Scholar
    • Export Citation
  • Marsigli, C., , Boccanera F. , , Montani A. , , and Paccagnella T. , 2005: The COSMO-LEPS mesoscale ensemble system: Validation of the methodology and verification. Nonlinear Processes Geophys., 12, 527536, doi:10.5194/npg-12-527-2005.

    • Search Google Scholar
    • Export Citation
  • Matsueda, M., 2011: Predictability of Euro-Russian blocking in summer of 2010. Geophys. Res. Lett.,38, L06801, doi:10.1029/2010GL046557.

  • Okamoto, K., , Iguchi T. , , Takahashi N. , , Iwanami K. , , and Ushio T. , 2005: The global satellite mapping of precipitation (GSMaP) project. Proc. 2005 IEEE Geoscience and Remote Sensing Symp., Vol. 5, Seoul, South Korea, IEEE, 3414–3416, doi:10.1109/IGARSS.2005.1526575.

  • Rawls, W. J., , Ahuja L. R. , , Brakensiek D. L. , , and Shirmohammadi A. , 1992: Infiltration and soil water movement. Handbook of Hydrology, D. R. Maidment, Ed., McGraw-Hill, 5.1–5.51.

    • Search Google Scholar
    • Export Citation
  • Sayama, T., , Ozawa G. , , Kawakami T. , , Nabesaka S. , , and Fukami K. , 2012: Rainfall-runoff- inundation analysis of the 2010 Pakistan flood in the Kabul River basin. Hydrol. Sci. J., 57, 298312, doi:10.1080/02626667.2011.644245.

    • Search Google Scholar
    • Export Citation
  • Shige, S., , Kida S. , , Ahiwake H. , , Kubota T. , , and Aonashi K. , 2013: Improvement of TMI rain retrieval in mountainous areas. J. Appl. Meteor. Climatol., 52, 242254, doi:10.1175/JAMC-D-12-074.1.

    • Search Google Scholar
    • Export Citation
  • Shiraishi, Y., , Fukami K. , , and Inomata H. , 2009: A proposal of correction method using the movement of rainfall area on satellite-based rainfall information by analysis in the Yoshino River basin (in Japanese with English abstract). Annu. J. Hydraul. Eng.,53, 385–390.

  • Shrestha, M. S., , Takara K. , , Kubota T. , , and Bajracharya S. R. , 2011: Verification of GSMaP rainfall estimates over the central Himalayas. J. Japan Soc. Civil Eng., Ser. B1,67, I-37–I-42.

  • Taniguchi, A., and Coauthors, 2013: Improvement of high-resolution satellite rainfall product for Typhoon Morakot (2009) over Taiwan. J. Hydrometeor., 14, 18591871, doi:10.1175/JHM-D-13-047.1.

    • Search Google Scholar
    • Export Citation
  • Thielen, J., , Bartholmes J. , , Ramos M.-H. , , and de Roo A. , 2009a: The European flood alert system—Part 1: Concept and development. Hydrol. Earth Syst. Sci., 13, 125140, doi:10.5194/hess-13-125-2009.

    • Search Google Scholar
    • Export Citation
  • Thielen, J., , Bogner K. , , Pappenberger F. , , Kalas M. , , del Medico M. , , and de Roo A. , 2009b: Monthly-, medium-, and short-range flood warning: Testing the limits of predictability. Meteor. Appl., 16, 7790, doi:10.1002/met.140.

    • Search Google Scholar
    • Export Citation
  • Tian, Y., , Peters-Lidard C. D. , , Adler R. F. , , Kubota T. , , and Ushio T. , 2010: Evaluation of GSMaP precipitation estimates over the contiguous United States. J. Hydrometeor., 11, 566574, doi:10.1175/2009JHM1190.1.

    • Search Google Scholar
    • Export Citation
  • United Nations, cited 2010: Pakistan Floods Emergency Response Plan. [Available online at http://ochaonline.un.org/humanitarianappeal/.]

  • Webster, P. J., and Coauthors, 2010: Extended-range probabilistic forecasts of Ganges and Brahmaputra floods in Bangladesh. Bull. Amer. Meteor. Soc., 91, 14931514, doi:10.1175/2010BAMS2911.1.

    • Search Google Scholar
    • Export Citation
  • Webster, P. J., , Toma V. E. , , and Kim H.-M. , 2011: Were the 2010 Pakistan floods predictable? Geophys. Res. Lett., 38, L04806, doi:10.1029/2010GL046346.

    • Search Google Scholar
    • Export Citation
  • WFP, 2010: Preliminary rapid damage assessment in the agriculture sector for flood-affected areas of Pakistan. World Food Program Rep., 27 pp. [Available online at https://www.cimicweb.org/cmo/pf2010/Documents/Assessments/AgricultureCluster_Preliminary.pdf.]

  • Zappa, M., and Coauthors, 2008: MAP D-PHASE: Real-time demonstration of hydrological ensemble prediction systems. Atmos. Sci. Lett., 9, 8087, doi:10.1002/asl.183.

    • Search Google Scholar
    • Export Citation
  • View in gallery

    Weather chart for 500-hPa height during the Pakistan flood. Color contours show geopotential height anomalies (m) with negative anomalies identified by dashed contours, and stream lines show airflow, averaged from 27 to 30 July 2010.

  • View in gallery

    Map of the model domain for the NWP and RRI model: (top) the outer frame at 20-km grid resolution and (bottom) the inner frame at 5-km grid resolution. The bold solid irregular line in the bottom panel shows the Kabul River basin with rivers and tributaries, with a small square showing the Peshawar valley. A dashed rectangle in the lower frame shows the area used to compute the rainfall in northern Pakistan. The thinner solid lines show the coast and borders, and the gray shading shows terrain height.

  • View in gallery

    Design of the lagged ensemble predictions. The wide horizontal arrows indicate forecast runs performed from different initial times from 0000 UTC 23 July to 0000 UTC 28 July. Four different runs per day were performed. The heavy rainfall occurred during 27–29 July.

  • View in gallery

    Schematic diagram of the RRI model.

  • View in gallery

    Total rainfall from 0300 UTC 27 July to 0300 UTC 30 July 2010, observed using ground rain gauges by the PMD. The color shading is computed by multiple Cressman objective analyses (in millimeters). The white star indicates the location of Peshawar. The bold broken line indicates the Kabul River basin.

  • View in gallery

    (top left to bottom right) Accumulated rainfall from 0000 UTC 23 July to 0000 UTC 28 July 2010 from NCEP GFS forecasting with various initial times. The result from 0600 UTC 23 July was omitted to simplify the display. The color shadings of dark blue, light blue, green, yellow-green, yellow, orange, red, and pink indicate 0–50, 50–100, 100–150, 150–200, 200–250, 250–300, 300–350, and over 350 mm, respectively. The area in each map enclosed by the red line indicates the Kabul River basin. The map titles above each panel are enclosed red when the rainfall in the Kabul River basin is forecasted to be above specified criteria (see text).

  • View in gallery

    As in Fig. 6, but with maps based on downscaled forecasts.

  • View in gallery

    Probability distribution of average rain rates within dashed square area in Fig. 2 (69°–76°N, 30°–36°N) by (a) GFS original forecasts and (b) downscaled forecasts. The gray boxes show the 25th to 75th percentiles and the bars show the minimum to maximum values. Dot–dashed lines and the dashed lines in both panels are the median and GSMaP rainfall corrected by the ground rain gauge, respectively.

  • View in gallery

    As in Fig. 8, but for the Kabul River basin.

  • View in gallery

    Simulated total rainfall vs forecast initial time for (a) the rectangle area in Fig. 2 and (b) the Kabul River basin. The light gray bars indicate the GFS original forecasts, and the dark gray bars indicate those of the downscaled forecasts. The horizontal line in each panel shows the corrected GSMaP rainfall.

  • View in gallery

    As in Fig. 8, but for simulated discharges at the outlet of the Kabul River basin by the RRI model.

  • View in gallery

    Threat score of area inundation as a function of initial time. The dashed line with the × marks shows the threat score of the inundated area based on GFS rainfall, and the solid line with the circle marks shows the threat score based on the downscaled rainfall by WRF. A horizontal line at a score of 0.615 is the score based on the corrected GSMaP rainfall.

  • View in gallery

    Probability distribution simulated by the RRI model based on GFS original rainfall. Probability distributions in discharge at the outlet of the Kabul River basin are shown, calculated from lagged ensemble forecasts composed by 13 runs in three time periods: (a) 0000 UTC 23 July to 0000 UTC 26 July, (b) 0000 UTC 24 July to 0000 UTC 27 July, and (c) 0000 UTC 25 July to 0000 UTC 28 July. The color shading shows accumulated probability density with higher probabilities in warm colors. The black line in each panel is the ensemble mean and the green line is rainfall forecasted by the corrected GSMaP. (d)–(f) Predicted inundation probability distribution corresponding to the three time periods. The location is the Peshawar valley near the outlet of the Kabul River basin (the small rectangle in Fig. 2, bottom). The color shading indicates the probability of inundation depth over 1 m. The black line indicates the inundation map from the observed indicator of the MODIS satellite.

  • View in gallery

    As in Fig. 13, but for streamflow simulation by downscaled rainfall.

All Time Past Year Past 30 Days
Abstract Views 0 0 0
Full Text Views 112 112 14
PDF Downloads 106 106 16

Numerical Simulation of 2010 Pakistan Flood in the Kabul River Basin by Using Lagged Ensemble Rainfall Forecasting

View More View Less
  • 1 International Center for Water Hazard and Risk Management, Public Works Research Institute, Tsukuba, Japan
© Get Permissions
Full access

Abstract

Lagged ensemble forecasting of rainfall and rainfall–runoff–inundation (RRI) forecasting were applied to the devastating flood in the Kabul River basin, the first strike of the 2010 Pakistan flood. The forecasts were performed using the Global Forecast System of the National Centers for Environmental Prediction (NCEP-GFS) and were provided four times daily. Dynamical downscaling was also applied to the forecasts by the Weather Research and Forecasting Model (WRF), a regional model. The forecasts of the rainfall and inundation area were verified by comparing rain gauge–corrected Global Satellite Mapping of Precipitation (GSMaP) data and the observed indicator of an inundation map based on Moderate Resolution Imaging Spectroradiometer (MODIS) satellite data. The GFS predicted a sign of heavy rainfall in northern Pakistan 4 days ahead of the onset. However, most of the forecasts predicted it in wrong places, and only those performed after the rainfall onset predicted it in the accurate location. Downscaling corrected the locations of the misplaced GFS forecasts and also underestimated or overestimated rainfall amount derived from GFS. Finally, downscaled forecasts predicted a reliable amount of rainfall in the Kabul River basin 1 day ahead of the rainfall onset and predicted a high probability of heavy rainfall 3 days ahead. Lagged ensemble forecasts of discharge and inundation distribution based on GFS rainfall predicted the probability of the actual discharge and inundation distribution, but in low reliability. The reliability substantially improved when downscaled rainfall was used. The reliability of the flood alert system combining NCEP-GFS, dynamical downscaling by WRF, and the RRI model was at an acceptable level in this study.

Corresponding author address: Tomoki Ushiyama, International Center for Water Hazard and Risk Management under the auspices of UNESCO, Public Works Research Institute, 1-6, Minamihara, Tsukuba, Ibaraki 305-8516, Japan. E-mail: ushiyama55@pwri.go.jp

Abstract

Lagged ensemble forecasting of rainfall and rainfall–runoff–inundation (RRI) forecasting were applied to the devastating flood in the Kabul River basin, the first strike of the 2010 Pakistan flood. The forecasts were performed using the Global Forecast System of the National Centers for Environmental Prediction (NCEP-GFS) and were provided four times daily. Dynamical downscaling was also applied to the forecasts by the Weather Research and Forecasting Model (WRF), a regional model. The forecasts of the rainfall and inundation area were verified by comparing rain gauge–corrected Global Satellite Mapping of Precipitation (GSMaP) data and the observed indicator of an inundation map based on Moderate Resolution Imaging Spectroradiometer (MODIS) satellite data. The GFS predicted a sign of heavy rainfall in northern Pakistan 4 days ahead of the onset. However, most of the forecasts predicted it in wrong places, and only those performed after the rainfall onset predicted it in the accurate location. Downscaling corrected the locations of the misplaced GFS forecasts and also underestimated or overestimated rainfall amount derived from GFS. Finally, downscaled forecasts predicted a reliable amount of rainfall in the Kabul River basin 1 day ahead of the rainfall onset and predicted a high probability of heavy rainfall 3 days ahead. Lagged ensemble forecasts of discharge and inundation distribution based on GFS rainfall predicted the probability of the actual discharge and inundation distribution, but in low reliability. The reliability substantially improved when downscaled rainfall was used. The reliability of the flood alert system combining NCEP-GFS, dynamical downscaling by WRF, and the RRI model was at an acceptable level in this study.

Corresponding author address: Tomoki Ushiyama, International Center for Water Hazard and Risk Management under the auspices of UNESCO, Public Works Research Institute, 1-6, Minamihara, Tsukuba, Ibaraki 305-8516, Japan. E-mail: ushiyama55@pwri.go.jp

1. Introduction

During the summer of 2010, Pakistan experienced the worst flood disaster in history. The westward shift of monsoonal rainfall caused extraordinarily severe large-scale floods, resulting in a devastating disaster over a wide area of the country. The series of floods killed more than 1700 people and affected 18 million people along the Indus River (United Nations 2010). The flooding first struck the northwestern part of Pakistan, particularly the province of Khyber Pakhtunkhwa (KPK). The floodwaters then traveled southward along the Indus River, where large-scale river floods caused severe damage to houses, crops, and livestock (WFP 2010). Torrential downpour in KPK caused severe floods in the Kabul River basin (92 605 km2) and also a large-scale flood inundation in the Peshawar valley, which was the beginning of the enormous disaster in Pakistan. At Peshawar, a 274-mm rainfall was recorded on 29 July, which was a record-breaking rainfall in this arid region. This daily rainfall corresponds to almost 80% of the annual rainfall (346 mm) at Peshawar. The floods killed 1156 people and affected 3.8 million people in KPK. As shown in Sayama et al. (2012), the river discharge and peak discharge timing from the Kabul River basin was influenced by inundated water in the basin, and this discharge and timing are important for flood prediction over the downstream Indus River. Therefore, we focused on simulation of the Kabul River basin flooding in this study.

The record rainfall occurred because of unusual climate conditions in the summer of 2010. As shown in Fig. 1, a blocking high had stayed over eastern Europe for nearly 2 months, since mid-June 2010, causing a Russian heat wave (Dole et al. 2011; Lau and Kim 2012). The blocking high formed a deep trough southward, reaching over northern Pakistan. The southward penetration of upper-level vorticity perturbations in this deep trough contributed to the development of heavy rainfall events over northern Pakistan and its vicinity. Monsoon surges triggered by an intraseasonal variation strike increased moisture transport simultaneously, which helped evolve those heavy rainfall events (Hong et al. 2011; Lau and Kim 2012).

Fig. 1.
Fig. 1.

Weather chart for 500-hPa height during the Pakistan flood. Color contours show geopotential height anomalies (m) with negative anomalies identified by dashed contours, and stream lines show airflow, averaged from 27 to 30 July 2010.

Citation: Journal of Hydrometeorology 15, 1; 10.1175/JHM-D-13-011.1

Flood forecasting is an effective tool to mitigate flood damage. Flood forecasting in the lower Indus River can be conducted through flood routing based on upstream discharge and water level. However, flood forecasting in upstream basins such as the Kabul River basin requires rainfall distribution monitoring and streamflow forecasting by using a rainfall runoff model; it is quite difficult to monitor rainfall in the vast Kabul River basin with an insufficient number of rain gauges (three gauges in the 92 605 km2 catchment). Sayama et al. (2012) attempted to overcome this difficulty by using satellite-observed rainfall, and they successfully simulated inundation distribution that was consistent with a Moderate Resolution Imaging Spectroradiometer (MODIS) inundation map. In addition, because these floods are characterized by a drastic increase in water level (from about 2000 m3 s−1 on 28 July to 15 000 m3 s−1 on 31 July) after rainfall, forecasting should be performed with a longer lead time. For this purpose, the application of numerical weather prediction (NWP) to streamflow forecasting is desired (Cloke and Pappenberger 2009; Cuo et al. 2011).

Coupling NWP with streamflow simulation has been attempted by advanced flood-forecasting centers and hydrologists in the past decade since the prediction capability of NWP has steadily increased as ensemble forecasting and data assimilation techniques have improved (Cloke and Pappenberger 2009; Cuo et al. 2011). However, the predictability of finescale rainfall distribution has not seen much improvement yet, even with a medium-range ensemble prediction system (EPS) developed by the European Centre for Medium-Range Weather Forecasts (ECMWF-EPS), which has one of the best forecasting scores among all of the operational weather forecasting centers. The applicability of finescale rainfall distribution is highly dependent on the scale or catchment size of a target basin because larger catchments round off the error of rainfall forecasts. Webster et al. (2010) demonstrated the effectiveness of flood forecasting systems with 10-day lead time by using bias-corrected ECMWF-EPS and streamflow simulation for the Brahmaputra River in Bangladesh.

For European rivers with much smaller catchments, smaller-scale limited area models are more adapted to the size of the basins. The limited-area ensemble prediction system (LEPS) of the Consortium for Small-Scale Modeling (COSMO-LEPS; Marsigli et al. 2005) is an example developed for regional probabilistic flood forecasting. LEPS downscales 16 selected members from ECMWF-EPS once a day in 7-km horizontal resolution. It is used in several European countries to provide rainfall forecasts in streamflow simulation models to accomplish a flood alert system (Cloke and Pappenberger 2009). The European Commission Joint Research Centre (JRC) developed the European Flood Awareness System (EFAS) for flood warning in transnational European river basins (Thielen et al. 2009a; Alfieri et al. 2013). Both ECMWF-EPS and COSMO-LEPS rainfall forecasting have been adapted to the EFAS model to form a flood alert system (Cloke and Pappenberger 2009). Thielen et al. (2009b) showed, in a case study of a Romanian flood, that adopting multiple EPSs into EFAS can provide a better skill score than deterministic forecasts and can give flood warning information 9–11 days in advance. As shown above, coupling of EPS with streamflow forecasts is increasingly introduced to operations and research to give added values to flood warning.

Webster et al. (2011) argued, after analyzing the probability distribution of rainfall predictions by using ECMWF-EPS, that the rainfall in northern Pakistan in the end of July 2010 was predictable 6–8 days earlier. More specifically, large-scale heavy rainfall distributions in northern Pakistan had been well predicted 4 days before the onset of the rainfall, and a good predictive skill of the average rainfall was found up to 6 days in advance. Alfieri et al. (2013) also showed the predictability of severe floods in the middle of the Indus River at least 3 days ahead of rainfall in their global ensemble streamflow forecasting.

Sayama et al. (2012) attempted to calculate flood inundation distribution in the Kabul River basin via the rainfall–runoff–inundation (RRI) model, using the rainfall data from the rain gauge–corrected Global Satellite Mapping of Precipitation (GSMaP) as forcing (Okamoto et al. 2005). They showed that the RRI model was capable of predicting flood inundation distribution that was consistent with the flood inundation extent derived from satellite-based data analysis by the United Nations Institute for Training and Research (UNITAR) Operational Satellite Applications Programme (UNOSAT) in the use of MODIS and was also consistent with the housing damage distribution by the United Nations Office for the Coordination of Humanitarian Affairs (OCHA). Even though it is known that satellite-derived rainfall includes uncertainty, their study quantified the usefulness of the rain gauge–corrected GSMaP rainfall and RRI model.

As Webster et al. (2011) and Sayama et al. (2012) suggested, it should be possible to forecast the Kabul River basin flood several days before the rainfall onset by coupling NWP rainfall and RRI model simulation. The purpose of the study, therefore, is to examine this hypothesis.

We also examine improvement in forecasting capability after downscaling the global NWP by using a regional weather forecasting model. The improvement is evaluated in terms of the location and amount of forecasted rainfall, as well as forecasted discharge. The prediction capability of the RRI simulation with NWP rainfall and downscaled global NWP for discharge and inundation distribution is discussed not in terms of a single forecast, but in terms of ensemble mean and variance. We chose lagged ensemble using the Global Forecast System (GFS) of the National Centers for Environmental Prediction (NCEP-GFS), since this type of ensemble is well suited for operational flood forecasting. NCEP-GFS is updated every 6 hours, while ECMWF operational forecast is updated every 12 h; therefore, NCEP-GFS is more suitable to make lagged ensemble with higher time resolution. It is computed with a relatively high horizontal resolution (about 27 km). Furthermore, lagged ensemble is composed of forecasts from updated initial conditions; thus, it can be performed in coordination with daily operations, which updates forecasts.

The scientific questions in this study are as follows:

  1. How well did the global NWP predict the heavy rainfall in the Kabul River basin?
  2. How well did a combination of the global NWP and a streamflow simulation predict the discharge and the inundation area in the lower Kabul River basin?
  3. How much did the downscaled forecasts improve the location and magnitude of rainfall?
  4. How much did the downscaled forecasts improve the reliability of the discharge and inundation forecasts estimated from the lagged ensemble?

Section 2 introduces a method for the numerical prediction of rainfall and RRI. Section 3 shows results in NWP and RRI. Section 4 discusses the results, and section 5 concludes the study.

2. Method

In this study, GFS rainfall forecasts and their downscaled version by the Weather Research and Forecasting Model (WRF), a regional weather prediction model, were used in the RRI model as input for flood forecasting. We evaluated the reliability of forecast runs with different initial times by means of lagged ensemble forecasting. The initial and boundary conditions were created from NCEP-GFS. The target was the Kabul basin flood that occurred from 28 July to 1 August 2010 because of the extraordinarily heavy rainfall from 27 to 29 July.

a. Design of dynamical downscaling

The numerical domain for the regional model for downscaling was defined with its center in the Kabul River basin to include all of Pakistan with 20-km horizontal resolution (Fig. 2). To simulate more realistic rainfall with our available resources, a two-way nested inner frame was defined to enclose the Kabul River basin with 5-km resolution. The 5-km horizontal resolution is considerably finer than operational global forecast models (17–27-km resolutions) and is good enough to compute a mesoscale cloud system, although it is not fine enough to compute severe cumulonimbus. The outer frame had a domain of about 4000 km × 3000 km, and the inner frame had a domain of about 1245 km × 1125 km, both with 28 layers. In both frames, the Kain–Fritsch cumulus parameterization (Kain and Fritsch 1993) was used to predict rainfall since the horizontal grid resolution was more than 5 km (Table 1). The WRF single-moment 3-class microphysics was used to predict water vapor, cloud water–ice, and rain–snow with a simple ice scheme. A thermal diffusion with a five-layer soil temperature model was used for the land surface process according to the characteristics of soil types selected from 24 categories of U.S. Geological Survey (USGS) land-use maps.

Fig. 2.
Fig. 2.

Map of the model domain for the NWP and RRI model: (top) the outer frame at 20-km grid resolution and (bottom) the inner frame at 5-km grid resolution. The bold solid irregular line in the bottom panel shows the Kabul River basin with rivers and tributaries, with a small square showing the Peshawar valley. A dashed rectangle in the lower frame shows the area used to compute the rainfall in northern Pakistan. The thinner solid lines show the coast and borders, and the gray shading shows terrain height.

Citation: Journal of Hydrometeorology 15, 1; 10.1175/JHM-D-13-011.1

Table 1.

Model setting for downscaling by WRF.

Table 1.

Figure 3 introduces a design of lagged ensemble forecasting by GFS and WRF. GFS originally conducted NWP with a horizontal resolution of 27 km (a technical implementation notice is available at http://www.emc.ncep.noaa.gov/GFS/doc.php), but NCEP provided coarser-resolution data in a horizontal resolution of 0.5° with 26 pressure levels and surface. Those data were used to determine the initial and boundary conditions for WRF. GFS data were provided four times daily, that is, at 0000, 0600, 1200, and 1800 UTC, at 3-h intervals up to 180 h of forecasting (7.5 days). Therefore, four different forecasts per day were calculated. Since the target rainfall occurred from 27 to 29 July, we calculated forecasts with the initial conditions from 0000 UTC 23 July to 0000 UTC 28 July, performing 21 forecast runs in total.

Fig. 3.
Fig. 3.

Design of the lagged ensemble predictions. The wide horizontal arrows indicate forecast runs performed from different initial times from 0000 UTC 23 July to 0000 UTC 28 July. Four different runs per day were performed. The heavy rainfall occurred during 27–29 July.

Citation: Journal of Hydrometeorology 15, 1; 10.1175/JHM-D-13-011.1

Lagged ensemble prediction is a simple way of ensemble forecasting compared to common EPS. Common EPS computes a number of forecasts started from a same point of time, while the lagged ensemble is a set of forecasts started from different points of time. Common EPS needs initial perturbations added to their initial conditions to create different growth modes of ensemble members. The initial perturbations are created as the number of ensemble members covers the possibilities of fast atmospheric growth modes. Large computational resources are required to create the initial perturbations as well as computation of the number of ensemble forecasts. Therefore, the common EPS is performed mainly in weather centers in developed countries (Kalnay 2003).

On the other hand, for the lagged ensemble forecasts perturbations are generated automatically from forecast errors (Hoffman and Kalnay 1983). It does not need further computational resources to make an ensemble. However, this type of ensemble prediction does not always cover initial perturbations of fast-growing modes, and the ensemble average may be tainted by older forecasts. However, lagged ensemble prediction can still work effectively, especially for cases with limited computational resources.

b. Rainfall–runoff–inundation model

The RRI model used in this study was developed and evaluated in Sayama et al. (2012). It is a distributed flood forecasting model that calculates not only river flow, but also inundation distribution, which typical distributed rainfall–runoff models cannot simulate. Since the effect of stream water loss due to flood inundation is not negligible in a simulation of large-scale river flooding, the RRI model, capable of calculating inundation distribution, is suited for flood forecasting for large-scale river flow with flooding such as the Kabul River flood (Sayama et al. 2012).

This paper only describes an overview and key components of the RRI model. Interested readers can find detailed descriptions in Sayama et al. (2012). Figure 4 shows a schematic diagram of the RRI model. The model deals with slopes and river channels separately. At a grid cell on which a river channel is located, the model assumes that both slope and river are positioned within the same grid cell. A channel is discretized as a single vector along its centerline of the overlying slope grid cell. The channel represents an extra flow path between grid cells lying over the actual river course. Lateral flows are simulated on slope cells on a two-dimensional basis. Slope grid cells on the river channel have two water depths: one for the channel and the other for the slope (or floodplain) itself. The inflow–outflow interaction between the slope and river is calculated based on different overflowing formulae depending on water-level and levee-height conditions.

Fig. 4.
Fig. 4.

Schematic diagram of the RRI model.

Citation: Journal of Hydrometeorology 15, 1; 10.1175/JHM-D-13-011.1

The equations to calculate lateral flows are based on a mass balance equation and momentum equations. They are modified under diffusion wave approximation combined with Manning’s equation. By using the modified momentum equations and the discretized mass balance equation, water depths and discharges are calculated for each grid cell for each time step. The RRI model calculates both surface and subsurface flows with the same algorithm made up with saturation excess overland flow and saturated subsurface flow. This model assumes the water surface slope as the hydraulic gradient, which is different from the kinematic wave models that assume that the hydraulic gradient is equal to the topographic slope.

The infiltration model used in this study is adjusted to the Kabul River basin, which has plateaus and valleys surrounded by mountains with steep and rugged mountain slopes maintained by strong physical weathering. The cover soil category is Leptosol, which is characterized by a very limited soil-forming process and hence a shallow soil depth. The unstable rocky slopes covered with weathered material are expected to allow rainwater to flow as quick subsurface flow or saturation excess overland flow. In contrast, the plains in the basin are covered with terrestrial and lacustrine sediments. The soil type is categorized as a Cambisol, which is characterized by incipient soil formation with the texture of sandy loam or finer. We added a simple infiltration component module based on the Green–Ampt infiltration model (Rawls et al. 1992) to consider vertical infiltration processes for a flood event simulation. For the distinction between mountains and plains, a slope of 0.05 is assumed as the threshold suggested from a geologic map and a topographic slope distribution map.

The Kabul River basin is estimated to drain an area of 92 605 km2, according to the topographic data of Hydrological Data and Maps Based on Shuttle Elevation Derivatives at Multiple Scales (HydroSHEDS; 30 s) (Lehner et al. 2008). The resolution was approximately 761 m × 924 m in this region, and the number of grid cells was 131 489 in the entire Kabul River basin. Since surveyed data on the river cross section were not available to us, we used the following simple regression equations (Coe et al. 2008): width (m) = 2.5A0.4 and depth (m) = 0.6A0.4, where A (km2) is the flow accumulation area. The coefficients of the river width equation are based on a satellite image of river width, and those of the depth equation are based on our experience. The levee height was set to zero since no apparent levees were found in the region. There is a large dam, Warsak Dam, at the Kabul River approximately 20 km upstream from Peshawar. We confirmed that the effect of the Warsak Dam for flood control is negligible, as this dam is for hydropower generation and the inflow and outflow is therefore nearly equal in flood time. The model was calibrated as it simulated inundation distribution consistent with the inundation map by UNOSAT. A more detailed description is found in Sayama et al. (2012).

c. Data for verification

We used GSMaP rainfall data and inundation distribution analysis released from UNOSAT based on the MODIS satellite for verification of rainfall and inundation distribution. GSMaP is a project to retrieve rainfall in high time and space resolutions solely from satellite observations (Okamoto et al. 2005). The data are provided at 0.1° resolution at 1-h intervals. The rainfall is retrieved from multifrequency microwave radiometer observations, based on a lookup table between brightness temperatures and optimum surface rain rates. These fine spatial and temporal resolutions are achieved by merging the forward-moving vectors of rainfall distributions estimated by infrared radiometers. GSMaP data are highly useful in sparsely rain gauged areas such as the study region in northern Pakistan and Afghanistan.

Despite the notable usefulness of satellite-based rainfall, some well-known disadvantages are the tendency to underestimate rainfall, particularly during severe storms or orographic rainfall (Kubota et al. 2009; Shiraishi et al. 2009; Shige et al. 2013; Taniguchi et al. 2013); the tendency to underestimate rainfall in continental high mountain regions (Shrestha et al. 2011); and low performance of quantitative precipitation estimates over the United States (Tian et al. 2010). In case of the rainfall in this study, the rain rates were significantly underestimated, as in Shrestha et al. (2011), and therefore, correction based on ground rain gauge data was required (Sayama et al. 2012). Ground rain gauge data were available from the Pakistan Meteorological Department (PMD) on a daily basis (e.g., Fig. 5). We corrected the GSMaP hourly rain rates based on the daily total ground rain gauge data from PMD, as shown in Fig. 5. The correction rates were determined for the respective locations. The corrected GSMaP rain rates agreed well with a Tropical Rainfall Measuring Mission (TRMM) Precipitation Radar (PR) rain-rate image.

Fig. 5.
Fig. 5.

Total rainfall from 0300 UTC 27 July to 0300 UTC 30 July 2010, observed using ground rain gauges by the PMD. The color shading is computed by multiple Cressman objective analyses (in millimeters). The white star indicates the location of Peshawar. The bold broken line indicates the Kabul River basin.

Citation: Journal of Hydrometeorology 15, 1; 10.1175/JHM-D-13-011.1

For verification of inundation distribution, a remote sensing inundation map based on the MODIS satellite data released from UNOSAT was employed. The inundation map by UNOSAT was made by analysis based on post-disaster images captured on 31 July and 1 August and pre-crisis images on 25–26 July. Since the UNOSAT analysis may involve uncertainty due to cloud interception, we compared it with the inundation extent analysis based on a RADARSAT-2 image captured on 31 July [obtained from Service Régional de Traitement d’Image et de Télédétection (SERTIT), http://sertit.u-strasbg.fr/]. The inundation distribution by the UNOSAT analysis covered all the inundation area by RADARSAT-2 analysis or even slightly overestimated. Thus, the UNOSAT inundation analysis is reliable as an observed indicator of maximum inundation extent.

3. Results

a. Rainfall distribution

This section describes rainfall distribution in the Kabul River basin derived from ground rain gauge observation and NWP. Figure 5 introduces rainfall accumulation from 0300 UTC 27 July to 0300 UTC 29 July 2010. It shows a high concentration of rainfall over 300 mm around the Kabul River basin. The rainfall recorded during the 3 days was close to the annual average, as mentioned in the introduction of this paper. The floods in the Peshawar Valley occurred from 30 July to 2 August because of a sudden increase of the Kabul River discharge (Sayama et al. 2012). In other words, the floods occurred 1 day after the rainfall, causing devastating disasters.

Figure 6 shows the forecasts of the accumulated rainfall from 0000 UTC 27 July to 0000 UTC 29 July by NCEP-GFS. The starting time of those forecasts varies from 0000 UTC 23 July to 0000 UTC 28 July, as shown in Fig. 3. Most of the forecasts show the approximate rainfall concentration in northern Pakistan but are different in rainfall amount and location. NCEP-GFS predicted less rainfall in forecast runs with earlier initial times and more rainfall in forecast runs with later initial times. In addition, although the forecasts for 25–26 July and 1200 UTC 27 July show a large amount of rainfall, the locations of the rainfall are different: they are mostly outside the Kabul River basin on 25 and 26 July and inside the basin on 27 July.

Fig. 6.
Fig. 6.

(top left to bottom right) Accumulated rainfall from 0000 UTC 23 July to 0000 UTC 28 July 2010 from NCEP GFS forecasting with various initial times. The result from 0600 UTC 23 July was omitted to simplify the display. The color shadings of dark blue, light blue, green, yellow-green, yellow, orange, red, and pink indicate 0–50, 50–100, 100–150, 150–200, 200–250, 250–300, 300–350, and over 350 mm, respectively. The area in each map enclosed by the red line indicates the Kabul River basin. The map titles above each panel are enclosed red when the rainfall in the Kabul River basin is forecasted to be above specified criteria (see text).

Citation: Journal of Hydrometeorology 15, 1; 10.1175/JHM-D-13-011.1

To improve the reliability and capability of forecasting heavy rainfall over the Kabul River basin, we downscaled GFS forecasts into a finer resolution by using the regional model (WRF) as shown in Figs. 2 and 3. Twenty-one GFS forecast runs with different initial times were used as the initial and boundary conditions for WRF to cover wide possibilities of the rain system growth (Fig. 3). Figure 7 introduces the downscaled distribution of rainfall accumulation during 27–29 July. This figure shows a finer structure of the rainfall than the one derived from the GFS original rainfall (Fig. 6). The rainfall amount and location indicated in each forecast are not always the same as those in Fig. 6. The rainfall amount around the Kabul River basin tended to increase in forecasts with later initial times: for example, a rainfall of over 250 mm is found over an wide area in the forecasts performed at 1200 UTC 25 July and those performed later than 1800 UTC 26 July, except the one at 1800 UTC 27 July. However, most of the forecasts underestimated the rainfall within the Kabul River basin compared with the ground rain gauge observation (Fig. 5). Furthermore, the rainfall concentration is located outside the Kabul River basin in many of the downscaled forecasts (e.g., 1200 UTC 25 July and 0600 UTC 27 July), which agrees with the results from GFS in Fig. 6. It is hard to evaluate the forecast reliability and capability from this figure alone. Therefore, we compared time series of area average rain rates in next figures.

Fig. 7.
Fig. 7.

As in Fig. 6, but with maps based on downscaled forecasts.

Citation: Journal of Hydrometeorology 15, 1; 10.1175/JHM-D-13-011.1

Figure 8 shows a probability distribution of time series of the average rain rates in the rectangle area in northern Pakistan (30°–36°N, 69°–76°E; same as Fig. 1 in Webster et al. 2011). The broken line in each panel is calculated from rain gauge–corrected GSMaP rainfall. Figure 8a shows GFS forecasts, of which a high probability of rain rate (boxes show 25th to 75th percentile) had two maxima on 28 and 29 July. The forecast rain-rate variation did not always agree with that of corrected GSMaP, but their amounts are comparable. This feature of the twin peaks was also found in the downscaled runs (Fig. 8b), but they show narrower width of probability than those of GFS original forecasts.

Fig. 8.
Fig. 8.

Probability distribution of average rain rates within dashed square area in Fig. 2 (69°–76°N, 30°–36°N) by (a) GFS original forecasts and (b) downscaled forecasts. The gray boxes show the 25th to 75th percentiles and the bars show the minimum to maximum values. Dot–dashed lines and the dashed lines in both panels are the median and GSMaP rainfall corrected by the ground rain gauge, respectively.

Citation: Journal of Hydrometeorology 15, 1; 10.1175/JHM-D-13-011.1

However, the rain rates inside the Kabul River basin were quite different from those of the rectangle-area average. Figure 9 shows a probability distribution of the average rain rates in the Kabul River basin. In Fig. 9a, the forecast probability of rain rates from GFS in the Kabul River basin underestimated considerably after 1200 UTC 28 July, although it was consistent with corrected GSMaP until 1200 UTC 28 July. The median of forecast rain rates was less than one-tenth of the corrected GSMaP after 1200 UTC 28 July, which suggests most of the forecasts were underestimated. On the other hand, Fig. 9b shows the rain-rate probability of the downscaled forecasts in the Kabul River basin. Its forecast probability still underestimated but improved considerably, where the median of rain rates after 1200 UTC 28 July was approximately a half of the corrected GSMaP.

Fig. 9.
Fig. 9.

As in Fig. 8, but for the Kabul River basin.

Citation: Journal of Hydrometeorology 15, 1; 10.1175/JHM-D-13-011.1

Table 2 summarizes the performance of each lagged ensemble forecast run. The successful members, shown by circles, were defined as the total area average rainfall that was larger than 85% of the corrected GSMaP. The members with total rainfall below the criteria were marked by crosses. For the forecast in the rectangle area, out of the 21 runs each performed based on the GFS original and downscaled data, 16 runs each were successful. Forecast members started on 24 July were mostly unsuccessful, suggesting the initial conditions on 24 July were not suitable.

Table 2.

Results of lagged ensemble members. Rectangle refers to the rectangle area average (30°–36°N, 70°–74°E) and Kabul basin refers to the Kabul basin area average. The circle marks denote that rainfall forecasts are greater than the criteria defined as 85% of the corrected GSMaP rainfall and the × marks are less than the criteria. The bold crosses denote misplaced forecasts and the double circles indicate rainfall as corrected by downscaling.

Table 2.

On the other hand, the lagged ensemble forecast members for the Kabul Basin area average differentiate the GFS original and downscaled runs. Only 6 of the 21 forecast runs for the GFS original forecasts met the successful criteria, and the remaining 15 runs underestimated the rainfall. As in Table 2, 11 of the 16 successful GFS forecast runs for the rectangle area were below the criteria for the Kabul River basin (indicated by the bold cross signs). This implies that GFS predicted fairly accurate rainfall for the 11 runs, but in wrong places outside the Kabul River basin, which resulted in the underestimation of the rainfall in the Kabul River basin. However, for downscaled runs, 10 of the 21 forecast runs met the successful criteria, which were about twice as many as the GFS forecast runs. Furthermore, for the 5 runs in which GFS underestimated, their downscaled runs recovered rainfall beyond the criteria (indicated by the double circle signs in Table 2). This suggests that the downscaling is effective in improving accuracy of rainfall forecasting in 5 out of 11 GFS misplaced forecasts. The downscaling corrected not only underestimated but also overestimated forecasts (Figs. 9a,b).

Figure 10 introduces total rainfall with respective runs (GFS original in light gray and downscaled in dark gray) as a function of the initial time. A horizontal line in each panel indicates the corrected GSMaP rainfall. Figure 10a shows total rainfall in the rectangle area as a function of the initial time. It simply shows that the total rainfall tends to increase as the forecast lead time decreases. However, the forecast runs from 0000 to 0600 UTC 23 July had rainfall over the corrected GSMaP rainfall, although they were from the longest forecast lead time. The forecast reliability fluctuates periodically with the lead time. There is no remarkable difference between GFS original and downscaled rainfall.

Fig. 10.
Fig. 10.

Simulated total rainfall vs forecast initial time for (a) the rectangle area in Fig. 2 and (b) the Kabul River basin. The light gray bars indicate the GFS original forecasts, and the dark gray bars indicate those of the downscaled forecasts. The horizontal line in each panel shows the corrected GSMaP rainfall.

Citation: Journal of Hydrometeorology 15, 1; 10.1175/JHM-D-13-011.1

Figure 10b shows total rainfall in the Kabul River basin as a function of the initial time. Overall, the downscaled forecasts (light gray bars) showed better results than the GFS original forecasts (dark gray bars) at most of the initial times. The prediction capability of the GFS forecasts varied with lead time: it was relatively better around 0600 UTC 24 July and 0000 UTC 26 July, with its peak at 1200 UTC 27 July. The downscaled rainfall also fluctuates with the lead time, but the downscaled forecasts are generally better than GFS forecasts, attaining stable reliability onward from 1800 UTC 26 July. The rainfall in the Kabul River basin also fluctuates periodically, with two peaks around 0600 UTC 24 July and 0000 UTC 26 July in both GFS and downscaled forecasts.

b. Streamflow simulation

As shown in section 3a, the GFS forecasts and their downscaled version would have predicted heavy rainfall before the actual rainfall began. This section describes streamflow simulation in the Kabul River basin by using the RRI model. Figure 11a shows predicted discharge probability at an outlet of the Kabul River basin based on the GFS forecast rainfall. As shown in Fig. 9a, the maximum discharge probabilities are larger than those from the corrected GSMaP (broken line), but 25th to 75th percentile probabilities underestimated those from the corrected GSMaP. On the other hand, the RRI simulation based on downscaled rainfall (Fig. 11b) shows a better forecast closer to those based on the corrected GSMaP. It is natural that reliable rainfall predictions produce reliable discharge predictions. Furthermore, the variability in time series of rain rate and also the differences in horizontal distribution were rounded off in the streamflow simulation. This streamflow simulation also showed that downscaled rainfall forecasts were better in the prediction capability and reliability of flood forecasting with longer lead times.

Fig. 11.
Fig. 11.

As in Fig. 8, but for simulated discharges at the outlet of the Kabul River basin by the RRI model.

Citation: Journal of Hydrometeorology 15, 1; 10.1175/JHM-D-13-011.1

To evaluate the simulated inundation distribution, we calculated threat score. The threat score is defined as
e1
where hits is the number of areas where the model successfully predicted inundated area, misses is the number of areas where the model could not predict inundated area, and false alarm is the number of areas where the model predicted inundation but there was no inundation in reality. A threat score of 1.0 means a perfect forecast. We adopted UNOSAT inundation distribution as the true inundated area. The criterion of 1-m depth is used to judge inundation area from simulated peak inundation depth, as in Sayama et al. (2012). Figure 12 shows the threat scores as function of forecast initial time. It shows that the scores based on GFS rainfall (dashed line) were quite low compared to those based on downscaled rainfall (solid line), which demonstrates the improvement by downscaling much clearer than Fig. 10b. The threat scores based on the downscaled rainfall increased as initial time approached the rainfall onset, and forecasts of initial time after 1800 UTC 26 July, except for 1800 UTC 27 July, had threat scores close to the score based on the corrected GSMaP rainfall (a horizontal line in Fig. 12). A forecast initialized at 0000 UTC 23 July surprisingly had a similar score of those latest forecasts.
Fig. 12.
Fig. 12.

Threat score of area inundation as a function of initial time. The dashed line with the × marks shows the threat score of the inundated area based on GFS rainfall, and the solid line with the circle marks shows the threat score based on the downscaled rainfall by WRF. A horizontal line at a score of 0.615 is the score based on the corrected GSMaP rainfall.

Citation: Journal of Hydrometeorology 15, 1; 10.1175/JHM-D-13-011.1

Figure 13 shows the prediction capability for discharge and inundation distribution at three different time points calculated from 13 different GFS rainfall forecast runs (i.e., lagged ensemble prediction by 13 members). The number of the forecast runs (13) was chosen because it is adequate enough for ensembles composed of initial time in three days. Figure 13a shows ensemble prediction on 26 July with 13 ensemble members initialized from 0000 UTC 23 July to 0000 UTC 26 July. The probability distribution (curvy belts in different colors) and the ensemble mean (black line) underestimated discharges compared with the prediction based on the corrected GSMaP rainfall (green line). The ensemble mean was less than half the GSMaP estimate. However, the forecasted probability distribution includes most of the discharge curve derived from the corrected GSMaP. This implies that some of the ensemble members predicted the right amount of discharge, even though their reliability was low. Figure 13b shows updated ensemble prediction on 27 July, which is more or less the same as the ensemble prediction on 26 July. Further updated ensemble prediction on 28 July in Fig. 13c shows a far wider probability distribution, but its ensemble mean is similar to the discharge by the corrected GSMaP. The overestimated rainfall in the last three forecast runs in Fig. 10b were the reason for the difference in probability distribution between Figs. 13b and 13c. In general, ensemble prediction composed of newer ensemble members produces more reliable prediction. However, it is hard to refer to the reliability level of this ensemble prediction because of too large a discharge probability.

Fig. 13.
Fig. 13.

Probability distribution simulated by the RRI model based on GFS original rainfall. Probability distributions in discharge at the outlet of the Kabul River basin are shown, calculated from lagged ensemble forecasts composed by 13 runs in three time periods: (a) 0000 UTC 23 July to 0000 UTC 26 July, (b) 0000 UTC 24 July to 0000 UTC 27 July, and (c) 0000 UTC 25 July to 0000 UTC 28 July. The color shading shows accumulated probability density with higher probabilities in warm colors. The black line in each panel is the ensemble mean and the green line is rainfall forecasted by the corrected GSMaP. (d)–(f) Predicted inundation probability distribution corresponding to the three time periods. The location is the Peshawar valley near the outlet of the Kabul River basin (the small rectangle in Fig. 2, bottom). The color shading indicates the probability of inundation depth over 1 m. The black line indicates the inundation map from the observed indicator of the MODIS satellite.

Citation: Journal of Hydrometeorology 15, 1; 10.1175/JHM-D-13-011.1

Figures 13d–f show predicted inundation distribution in the Peshawar Valley with an observed indicator (the boundary of the flood inundation) based on the MODIS satellite data. Figures 13d and 13e show a high inundation probability along the river, and the probability decreases away from the river. The inundation probability shown by color shading expands close to the observed indicator. Figure 13f shows the inundation probability expands beyond the observed indicator, probably as a result of the overestimated rainfall and discharge in the ensemble prediction in Fig. 13c.

Figure 14 shows the prediction capability for discharge and inundation distribution, which are the same as Fig. 13, except for the calculation from the downscaled rainfall. This figure illustrates significant improvement in prediction compared with that based on the GFS original rainfall. Figure 14a, an ensemble forecast on 26 July, shows discharge probabilities close together with one another, and its ensemble mean was closer to that from the corrected GSMaP rainfall, compared with that of GFS (Fig. 13a). Figure 14b shows more reliable prediction than Fig. 14a, in which the prediction based on the corrected GSMaP (green line) corresponded to a larger probability of discharge prediction (yellow belt). It implies that the forecast reliability improved because of the ensemble prediction from 26 to 27 July (from Figs. 14a to 14b). Figure 14c shows further improvement in reliability. The ensemble mean was much closer to that from the corrected GSMaP, and it shows in the area a higher discharge probability based on the corrected GSMaP. Further, the probability distribution still stays relatively narrow, whereas that calculated from the GFS rainfall (Fig. 13c) diverged significantly. In Figs. 14a and 14b, the ensemble means before the onset of the rainfall predicted an about 80% discharge (a peak discharge of about 12 000 m3 s−1 in 15 000 m3 s−1 by the corrected GSMaP). In Fig. 14c, the ensemble mean after the rainfall onset predicted about 86% of the estimation from the corrected GSMaP.

Fig. 14.
Fig. 14.

As in Fig. 13, but for streamflow simulation by downscaled rainfall.

Citation: Journal of Hydrometeorology 15, 1; 10.1175/JHM-D-13-011.1

Figures 14d–f illustrate inundation probability distribution calculated based on the downscaled rainfall. They predicted quite a high probability of inundation over a wide area around the river, and a high probability extends within the observed indicator by UNOSAT. All three inundation probability predictions (Figs. 14d–f) corresponded well with the observed indicator by UNOSAT, except the inundation in the northwestern tributary. Those inundation forecasts significantly improved compared with those based on the GFS rainfall. This reflects the results shown in Fig. 12.

The calculation time was also measured. For downscaling GFS into a finer resolution, it took 8 h to calculate for 6-day forecasts with a newer personal computer (Intel Core i7 Sandy bridge architecture, six cores). The RRI model calculation needed about 1 h for the same forecasting. On the other hand, GFS forecasts took about 8 h after the initial forecasting time. Hence, with a decent number of personal computers, it should be possible to produce a inundation forecast derived from downscaled rainfall in 17 h after the initial time.

4. Discussion

Webster et al. (2011) argued that the rainfall that caused the 2010 Pakistan flood could have been predicted by using ECMWF-EPS; that is, the large-scale heavy rainfall in northern Pakistan could have been predicted 4 days before the onset of the rainfall. Alfieri et al. (2013) also showed the predictability of a severe flood in the middle of the Indus River at least 3 days before rainfall in their global ensemble streamflow forecasting. In this study, we conducted NCEP-GFS deterministic forecasting instead of using ECMWF-EPS. Nevertheless, NCEP-GFS predicted the heavy rainfall in northern Pakistan 4 days before the rainfall onset (Table 2, Fig. 10). However, most of the NCEP-GFS forecasts predicted the rainfall outside the Kabul River basin, and only the forecasts performed after the rainfall onset showed the right location of the rainfall (Table 2). On the other hand, the downscaled forecasts predicted the rainfall inside the Kabul River basin even before the rainfall onset (Table 2). This study found that global forecasts were far less reliable than regional forecasts in the Kabul River basin case. It should be remembered that we used deterministic forecasts, whereas Webster et al. (2011) used EPS. Since the forecasting method used by Webster et al. (2011) was also a type of global forecasting, their argument may not be valid, especially in terms of accuracy in locating rainfall events.

In this study, the horizontal resolution of NCEP-GFS was downscaled from 27 to 5 km. The finer the resolution of the NWP model is, the more realistically the model calculates air motion and resulting rainfall according to detailed topography and land use. The Kabul River basin is surrounded by high, complex mountain ranges, especially in the northern edge of the catchment (Fig. 2). The finer resolution of the geography may be a reason for the significant improvement in rainfall forecasting. The 5-km resolution may have increased the level of forecasting significantly.

The predictions of the area-averaged rainfall and discharge were successful in some of the downscaled forecasts in this study. However, in past studies, flood forecasting with NWP was not always promising, even in sophisticated European flood forecasting with EPS. COSMO-LEPS was sensitive to the length of lead time and was not always reliable in the Mulde River experiment with a catchment of 7400 km2 (Dietrich et al. 2008). It also had shortcomings in forecasting for a very small catchment (186 km2) because of the coarse model resolution in southern Switzerland (Alfieri et al. 2012), even though Zappa et al. (2008) showed successful hydrological forecasting with the use of COSMO-LEPS in the same small catchment (186 km2). They suggested that the rainfall forecasts, when misplaced, affect the rainfall over the target area, seriously in the case of a small catchment. In this study, the catchment size was 92 605 km2, much larger than the European case, which helped to reduce the uncertainty of the predicted rainfall and discharge because larger catchments offset uncertainty by averaging misplaced rainfall and magnitude. This may be a reason why our forecasting was within the acceptable level. In other words, the balance between catchment size and model resolution was appropriate in this study. In some cases, however, misplaced rainfall forecasts fail proper forecasting of discharge in tributaries or arrival time of water flow even in a large catchment (e.g., in a northwest tributary in Figs. 14d–f). Nonetheless, the inundation forecast in the Peshawar Valley in this study is still plausible, considering the high level of agreement with the inundation area derived from the MODIS data (Figs. 14d–f).

In terms of the prediction capability of NWP for heavy rainfall, the predictability of the blocking high in Russia (Lau and Kim 2012) can be a key to control the forecasting performance. Some EPS forecasting actually predicted the blocking 9 days before the rainfall (Matsueda 2011). All the NCEP-GFS forecast runs after 23 July predicted the blocking, but the forecasts varied in the strength of the predicted trough (no figures shown). In addition, the moisture advection from the monsoon surge affected the heavy rainfall, and it was exerted by an arrival of intraseasonal oscillation (Hong et al. 2011). Therefore, if the model had predicted the intraseasonal oscillation, it could have predicted a sign of the heavy rainfall. The heavy rainfall events causing the Kabul River basin flood were closely connected with synoptic-scale events with a time scale of more than a week. Therefore, such resulting events as heavy rainfall could be predictable with a lead time of a week, as long as the model predicts such causing events as synoptic-scale events.

As this study shows, lagged ensemble forecasting can work very effectively. It is very simple to perform and, unlike modern EPS, it requires no special generation of perturbations since they are automatically given to ensemble forecasting from model errors with multiple initial times. Therefore, it is suitable for short-range forecasts with limited time (e.g., Dietrich et al. 2008) or for daily operations with limited computer resources. There are, however, also disadvantages. A large lagged ensemble forecast includes extremely old forecasts, and systematic observational errors may not be filtered out as they carry on from one initial condition to another. There is a limited number of possibilities of atmospheric disturbance growths as compared to full EPS.

In this study the lagged ensemble forecast worked in acceptable reliability. It is worth considering applying this method to flood forecasting under the limited conditions. Nevertheless, lagged ensemble forecasts from different global forecasts can be a better option to involve more possibilities. Also, lagged ensemble forecasts using different selection of parameterizations can add more possibilities of disturbance growth. Those trials in future studies could improve the forecast reliability.

5. Conclusions

The Kabul River basin flood in the summer of 2010 in Pakistan was examined by means of lagged ensemble rainfall forecasting, which used NCEP-GFS deterministic forecasts and their downscaled versions by WRF. Since NCEP-GFS provided forecasts for 7.5 days and the heavy rainfall lasted for 3 days, the forecasts of the lead time on days 0–4 were employed for analysis. By using the forecasted rainfall, the river discharge and inundation distribution were simulated with the RRI model. The rainfall distribution was validated by the GSMaP rainfall corrected based on the ground rain gauge data, and the inundation distribution was validated by the observed indicator derived from the MODIS satellite.

The rainfall forecasts from NCEP-GFS showed a sign of heavy rainfall in northern Pakistan with a lead time of a few days. However, most of them underestimated the rainfall within the Kabul River basin because of the misplacement of the rainfall events. Only NCEP-GFS forecast runs performed after the rainfall onset predicted the rainfall in the accurate location, though overestimating it significantly. Downscaling by a regional model corrected the original GFS-based location of the rainfall and predicted a more accurate amount of rainfall in the Kabul River basin 2 days before the rainfall. Those results also were evaluated by threat score analysis. It is confirmed that downscaling not only corrected the location of the rainfall but also corrected the underestimated or overestimated rainfall.

Three lagged ensembles using 13 forecast runs, which were obtained 2, 1, and 0 days before the rainfall onset, were composed in terms of discharge and inundation distribution. Lagged ensemble forecasting by using NCEP-GFS and the RRI model showed some signs of inundation prior to the rainfall, but its reliability was low. The ensemble with zero lead time showed a large probability of inundation but had too large a variance. On the other hand, the lagged ensemble by using the downscaled rainfall and the RRI model showed a reliable inundation probability 2 days ahead, and the reliability increased as the lead time decreased.

NCEP-GFS and its downscaled version successfully predicted the heavy rainfall that caused the Kabul River basin flood a few days before the rainfall. There are several reasons for this successful forecasting: 1) the size of the catchment was large enough to offset uncertainty, 2) the size of the catchment and the NWP resolution were balanced, and 3) the rainfall event was an easy case to reproduce by the coarse NWP model. The Kabul River basin case was a good example in that GFS, downscaling, and the RRI model can be used to successfully predict a severe flood event a few days ahead of the rainfall. However, our study is only a case study for a specific location. We need further research to examine whether the aforementioned reasons are correct or not. Moreover, to apply this forecasting system to other flood events or daily operations, further research is necessary to examine its validity with various flood events in the same and other locations.

Acknowledgments

This work was conducted under the support of grants for operating expenses of the Public Works Research Institute funded by the Ministry of Land, Infrastructure, Transport and Tourism. The NCEP GFS data were provided from the National Centers for Environmental Prediction through the National Climatic Data Center. The regional model WRF was developed by the National Center for Atmospheric Research (NCAR), which is operated by the University Corporation for Atmospheric Research (UCAR) in the United States. The Grid Analysis and Display System (GrADS) and GFD Dennou Club Library were used to draw figures.

REFERENCES

  • Alfieri, L., , Thielen J. , , and Pappenberger F. , 2012: Ensemble hydro-meteorological simulation for flash flood early detection in southern Switzerland. J. Hydrol., 424–425, 143153, doi:10.1016/j.jhydrol.2011.12.038.

    • Search Google Scholar
    • Export Citation
  • Alfieri, L., , Burek P. , , Dutra E. , , Krezeminski B. , , Murato D. , , Thielen J. , , and Pappenberger F. , 2013: GloFAS—Global ensemble streamflow forecasting and flood early warning. Hydrol. Earth Syst. Sci., 17, 11611175, doi:10.5194/hess-17-1161-2013.

    • Search Google Scholar
    • Export Citation
  • Cloke, H. L., , and Pappenberger F. , 2009: Ensemble flood forecasting: A review. J. Hydrol., 375, 613626, doi:10.1016/j.jhydrol.2009.06.005.

    • Search Google Scholar
    • Export Citation
  • Coe, M. T., , Costa M. H. , , and Howard E. A. , 2008: Simulating the surface waters of the Amazon River basin: Impacts of new river geomorphic and flow parameterizations. Hydrol. Processes, 22, 25422553, doi:10.1002/hyp.6850.

    • Search Google Scholar
    • Export Citation
  • Cuo, L., , Pagano T. C. , , and Wang Q. J. , 2011: A review of quantitative precipitation forecasts and their use in short- to medium-range streamflow forecasting. J. Hydrometeor., 12, 713728, doi:10.1175/2011JHM1347.1.

    • Search Google Scholar
    • Export Citation
  • Dietrich, J., , Trepte S. , , Wang Y. , , Shumann A. H. , , Voß F. , , Hesser F. B. , , and Denhard M. , 2008: Combination of different types of ensembles for the adaptive simulation of probabilistic flood forecasts: Hindcasts for the Mulde 2002 extreme event. Nonlinear Processes Geophys., 15, 275286, doi:10.5194/npg-15-275-2008.

    • Search Google Scholar
    • Export Citation
  • Dole, R., and Coauthors, 2011: Was there a basis for anticipating the 2010 Russian heat wave? Geophys. Res. Lett., 38, L06702, doi:10.1029/2010GL046582.

    • Search Google Scholar
    • Export Citation
  • Hoffman, R. N., , and Kalnay E. , 1983: Lagged average forecasting, and alternative to Monte Carlo forecasting. Tellus, 35A, 100118, doi:10.1111/j.1600-0870.1983.tb00189.x.

    • Search Google Scholar
    • Export Citation
  • Hong, C.-C., , Hsu H.-H. , , Lin N.-H. , , and Chiu H. , 2011: Roles of European blocking and tropical-extratropical interaction in the 2010 Pakistan flooding. Geophys. Res. Lett., 38, L13806, doi:10.1029/2011GL047583.

    • Search Google Scholar
    • Export Citation
  • Kain, J. S., , and Fritsch J. M. , 1993: Convective parameterization for mesoscale models: The Kain–Fritsch scheme. The representation of cumulus convection in numerical models. The Representation of Cumulus Convection in Numerical Models, Meteor. Monogr., No. 46, Amer. Meteor. Soc., 165170.

    • Search Google Scholar
    • Export Citation
  • Kalnay, E., 2003: Atmospheric Modeling, Data Assimilation and Predictability. Cambridge University Press, 341 pp.

  • Kubota, T., , Ushio T. , , Shige S. , , Kida S. , , Kachi M. , , and Okamoto K. , 2009: Verification of high-resolution satellite-based rainfall estimates around Japan using a gauge-calibrated ground-radar dataset. J. Meteor. Soc. Japan, 87A, 203222, doi:10.2151/jmsj.87A.203.

    • Search Google Scholar
    • Export Citation
  • Lau, W. K. M., , and Kim K.-M. , 2012: The 2010 Pakistan flood and Russian heat wave: Teleconnection of hydrometeorologic extremes. J. Hydrometeor., 13, 392403, doi:10.1175/JHM-D-11-016.1.

    • Search Google Scholar
    • Export Citation
  • Lehner, B., , Verdin K. , , and Jarvis A. , 2008: New global hydrography derived from spaceborne elevation data. Eos, Trans. Amer. Geophys. Union, 89, 9394, doi:10.1029/2008EO100001.

    • Search Google Scholar
    • Export Citation
  • Marsigli, C., , Boccanera F. , , Montani A. , , and Paccagnella T. , 2005: The COSMO-LEPS mesoscale ensemble system: Validation of the methodology and verification. Nonlinear Processes Geophys., 12, 527536, doi:10.5194/npg-12-527-2005.

    • Search Google Scholar
    • Export Citation
  • Matsueda, M., 2011: Predictability of Euro-Russian blocking in summer of 2010. Geophys. Res. Lett.,38, L06801, doi:10.1029/2010GL046557.

  • Okamoto, K., , Iguchi T. , , Takahashi N. , , Iwanami K. , , and Ushio T. , 2005: The global satellite mapping of precipitation (GSMaP) project. Proc. 2005 IEEE Geoscience and Remote Sensing Symp., Vol. 5, Seoul, South Korea, IEEE, 3414–3416, doi:10.1109/IGARSS.2005.1526575.

  • Rawls, W. J., , Ahuja L. R. , , Brakensiek D. L. , , and Shirmohammadi A. , 1992: Infiltration and soil water movement. Handbook of Hydrology, D. R. Maidment, Ed., McGraw-Hill, 5.1–5.51.

    • Search Google Scholar
    • Export Citation
  • Sayama, T., , Ozawa G. , , Kawakami T. , , Nabesaka S. , , and Fukami K. , 2012: Rainfall-runoff- inundation analysis of the 2010 Pakistan flood in the Kabul River basin. Hydrol. Sci. J., 57, 298312, doi:10.1080/02626667.2011.644245.

    • Search Google Scholar
    • Export Citation
  • Shige, S., , Kida S. , , Ahiwake H. , , Kubota T. , , and Aonashi K. , 2013: Improvement of TMI rain retrieval in mountainous areas. J. Appl. Meteor. Climatol., 52, 242254, doi:10.1175/JAMC-D-12-074.1.

    • Search Google Scholar
    • Export Citation
  • Shiraishi, Y., , Fukami K. , , and Inomata H. , 2009: A proposal of correction method using the movement of rainfall area on satellite-based rainfall information by analysis in the Yoshino River basin (in Japanese with English abstract). Annu. J. Hydraul. Eng.,53, 385–390.

  • Shrestha, M. S., , Takara K. , , Kubota T. , , and Bajracharya S. R. , 2011: Verification of GSMaP rainfall estimates over the central Himalayas. J. Japan Soc. Civil Eng., Ser. B1,67, I-37–I-42.

  • Taniguchi, A., and Coauthors, 2013: Improvement of high-resolution satellite rainfall product for Typhoon Morakot (2009) over Taiwan. J. Hydrometeor., 14, 18591871, doi:10.1175/JHM-D-13-047.1.

    • Search Google Scholar
    • Export Citation
  • Thielen, J., , Bartholmes J. , , Ramos M.-H. , , and de Roo A. , 2009a: The European flood alert system—Part 1: Concept and development. Hydrol. Earth Syst. Sci., 13, 125140, doi:10.5194/hess-13-125-2009.

    • Search Google Scholar
    • Export Citation
  • Thielen, J., , Bogner K. , , Pappenberger F. , , Kalas M. , , del Medico M. , , and de Roo A. , 2009b: Monthly-, medium-, and short-range flood warning: Testing the limits of predictability. Meteor. Appl., 16, 7790, doi:10.1002/met.140.

    • Search Google Scholar
    • Export Citation
  • Tian, Y., , Peters-Lidard C. D. , , Adler R. F. , , Kubota T. , , and Ushio T. , 2010: Evaluation of GSMaP precipitation estimates over the contiguous United States. J. Hydrometeor., 11, 566574, doi:10.1175/2009JHM1190.1.

    • Search Google Scholar
    • Export Citation
  • United Nations, cited 2010: Pakistan Floods Emergency Response Plan. [Available online at http://ochaonline.un.org/humanitarianappeal/.]

  • Webster, P. J., and Coauthors, 2010: Extended-range probabilistic forecasts of Ganges and Brahmaputra floods in Bangladesh. Bull. Amer. Meteor. Soc., 91, 14931514, doi:10.1175/2010BAMS2911.1.

    • Search Google Scholar
    • Export Citation
  • Webster, P. J., , Toma V. E. , , and Kim H.-M. , 2011: Were the 2010 Pakistan floods predictable? Geophys. Res. Lett., 38, L04806, doi:10.1029/2010GL046346.

    • Search Google Scholar
    • Export Citation
  • WFP, 2010: Preliminary rapid damage assessment in the agriculture sector for flood-affected areas of Pakistan. World Food Program Rep., 27 pp. [Available online at https://www.cimicweb.org/cmo/pf2010/Documents/Assessments/AgricultureCluster_Preliminary.pdf.]

  • Zappa, M., and Coauthors, 2008: MAP D-PHASE: Real-time demonstration of hydrological ensemble prediction systems. Atmos. Sci. Lett., 9, 8087, doi:10.1002/asl.183.

    • Search Google Scholar
    • Export Citation
Save