Skill of a global seasonal streamflow forecasting system, relative roles of initial conditions and meteorological forcing

Authors


Abstract

[1] We investigate the relative contributions of initial conditions (ICs) and meteorological forcing (MF) to the skill of the global seasonal streamflow forecasting system FEWS-World, using the global hydrological model PCRaster Global Water Balance. Potential improvement in forecasting skill through better climate prediction or by better estimation of ICs through data assimilation depends on the relative importance of these sources of uncertainty. We use the Ensemble Streamflow Prediction (ESP) and reverse ESP (revESP) procedure to explore the impact of both sources of uncertainty at 78 stations on large global basins for lead times upto 6 months. We compare the ESP and revESP forecast ensembles with retrospective model simulations driven by meteorological observations. For each location, we determine the critical lead time after which the importance of ICs is surpassed by that of MF. We analyze these results in the context of prevailing hydroclimatic conditions for larger basins. This analysis suggests that in some basins forecast skill may be improved by better estimation of initial hydrologic states through data assimilation; whereas in others skill improvement depends on better climate prediction. For arctic and snowfed rivers, forecasts of high flows may benefit from assimilation of snow and ice data. In some snowfed basins where the onset of melting is highly sensitive to temperature changes, forecast skill depends on better climate prediction. In monsoonal basins, the variability of the monsoon dominates forecasting skill, except for those where snow and ice contribute to streamflow. In large basins, initial surface water and groundwater states are important sources of skill.

1. Introduction

[2] Forecasting of water availability and scarcity is a prerequisite for managing the risks and opportunities caused by the interannual variability of streamflow. Reliable seasonal streamflow forecasts are necessary to prepare for an appropriate response in disaster relief, management of hydropower reservoirs, water supply, agriculture, and navigation. Seasonal hydrological forecasting on a global scale could be valuable, especially for developing regions of the world, where effective hydrological forecasting systems are scarce. Furthermore, global seasonal forecasts may provide spatially consistent predictions of streamflow anomalies. These may provide information to disaster management organizations operating at global scale to prepare for response, as well as to the energy market about the regional availability of hydropower in the coming months.

[3] Several studies demonstrated the capability of global hydrological models to predict streamflow, such as the WaterGap [Alcamo et al., 2003; Döll et al., 2003], LaD (The Land Dynamics Model) [Milly and Schmakin, 2002], VIC (Variable Infiltration Capacity Model) [Nijssen et al., 2001], WBM (Water Balance Model) [Vörösmarty et al., 2000; Fekete et al., 2002], Macro-PDM (Probability Distributed Moisture Model) [Arnell, 1999, 2004], and PCRaster Global Water Balance (PCR-GLOBWB) [Sperna-Weiland et al., 2010; Van Beek et al., 2011]. Candogan Yossef et al. [2012] assessed the skill of the global hydrological model PCR-GLOBWB in reproducing past discharge extremes in 20 large rivers of the world, as a first step toward developing and assessing a global seasonal hydrological forecasting system. This preliminary assessment in hindcast quantified skill using a meteorological forcing (MF) based on observations and concluded that the prospects for seasonal forecasting with PCR-GLOBWB or comparable models are positive. Note that this study did not include actual forecasts. Thus, the meteorological forcing errors due to uncertainty from numerical weather prediction models were not assessed.

[4] In an actual forecasting setup, the predictive skill of a hydrological forecasting system is affected by errors in model structure and parameterization, MF, and initial conditions (ICs), most importantly soil moisture, groundwater and snow. Skill of seasonal hydrological forecasts can thus be improved on the one hand by better prediction of future climate and on the other hand by better estimation of initial hydrologic states through assimilation of independent hydrological observations such as soil moisture and snow data from earth observation. The improvement in the overall predictability that may be attained depends on the relative importance of these two sources of uncertainty, which varies considerably according to location, season and lead time [Bierkens and van den Hurk, 2007; Bierkens and van Beek, 2009; Shukla and Lettenmaier, 2011; Shukla et al., 2011]. Therefore, determining the role of each factor is helpful in deciding which skill improvement methods are more promising [Paiva et al., 2012].

[5] The theoretical framework for quantifying the contributions of boundary forcing and initial conditions to predictability was developed in atmospheric sciences by Collins and Allen [2002]. Wood and Lettenmaier [2008] and Wood et al. [2002, 2005] adopted this approach in hydrological forecasting. They presented an Ensemble Streamflow Prediction (ESP) and reverse Ensemble Streamflow Prediction (revESP) approach and evaluated the relative roles of MF and ICs in seasonal hydrologic prediction in two western US basins. The ESP/revESP framework contrasts the forecast variance arising from a forecast ensemble based on perturbations of the initial states, and the forecast variance arising from an ensemble of meteorological forcing, to the internal, climatological variance. Li et al. [2009] used a similar approach for the Ohio River basin and the southeastern US to investigate the varying roles of ICs and MF in seasonal hydrologic forecasting. The ESP/revESP approach was applied to seasonal forecasts in the US by Shukla and Lettenmaier [2011]. Shukla et al. [2011] used the same approach to evaluate cumulative runoff and soil moisture forecasts on a global scale. Paiva et al. [2012] applied the ESP/revESP approach to the Amazon River basin.

[6] In this study, we apply the ESP/revESP approach on a global scale to examine the relative contributions of ICs and MF to the skill of seasonal streamflow forecasts. We investigate the roles of both sources of uncertainty in the skill of the global seasonal streamflow forecasting system FEWS-World, using the global hydrological model PCR-GLOBWB. FEWS-World has been setup within the European Commission 7th Framework Programme project Global Water Scarcity Information Service (GLOWASIS). The assessment is based on the ESP/revESP procedure outlined by Wood and Lettenmaier [2008]. We simulate global monthly streamflow with lead times ranging from 1 to 6 months for a historical period of 30 years (1981–2010). We analyze the impact of both sources of uncertainty at 78 stations on large river basins across the globe, for all the months of the year and for lead times up to 6 months. These 78 stations have previously been selected for analysis within the GLOWASIS project to represent different hydroclimatic conditions and all continents, but in the same time considering the availability of discharge data. In this study, we compare the ESP and revESP forecast ensembles with retrospective model simulations driven by meteorological observations, and not with direct hydrological observations. The advantage of comparing against simulations instead of observations is that in the former case model errors are eliminated and predictability is related only to knowledge of ICs and the uncertainty in future MF. Note that in previous work [Candogan Yossef et al., 2012] we investigated the model errors of the forecasting system.

[7] The remaining part of this paper is set up as follows. Section 'Materials and Methods' describes the global seasonal hydrological forecasting system, FEWS-World, the global hydrological model PCR-GLOBWB and the meteorological forcing data. Section 'Results' describes the hydrological simulations and the skill measures. Results are presented in section 'Discussion of Results', followed by discussion in section 'Conclusions' and conclusions in the last section.

2. Materials and Methods

2.1. Global Hydrological Forecasting System FEWS-World

[8] FEWS-World is a global hydrological forecasting system configured within the forecasting environment Delft-FEWS (Flood Early Warning System). Delft-FEWS is an open shell for managing, data handling and guiding of forecasting processes [Werner et al., 2013]. It is used by a large number of operational forecasting centers and agencies around the world for various purposes such as forecasting hydrological storm surges, river flows, reservoir management and water quality. FEWS-World has been built as part of the GLOWASIS project. The FEWS-World system consists of a Master Controller, a Postgres database and 18 forecasting shells (i.e., computational cores) for efficient handling of ensemble forecasts and data processing. Within FEWS-World several workflows have been setup for running the global hydrological model PCR-GLOBWB using the precipitation, temperature and potential evaporation fields from the ERA Interim-Global Precipitation Climatology Project (GPCP) data set [Balsamo et al., 2010]. Further descriptions of the meteorological forcings are given in section 'Meteorological Forcing Data'.

[9] PCR-GLOBWB simulates the terrestrial part of the global water cycle [van Beek et al., 2011; van Beek and Bierkens, 2009]. It is coded in the high-level computer language PCRaster for constructing environmental models [Wesseling et al., 1996]. The model is fully distributed and operates on a regular grid with a cell size of 0.5 × 0.5° on a daily time step. Meteorological forcing is assumed to be constant over the grid cell. Subgrid variability of hydrological processes is taken into account in the representation of short and tall vegetation, open water, different soil types, saturated area, surface runoff, interflow and groundwater discharge.

[10] PCR-GLOBWB calculates the water balance for every grid cell by tracking the transfer of water between the atmosphere and the cell, through stores within each cell, and laterally, as discharge, from one cell to the downstream neighbor. The model calculates the storages and fluxes of water, simulates the generation of runoff and its propagation as discharge through the river network. Precipitation falls either as snow or rain depending on atmospheric temperature. It can be intercepted by vegetation and added to the finite canopy storage, which is subject to open water evaporation. Snow is accumulated when the temperature is lower than 0°C and melts when it is higher. Snow melt is added to rain and throughfall; it is either stored in the available pore space in the snow cover, or it infiltrates into the top soil layer. Part of this water is transformed into surface runoff and the remainder infiltrates into the soil through two vertically stacked soil layers and an underlying groundwater layer. Water is exchanged between these layers following Darcy's law and the resulting soil moisture is subject to evapotranspiration. The remaining water contributes to lateral drainage as interflow from the soil layers or base flow from the groundwater reservoir. The total drainage, consisting of surface runoff, interflow and base flow is routed through the drainage network of rivers, lakes, wetlands and reservoirs based on DDM30 [Döll and Lehner, 2002], using the kinematic wave approach. An extensive description of PCR-GLOBWB can be found in Van Beek and Bierkens [2009].

2.2. Meteorological Forcing Data

[11] The meteorological variables required to force PCR-GLOBWB are daily values of precipitation, evapotranspiration and temperature. In the absence of direct estimates of actual evapotranspiration, the model can be forced with values of potential evapotranspiration calculated from temperature, radiation, cloud cover, vapor pressure and wind speed.

[12] We force PCR-GLOBWB with the ERA Interim-GPCP data set [Balsamo et al., 2010]. This is a global meteorological data set, which is a combination of the ERA Interim reanalysis [Dee et al., 2011] and GPCP monthly rainfall observations [Huffman and Bolvin, 2011; Huffman et al., 2009]. ERA-Interim is the latest global atmospheric reanalysis produced by the European Centre for Medium-Range Weather Forecasts. It is an “interim” reanalysis initially started from year 1989; later extended back to the year 1979; and continues to be updated forward in time. ERA-Interim reanalysis was produced as a part of the next-generation extended reanalysis intended to replace ERA-40. The GPCP is part of the Global Energy and Water Cycle Experiment of the World Climate Research program. The GPCP provides global precipitation estimates by merging infrared and microwave satellite estimates with rain gauge data from more than 6000 stations.

[13] Potential evaporation has been estimated from ERA Interim as well. We estimated the monthly values of potential evaporation by application of the Penman-Monteith equation [Monteith, 1981; Penman, 1948] for a reference grass canopy, according to the Food and Agriculture Organization methodology [Allen et al., 1998]. Reference potential evaporation is multiplied by a monthly crop factor to obtain land cover specific potential evaporation.

2.3. Hydrological Simulations

[14] PCR-GLOBWB is run at a daily time step to produce ESP/revESP forecast ensembles as well as the control simulation. The model spin-up is carried out over the period 1979–1984 using ERAInterim-GPCP data set. Subsequently, the hydrological states at the end of this 5 year spin-up are used as initial states for the 30 year historical simulation covering the historical period from 1981 to 2010, with an extra 2 years spin-up period from 1979 to 1981.

[15] The control run is a single simulation covering the whole 30 years period. Daily discharge values are aggregated into monthly totals. Monthly aggregation provides a more appropriate forecast at the seasonal scale and a proxy of the underlying distribution. Hydrologic states, as well as monthly discharge totals are saved at the end of each month. These states are used for both running the ESP forecasts and producing resampled ICs for each month of the year for the revESP forecasts.

[16] The ESP and revESP forecast ensembles are produced with ESP and revESP workflows. In the ESP workflow, input ensembles of MF are created from the 32 year input data series (1979–2010). PCR-GLOBWB model runs are initialized on the first day of each month for the period 1981–2010 using the stored ICs. In the revESP workflow, stored ensembles of ICs for 12 × 30 = 360 months and the meteorological input data are used to start a PCR-GLOBWB run on the first day of each month for the period 1981–2010. This results in 360 ESP runs, each run containing 32 members and 360 revESP runs, each run containing 30 members. The runs are carried out in batch using the FEWS-World forecasting system. Each run continues for 6 months ahead and produces an ensemble of monthly discharge values for 6 lead times, from 1 to 6 months.

[17] The skill of ESP forecasts comes from ICs and the ensemble spread is caused by MF uncertainty. RevESP on the other hand, uses an ensemble of ICs resampled from the hydrologic states for the same day of the year and the model is forced by observed MF for the given forecast period. RevESP skill results from the MF and the ensemble spread is caused by uncertainty in ICs.

2.4. Measures of Skill

[18] We assess the relative contribution of ICs and MF to forecast skill in 78 stations on large global basins. To quantify each contribution to predictability, we use a ratio of variances framework that compares the skill obtained from the ESP and revESP with the skill resulting when climatology is used as an ensemble prediction, which is equivalent to the climatological variance [Wood and Lettenmaier, 2008]

[19] We calculate the mean squared error (MSE) values of the ESP and revESP as well as of the climatology for 78 global locations, for 12 months of the year and for 6 lead times. The MSE values for a given month and lead time are calculated according to the following formulas:

display math
display math
display math

where,

  • hsm(t)= streamflow for initial state s, meteorological forcing m
  • S = number of years in the historical period
  • M = number of ensemble members, i.e., number of meteorological forcing values

[20] We then calculate the ratios of MSE of both ESP and revESP to the MSE of climatology for all locations, months and lead times.

[21] When the ratio of the MSE of either forecast ensemble to the MSE of the climatology is equal to one, the forecast skill is equal to that of a climatological forecast. Ratios smaller than one indicate a forecast that is more skillful than the mere climatology; whereas ratios greater than one indicate less skill than the climatology. The ratio approaches zero for a perfect forecast.

[22] When calculating the MSE ratios for a given month and a given lead time, we use the forecast ensembles that predict the total monthly discharge generated during that given month. In other words, we use the discharge ensembles resulting from the simulations which start at time t0 and end at time tn, with a lead time of n months, where t0 is prior to the end of the given forecast month by n months. Thus, for the month of May and for 1 month lead time, n = 1, t0 is the 1st of May and tn is the 31st of May. For 2 months lead time, n = 2, t0 is the 1st of April and tn is again the 31st of May.

3. Results

3.1. Results of the Hydrological Simulations

[23] The results of the 360 ESP runs and 360 revESP runs at each location, each one covering a period of 6 months, are combined in 12 discharge time series, 6 for ESP and 6 for revESP, each 360 months long. These time series represent the chained discharge forecasts at the 6 monthly lead times of interest. To demonstrate the typical behavior of the ESP and revESP, the discharge time series in the Amazon and the Murray are presented in Figure 1. For better visual inspection the results of only the first 24 months period are shown. The ensemble results as well as the results of the control run are presented. Discharge time series for the ESP and revESP runs for all 78 locations can be seen in supporting information Figure 1 in the online supporting information data file 2013WR013487-fs01.doc.

Figure 1.

Chained discharge time series of ESP and revESP for the Amazon and the Murray.

[24] The ESP discharge time series in the Amazon (Figure 1a) show that the ensemble spread of the ESP simulations increases with increasing lead time as the relative importance of ICs diminish. For the revESP simulations at the same location (Figure 1b), the ensemble spread decreases with increasing lead time, converging toward the control run as the relative role of the MF becomes more dominant.

[25] In the Murray basin the ESP ensemble spreads much more rapidly with increasing lead time (Figure 1c) than the discharge series in the Amazon, while the revESP ensemble spread collapses to the value of the control run very rapidly (Figure 1d). This means that the expected skill due to IC memory is much smaller in the Murray than in the Amazon, while the expected skill due to MF is much larger. The relative importance of ICs and MF is further discussed in section 'Discussion of Results'.

3.2. Skill for ESP and Reverse ESP

[26] Ratios of MSE of the ensemble simulations to the MSE of climatology for all 78 locations, 12 months of the year and 6 lead times are documented in the supporting information Table 1 that can be found online in the supporting information data file 2013WR013487-ts01.xls.

[27] As explained in section 'Hydrological Simulations', the MSE ratios for a given month are calculated using the results of the simulations starting at a prior time t0, and end at the end of the considered month. Therefore the month mentioned in the MSE tables denote the month for which the forecast is made, not the month at which the forecast starts.

3.3. Changes in Skill With Increasing Lead Time

[28] The MSE ratios for the ESP and revESP are plotted on bar charts. As two contrasting examples, the skill charts for the Nile and Ob basins are presented in Figure 2. Skill charts for all 78 locations are presented in supporting information Figure 2 that can be found online in supporting information data file 2013WR013487-fs02.doc.

Figure 2.

ESP and revESP skill charts for the Nile and the Ob.

[29] The charts in Figure 2 demonstrate that the MSE ratios of the ESP increase with increasing lead time, while the MSE ratios of the revESP decrease. In the Nile basin, the MSE ratios of the ESP for all lead times are smaller than 1, indicating that the ESP is more skillful than climatology over the full lead time considered. In contrast, in the Ob the ESP is skillful only upto 2 months lead time. As the relative importance of the ICs diminishes with increasing lead time and the importance of MF becomes more dominant, the skill of the revESP increases, hence the MSE ratios decrease. It can be observed that for the Ob, the graphs of the ESP and revESP intersect each other at a point. This point denotes the time within the 6 months lead time range, after which the relative importance of MF surpasses that of ICs. In the Nile however, there is no such intersection point, since the ICs dominate throughout the 6 months range. The reasons for these differences are discussed at global scale in section 'Conclusions'.

3.4. Seasonal and Geographical Distribution of Skill

[30] Skill maps are prepared to present the seasonal and geographical distribution of skill. The maps indicate skill of the simulations for each month of the year at each location. Separate maps are prepared for the ESP and revESP runs, as well as for each lead time. MSE ratios for each basin are shown on each map for a given lead time and a given month of the year. Maps for the month of January for all lead times are presented in Figure 3. Maps for the ESP and revESP runs for all months of the year are presented in supporting information Figure 3 that can be found online in supporting information data file 2013WR013487-fs03.doc. The 78 stations are marked with a circle, colored according to a skill scale from 0 to 1.5, 0 indicating perfect skill, 1 indicating skill equivalent to that of the climatology, and values above 1 indicating less skill than the climatology. The sizes of the circles reflect basin size.

Figure 3.

World maps indicating ESP and revESP skill scores in January.

[31] The ESP maps for January and for all 6 lead times (Figure 3a) show in which basins and up to what lead times, forecasts for the given month may be skillful, with the knowledge of ICs only and no information besides past observed climatology on the future MF. The revESP maps (Figure 3b) show on the other hand, the forecast skill that may be attained for a given month, with perfect knowledge of future MF, but no information on ICs.

3.5. Critical Lead Times

[32] Combining the skill measures of the ESP and revESP simulations, we calculate critical lead times (CLT) for each station and for each month of the year. CLT is defined as the lead time in months after which the importance of ICs is surpassed by that of MF. The CLT are plotted on 12 maps for each month and presented in Figure 4.

Figure 4.

World maps indicating the Critical Lead Time (CLT) in months.

[33] The maps show that there are some basins, where in certain months of the year, it is not likely to make skillful forecasts on any lead time, with only the knowledge of ICs and none on the future MF. These basins are marked with 0. There are also some basins where without any information on future MF, perfect knowledge of ICs allows forecasts for certain months up to 6 months ahead. It should be noted that CLT is only meaningful when there is skill in ESP and revESP. When both ESP and revESP are of poor skill, the resulting CLT may be high but this does not mean that forecasts are skillful. In most basins, this critical lead time changes significantly with the season. These issues are discussed in section 'Conclusions'.

4. Discussion of Results

[34] Our results show that the relative roles of ICs and MF in hydrological forecasting skill vary both seasonally and geographically. In this section, we discuss the results for several larger basins in the context of prevailing hydroclimatic conditions.

4. 1. Tropical, Monsoon-Dominated Basins

[35] Results indicate that in the tropical, monsoon-dominated South American basins, such as the Amazon and the Parana, skill due to ICs is limited to 1–2 months lead time during most of the year. Hydrological forecasts beyond 1–2 months are dominated by MF. In the Amazon basin, forecasts for the spring months August until November have a higher CLT of 3–4 months. Hence knowledge of ICs contributes to the skill of forecasts, initiated 3–4 months ahead. This shows that better estimates of ICs during winter months (June, July, and August) may improve the spring forecasts in the Amazon. Our conclusion is in agreement with the conclusions of the study of Paiva et al. [2012] where they demonstrated that in the Amazon River uncertainty on ICs play an important role for hydrological predictability for up to 90 days. Paiva et al. [2012] conclude that ICs are more important especially during high flow conditions (March, April, and May) and recession (June, July, and August). Our results emphasize the importance of ICs during the recession period (June, July, and August), when the increased groundwater recharge plays an important role. Our findings, as well as the findings of Paiva et al. [2012] are in disagreement with Shukla et al. [2011], who found that MF uncertainty dominates the forecasts in the Amazon, even for shorter lead times. The reason for the disagreement may be the importance of routing in large basins such as the Amazon where long travel times are involved. Extensive floodplains store water and release it slowly, so that the outflow has already been in the river for a number of months. Therefore the knowledge of surface water ICs is an important source of forecast skill. Shukla et al. [2011] studied cumulative runoff, without flow routing, whereas our study as well as the study by Paiva et al. [2012] include routing.

[36] In the large rivers of the Indian subcontinent the relative roles of ICs and MF follow the seasonal patterns of the monsoon. In the Ganges and Brahmaputra, forecast skill is strongly dominated by MF during the monsoon season. At the onset of the monsoon, the ICs are always more or less the same, which reduces the impact of the ICs on model skill. Snowmelt coincides with the monsoon and the contribution of snowmelt to streamflow is very low compared to rainfall. The interannual variability of the monsoon is the determining factor in forecasts during the wet season. Maps in supporting information Figure 3 (online supporting information data file 2013WR013487-fs03.doc) indicate that in the Brahmaputra skill of the ESP is below the climatology from April to September even for lead times of 1 month; hence the forecasts are dominated by the MF for all lead times. In the period from October to March, the knowledge of ICs contributes to skill for lead times up to 6 months. The results for the Ganges show that during the monsoon season the ICs play a more important role than the Brahmaputra. Only June and July forecasts are dominated by the MF for all lead times. Skill derived from ICs is similar to the Brahmaputra for the rest of the year. The hydrographs for the Ganges and Brahmaputra are characterized by double peaks during the wet season. The second peak due to the reverse Northeastern monsoon in June is more pronounced in the Brahmaputra. Whereas the ICs prior to the first peak are generally the same every year, the ICs before the second peak are strongly conditioned by the strength of the monsoon in the first peak. Therefore, ICs are more important for the skill in the second peak. The decreased skill of the revESP for June to September can be seen on the maps in supporting information Figure 3 (online supporting information data file 2013WR013487-fs03.doc).

[37] In the Indus, where snow and ice have a larger contribution to streamflow [Immerzeel et al., 2012], ICs play a relatively more important role. Snow at lower altitudes begins to melt in March to April, whereas glaciers and snow at higher altitudes begin to melt in June. Therefore June and July forecasts in the Indus are dominated by the ICs for 4 months lead times. The effect of ICs in June on the flow after July is less however because the ice and snowpack at higher altitudes is more or less consistent from year to year. Flow is low during the winter months when most precipitation is stored as snow and ice. The effect of MF is relatively low from November to January, causing a higher CLT of up to 4 months. For the rest of the year forecasts beyond 1 month are dominated by MF.

[38] Our results for the large rivers of China, such as the Yangtze and the Yellow River indicate that hydrologic forecasts are dominated by MF beyond 1 month lead time except for the low flow period in winter. High flows extend from May to October in the Yellow River and from April to September in the Yangtze. Summer flows are monsoon dominated for the most part and the onset of the monsoon coincides with the melting season of snowpack and glaciers, just as for the Indian basins. The variability in MF during these months dominates forecast skill. ICs do not contribute to skill of forecasts for the wet period longer than 1 or 2 months ahead. For the dry period from November through February, the relative importance of MF decreases and the CLT increases to 2, 3, or 4 months lead time.

[39] The results for the Mekong are similar to the Yangtze and the Yellow River. The rainy months in the Mekong are May to October and highest discharges occur from July to October. During the wet season MF dominates the forecasts beyond 1 month lead time. ICs become more important during the dry months. Forecasts for the months February through April derive their skill from ICs of up to 6 months earlier.

4.2. Arctic Basins

[40] In the North American arctic rivers, such as the McKenzie, Yukon and Nelson, forecasts are dominated by ICs for lead times up to 6 months. The importance of ICs is due to the large memory of these arctic river systems. The temperature in the McKenzie is below freezing for about 300 days during the year. The rainy months are July to September. Discharge is lowest in March, and peaks from May to September. The Yukon basin is frozen for almost half of the year with little or no flow. Discharge peaks in May to June, following snowmelt and declines until November when the basin freezes again. Both snowpack and groundwater play an important role in these arctic rivers. These processes have a long memory, causing the large importance of ICs in forecast skill. The forecasts for the month of March in the McKenzie for instance have a CLT of 6 months, which reduces to 2 months by July. In the lake area the large memory introduced by the lakes continues through the summer months (June, July and August) as well.

[41] The relative importance of ICs and MF in Asian arctic rivers follow seasonal patterns similar to the North American arctic rivers. The Ob, Yenisey, and Lena are ice bound from October to April. Forecasts for the colder months are dominated by ICs for lead times of 5–6 months. Following snowmelt, the discharge peaks in all three rivers in the beginning of June. The results show that the skill in the Yenisey basin derives from ICs for up to 6 months for March forecasts, whereas April forecasts are dominated by MF even for a lead time of 1 month. This sudden increase in the importance of MF can be attributed to the effect of temperature in April that determines the start of the melting season. Similar decrease in CLT occurs in the Ob from 6 months for March forecasts to 0 for April forecasts. During the summer, which is the rainy season, ICs play a less important role in these rivers. In the Lena a CLT of 4 months persist into the forecasts for April and May, which again decreases to 0 for July. The Lena basin has a more continental climate than the Ob and Yenisey, with a later onset of the snowmelt in June. This explains the time lag in Lena compared to Ob and Yenisey.

4.3. Temperate Regions

[42] European rivers display quite uniform results in the relative roles of ICs and MF in forecast skill. In general MF dominates beyond lead times of 1–2 months throughout the year. The forecasts for the months of July to October are dominated by MF even for 1 month lead time in Western Europe. In the Rhine basin, skill of the ESP forecasts for these months initiated 1 month ahead is below the climatology. The results agree with the fact that forecast skill in the Rhine is limited to a maximum of 2 weeks during the rain dominated season. Therefore improvement in seasonal hydrological forecasts in these basins depends on better climate forecasts.

[43] In the Danube and Volga, relative skill due to ICs and MF follows the same seasonal pattern as the Rhine but the role of ICs is relatively higher, especially in the Volga where skill due to ICs extends back up to 4 months lead time. Snowmelt dominates the flows in April and May more than it does in the Rhine and groundwater is important during the low flows in winter when most precipitation falls as snow, therefore reducing the importance of MF.

[44] In the western United States Shukla and Lettenmaier [2011] showed that skill due to ICs is high during spring and summer months, mainly June. Our results for the Columbia and Colorado basins agree with the findings of this study. Discharge (80–90%) of the Colorado River originates from snowmelt and the rest from groundwater base flow and summer monsoon rain. Snowmelt begins in April, peaks in May to June, and finishes by July to August. ESP skill maps presented in the online supporting information data file 2013WR013487-fs03.doc demonstrate high skill for up to 6 months lead time in the Colorado basin in spring and summer. Our findings support the conclusions of Shukla and Lettenmaier [2011] that spring and summer forecasts for the western US regions could benefit from improvements in knowledge of ICs during winter and spring months.

[45] For the eastern US, the results of Shukla and Lettenmaier [2011] show that ESP forecasts initialized from December to April are skillful only for 1–2 months lead times. Our results for the St. Lawrence River are in disagreement with these findings. Skill maps presented in supporting information Figure 3 (online supporting information data file 2013WR013487-fs03.doc) indicate that skill due to knowledge of ICs in this basin is higher than the climatology for up to 5–6 months. Peak flows in the St. Lawrence is due to spring and summer snowmelt accompanied by rain. Forecast skill in the spring therefore depends largely on the snowpack accumulated during the previous winter months. The disagreement between our results and those of Shukla and Lettenmaier [2011] is probably due to errors in one or both of the modeling applications in the estimation of snow accumulation. The presence of lakes imply that the routing in our study may also have an affect. The reasons for the different results may be clarified by a further comparative study involving both applications.

[46] For the southeastern rivers such as Mississippi, Missouri and Arkansas our results also agree with those of Shukla and Lettenmaier [2011], showing that skill due to ICs diminishes after 1–2 months lead time. At Vicksburg station on the Mississippi peak flow in spring depends not only on the winter ICs of snowpack, but also on the MF which determines the timing of snowmelt and on rains in the Great Plains and the lower valley. Our results thus confirm the conclusion of Shukla and Lettenmaier [2011] that forecasts in the eastern US would benefit most from improvements in MF throughout the year.

4.4. Semiarid Regions

[47] The results for the Australian basins such as the Murray, Darling, Fitzroy and Coopers Creek show that the relative importance of ICs is the lowest in this continent. Streamflow is not seasonal and highly sensitive to precipitation. Year-round rain is highly erratic with large interannual variability. Different precipitation patterns exist for different parts of the Murray basin. In summer, water from subtropical mountainous regions is lost on the way downstream due to high evaporation. Flow is not fed by groundwater and therefore knowledge of ICs does not contribute to the skill. MF dominates the forecast skill even for 1 month lead time throughout the year. Figure 4 shows that only for July forecasts in the Murray CLT is 2 months. However the skill due to knowledge of ICs for 2 months lead time is below the climatology therefore the CLT is not a meaningful measure in this case. Improvement of hydrological forecasts for Australian basins therefore depends on better climate forecasts.

[48] The interannual variability of rainfall is also high in the semiarid Orange River basin in Africa. The runoff coefficient in this basin is very sensitive to rainfall variability; therefore the ICs do not contribute much to skill. Rainfall after a very dry period causes high unpredictability in the Zambezi as well. While skill of the ESP for November forecasts is high up to 6 months ahead, skill due to ICs is very low for February and March forecasts, which is the season when large floods occur. This is due to the peak inflows into the reservoir Cahora Bassa in February causing the reservoir to either spill or not spill, which in return depends very sensitively on the MF. The results for the Niger basin show the strongest seasonality in the relative importance of ICs and MF. Knowledge of ICs is the main source of skill for forecasts for the months of January to July initiated up to 6 months ahead. Groundwater outflow is the dominant factor for this season. For the rain dominated period from August to December on the other hand, the importance of ICs is limited to lead times of 1–2 months, after which MF becomes more important. In the Congo where there is no rain during the months of May, June and July, ICs play an important role in the forecasts for the months of July and August for up to 4 months lead time. From October to December MF dominate beyond 1 month, while for the rest of the year, IC's contribute to skill for 2–3 months lead time. The Nile stands out in Africa, with CLTs of 5–6 months. Since both stations on the Nile are downstream of the High Aswan Dam, skill of the ESP forecasts is very high throughout the year, assuming that the release strategy of the reservoir is known.

5. Conclusions

[49] We investigated the relative contributions of ICs and MF to the forecasting skill of the global seasonal streamflow forecasting system FEWS-World. Potential improvement in forecasting skill through better climate prediction or by better estimation of initial conditions through data assimilation depends on the relative importance of these two sources of uncertainty. We explored the impact of both sources of forecast uncertainty at large river basins across the globe using the ESP/revESP procedure. Global monthly streamflow was simulated with lead times of 1–6 months for a historical period of 30 years (1981–2010). The ESP and revESP forecast ensembles were compared with retrospective model simulations driven by meteorological observations, thus model errors are eliminated and predictability is related only to knowledge of ICs and the uncertainty in future MF. We compared the variance of the ESP and revESP forecast ensembles to the climatological variance by calculating the ratios of the MSE of both ESP and revESP to the MSE of the climatology for 78 locations, for 12 months of the year and for 6 lead times. We also calculate for each basin and for each month of the year, the CLT after which the importance of ICs is surpassed by that of MF.

[50] Skill maps for the ESP and revESP as well as the CLT values indicate that the contribution of ICs and MF to hydrological forecasting skill varies considerably according to location, season and lead time. We analyzed these results in the context of prevailing hydroclimatic conditions for several larger basins. This analysis suggests that in some basins forecast skill may be improved by better estimation of initial hydrologic states through assimilation of snow, soil moisture or surface water data; whereas in others improvement of forecast skill depends on more accurate seasonal climate prediction. The conclusions can be summarized as follows:

[51] 1. For arctic rivers as well as for rivers fed by snow and ice from mountainous regions, such as the Volga and Colorado, forecasts of high flows during the melt season depend largely on the ice and snowpack, especially where these have a high interannual variability. These forecasts may thus benefit from assimilation of data on the snow and ice accumulated during the cold season.

[52] 2. In some snowfed basins such as the Yenisey and the Mississippi, the onset of ice and/or snowmelt and consequently the timing of peak flow are highly sensitive to temperature changes at the end of the cold season. The importance of ICs diminishes in these cases and improvement of forecast skill in these cases depends more heavily on better climate prediction

[53] 3. In monsoonal basins, the interannual variability of the monsoon is the main factor determining the skill of hydrological forecasts for the wet period. In basins such as the Brahmaputra and the Yangtze where the onset of the thawing of snowpack and glaciers coincides with the start of the monsoon season, forecasts of high flows are dominated by the MF and skill improvement depends on prediction of the monsoon. ICs play a more important role in basins like the Indus where snow and ice have a larger contribution to streamflow, especially when the ice and snowpack is variable from year to year. Better estimation of initial snow/ice states is likely to improve forecast skill during the wet season.

[54] 4. In large basins like the Amazon with extensive flood plains and large travel times of surface water, knowledge of ICs of surface water is an important source of skill for high flow forecasts on lead times of 2–3 months. The role of initial groundwater states also gains importance during the recession stage, when the groundwater discharge plays an important role.

[55] The results of this study show the relative contributions of initial conditions and meteorological forcing to the potential skill of the global seasonal streamflow forecasting system FEWS-World. In an actual forecast, both the ICs and the MF will be uncertain. Therefore as a next step, the actual forecasting skill of the system should be assessed in a real forecasting mode, using probabilistic seasonal meteorological forecasts and comparing the ESP results to actual discharge observations. Since model error cannot be excluded in actual forecasting mode, the resulting skill should be further improved by bias-correcting the meteorological input as well as by better estimation of ICs through data assimilation where it is indicated by this study to be potentially useful.

Acknowledgments

[56] The forecast system, used in this research, has been set up in the 7th Framework Programme project Global Water Scarcity Information Service (GLOWASIS). We acknowledge the 7th Framework Programme of the European Commission for the financial support. Furthermore we acknowledge the European Centre for Medium-ranged Weather Forecasts for making available the ERA Interim—GPCP data set, used in this study as meteorological forcing.