1. Background and motivation
The hydrologic cycle of the Amazon Basin has global importance. It sustains the largest tropical rain forest worldwide, which provides habitat to a huge diversity of species and significantly impacts the balance of carbon dioxide in the atmosphere. Furthermore, it influences hydrometeorological dynamics in neighboring and remote areas. At the continental scale, the Amazon Basin supplies moisture to the La Plata watershed by means of a low-level jet east of the Andes that persists year-round (Berbery and Barros 2002). At a larger scale, the latent heat released in association with the abundant Amazonian rainfall fuels global atmospheric circulation (Costa and Foley 1999).
The Amazon Basin’s hydrology is intimately linked to its land cover, which influences energy and water fluxes at the land surface as well as cloud formation (Chagnon et al. 2004; Chagnon and Bras 2005). The effects and repercussions of accelerating deforestation in this basin on hydrometerological dynamics locally and beyond can only be understood and predicted if its water budget is well quantified [see Laurance et al. (2001) for an overview of Amazonian deforestation].
Pioneering work was carried out by Rasmusson (1967, 1968) in which he used raw radiosonde measurements of atmospheric humidity and wind speed to outline the patterns of atmospheric moisture flux convergence over North America. More recent studies have come to rely on reanalysis products to augment, or in lieu of, raw atmospheric observations. Reanalysis systems assimilate archived observations into an operational forecast model, keeping both the assimilation algorithm and the atmospheric circulation model constant, and thus avoiding artificial shifts and trends in their results (Trenberth et al. 2001). Short-term numerical weather forecasts are adjusted in the direction of available observations to produce “analyses” of atmospheric fields such as upper-air temperature, wind velocity, and humidity. These in turn are used to initialize subsequent forecasts, and so forth. Thus, observations affect the analyses of vertically resolved specific humidity and wind velocity by constraining both the initial conditions of the short-term forecasts and their result. Today’s reanalyses rely on a large database of observations, including rawinsonde, aircraft, and surface marine data, in addition to data from satellite-borne infrared and microwave sensors (Kalnay et al. 1996; Uppala et al. 2005). Their numerical weather models fill in the gaps to produce gridded data at regular time intervals. They thus allow higher spatial and temporal resolution than can be obtained simply from radiosondes.
Studies carried out on the atmospheric water budget of the Mississippi River basin in the United States have shown that reanalysis datasets successfully represent monthly and annual-scale variability in net atmospheric water vapor flux convergence over regions of 105 km2 area or greater (Gutowski et al. 1997; Yeh et al. 1998; Seneviratne et al. 2004). This conclusion, though, cannot be extended to the earth’s tropical regions where observations of atmospheric fields are scarce, and hence the numerical weather models are less constrained. Roads (2003) found large differences between patterns of atmospheric moisture flux convergence over the global tropics derived from the National Centers for Environmental Prediction–National Center for Atmospheric Research (NCEP–NCAR) and NCEP–Department of Energy (DOE) reanalysis products. He attributes these to the fact that observations in these regions are scarce, and hence the estimates of atmospheric moisture and wind velocity are highly model-dependent. The modifications to model parameterizations in the NCEP–DOE reanalysis relative to its predecessor, the NCEP–NCAR reanalysis, apparently significantly impact its simulation of tropical hydrometeorology (Roads 2003).
Our aim in this paper is to characterize and quantify the uncertainty in estimates of the net convergence of atmospheric water vapor flux over the Amazon Basin provided by today’s cutting-edge global reanalysis systems: the NCEP–NCAR reanalysis (Kalnay et al. 1996; henceforth NCEP-1), the NCEP–DOE reanalysis (Kanamitsu et al. 2002; henceforth NCEP-2), and the 40-yr European Centre for Medium-Range Weather Forecasts (ECMWF) Re-Analysis (ERA-40; Uppala et al. 2005). This approach recognizes that there are important differences between these reanalyses that impact simulated atmospheric moisture transport, and there is inadequate evidence to establish the superiority of one over another in relation to its accuracy in depicting water vapor flux convergence over the Amazon region.
2. Methods and datasets
Our study area is the subbasin of the Amazon River, which outlets at Obidos (1°56′S, 55°30′W), about 800 km from the Atlantic Ocean, where the farthest downstream gauging station on the river is located. This watershed constitutes 70% of the entire Amazon Basin, and hence our results can be assumed to be representative of the larger basin. As will become clear later, this control region was selected so as to allow us to utilize the reliable river discharge data collected at the Obidos station in our analysis, rather than relying on uncertain extrapolations of river flow at the Amazon’s mouth. We compare estimates of the net convergence of atmospheric water vapor flux over the Obidos subbasin ([C]) derived from ERA-40, NCEP-1, and NCEP-2. The square brackets used here indicate spatial averaging over our control region. We use specific humidity, wind velocity, and surface pressure data distributed on the reanalysis models’ native horizontal and vertical grids. The horizontal grid is Gaussian for all three models, though the resolution differs between ERA-40 and the two NCEP reanalyses (Table 1). The vertical grid consists of terrain-following sigma levels (σ) in NCEP-1 and NCEP-2, and of a hybrid grid that uses both sigma and pressure levels in ERA-40 (Table 1). All three reanalyses have a temporal resolution of 6 h, providing analyses of atmospheric fields for 4 times per day at 0000, 0600, 1200, and 1800 UTC.
Monthly averages of Q are computed, as we are primarily interested in the seasonal signature in the Amazon Basin’s water budget. Since the pressure at the sigma levels is time varying, depending on surface pressure, the vertical integral must be computed prior to temporal averaging (Trenberth et al. 2002). The spatially distributed divergence of Q is then computed by a centered difference algorithm on the reanalyses’ Gaussian grid.
In the case of ERA-40, we use two of its archived data products: 1) monthly means of vertically integrated atmospheric moisture flux Q and 2) their divergence, both spatially distributed fields. The archived divergence field is computed in the model’s spectral space, substituting multiplication for derivation and producing a more accurate result than a finite-difference approximation of the derivatives on the Gaussian grid (Seneviratne et al. 2004).
For each of the three data sources, we compute the spatial integral of the divergence of Q over our control region and divide it by the region’s area. Taking the negative of the result yields the monthly net convergence of atmospheric water vapor flux over the region, [C], in units of mm day−1.
In our work, the Obidos subbasin constituting our study region is defined using the Total Runoff Integrating Pathways (TRIP) dataset at 0.5° resolution created by Oki and Sud (1998) based on the 5-minute gridded elevations/bathymetry for the world (ETOPO5) global digital elevation model (DEM), which has a 5′ × 5′ horizontal resolution (Fig. 1). The area of this watershed obtained from the TRIP dataset (4.784 × 106 km2) compares well with the area cited in Callede et al. (2002) of 4.676 × 106 km2.
Monthly averages of Amazon river discharge at the Obidos gauging station for the period 1980–2001 were kindly provided by Jacques Callede, who is affiliated with the Hydrology and Geochemistry of the Amazon (HyBAm) project through the French “Institut de Recherches pour le Développement” (IRD) (Callede et al. 2002). Work by Callede et al. (2001, 2002) improved the state discharge relationship for this station based on detailed and long-term discharge measurements using an acoustic Doppler current profiler. The resulting relationship reduced the mean dispersion between stage-derived discharge and directly measured discharge to 2.9%.
3. Quality of and uncertainty in reanalysis estimates of [C]
Figure 2 shows monthly [C] for the period 1997–2001, derived from the three reanalyses. Two ERA-40–based time series are plotted: the first is based on the archived divergence of the Q product, which is computed in the model’s spectral space, and the second is derived from the archived Q product using center difference on the model’s Gaussian grid, as was done in the case of NCEP-1 and NCEP-2 data (see section 2). The climatology of [C] for this 5-yr period according to the different data sources is shown in Fig. 3. The seasonal cycles simulated by the three reanalyses are similar, with a minimum in July–September and a peak in January–March. However, large discrepancies between monthly [C] estimates produced by the different models are evident (Fig. 2).
Figure 4 shows the monthly anomalies of [C], which were derived from each of the [C] time series by subtracting its mean annual cycle over the 5-yr period. The correlations between the resulting estimates are presented in Table 2. NCEP-1 and NCEP-2 agree well in depicting monthly and interannual variability in [C], while the correlation between their results and those of ERA-40 is poor. This is expected given the greater similarity between the underlying models of the two NCEP reanalyses.
We find that σm is 0.72 mm day−1, significantly exceeding σpp, which is 0.15 mm day−1. Figure 3 shows that the average range of [C] over a climatological annual cycle is about 5 mm day−1. Hence, σm is more than 14% of this range. A closer look at the error introduced by the numerical approximation of divergence shows that this approximation leads consistently to an underestimation of monthly [C], by 0.13 mm day−1 on average (Fig. 5). Hence, a bias correction applied to [C] based on an independent estimate of its long-term average (such as average river discharge at Obidos) will reduce the absolute error attributed to the discretization of divergence by about half, to an average of about 0.06 mm day−1. This error would then be an order of magnitude smaller than the model-associated uncertainty in [C] estimates (σm = 0.72 mm day−1), and would thus be negligible.
From the preceding investigation of reanalysis-derived estimates of [C] for the Obidos subbasin over the limited period of 1997–2001, we conclude that they exhibit significant uncertainty at the monthly and interannual time scales. The model-associated component of this uncertainty overwhelms the postprocessing errors introduced by the finite difference approximation of divergence. In the following sections we analyze longer time series of [C] derived from each reanalysis, with the aim of assigning to each an error estimate based on comparison to independent, more reliable data.
4. Discontinuity in ERA-40 [C] coinciding with its incorporation of SSM/I data
Figure 6 shows annual [C] derived from ERA-40, NCEP-1, and NCEP-2, as well as annual basin-averaged runoff [R] for the 22-yr period 1980–2001. Here [R] is computed by dividing annual discharge observed at the Obidos gauging station by the area of the subbasin used as our control region. A large bias in the [C] estimates from all three reanalyses is evident, since long-term averages of [C] and [R] should equal each other so as to close the basin’s water balance.
Figures 7a–c allow a closer look at the time series of annual [C] derived from the three reanalyses. There is a clear change in the record derived from ERA-40 between the years of 1987 and 1988 (Fig. 7c). It appears that the annual [C] values following that point have a positive bias relative to those preceding it. A similar transition in behavior is not apparent in the NCEP-1 and NCEP-2 data (Figs. 7a,b). A plot of monthly averages of [C] over this 22-yr period shows further evidence of a transition in the nature of the ERA-40 record around 1987: the amplitude of the seasonal cycles is consistently greater after that year (Fig. 8c). In contrast, NCEP-1 and NCEP-2 data do not exhibit a similar change in seasonal range (Figs. 8a,b).
This transition in estimates of Amazonian atmospheric moisture flux convergence by ERA-40 has not been explicitly described in the literature. It appears to coincide with the time when Special Sensor Microwave Imager (SSM/I) data became available for assimilation in the ECMWF reanalysis (starting in June 1987). Although reanalysis models are frozen, the data assimilated changes over the years as new sources of data become available, which has a major impact on the resulting analyses (Betts et al. 2005). SSM/I radiances are assimilated for the analysis of total column water vapor over the ocean. They present an important addition to the assimilated satellite data, which were previously limited to Television and Infrared Observation Satellite (TIROS) Operational Vertical Sounder (TOVS) data relating to atmospheric humidity and temperature profiles, and cloud motion wind data from geostationary satellites (Betts et al. 2005). Note that the NCEP-1 and NCEP-2 models similarly rely on TOVS and cloud motion winds data but do not assimilate SSM/I data.
Betts et al. (2005) identified a potentially related transition in the atmospheric water budget simulated by ERA-40 over the Amazon Basin. They found that ERA-40 exhibits large analysis increments1 for total precipitable water (TPW; equivalent to total atmospheric column water vapor) over this region during the period 1973–87. These analysis increments drop rapidly between 1987 and 1988, coinciding with the rise in [C] evident in our results. Betts et al. (2005) do not explicitly state the connection between the assimilation of SSM/I radiances and the observed shift in the pattern of TPW analysis increments. Nevertheless, their results support ours in suggesting an important influence of SSM/I data assimilation on ERA-40 analyses relating to atmospheric moisture transport over the Amazon Basin.
Sudradjat et al. (2005) did a global comparison of monthly averages of TPW derived from ERA-40, NCEP-1, and NCEP-2, amongst each other and to estimates produced by NASA’s Water Vapor Project (NVAP) dataset, over the period January 1988–December 1999. Over the tropical oceans, they find better agreement between the ERA-40 and NVAP TPW fields, relative to those simulated by the NCEP reanalyses, which they attribute to a common dependence on SSM/I observations. They conclude that the assimilation of SSM/I data by ERA-40 has a significant effect on the long-term mean of the TPW analysis over the tropical oceans as well as on its variability at the monthly/seasonal and interannual time scales. As atmospheric water vapor is transported into the Amazon Basin from the tropical Atlantic (Costa and Foley 1999), the TPW analyses over the ocean are expected to have an important influence on the magnitude of simulated atmospheric moisture flux convergence over this basin.
The effect of the introduction of SSM/I data on the ERA-40 analyses of upper-air fields, both on their long-term means and their variability at higher temporal resolutions, has not been quantified. Hence, it is difficult to formulate an adequate correction that assures continuity in the ERA-40 estimates of [C] over the period 1980–2001. For this reason, our subsequent investigation is limited to the period 1988–2001 to avoid the artificial discontinuity in the ERA-40 record, likely related to the start of assimilation of SSM/I radiances in TPW analyses.
5. Bias error versus random error in [C] estimates
We begin by correcting the bias in [C] data derived from the three reanalyses, ERA-40, NCEP-1, and NCEP-2. This is done by adjusting the full-length average of each time series to match that of discharge at Obidos, our subbasin’s outlet. Multiyear averages of [C] and [R] must agree, since over windows of several years, changes in water storage on land and in the atmosphere can be neglected, and the net convergence of atmospheric water vapor flux into a region translates completely to surface and groundwater flow out of its boundaries.
In all subsequent work, we use water years instead of calendar years in computing the hydrologic budget so as not to split the rainy season in two and to account for the lag between atmospheric moisture flux convergence over the basin and river discharge at its mouth. The climatological annual cycles of both [C] and spatially averaged precipitation over our control region ([P]) reach their minimum in August (Fig. 9). The response in discharge at Obidos follows 3 months later, with minimum [R] occurring in November (Fig. 9). Hence, we define the hydrological year for Amazonian rainfall and atmospheric moisture flux convergence to extend between 1 September and 31 August of the following year, agreeing with Marengo (2005) and Betts et al. (2005), while that for river discharge extends between 1 December and 30 November of the following year. The time series of monthly [C] and [P] used in this analysis begin in September 1987 and end in August 2001, while that of [R] extends between December 1987 and November 2001. All annual averaging is done in terms of hydrologic years, implicitly accounting for the lag between [C] and [R].
The 14-yr averages of [C] estimates derived from the various reanalyses and their bias errors are listed in Table 3. All three reanalyses give artificially low estimates of [C], with biases ranging from about 50% (ERA-40 and NCEP-1) to 74% (NCEP-2) of mean runoff. Others have similarly identified a large negative bias in reanalysis-derived estimates of net atmospheric moisture flux convergence over the Amazon Basin. Roads (2002) and Marengo (2005) found that the negative bias in estimates of this field based on the NCEP–NCAR reanalysis is about 50% of the long-term mean discharge of the Amazon River, matching our results. Zeng (1999) found that this bias amounts to 73% of river discharge when data from the Goddard Earth Observing System (GEOS-1) reanalysis are used.
Making the zero-order assumption that the observed bias error is uniformly distributed over our 14-yr time series, we subtract a constant bias error from the original monthly estimates of [C]. Such an approach has been adopted in many other studies of regional atmospheric water budgets (e.g., Rasmusson 1971; Marengo 2005). The bias-corrected estimates of annual [C] from the three reanalyses for the hydrological years 1987/88–2000/01 are presented in Fig. 10, alongside annual basin-averaged runoff.
Figure 10 shows that following bias correction, the agreement between estimates of [C] derived from different reanalyses remains poor, even for annual means of this variable. This is not unexpected. The availability of radiosonde observations of atmospheric fields is poor in the Amazon region, and thus these “conventional” observations only weakly constrain the numerical weather prediction (NWP) models of the reanalyses (Marengo 2005; Roads 2003). Furthermore, there are important differences in the types of satellite data related to upper-air fields assimilated by the different reanalyses, as well as in the methods of assimilation employed. While ERA-40 assimilates both TOVS and SSM/I radiances, the U.S. reanalyses do not utilize the SSM/I data (Kalnay et al. 1996; Betts et al. 2005; Sudradjat et al. 2005). Moreover, the U.S. reanalyses do not utilize TOVS-derived water vapor information and only assimilate the vertical temperature soundings retrieved from TOVS sensors by the National Environmental Satellite, Data, and Information Service (NESDIS) (Kistler et al. 2001; Trenberth and Guillemot 1998). These temperature soundings are derived using observations from the three TOVS instruments: the High Resolution Infrared Radiation Sounder (HIRS/2), the Stratospheric Sounding Unit (SSU), and the Microwave Sounding Unit (MSU) (Kidwell 1998). In contrast, ERA-40 uses three-dimensional variational data assimilation to directly assimilate TOVS radiances (Uppala et al. 2005; Sudradjat et al. 2005). As TOVS radiances depend strongly on both atmospheric temperature and humidity, the analysis of both variables is affected. Other important differences between the U.S. reanalyses and ERA-40 are related to their NWP models. Differences in their vertical and horizontal resolutions have important effects on simulated upper-air fields (Sudradjat et al. 2005). Moreover, the three reanalyses differ in the physical parameterizations they employ, including their convective and boundary layer parameterizations, which affect moisture transport in the models. NCEP-2 uses different boundary layer, shortwave radiation, and convective parameterizations from those used in NCEP-1, and thus shows important differences in its atmospheric humidity and transport patterns (Sudradjat et al. 2005; Roads 2003; Kanamitsu et al. 2002).
A particularly striking feature in the NCEP-1 and NCEP-2-based time series of [C] plotted in Fig. 10 is that they remain persistently below their mean value between the years of 1992 and 1998. In contrast, the time series of [R] shows lower-than-average values in the years 1992 and 1998, but not for the whole period in between; the same can be said about the [C] time series derived from ERA-40. Biases in tropical climate simulated by the NCEP/NCAR reanalysis between the years 1992 and 1998 have been identified by others, particularly in relation to tropospheric temperatures. Several studies have described a negative bias in tropospheric temperatures simulated by NCEP-1 during this period, relative to tropospheric temperatures derived from MSU radiances (Basist and Chelliah 1997; Chelliah and Ropelewski 2000; Stendel et al. 2000). While these two datasets of tropospheric temperatures are not entirely independent, the TOVS temperature retrievals assimilated by the NCEP reanalysis incorporate HIRS/2 and SSU observations, in addition to MSU data, and use a different algorithm for processing MSU radiances (Basist and Chelliah 1997). The existence of this bias in NCEP-1 tropospheric temperatures was further corroborated by comparison to a surface-based observational dataset (Chelliah and Ropelewski 2000). While the reasons for it remain unresolved, there are two dominant explanations, both relating to biases in the TOVS temperature soundings assimilated by the NCEP reanalysis (Basist and Chelliah 1997; Chelliah and Ropelewski 2000). The most cited reason, first presented by Basist and Chelliah (1997), relates to a change in the algorithms used by NESDIS for retrieving temperature soundings from TOVS sensors over cloudy areas, implemented in April 1992 (Basist and Chelliah 1997; Chelliah and Ropelewski 2000; Trenberth et al. 2001). The second reason presented by Basist and Chelliah (1997) relates to the interference of aerosols produced by the Mount Pinatubo eruption in June 1991 with observed radiances from HIRS/2. In the NESDIS temperature retrievals, corrections to the HIRS/2 radiances were implemented to account for the aerosol effects using collocated radiosonde observations (Stubenrauch et al. 2006). However, where radiosonde observations are scarce, it is likely that these corrections were inadequate (Basist and Chelliah 1997).
The negative bias in tropospheric temperatures simulated by NCEP-1 during the 1990s was observed in data-scarce regions, particularly in the tropics and areas of the former Soviet Union, and was absent in data-rich regions such as North America and Australia (Basist and Chelliah 1997; Stendel et al. 2000). This is expected, given that satellite-based observations are only given significant weight in the reanalysis where more conventional data, such as radiosonde, aircraft, ship, and buoy data, are limited (Basist and Chelliah 1997).
Note that the negative bias in NCEP-1 tropical tropospheric temperatures described in the studies reviewed above is consistent with a negative bias in simulated atmospheric water vapor concentrations, following the Clausius–Clapeyron relationship (Sudradjat et al. 2005), and hence is also consistent with the negative bias in simulated atmospheric moisture flux convergence suggested by our study. Therefore, our results corroborate previous findings concerning a biased NCEP-1 tropical climate in the 1990s, and further show that this bias is not corrected and appears even more pronounced in the NCEP-2 reanalysis (Fig. 10). ERA-40 may have been buffered from such a bias by its assimilation of SSM/I data in addition to TOVS radiances.
In conclusion, inadequate model parameterizations and scarce and/or inaccurate observations produce error in [C] estimates derived from available reanalysis datasets. The component of this error that persists following bias correction can be viewed as random error. This random error cannot be estimated by tracking the errors associated with the atmospheric humidity and wind speed fields that constitute C, as there are no adequate independent observations of these fields, particularly in the Amazon Basin. It must thus be estimated by relating the estimates of [C] to other data that are less uncertain, particularly river discharge and precipitation records.
a. Estimating random error in [C] time series by comparison to river discharge
Callede et al. (2004) show that there is no significant autocorrelation in the annual river discharge record at the Obidos gauging station over the period 1903–99. This implies that net changes in terrestrial water storage at the interannual time scale integrated over the contributing basin’s area are minimal. Thus, over any given 5-yr period, the total convergence of atmospheric water vapor flux over our control region should equal discharge at its outlet.
b. A comparison of [C] and basin-averaged rainfall
In evaluating a region’s atmospheric water budget, surface rainfall is undeniably the hydrologic component for which observations are usually most abundant, and that can be most confidently quantified by available datasets. In this section, we carry out a qualitative comparison of domain-averaged rainfall over the Obidos subbasin ([P]) and net atmospheric moisture flux convergence, to extract further information on the reliability of the latter field as simulated by the various reanalyses.
Figure 12 presents the anomalies in annual averages of [C] and [P] relative to their respective 14-yr means, computed using atmospheric data from each reanalysis and rainfall data from the Global Precipitation Climatology Project (GPCP) Combined Precipitation Dataset (Version 2) (Huffman et al. 1997). The time series of [P] anomalies show the characteristic signatures of El Niño and La Niña events. In the Amazon Basin, El Niño events are associated with negative anomalies in [P], while the reverse is observed for La Niña events (Marengo 2005). The strong El Niño events of 1991/92 and 1997/98 show up as distinctive negative anomalies in annual [P]. The positive [P] anomalies can all be explained by the occurrence of La Niña events, except for that at 1993/94.
The various El Niño and La Niña events are reflected to different extents in the three time series of [C] anomalies (Fig. 12). Most obvious is the large negative bias in NCEP-1 and NCEP-2 estimates between 1992 and 1998. The excessive negative anomalies in [C] data from the two NCEP reanalyses during this period are neither paralleled in the precipitation data nor in the river discharge record (Figs. 12, 10, 11). This corroborates the conclusion that these anomalies are artificial.
6. Conclusions
Our study shows that reanalysis-derived estimates of the net convergence of atmospheric moisture flux over the Amazon Basin are highly uncertain. This uncertainty is primarily “model-associated,” originating in the reanalysis data products because of inadequacies in underlying parameterizations of hydrometeorological processes and data assimilation algorithms. A close look at time series of [C] for the Obidos subbasin covering the period 1980–2001, derived from each of NCEP-1, NCEP-2, and ERA-40, revealed distinctive error patterns that may be traced to satellite-based data assimilated by these reanalyses. A shift in ERA-40 [C] estimates is apparent around 1987, such that estimates following this year are, on average, greater than those preceding it. This shift coincides with the start of assimilation of SSM/I data in this reanalysis, and may be explained by the impact of this additional dataset on the model. A similar transition is not obvious in the [C] time series derived from NCEP-1 and NCEP-2, which do not assimilate SSM/I data. In these latter time series, a pronounced drop is evident between 1991 and 1992, which persists with artificially low values of [C] for several years. This negative bias likely originates in biased TOVS data assimilated by the NCEP models. It is consistent with a negative bias in tropospheric temperatures simulated by NCEP-1 between 1992 and 1998, identified by other investigators.
A quantitative description of the error in the time series of [C] was carried out for the period covering 1988–2001. Large bias error was identified in estimates of this field derived from all three reanalyses upon comparison to river discharge data. This bias is expected, as it has been identified in several other studies, and is easily corrected for if one assumes that it is uniformly distributed in time. Random error in each of the time series was estimated at the annual time scale, also using river discharge data as reference. While this measure of error provides no information on the subannual accuracy of the [C] estimates, it is assumed to be an indicator of the relative reliability of [C] estimates produced by the different reanalyses, regardless of temporal scale. For the 14-yr study period our random error estimates suggest that ERA-40 simulates Amazonian atmospheric moisture flux convergence more accurately than either of the NCEP reanalyses. The bias in the NCEP time series between 1992 and 1998, described above, appears to contribute most to their elevated “random” error values. Hence, a sound correction of this bias, not attempted here, would increase the accuracy of NCEP-1 and NCEP-2 estimates of atmospheric moisture flux convergence over the Amazon Basin.
Finally, our study highlights the importance of carefully investigating the accuracy of reanalysis data prior to utilizing them in water budget studies. An understanding of the changes in the observations assimilated by a given reanalysis and their effects on simulated climate is especially helpful in identifying artificial shifts and biases in its data products. Furthermore, comparison to independent runoff and rainfall data is very useful, particularly for studies of regional water budgets. In regions where unbiased radiosonde observations are scarce, such as the Amazon Basin and other tropical regions, reanalyses become particularly sensitive to errors in their atmospheric circulation models as well as biases in assimilated satellite data (Trenberth et al. 2001). In hydrological studies of these regions, it is especially important to avoid relying on one reanalysis as a source of atmospheric data, and rather to make use of and compare between multiple available data sources.
Acknowledgments
Grant funding from TRMM/NASA (NAG5-13638) supported this work. Data on discharge of the Amazon River at the Obidos gauging station were provided by Jacques Callede of the French “Institut de Recherches pour le Développement” (IRD). The Ferret program was used for data analysis and some graphics production. Ferret is a product of NOAA’s Pacific Marine Environmental Laboratory. Information is available online (http://ferret.pmel.noaa.gov/Ferret/). We are very grateful to Dr. Jingfeng Wang for his comments on the draft of this paper and to Daniel Sheehan for his guidance in using GIS software for watershed delineation. We are also thankful for the suggestions made by three anonymous reviewers toward improving this paper in content and presentation.
REFERENCES
Basist, A. N., and Chelliah M. , 1997: Comparison of tropospheric temperatures derived from the NCEP/NCAR reanalysis, NCEP operational analysis, and the Microwave Sounding Unit. Bull. Amer. Meteor. Soc., 78 , 1431–1447.
Berbery, E. H., and Barros V. R. , 2002: The hydrological cycle of the La Plata basin in South America. J. Hydrometeor., 3 , 630–645.
Betts, A. K., Ball J. H. , Viterbo P. , Dai A. G. , and Marengo J. , 2005: Hydrometeorology of the Amazon. J. Hydrometeor., 6 , 764–774.
Callede, J., Kosuth P. , and de Oliveira E. , 2001: Establishment of the stage-discharge relationship of the River Amazon at Obidos: Normal difference in level method using variable geometry. Hydrol. Sci. J., 46 , 451–463.
Callede, J., Guyot J. L. , Ronchail J. , Molinier M. , and De Oliveira E. , 2002: The River Amazon at Obidos (Brazil): Statistical studies of the discharges and water balance. Hydrol. Sci. J., 47 , 321–333.
Callede, J., Guyot J. L. , Ronchail J. , L’Hote Y. , Niel H. , and de Oliveira E. , 2004: Evolution of the River Amazon’s discharge at Obidos from 1903 to 1999. Hydrol. Sci. J., 49 , 85–97.
Chagnon, F. J. F., and Bras R. L. , 2005: Contemporary climate change in the Amazon. Geophys. Res. Lett., 32 .L13703, doi:10.1029/2005GL022722.
Chagnon, F. J. F., Bras R. L. , and Wang J. , 2004: Climatic shift in patterns of shallow clouds over the Amazon. Geophys. Res. Lett., 31 .L24212, doi:10.1029/2004GL021188.
Chelliah, M., and Ropelewski C. F. , 2000: Reanalyses-based tropospheric temperature estimates: Uncertainties in the context of global climate change detection. J. Climate, 13 , 3187–3205.
Costa, M. H., and Foley J. A. , 1999: Trends in the hydrologic cycle of the Amazon basin. J. Geophys. Res., 104 , 14189–14198.
Gutowski W. J. Jr., , Chen Y. , and Ötles Z. , 1997: Atmospheric water vapor transport in NCEP–NCAR reanalyses: Comparison with river discharge in the central United States. Bull. Amer. Meteor. Soc., 78 , 1957–1969.
Huffman, G. J., and Coauthors, 1997: The Global Precipitation Climatology Project (GPCP) combined precipitation dataset. Bull. Amer. Meteor. Soc., 78 , 5–20.
Kalnay, E., and Coauthors, 1996: The NCEP/NCAR 40-Year Reanalysis Project. Bull. Amer. Meteor. Soc., 77 , 437–471.
Kanamitsu, M., Ebisuzaki W. , Woollen J. , Yang S. , Hnilo J. J. , Fiorino M. , and Potter G. L. , 2002: NCEP–DOE AMIP-II Reanalysis (R-2). Bull. Amer. Meteor. Soc., 83 , 1631–1643.
Karam, H. N., and Bras R. L. , 2008: Climatological basin-scale Amazonian evapotranspiration estimated through a water budget analysis. J. Hydrometeor., 9 , 1048–1060.
Kidwell, K. B., 1998: NOAA polar orbiter data user’s guide. NOAA/NESDIS, November 1998 Revision. [Available online at http://www2.ncdc.noaa.gov/docs/podug/cover.htm.].
Kistler, R., and Coauthors, 2001: The NCEP–NCAR 50-Year Reanalysis: Monthly means CD-ROM and documentation. Bull. Amer. Meteor. Soc., 82 , 247–267.
Laurance, W. F., Albernaz A. K. M. , and Da Costa C. , 2001: Is deforestation accelerating in the Brazilian Amazon? Environ. Conserv., 28 , 305–311.
Marengo, J. A., 2005: Characteristics and spatio-temporal variability of the Amazon River Basin Water Budget. Climate Dyn., 24 , 11–22.
Oki, T., and Sud Y. C. , 1998: Design of Total Runoff Integrating Pathways (TRIP)—A global river channel network. Earth Interactions, 2 .[Available online at http://EarthInteractions.org.].
Rasmusson, E. M., 1967: Atmospheric water vapor transport and the water balance of North America. Part I: Characteristics of the water vapor flux field. Mon. Wea. Rev., 95 , 403–426.
Rasmusson, E. M., 1968: Atmospheric water vapor transport and the water balance of North America. Part II: Large-scale water balance investigations. Mon. Wea. Rev., 96 , 720–734.
Rasmusson, E. M., 1971: A study of the hydrology of eastern North America using atmospheric water vapor flux data. Mon. Wea. Rev., 99 , 119–135.
Roads, J., 2002: Closing the water cycle. GEWEX News, No. 12, International GEWEX Project Office, Silver Spring, MD, 1–8.
Roads, J., 2003: The NCEP–NCAR, NCEP–DOE, and TRMM tropical atmosphere hydrologic cycles. J. Hydrometeor., 4 , 826–840.
Seneviratne, S. I., Viterbo P. , Lüthi D. , and Schär C. , 2004: Inferring changes in terrestrial water storage using ERA-40 reanalysis data: The Mississippi River basin. J. Climate, 17 , 2039–2057.
Stendel, M., Christy J. R. , and Bengtsson L. , 2000: Assessing levels of uncertainty in recent temperature time series. Climate Dyn., 16 , 587–601.
Stubenrauch, C. J., Chédin A. , Rädel G. , Scott N. A. , and Serrar S. , 2006: Cloud properties and their seasonal and diurnal variability from TOVS Path-B. J. Climate, 19 , 5531–5553.
Sudradjat, A., Ferraro R. R. , and Fiorino M. , 2005: A comparison of total precipitable water between reanalyses and NVAP. J. Climate, 18 , 1790–1807.
Trenberth, K. E., and Guillemot C. J. , 1998: Evaluation of the atmospheric moisture and hydrological cycle in the NCEP-NCAR reanalyses. Climate Dyn., 14 , 213–231.
Trenberth, K. E., Stepaniak D. P. , Hurrell J. W. , and Fiorino M. , 2001: Quality of reanalyses in the Tropics. J. Climate, 14 , 1499–1510.
Trenberth, K. E., Stepaniak D. P. , and Caron J. M. , 2002: Accuracy of atmospheric energy budgets from analyses. J. Climate, 15 , 3343–3360.
Uppala, S. M., and Coauthors, 2005: The ERA-40 re-analysis. Quart. J. Roy. Meteor. Soc., 131 , 2961–3012.
Yeh, P. J. F., Irizarry M. , and Eltahir E. A. B. , 1998: Hydroclimatology of Illinois: A comparison of monthly evaporation estimates based on atmospheric water balance and soil water balance. J. Geophys. Res., 103 , 19823–19837.
Zeng, N., 1999: Seasonal cycle and interannual variability in the Amazon hydrologic cycle. J. Geophys. Res., 104 , 9097–9106.
Outline (in dark gray) of the Amazonian subbasin that outlets at the Obidos gauging station (1.9°S, 55.5°W, marked with the black triangle). The entire Amazon Basin is colored in lighter gray. The delineation of these was done based on the TRIP 0.5° dataset using ArcGIS software. Also shown is the Amazon River network derived from the TRIP dataset using ArcGIS software.
Citation: Journal of Hydrometeorology 9, 5; 10.1175/2008JHM887.1
Outline (in dark gray) of the Amazonian subbasin that outlets at the Obidos gauging station (1.9°S, 55.5°W, marked with the black triangle). The entire Amazon Basin is colored in lighter gray. The delineation of these was done based on the TRIP 0.5° dataset using ArcGIS software. Also shown is the Amazon River network derived from the TRIP dataset using ArcGIS software.
Citation: Journal of Hydrometeorology 9, 5; 10.1175/2008JHM887.1
Outline (in dark gray) of the Amazonian subbasin that outlets at the Obidos gauging station (1.9°S, 55.5°W, marked with the black triangle). The entire Amazon Basin is colored in lighter gray. The delineation of these was done based on the TRIP 0.5° dataset using ArcGIS software. Also shown is the Amazon River network derived from the TRIP dataset using ArcGIS software.
Citation: Journal of Hydrometeorology 9, 5; 10.1175/2008JHM887.1
Monthly net convergence of atmospheric water vapor flux over the Obidos subbasin ([C]) between January 1997 and December 2001, derived from ERA-40, using the ∇ · Q field computed in spectral space (white squares); ERA-40 using the finite-difference computation of ∇ · Q (black line), NCEP-1 (black circles), and NCEP-2 (gray triangles). Negative values indicate net divergence over the basin.
Citation: Journal of Hydrometeorology 9, 5; 10.1175/2008JHM887.1
Monthly net convergence of atmospheric water vapor flux over the Obidos subbasin ([C]) between January 1997 and December 2001, derived from ERA-40, using the ∇ · Q field computed in spectral space (white squares); ERA-40 using the finite-difference computation of ∇ · Q (black line), NCEP-1 (black circles), and NCEP-2 (gray triangles). Negative values indicate net divergence over the basin.
Citation: Journal of Hydrometeorology 9, 5; 10.1175/2008JHM887.1
Monthly net convergence of atmospheric water vapor flux over the Obidos subbasin ([C]) between January 1997 and December 2001, derived from ERA-40, using the ∇ · Q field computed in spectral space (white squares); ERA-40 using the finite-difference computation of ∇ · Q (black line), NCEP-1 (black circles), and NCEP-2 (gray triangles). Negative values indicate net divergence over the basin.
Citation: Journal of Hydrometeorology 9, 5; 10.1175/2008JHM887.1
Climatology of [C] for the period 1997–2001, derived from ERA-40 using the ∇ · Q field computed in spectral space (white squares), NCEP-1 (black circles), and NCEP-2 (gray triangles). Negative values indicate net divergence.
Citation: Journal of Hydrometeorology 9, 5; 10.1175/2008JHM887.1
Climatology of [C] for the period 1997–2001, derived from ERA-40 using the ∇ · Q field computed in spectral space (white squares), NCEP-1 (black circles), and NCEP-2 (gray triangles). Negative values indicate net divergence.
Citation: Journal of Hydrometeorology 9, 5; 10.1175/2008JHM887.1
Climatology of [C] for the period 1997–2001, derived from ERA-40 using the ∇ · Q field computed in spectral space (white squares), NCEP-1 (black circles), and NCEP-2 (gray triangles). Negative values indicate net divergence.
Citation: Journal of Hydrometeorology 9, 5; 10.1175/2008JHM887.1
Monthly anomalies of Amazonian [C] relative to the mean annual cycle for January 1997–December 2001, computed from: ERA-40 (white squares), NCEP-1 (black circles), and NCEP-2 (gray triangles).
Citation: Journal of Hydrometeorology 9, 5; 10.1175/2008JHM887.1
Monthly anomalies of Amazonian [C] relative to the mean annual cycle for January 1997–December 2001, computed from: ERA-40 (white squares), NCEP-1 (black circles), and NCEP-2 (gray triangles).
Citation: Journal of Hydrometeorology 9, 5; 10.1175/2008JHM887.1
Monthly anomalies of Amazonian [C] relative to the mean annual cycle for January 1997–December 2001, computed from: ERA-40 (white squares), NCEP-1 (black circles), and NCEP-2 (gray triangles).
Citation: Journal of Hydrometeorology 9, 5; 10.1175/2008JHM887.1
Error in monthly [C] computed from ERA-40 data using a finite-difference approximation of ∇ · Q, relative to its computation in spectral space, also from ERA-40 data. The dashed line is drawn at the average error, −0.13 mm day−1. Negative values indicate underestimation by the finite-difference algorithm.
Citation: Journal of Hydrometeorology 9, 5; 10.1175/2008JHM887.1
Error in monthly [C] computed from ERA-40 data using a finite-difference approximation of ∇ · Q, relative to its computation in spectral space, also from ERA-40 data. The dashed line is drawn at the average error, −0.13 mm day−1. Negative values indicate underestimation by the finite-difference algorithm.
Citation: Journal of Hydrometeorology 9, 5; 10.1175/2008JHM887.1
Error in monthly [C] computed from ERA-40 data using a finite-difference approximation of ∇ · Q, relative to its computation in spectral space, also from ERA-40 data. The dashed line is drawn at the average error, −0.13 mm day−1. Negative values indicate underestimation by the finite-difference algorithm.
Citation: Journal of Hydrometeorology 9, 5; 10.1175/2008JHM887.1
Annual averages of [C] and [R] between 1980 and 2001. The [R] was computed from discharge observations at Obidos (black line) and [C] was computed from ERA-40 (white squares), NCEP-1 (black circles), and NCEP-2 (gray triangles).
Citation: Journal of Hydrometeorology 9, 5; 10.1175/2008JHM887.1
Annual averages of [C] and [R] between 1980 and 2001. The [R] was computed from discharge observations at Obidos (black line) and [C] was computed from ERA-40 (white squares), NCEP-1 (black circles), and NCEP-2 (gray triangles).
Citation: Journal of Hydrometeorology 9, 5; 10.1175/2008JHM887.1
Annual averages of [C] and [R] between 1980 and 2001. The [R] was computed from discharge observations at Obidos (black line) and [C] was computed from ERA-40 (white squares), NCEP-1 (black circles), and NCEP-2 (gray triangles).
Citation: Journal of Hydrometeorology 9, 5; 10.1175/2008JHM887.1
Annual [C] for 1980–2001 and its 22-yr average (dashed line), derived from (a) NCEP-1, (b) NCEP-2, and (c) ERA-40. Note the difference in scales on the y axis.
Citation: Journal of Hydrometeorology 9, 5; 10.1175/2008JHM887.1
Annual [C] for 1980–2001 and its 22-yr average (dashed line), derived from (a) NCEP-1, (b) NCEP-2, and (c) ERA-40. Note the difference in scales on the y axis.
Citation: Journal of Hydrometeorology 9, 5; 10.1175/2008JHM887.1
Annual [C] for 1980–2001 and its 22-yr average (dashed line), derived from (a) NCEP-1, (b) NCEP-2, and (c) ERA-40. Note the difference in scales on the y axis.
Citation: Journal of Hydrometeorology 9, 5; 10.1175/2008JHM887.1
Monthly [C] for the period January 1980–December 2001, derived from (a) NCEP-1, (b) NCEP-2, and (c) ERA-40.
Citation: Journal of Hydrometeorology 9, 5; 10.1175/2008JHM887.1
Monthly [C] for the period January 1980–December 2001, derived from (a) NCEP-1, (b) NCEP-2, and (c) ERA-40.
Citation: Journal of Hydrometeorology 9, 5; 10.1175/2008JHM887.1
Monthly [C] for the period January 1980–December 2001, derived from (a) NCEP-1, (b) NCEP-2, and (c) ERA-40.
Citation: Journal of Hydrometeorology 9, 5; 10.1175/2008JHM887.1
Climatological annual cycle of [P] (solid black line), [R] (thick gray line), and [C] derived from NCEP-1 (black circles), NCEP-2 (gray triangles), and ERA-40 (white squares), computed using monthly data for the period September 1987–August 2001.
Citation: Journal of Hydrometeorology 9, 5; 10.1175/2008JHM887.1
Climatological annual cycle of [P] (solid black line), [R] (thick gray line), and [C] derived from NCEP-1 (black circles), NCEP-2 (gray triangles), and ERA-40 (white squares), computed using monthly data for the period September 1987–August 2001.
Citation: Journal of Hydrometeorology 9, 5; 10.1175/2008JHM887.1
Climatological annual cycle of [P] (solid black line), [R] (thick gray line), and [C] derived from NCEP-1 (black circles), NCEP-2 (gray triangles), and ERA-40 (white squares), computed using monthly data for the period September 1987–August 2001.
Citation: Journal of Hydrometeorology 9, 5; 10.1175/2008JHM887.1
Annual [R] and bias-corrected [C] for the hydrological years 1987/88–2000/01. The [R] was computed from discharge observations at Obidos (black line). The [C] was computed from NCEP-1 (black circles), NCEP-2 (gray triangles), and ERA-40 (white squares). The data points at 1988 correspond to the water year 1987/88, etc. The dashed line is drawn at the long-term mean [R] of 3.11 mm day−1.
Citation: Journal of Hydrometeorology 9, 5; 10.1175/2008JHM887.1
Annual [R] and bias-corrected [C] for the hydrological years 1987/88–2000/01. The [R] was computed from discharge observations at Obidos (black line). The [C] was computed from NCEP-1 (black circles), NCEP-2 (gray triangles), and ERA-40 (white squares). The data points at 1988 correspond to the water year 1987/88, etc. The dashed line is drawn at the long-term mean [R] of 3.11 mm day−1.
Citation: Journal of Hydrometeorology 9, 5; 10.1175/2008JHM887.1
Annual [R] and bias-corrected [C] for the hydrological years 1987/88–2000/01. The [R] was computed from discharge observations at Obidos (black line). The [C] was computed from NCEP-1 (black circles), NCEP-2 (gray triangles), and ERA-40 (white squares). The data points at 1988 correspond to the water year 1987/88, etc. The dashed line is drawn at the long-term mean [R] of 3.11 mm day−1.
Citation: Journal of Hydrometeorology 9, 5; 10.1175/2008JHM887.1
Time-averaged [C] and [R] over a moving 5-yr window. The x axis lists the initial year of each 5-yr interval; that is, the [C] estimate associated with 1988 is an average over the period covering hydrological years 1987/88–1991/92. The [R] was computed from discharge observations at Obidos (black line). The [C] was computed from NCEP-1 (black circles), NCEP-2 (gray triangle), and ERA-40 (white squares).
Citation: Journal of Hydrometeorology 9, 5; 10.1175/2008JHM887.1
Time-averaged [C] and [R] over a moving 5-yr window. The x axis lists the initial year of each 5-yr interval; that is, the [C] estimate associated with 1988 is an average over the period covering hydrological years 1987/88–1991/92. The [R] was computed from discharge observations at Obidos (black line). The [C] was computed from NCEP-1 (black circles), NCEP-2 (gray triangle), and ERA-40 (white squares).
Citation: Journal of Hydrometeorology 9, 5; 10.1175/2008JHM887.1
Time-averaged [C] and [R] over a moving 5-yr window. The x axis lists the initial year of each 5-yr interval; that is, the [C] estimate associated with 1988 is an average over the period covering hydrological years 1987/88–1991/92. The [R] was computed from discharge observations at Obidos (black line). The [C] was computed from NCEP-1 (black circles), NCEP-2 (gray triangle), and ERA-40 (white squares).
Citation: Journal of Hydrometeorology 9, 5; 10.1175/2008JHM887.1
Annual anomalies of [C] and [P] relative to their 14-yr means, for water years 1987/88–2000/01. The [P] was computed from GPCP data (black line). The [C] was computed from NCEP-1 (black circles), NCEP-2 (gray triangles), and ERA-40 (white squares). Water years are used instead of calendar years; a data point plotted at 1988 is associated with the water year September 1987–August 1988.
Citation: Journal of Hydrometeorology 9, 5; 10.1175/2008JHM887.1
Annual anomalies of [C] and [P] relative to their 14-yr means, for water years 1987/88–2000/01. The [P] was computed from GPCP data (black line). The [C] was computed from NCEP-1 (black circles), NCEP-2 (gray triangles), and ERA-40 (white squares). Water years are used instead of calendar years; a data point plotted at 1988 is associated with the water year September 1987–August 1988.
Citation: Journal of Hydrometeorology 9, 5; 10.1175/2008JHM887.1
Annual anomalies of [C] and [P] relative to their 14-yr means, for water years 1987/88–2000/01. The [P] was computed from GPCP data (black line). The [C] was computed from NCEP-1 (black circles), NCEP-2 (gray triangles), and ERA-40 (white squares). Water years are used instead of calendar years; a data point plotted at 1988 is associated with the water year September 1987–August 1988.
Citation: Journal of Hydrometeorology 9, 5; 10.1175/2008JHM887.1
Horizontal and vertical resolutions of NCEP-1, NCEP-2, and ERA-40.
Correlations between time series of monthly [C] anomalies derived from different reanalyses covering the period 1997–2001.
Average [C] over the period September 1987–August 2001 according to each of the reanalyses, and associated bias error derived by comparing it to mean runoff over this period ([R]) computed from data on river discharge at the Obidos gauging station.
Analysis increments are the adjustments to model forecasts toward assimilated observations.
