Balancing global moisture budgets is a difficult task that is even more challenging at regional scales. Atmospheric water budget components are investigated within five tropical (15°S–15°N) ocean regions, including the Indian Ocean, three Pacific regions, and one Atlantic region, to determine how well data products balance these budgets. Initially, a selection of independent observations and a reanalysis product are evaluated to determine overall closure, between 1998 and 2007. Satellite-based observations from SeaFlux evaporation and Global Precipitation Climatology Project (GPCP) precipitation, together with Interim ECMWF Re-Analysis (ERA-Interim) data products, were chosen. Freshwater flux (evaporation minus precipitation) observations and reanalysis atmospheric moisture divergence regional averages are assessed for closure. Moisture budgets show the best closure over the Indian Ocean with a correlation of 89% and an overall imbalance of −3.0% of the anomalies. Of the five regions, the western Pacific Ocean region produces the worst atmospheric moisture budget closure of −21.1%, despite a high correlation of 93%. Average closure over the five regions is within 8.1%, and anomalies are correlated at 83%. ERA-Interim and Modern-Era Retrospective Analysis for Research and Applications (MERRA) evaporation rates are 29 and 19 mm month−1 greater than SeaFlux, respectively. To diagnose the differences, wind speed and humidity gradients of the three products are compared utilizing the bulk formula for evaporation. SeaFlux wind speeds are higher, but sea–air humidity gradients are lower. Higher humidity gradients in the reanalyses are due to much dryer near-surface air in ERA-Interim, and the same to a lesser degree in MERRA. These differences counteract each other somewhat, but overall humidity biases exceed wind biases. This is consistent with buoy observations.
Redistribution of energy and water is essential in structuring earth’s climate. The energy excess at the surface and the deficit in the atmosphere establish energy imbalances, which are then balanced by the hydrological cycle. Specifically, energy is removed from the surface via evaporation, while energy is added to the atmosphere by subsequent condensation. The transport of water vapor, in the meantime, can serve to transport energy in the atmosphere.
Recently, there has been increased interest in understanding and parameterizing the global energy budgets, in order to further our understanding of the current climate state. Climate change affects global energy and moisture fluxes, with any changes to global energy budgets also affecting the hydrological cycle. Examining the closure of energy or moisture budgets tests the fundamental quality of the data products themselves. Additionally, closure provides a test of trends in individual components and an indication of how the system adapts to changes. The water and moisture components of two recent global energy budgets display variability that indicates the uncertainty surrounding the quantification of moisture fluxes. Surface latent heat flux estimates of 80 and 88 W m−2 were suggested by Trenberth et al. (2009) and Stephens et al. (2012), respectively, and 17 and 24 W m−2 for the sensible heat flux. Estimates of the latent heat flux equate to a difference of approximately 10%, but both are higher than the 76 W m−2 given as the best estimate of global precipitation by the Global Precipitation Climatology Project (GPCP; Adler et al. 2012) product. Observational studies by Gastineau and Soden (2011) noted weakening of atmospheric circulation and tropical wind extremes as increases in tropical hydrological cycles and heavy precipitation occurred. However, some contradictory studies have shown that circulation intensifies in reanalyses and observations (Zahn and Allan 2011; Sohn and Park 2010). Thermodynamic assumptions also do not explain hydrological changes that occurred in the Last Glacial Maximum (Boos 2012).
The thermodynamic relationship between warming and moisture is generally determined according to the Clausius–Clapeyron relation (Held and Soden 2006). The Clausius–Clapeyron relation dictates that when relative humidity is invariant warming increases water vapor by 7% K−1. Such changes have been observed over oceans using passive microwave sensors by Wentz and Schabel (2000). Changes in evaporation rates and precipitation, however, are more difficult to predict. While Allan et al. (2013) suggest that precipitation should increase by 2%–3% K−1 due to energetic constraints, Wentz et al. (2007) showed results indicating a trend similar to the water vapor of 7% K−1.
A number of climate modeling studies have shown that warming leads to increases in the hydrological cycle in the 2%–3% K−1 range and a weakening of the large-scale circulation (Held and Soden 2006; Vecchi and Soden 2007; Gastineau et al. 2009). The Vecchi and Soden study points out that both the Walker and Hadley circulations will weaken as warming occurs. Hadley cells may widen (Schneider et al. 2010) and reduce the amount moisture being transported to higher latitudes (Seager et al. 2010). Observational studies have also shown a widening of the Hadley cell (Seidel et al. 2008; Zhou et al. 2011). A “wet get wetter and dry get drier” precipitation pattern has been associated with a warming climate (Meehl et al. 2007; Zhou et al. 2011; Allan et al. 2013).
This study will focus on observations of the hydrologic parameters over equatorial oceanic regions. Tropical ocean regions are particularly important in the global context as insolation, evaporation, water vapor transport, and precipitation are all large. Unfortunately, none of the parameters is measured directly. Care must therefore be taken to understand the actual measurements and their interpretation. Evaporation serves as good example. While it is possible to measure turbulent fluxes directly using eddy correlation measurements (Kessomkiat et al. 2013; Liu et al. 2013), these are limited to land areas. As evaporation is driven dynamically by winds and thermodynamically by sea surface temperature (SST) and specific humidity gradients, the complexities of observing evaporation directly are such that bulk parameterizations are typically used instead to quantify evaporation over oceans (Brunke et al. 2003, 2011). Bulk formulations exploit available temperature, humidity, and wind speed observations. The bulk aerodynamic formula for evaporation (E) is
where the variables are air density (ρa), transfer coefficient (Cq), 10-m wind (ux), surface specific humidity (q0), and specific humidity at 10 m (qa).
Because the parameterizations depend primarily on satellite-derived SST, near-surface wind speed, and near-surface humidity, it is not surprising that evaporation products are highly correlated with those input fields. Buoys and ships provide the largest networks of ocean surface observations that validate observations used in such parameterizations (Kumar et al. 2012; Clayson et al. 2013, manuscript submitted to Int. J. Climatol.). Comparisons of oceanic evaporation averages from reanalysis models and observations indicate that these parameterization-derived data sources produce a large range of values (Curry et al. 2004; Andersson et al. 2011; Bosilovich et al. 2011; Kumar et al. 2012). The differences, fortunately, can generally be traced back to one or more of the bulk parameters.
Wind speed is the more easily estimated parameter of the two evaporative sea–air gradients, and therefore its relationship with evaporation is clearer. Satellite observations show that wind speeds have increased over oceans (Wentz et al. 2007; Wentz and Ricciardulli 2011; Young et al. 2011). This change in wind speeds is inconsistent with models that show circulation to be weakening, but the reason for this discrepancy is unknown. Declining wind speeds over land have been related to decreases evaporative rates, and evaporation was most sensitive to changes in wind speed, as opposed to other meteorological variables (McVicar et al. 2012). Trends in observed variables can have their origin in measurements drifts while trends in reanalyses are often due to inhomogeneities and trends in ancillary data (Bosilovich et al. 2011; Large and Yeager 2012). Oceanic evaporation, precipitation and water vapor fluxes show the output from different reanalysis models disagree both in magnitude and trends, and consequently they should be used with caution (Trenberth et al. 2011).
Balancing moisture and energy budgets at regional scales is even more challenging than globally. The advantage of balancing budgets at global scales is that the overall radiative imbalances must be balanced by the hydrologic cycle as the Earth system has very little ability to store excess energy. When specific regions are evaluated, the transport of water vapor must be accurately quantified to achieve closure. While water vapor can be measured quite well over oceans, the movement of atmospheric moisture is difficult to accurately quantify from limited three-dimensional wind speed observations.
This study aims to determine how well independently generated datasets balance atmospheric moisture budgets in the tropics, and to evaluate which factors are contributing to their variability. Moisture balance equations will be used to evaluate observation-based and reanalysis model data products in five tropical ocean regions. The tropics are selected as energetic and hydrological processes are large in this region, and consequently analyzing the processes in this region should enhance inadequacies that may exist in the data products. Figure 1 shows the tropical oceanic regions that are used to assess the moisture budgets. Five regions were chosen because of their tropical oceanic location, sufficient areal extent to allow moisture budgets to be robust, and representativeness of the different regional hydrological processes that occur in the tropics. The regions will be assessed for how well the moisture budgets close, what we can learn from the covariance of the moisture parameters, and whether they show trends.
Initially, a brief description of the data products being analyzed is provided in section 2. The primary moisture budget components of selected products are investigated for closure in the five regions, and their relationship with supplementary evaporative parameters is established, in section 3. The fourth section follows with precipitation, evaporation, and atmospheric moisture divergence being individually assessed to explore the differences between reanalysis and observation-based data products. These are also presented and assessed for each of the five tropical regions. Section 5 further examines the differences between individual components of the bulk aerodynamic formula for evaporation using three data products, and compares them with buoy observations. Finally, a summary and discussion of the findings are presented in section 6.
2. Atmospheric moisture budgets
Vertically integrated atmospheric moisture budgets are conserved according to the following equation:
where q is specific humidity, is vertically integrated moisture transport or moisture divergence (∇Q), E is surface evaporation, and P is surface precipitation (Peixoto and Oort 1992). When averaged over longer periods of time the temporal component of the above equation is small and therefore ∇Q equals E − P (freshwater flux). This relationship has provided the basis of many analyses of atmospheric water budgets (Trenberth et al. 2011; Wong et al. 2011; Newman et al. 2012) and provides a useful tool in evaluating the closure of moisture budget components within and between data sources.
Many datasets provide values for the evaporation, precipitation, and ∇Q components of atmospheric moisture budgets. Reanalysis products produce assimilations that quantify all the elements of these budgets. Two reanalysis datasets are selected to indicate how models redistribute moisture, and to utilize their comprehensive and complete set of parameters for diagnostic purposes. In comparison, observation-based datasets produce individual moisture budget components, such as precipitation and evaporation. The datasets that are used here to calculate atmospheric moisture budgets are listed in Table 1, and are further described in the following sections.
One of the most recent ocean evaporation datasets produced is SeaFlux (Clayson et al. 2013, manuscript submitted to Int. J. Climatol.; see also http://seaflux.org/). Satellite data are used to estimate the bulk flux variables. SeaFlux uses a neural network version of the Coupled Ocean–Atmosphere Response Experiment (COARE 3.0) bulk parameterization of an air–sea fluxes algorithm (Fairall et al. 2003) to calculate surface turbulent fluxes (Clayson et al. 2013, manuscript submitted to Int. J. Climatol.). The resultant 3-hourly 0.25° × 0.25° gridded dataset was downloaded from http://seaflux.org/. Wind speed, SST, and near-surface specific humidity data are also available from SeaFlux, and are used for diagnostic studies in subsequent sections. An additional evaporative dataset, the objectively analyzed air–sea fluxes product (OAFlux; Yu et al. 2008), is also assessed. OAFlux provides an intermediate evaporation product as it is based on a hybrid model of observations and reanalysis data. The COARE 3.0 algorithm is also used by OAFlux to determine evaporation. OAFlux produces daily data, at 1.0 ° × 1.0 ° resolution (http://oaflux.whoi.edu).
The GPCP One-Degree Daily (1DD) Precipitation Data Set version 1.2 (Huffman et al. 2001, 2012) is one of the most widely available and utilized satellite-gauge merged precipitation products. The data are produced by aggregating satellite products into monthly composites, blending them with monthly gauge products at that scale, and subsequently using any biases between the satellite-only and the satellite-gauge composite to calibrate the gridded daily satellite-based observations. Ocean data are primarily satellite-derived, as few gauge data exist in oceanic regions (GPCP data were obtained from http://precip.gsfc.nasa.gov/).
Two reanalysis products are included, the Interim European Centre for Medium-Range Weather Forecasts (ECMWF) Re-Analysis (ERA-Interim; Dee et al. 2011) and the National Aeronautics and Space Administration (NASA) Modern-Era Retrospective Analysis for Research and Applications (MERRA; Rienecker et al. 2011). ERA-Interim incorporates observations and satellite data into 12-hourly analysis cycles with a four-dimensional variational assimilation scheme. The forecasts from the analysis cycles are combined with observations to produce an evolving model that observations constrain, with analysis increments being produced from this adjustment. Analysis increments are considered to be the net response of the assimilation to observations (Dee et al. 2011). Surface parameters are provided at 3-hourly resolution and pressure level parameters are 6-hourly. This analysis uses 6-hourly products for consistency among surface and three-dimensional fields. The ERA-Interim full-resolution 0.75° × 0.75° gridded data were accessed from http://apps.ecmwf.int/datasets/.
MERRA uses the Goddard Earth Observing System version 5 (three-dimensional variational) data assimilation system with observation and satellite data (Rienecker et al. 2011). An incremental analysis update procedure (Bloom et al. 1996) smooths the forecast initialization and produces an additional term to indicate the effect of observations on the analysis (Rienecker et al. 2011). The inclusion of the increment analysis term also allows MERRA budgets to balance. MERRA data are on a 0.5° latitude by 0.67° longitude grid at 3-hourly resolution for the three-dimensional fields and at hourly resolution for surface fields (https://gmao.gsfc.nasa.gov/).
Of the three variables in the moisture equation [Eq. (2)], ∇Q is most easily available from reanalysis model output. Methods that derive moisture divergence from satellite observations are available (Hilburn 2009; Xie et al. 2008), but the development of such data is in its infancy and little independent validation has been performed. Therefore, reanalysis model output is used here. Atmospheric moisture divergence is calculated according to Eq. (2), using u wind, υ wind, and specific humidity from 1000 hPa up to 200 hPa from ERA-Interim. Little moisture exists above 200 hPa. While both reanalyses produce an atmospheric moisture divergence term, it was only calculated using Eq. (2) for ERA-Interim in anticipation of future work involving moisture at different levels of the atmosphere.
Because of the limited temporal extent of some of the observed data, all the datasets were limited to the period from 1998 to 2007. While each dataset has different spatial and temporal resolutions (Table 1), it is assumed that these differences will be minimal because of the spatial averaging used. Areal averaging is performed by calculating the grid point averages and then averaging the data over the time period chosen. Dataset-specific land masks were applied where they were available, or masks of similar resolution were used.
3. Tropical water balance
Monthly averages (1998–2007) of the five tropical region’s atmospheric water budget components are shown in Figs. 2–6. SeaFlux E and GPCP P time series provide the primary elements of atmospheric moisture cycling in these figures. The inclusion of SeaFlux E−GPCP P and ERA-Interim divergence complete the major components of atmospheric moisture conservation. Because of inhomogeneities affecting MERRA data (Rienecker et al. 2011) they are not used in this initial assessment of the regional moisture balances. Reynolds sea surface temperatures (Reynolds et al. 2007) are also shown to provide context for the thermodynamic environment of the tropical regions.
To complement the figures of the water budget components, a correlation matrix based on monthly anomalies of the components and two bulk formula parameters is included (Table 2). The bulk aerodynamic formula for evaporation parameters are represented by Δu and Δq from ERA-Interim. For ERA-Interim 2-m qa (10-m data are not produced in this reanalysis) and q0 specific humidity values are calculated according to Buck (1981). The surface specific humidity component of the bulk aerodynamic formula [Eq. (1)] is closely related to SST from which q0 is determined, after applying a salinity adjustment of 0.98. The value of Δu is calculated as the difference between E from the bulk formula and where the climatological average of Δq () is used [Eq. (3)]. In contrast, Eq. (4) describes how Δq is calculated using the climatological average Δu ():
The climatologies of the tropical regions will now be described based on their atmospheric moisture budget components (Figs. 2–6) and the correlation matrix (Table 2). Each region is assessed for closure, parameter covariance, and trends. Trends are visually evaluated so that temporal effects are negated, while closure is assessed using the correlation and the imbalance between E − P and ∇Q. This imbalance is quantified relative the regional mean evaporation and precipitation from the SeaFlux and GPCP data, which is a comparable method to that of Andersson et al. (2011). In the context of this work weak divergence–convergence is defined as averaging less than 50 mm month−1 and strong divergence–convergence as averaging over 100 mm month−1.
a. Tropical Indian
The tropical Indian region features a very close relationship between E − P from observations and ∇Q from ERA-Interim (Fig. 2) that is reinforced by a correlation of 89% (Table 2). Over the period 1998–2007, the average difference between ∇Q and E − P is −3.4 mm month−1 at the monthly scale, or −3.0% of average E and P values. Monthly averages of E − P and ∇Q show little difference over the whole 10-yr period, which implies that this region’s water balance can be closed using three independent data sources.
SST values in the Indian Ocean exhibit pronounced seasonality, and monthly SST anomalies are inversely related to Δu (Fig. 2, Table 2). Changes in average wind speed cause more variability in E than specific humidity, and consequently variability in E shows a muted response to changes in SST. Note that P is larger and more variable than E on average, which results in the region being an area of weak convergence on average. The convergence indicated from the negative values of ERA-Interim divergence and E − P is well correlated, while being anticorrelated with P (Table 2). Observed evaporation and precipitation show increasing trends, despite no change in SST. Both E − P and ∇Q show no trends in the amount of convergence occurring in the Indian Ocean.
b. Tropical western Pacific
Observed freshwater fluxes are lower than ∇Q by almost 25 mm month−1 on average, in the western Pacific Ocean. Precipitation is high in this region, and the imbalance between the observed and reanalysis data is considerable, at −21.1%. While the monthly fluctuations in these time series exhibit considerable consistency between E − P and ∇Q, as indicated by a correlation of 93%, the increasing trend in the difference between the two measures is apparent in Fig. 3. This trend indicates that the one of the three products used in the closure study is drifting away from the other two measurements in this basin. Unfortunately, this analysis does not allow one to pinpoint which product is responsible for this trend.
A weak semiannual seasonality exists for SSTs in the tropical western Pacific (Fig. 3), with the western Pacific warm pool contributing to their consistently high temperatures. The first six years have increasing in SSTs of approximately 1.0°C, after which temperatures show little change. There is no evidence of changes in SST and Δq affecting E, as these are uncorrelated (Table 2). In contrast, the variability associated with Δu is closely related to E, as these parameters are correlated at over 60%.
The increasing trend in P prior to 2003 equates to an overall increase of 100 mm month−1, with only a slight decrease in the second half of the time series. Convergence dominates the region as P exceeds E, and strongly negative atmospheric moisture divergence values persist throughout the year. Changes in P dominate the variability of E − P and divergence, causing the two series to be anticorrelated with P.
Large trends exist in many of the parameters associated with the atmospheric moisture budget of the tropical western Pacific. Increases in SSTs produce similar trends in SeaFlux E, but not in ERA-Interim E. Observed precipitation data exhibit a strong increase, and consequently E − P and ∇Q show strong decreasing trends. This region produces strong convergence that continues to increase by over 7 mm month−1 during the 10 yr analyzed.
c. Tropical central Pacific
Freshwater fluxes show little difference when compared to ∇Q, suggesting that the atmospheric moisture budget almost achieves closure in the tropical central Pacific. An average difference of −8.2 mm month−1 (−7.2%) is due to the observed fluxes being slightly higher than moisture divergence from ERA-Interim, with the two being highly correlated (Table 2). However, although there are periods where E − P and ∇Q are close, some considerable separations also occur (Fig. 4). These separations are associated with periods when GPCP precipitation is lower than 90 mm month−1. El Niño conditions in 1998 do not affect the closure achieved, despite the large increase in precipitation that occurs.
The presence of the El Niño event in 1998 is the most prominent feature in the central Pacific time series in Fig. 4. With the exception of the high SSTs occurring during the El Niño, SSTs show similar increases to those described in the tropical western Pacific. Between 1999 and 2003 SSTs warmed by 1°C, before plateauing and then cooling slightly. Correlations in Table 2 indicate that Δq is heavily dependent on SSTs, Δq and Δu are weakly inversely related, and Δu is highly correlated with SeaFlux E. In comparison, ERA-Interim E is more closely related to Δq.
Despite the strong influence of the 1998 El Niño on SSTs, there is only a subtle influence on E. El Niño conditions have a strong effect on P in the central Pacific, with P increasing dramatically by 200 mm month−1. Consequently, both E − P and ERA-Interim divergence both indicate that strong regional convergence is occurring during 1998. However, the time series from 1999–2007 indicates that this region is generally characterized by weak divergence, due to E exceeding P. The relationship between ERA-Interim divergence and E − P shows some small systematic differences, with E − P being lower than divergence during the 1999–2001 and 2005–08 periods.
Many of the variables in Fig. 4 feature trends after the El Niño event. An increasing trend in SST contributes to a large increasing trend seen in evaporation of approximately 1.5 mm month−1. Precipitation values also show an increasing trend in the ERA-Interim data only. Comparable increases in ∇Q are also evident due to changes in reanalysis evaporation and precipitation, while observations of E − P are unchanged.
d. Tropical eastern Pacific
Moisture budgets in the tropical eastern Pacific do not appear to close very well (Fig. 5), in spite of the small average difference between the observed and reanalysis data of −3.7%. The correlation between E − P and ∇Q is lower than the two other Pacific regions at only 77%, indicating there are regional differences in how well the three products observe the system. Higher variability in E − P is predominantly associated with stronger seasonal cycles in the satellite-based precipitation product. As was noted in the TCP region, the discrepancies between E − P and ∇Q are highest when monthly precipitation averages are low. However, unlike comparisons in other regions, E − P regularly exceeds ∇Q, associated with the seasonal peaks in precipitation.
Pronounced seasonality exists in the tropical eastern Pacific (TEP) SSTs, due to peaks in warming occurring during the Northern Hemisphere (NH) spring (Fig. 5). The 1998 El Niño only causes a slight increase in SST. SSTs dominate Δq variability, which in turn are well correlated with ERA-Interim E, while winds are better correlated with SeaFlux E (Table 2). The value of E is slightly higher during the El Niño, and there is only a small seasonal fluctuation in the time series.
Strong P seasonality fluctuates between 17 and 136 mm month−1, excluding the El Niño year. This seasonality in P lags that of SST by 3 months to peak in the NH summer. Intercorrelation exists among Δq, E, and P, although the variability of the latter is much larger. Weak divergence occurs throughout most of the year in the TEP, although E − P is more variable than divergence.
Excluding the 1998 El Niño year, SeaFlux evaporation exhibits a slight increasing trend. No such increase is present in precipitation data between 1999 and 2007. Increased divergence in ERA-Interim is indicated though the positive trends in ∇Q, while no change is observed in E − P.
e. Tropical Atlantic
In the tropical Atlantic region, the E − P and ∇Q budgets do not balance. The small average difference between the two budget parameters of 6.3 mm month−1 (5.5%) belies their variability, and contributes to a correlation of only 62% (Table 2). Prior to 2003 there are some very large differences of up to 50 mm month−1 between E − P and ∇Q. These differences coincide with dips in GPCP, as changes in evaporation are comparatively small. After 2003, there is greater correspondence between E − P and ∇Q, and the moisture budgets almost close.
Tropical Atlantic SSTs have a semiannual seasonal cycle that is strongest at the beginning of the year, as shown in Fig. 6. In Table 2, Δu is correlated with E at 57%, but Δq is well correlated with ERA-Interim E only. A lack of seasonality in E is countered by a strong seasonal pattern in GPCP P, which peaks during the NH summer. The weakest correlations between GPCP P and ERA P, and consequently between E − P and ERA-Interim divergence, occur in this region. Some discrepancies of up to 40 mm month−1 exist between E − P and divergence and these discrepancies are more common during the first four years of the data period. The tropical Atlantic shows the weakest intercorrelations between all the different atmospheric moisture balance parameters. This suggests that either the region is distinctive climatologically compared to the other regions or that its parameters are not well quantified.
An increase in evaporation occurs in the SeaFlux data. In comparison, no trends in precipitation are present over the 10-yr period examined. The effect of these differences in trend results in a positive trend in the observed freshwater flux, but there is no evidence of a change in the reanalysis ∇Q values.
f. Tropical regions summary
The closure of independently generated observed and modeled atmospheric moisture budgets over tropical oceans varies regionally. Correspondence between SeaFlux and GPCP E − P and ERA-Interim ∇Q in the tropical Indian region suggests that the moisture transport processes are well captured here. All four other regions indicate that there are problems resolving how moisture cycling occurs when comparing observations and the reanalysis model. However, only the imbalance present in the tropical western Pacific exceeds the bias in GPCP of ±8% determined by Adler et al. (2012).
A consistent bias is present in the tropical western Pacific, as the convergence indicated by ERA-Interim is larger than from satellite observations. Consequently the atmospheric moisture balance is not close to balancing during the majority of the 1998–2007 period. The trend in difference between observed E − P and reanalysis ∇Q (Fig. 3) indicates that the moisture imbalance is also increasing. As the ERA-Interim divergence term does not simply equate to the reanalysis E − P it is difficult to determine what is driving this trend. However, it appears that the increase in SeaFlux evaporation data contributes to the trend in E − P and ∇Q.
Moisture budgets are similar in the three regions characterized by divergence in atmospheric moisture (central Pacific, eastern Pacific, and Atlantic). The largest deviations from closure of E − P and ∇Q are due to the periods of low precipitation in GPCP, while no coincident drop in P and water vapor convergence from ERA-Interim occurs. These deviations are likely to be associated with weak rainfall that is widespread and persistent in models (Pfeifroth et al. 2013; Kim and Alexander 2013). Conversely, deviations could also be associated with the satellite data in GPCP missing areas of light rain (Berg et al. 2010).
Higher correlations associated with observed and reanalysis P indicate that the processes leading to P are parameterized better than for E. Monthly tropical Indian and Pacific precipitation variability in GPCP P and ERA-Interim P is generally similar, and therefore highly correlated. In the tropical Atlantic the differences between GPCP P and ERA P indicate that this variable is not as well known. Problems with tropical Atlantic bias in GCMs are well known (Patricola et al. 2012; Richter et al. 2012) and these discrepancies are present in ERA-Interim (Richter et al. 2012).
The variability associated with E is smaller than that of P and therefore dominates the variability of ∇Q and the characteristics of the moisture regimes in the five regions. The tropical Indian and the western Pacific Ocean regions are regions with the highest SSTs and where convergence occurs. In these convergence zones moist air is being transported into the regions, especially during the monsoon and when the Indo-Pacific warm pool is most active. Central and eastern Pacific and Atlantic tropical ocean regions feature the opposite patterns with divergence occurring on average, and moist air being transported out of these regions. Pacific moisture is primarily transported westward as a result of Walker circulation–driven winds.
4. Atmospheric moisture fluxes from observations and reanalyses
Regional differences in tropical atmospheric moisture flux components are shown in Figs. 7 and 8. Average SeaFlux, OAFlux, ERA-Interim, and MERRA evaporation rates (Fig. 7) indicate the reanalysis data products produce the highest average evaporation of 141 and 131 mm month−1 for ERA-Interim and MERRA, respectively. In contrast, evaporation is 122 mm month−1 from OAFlux and 112 mm month−1 from SeaFlux. These two observation-based products produce less evaporation than the reanalyses by 19 mm month−1, or roughly 19% on average. OAFlux evaporation data are also not considered further here, as they merely provide an additional product that falls between observations and reanalyses.
Differences in average precipitation from GPCP, ERA-Interim, and MERRA are apparent in Fig. 7. Precipitation from ERA-Interim is 30 mm month−1 higher than GPCP when averaged over the five regions, while MERRA is 20 mm month−1 higher than the observations. Previous studies have shown excessive precipitation occurs in ERA-Interim, although MERRA is an exception (Trenberth et al. 2011). However, an increasing trend in MERRA precipitation data has resulted in the tropical precipitation averages being similar to ERA-Interim by the end of the data period (not shown). This suggests that both products are producing excessive moisture cycles during this period. Recent studies have shown biases due to persistent rainfall in MERRA (Kim and Alexander 2013).
The time tendency term (∂W/∂t) in Eq. (1) was confirmed to have a small or negligible effect in each region, compared to the other terms, validating the absence of this term when calculating the moisture budgets. Average divergence values (1998–2007) for ERA-Interim and MERRA are shown in Fig. 7. The five regions show differences of up to 37 mm month−1 in ∇Q between ERA-Interim and MERRA. ERA-Interim moisture divergence is lower on average than for MERRA.
Reanalysis and observed data products reproduce the primary spatial features of tropical moisture flux, with differences in magnitude being attributed to the different types of data products (Fig. 8). It should also be noted that the atmospheric moisture budgets of the reanalysis models do not achieve closure without an additional term, such as the analysis increment tendency in MERRA (Bosilovich et al. 2011). The analysis increment provides a forcing term linking the analysis model to observations.
While not the focus of this paper, it is nonetheless instructive to examine the degree of closure in the analyses themselves. Table 3 lists the bias between ∇Q and E − P for ERA-Interim and MERRA, as well as their correlation. Bias or imbalance in the reanalyses is calculated as a percentage relative to the region’s observed E − P (i.e., SeaFlux evaporation minus GPCP), to be consistent with earlier differences. These biases essentially equate to the analysis increment in the models and a small time tendency effect. The five regions show a lot of variability in their biases from ERA-Interim, and especially in MERRA. Imbalances of up to ±24% are produced by the MERRA in the five regions. ERA-Interim imbalances are of a similar range to those between the observed E − P and ERA-Interim ∇Q, excluding the tropical western Pacific. Despite the regional imbalances, some very high correlations of up to 97% occur between ∇Q and E − P monthly anomalies for the individual reanalyses. ERA ∇Q and E − P produces higher correlations than MERRA ∇Q and E − P.
MERRA precipitation and surface fluxes have been noted to be poorly constrained, and consequently exhibit sensitivity to observing system changes (Rienecker et al. 2011). Bosilovich et al. (2011) noted that MERRA’s large changes in precipitation in 1999 and 2001 are due to including Advanced Microwave Sounding Unit (AMSU) instruments. This inhomogeneity is primarily a tropical oceanic phenomenon (Robertson et al. 2011). Because of the problems associated with precipitation inhomogeneities in MERRA during the period analyzed, this dataset is considered as secondary to ERA-Interim in this work. While AMSU observations were also included in ERA-Interim (Dee and Uppala 2009), its influence is less pronounced. Inhomogeneities associated with assimilating other datasets are also likely to affect data but to a lesser extent.
The balance of evidence compiled thus far from the observations and reanalyses indicates that reanalysis products are producing too much evaporation and precipitation, but they produce an adequate overall moisture flux. Therefore, as evaporation is the primary source of moisture it is likely that this term is important in creating this overestimation. Comparisons between the different bulk formula parameters of these different products would elucidate what may be contributing to evaporative differences.
5. Comparison of bulk formula parameters
a. SeaFlux, ERA-Interim, and MERRA bulk formula parameters
SeaFlux data consistently produce the lowest evaporation amounts of the data products examined here. To establish what produces different evaporation rates in data products the bulk formula parameters [E = ρaCqux(q0 − qa); see Eq. (1)] of SeaFlux, ERA-Interim, and MERRA will be evaluated. Wind speed, SST, and qa are produced as supplementary variables to the SeaFlux dataset, and also by the reanalyses. Consequently, the meteorological differences among SeaFlux, ERA-Interim, and MERRA variables that produce evaporation can be investigated.
It should be noted that while these datasets are generated independently of each other, they also utilize some of the same data observations in their inception. The neural network algorithm (Roberts et al. 2010) that generates SeaFlux’s surface humidity is based on Special Sensor Microwave Imager (SSM/I) brightness temperatures, wind speeds are based on Cross-Calibrated Multi Platform (CCMP) ocean surface wind components (SSM/I, AMSR-E, TMI, QuikSCAT, and SeaWinds) data (Atlas et al. 2011), and the Reynolds Optimally Interpolated SST version 2 uses AVHRR-only SSTs (Reynolds et al. 2007). ERA-Interim and MERRA utilize many of the same satellite products and observations as SeaFlux in their assimilation procedure.
SeaFlux and ERA-Interim’s bulk formula parameters exhibit considerable spatial consistency in the average tropical wind speed and humidity patterns (Fig. 9). Since ERA-Interim and SeaFlux represent the high and low evaporation values, MERRA is not included in this figure. Wind speeds are higher in SeaFlux data, and the areas with slowest and fastest wind speeds show the largest differences. For SSTs, there are only small, subtle differences between SeaFlux and ERA-Interim. This similarity is not unexpected as while these datasets may use different SST data products, they are based some of the same observations. Near-surface humidity fields feature the largest differences, and the higher values are associated with ERA-Interim data.
In the five tropical regions, the differences between bulk formula parameters from SeaFlux, ERA-Interim, and MERRA are evident (Fig. 10). MERRA has a higher Δq than ERA-Interim, and it is 1.16 g kg−1 (23%) higher on average than for SeaFlux. These larger values of Δq are associated with lower near-surface humidity averages in the reanalyses as SSTs are comparable. However, ERA-Interim is the only near-surface specific humidity provided at 2 m, whereas a 10-m variable that matches the other two datasets would produce lower humidities and further increase Δq. In contrast, the lower Δu values of −0.37 m s−1 (6%) and −0.53 m s−1 (10%), for ERA-Interim and MERRA respectively compared to SeaFlux, counteract the higher Δq. Therefore the overall effect of Δq and Δu differences on evaporation is smaller.
Near-surface q is the variable that shows the largest differences between the reanalyses and SeaFlux. These differences are most pronounced in the Indo-Pacific warm pool where qa is lower in the reanalysis models, signifying the air is drier. Drier air increases the near-surface moisture gradient, which results in ERA-Interim’s and MERRA’s Δq being higher in these regions. This dry bias has been noted previously and is also associated with a cold bias in air temperatures (Brunke et al. 2011; Kumar et al. 2012). Consequently, high Δq in ERA-Interim, and to a lesser extent in MERRA, produces higher evaporation rates compared to SeaFlux.
The effective transfer coefficient in the bulk aerodynamic formula may also have an effect on evaporation. By assuming a constant air density of 1.3 kg m−3 the effective transfer coefficient (ETC) can calculated [i.e., ETC = 1.3ux(q0 − qa)/E]. For MERRA, the opposing effects of slower winds and much higher near-surface humidity gradients compared to SeaFlux contribute to higher evaporation rates, but a higher transfer coefficient than SeaFlux’s is required to match MERRA’s evaporation rates. To produce the high evaporation rates of ERA-Interim either a higher ETC or a higher Δq is needed. Adjusting ERA-Interim qa to 10 m, using MERRA 2-m and 10-m qa ratios, increases Δq and reduces the ETC to a comparable value to that of SeaFlux. However, attributing the effect of each bulk formula parameter on evaporation is inherently dependent on the accuracy of gridded products.
b. Gridded bulk formula parameters and buoy data
A final assessment of SeaFlux, ERA-Interim, and MERRA bulk aerodynamic formula parameters is made using buoy observations for comparison. It is pertinent to note, however, that buoy observations used to train and validate the neural network algorithm (Roberts et al. 2010) in SeaFlux are also ingested in the reanalyses. Therefore the datasets cannot be considered independent of the buoy data. However, buoy observations are likely to have a greater effect on SeaFlux than ERA-Interim and MERRA.
Buoy data were obtained from the Global Tropical Moored Buoy Array for Research Moored Array for African–Asian–Australian Monsoon Analysis and Prediction (RAMA) Indian, Tropical Atmosphere Ocean–Triangle Trans-Ocean Buoy Network (TAO/TRITON) Pacific, and Prediction and Research Moored Array in the Tropical Atlantic (PIRATA) Atlantic Ocean arrays. The buoys are matched with the nearest grid point in SeaFlux, ERA-Interim, and MERRA data to determine how well these three data products matched observations. A total of 83 buoys were used for comparisons, ranging from only 5 in the tropical Indian region to 31 in the tropical western Pacific region.
The average differences between the buoys and the three gridded datasets are shown in Fig. 11. As posited previously, SeaFlux wind speeds and near-surface humidities are more accurate than those of ERA-Interim and MERRA. Surface specific humidity comparisons show that the SST differences between buoy and gridded datasets are very small. Buoy data in the tropical Indian Ocean are also problematic as fewer data exist than in the other regions and the existing data show erratic changes in all three parameters. Therefore Indian Ocean buoy comparisons should not be considered.
Figure 11 clearly shows that ERA-Interim reanalysis wind speeds are approximately 9% lower on average than the buoys, with MERRA winds being 15% lower. SeaFlux wind speeds are only 3% lower. Excluding the tropical Indian region, the product’s wind speeds exhibit consistent deviations from each other, relative to the buoys. Higher wind speeds in SeaFlux and MERRA are consistent with the wind speed differences in Fig. 10.
SeaFlux near-surface specific humidities have a less consistent relationship relative to the buoys than winds. While SeaFlux qa is generally slightly lower than buoy qa, it is higher in the tropical west Pacific. ERA-Interim qa data are again lower than for SeaFlux and buoys, reinforcing the evidence of a dry bias noted earlier. Near-surface humidities in MERRA are 7% lower than the buoys on average and are comparable to the average difference between SeaFlux and the buoys. The standard deviations of the gridded data and buoy differences (not shown) are large for wind speed and qa, and relatively small for SST, reflecting the uncertainty associated with each term.
Similar differences between ERA-Interim and buoys were found by Kumar et al. (2012) for wind speed, SST, and qa, although the qa differences were larger. Ship observations also indicate that qa in ERA-Interim is too dry, but little systematic difference is seen for winds and SST (Brunke et al. 2011). In contrast, MERRA was found to produce a slightly higher qa in comparison to ship observations, but no other systematic differences. The sparse nature of the ship and buoy data provides additional variability when comparing these with the nearest grid points. The differences between gridded and buoy data produce variability that is of higher magnitude than the grid–buoy differences. Therefore, while the buoy differences are consistent with what would be expected from previous studies, the variability of datasets precludes making definitive conclusions.
6. Summary and discussion
Evaporation, precipitation, and atmospheric moisture divergence from SeaFlux, OAFlux, GPCP, ERA-Interim, and MERRA have been compared. The results indicate that there is significant variability between the amount of evaporation and precipitation from different data sources. ERA-Interim reanalysis data produce the most evaporation and precipitation in the tropics, while observations (primarily from satellites) produce the least. Here, observed evaporation and precipitation are considered to be better data products than those produced by the reanalyses.
The closure between observed freshwater fluxes from evaporation and precipitation and atmospheric moisture divergence from ERA-Interim reanalysis data is quite good generally. Closure implies that the reanalysis model is able to produce a reasonable state variable assimilation, but the data assimilation systems perform less favorably when compared to observed evaporation and precipitation. Trenberth et al. (2011) note that E − P is very stable over time and across analyses, and despite the large differences in the magnitude of these two parameters, SeaFlux-GPCP E − P and ERA-Interim atmospheric moisture divergence show remarkable correspondence. This correspondence suggests that while the exact amount of moisture that is being transported in the atmosphere is uncertain, there is more confidence in how it is being transported.
In the tropics, discrepancies between the different datasets appear to be a product of scaling of the moisture fluxes. This scaling could be associated with the time scales that evaporation and precipitation occur over, reflecting the processes that produce them. Evaporation is primarily due to large-scale processes on spatial and temporal scales, while precipitation is associated with finer-scale, fast processes. The ability of reanalysis products to realistically parameterize such features is challenging. Problems with excessive evaporation in the ERA-Interim reanalysis produce the high freshwater flux values and contribute to excessive cycling and a short lifetime of moisture (Trenberth et al. 2011).
Observations may also be underestimating the freshwater fluxes due to the scarcity of oceanic measurements requiring the dependence on satellite data. Estimated bias errors of up to 20% are present in GPCP for the tropical oceans (Adler et al. 2012) and these could also contribute to the differences between reanalyses and observations. However, the consistent relationship between freshwater flux evaporation and precipitation from SeaFlux and GPCP with ERA-Interim moisture divergence provides some validation for the accuracy of GPCP precipitation.
Changes in precipitation dominate the fluctuations in atmospheric moisture divergence in the tropics, as evaporation displays comparatively little variability. The Indo-Pacific warm pool is the most extensive region of convergence and precipitation rates are especially high in the western Pacific. Differences between the different precipitation data sources are the highest in this region. It is well known that ERA-Interim reanalysis has a stronger hydrological cycle over the western Pacific (Trenberth et al. 2011).
The poorest balance between atmospheric moisture divergence and observed freshwater fluxes is seen in the western Pacific. This lack of balance is created by the much larger difference (approximately 20 mm month−1) between GPCP and ERA-Interim precipitation in the region, relative to SeaFlux and ERA-Interim evaporation. As the western Pacific has high and extensive precipitation occurring, the moisture cycling problems of the ERA-Interim reanalysis model are exacerbated.
Uncertainties associated with determining evaporation appear to be greater than those for precipitation. Evaporation is computed from bulk formulas that accumulate errors from wind speed, SST, and near-surface humidity measurements as well as transfer coefficients that incorporate atmospheric stability among other parameters. Evaporation initializes the hydrological cycle and therefore it is inherently linked the other components assessed here. Differences in the absolute magnitude of evaporation between datasets are a product of disparities in wind speed and near-surface humidity, particularly the latter. The deviations from SeaFlux wind and specific humidity parameters by MERRA and cancelled out to a large degree, but for ERA-Interim low near-surface humidities are the primary cause its high evaporation. The temporal resolution and transfer coefficients of the datasets may also effect evaporation. SeaFlux produces a lower latent heat flux and higher sensible heat flux than other ocean products (Clayson et al. 2013, manuscript submitted to Int. J. Climatol.). This indicates that energy partitioning between the latent and sensible heat fluxes is also a differentiating component.
Comparisons with the buoy observations indicate that the near-surface environment used to calculate evaporation is more accurately parameterized by SeaFlux and MERRA than by ERA-Interim. SeaFlux’s higher spatial and temporal resolution enables better parameterizations than ERA-Interim, especially of diurnal cycles. SSTs have a diurnal cycle that is present in SeaFlux data (Clayson et al. 2013, manuscript submitted to Int. J. Climatol.), but the diurnal cycle in MERRA and ERA-Interim SSTs is almost nonexistent (Brunke et al. 2011). Clayson and Bogdanoff (2013) showed these diurnal variations to be important for air–sea fluxes in the tropics.
The use of COARE 3.0 by SeaFlux is beneficial, as it has been consistently shown to be the best performing algorithm for determining surface fluxes (Brunke et al. 2003, 2011; Iwasaki et al. 2010). SeaFlux also has the advantage that it has been specifically calculated for these fluxes only, and is validated with observations. ERA-Interim and MERRA, however, have very complex physically parameterized reanalyses from which the fluxes are somewhat of a by-product of the model. Brunke et al. (2011) found that ERA-Interim’s greatest latent heat flux uncertainties were from algorithm problems and measurement uncertainties, whereas MERRA’s were from bulk variable differences.
Water and energy data products are constantly being updated and improved to include new datasets and scientific advances. Currently, GPCP and SeaFlux are two state-of-the-art observation-based data products. Although the use of these two datasets does not fully close atmospheric moisture budgets, they produce a freshwater flux that corresponds well with ERA-Interim atmospheric moisture divergence, despite the latter’s excessive hydrological cycle. Further developments to produce more accurate reanalyses will increase their correspondence with observational datasets and improve closure of atmospheric moisture budgets.
This project was funded under NASA Grant NNX13AG31G. The 1DD data were provided by the NASA/Goddard Space Flight Center’s Mesoscale Atmospheric Processes Laboratory, which develops and computes the 1DD as a contribution to the GEWEX Global Precipitation Climatology Project. We also thank Rebecca Smith for providing vertically integrated moisture divergence code and Carol Anne Clayson for providing the SeaFlux data, as well as advice on this dataset.