2.1. Conventional Water Balances
 The annual water balance of a closed hydrologic system can be described by the apportionment of mean annual water input by precipitation (P),
where R and ET represent area-standardized estimates of mean river outflow (i.e., runoff) and total evaporation flux to the atmosphere, respectively, and ΔS is the inter-annual change in the proportion of P stored in a watershed. On the basis of measurements of P and R over an extended period of time (years to decades), ΔS can be assumed to be a constant, implying that annual water input by P is balanced by outputs via ET and R. Consequently, (P − R) represents an approximation of ET, the collective flux of moisture by direct evaporation flux from soils, water bodies, and vegetated surfaces, and plant transpiration.
 Using ArcGIS™ software and monthly precipitation data from a 10′ gridded terrestrial climatology for the period 1961–1990 [New et al., 2000], area-weighted estimates of P were defined by the grid points completely contained by the boundaries of each watershed. Except for the watersheds in New Guinea and South America, R was based on drainage areas and historical records of river flow that were acquired from the online Global River Discharge Database of the Center for Sustainability and the Global Environment, University of Wisconsin-Madison (http://www.sage.wisc.edu/riverdata), B. M. Fekete et al. (University of New Hampshire/Global Runoff Data Centre data, 2000, available at http://www.grdc.sr.unh.edu/), and Vörösmarty et al. . Other compilations of river flow data exist [Perry et al., 1996; Dai and Trenberth, 2002], but these databases do not include the primary, annual flow data for tributaries of large rivers, such as the Volta River and Niger River, nor did they include data for the Piracicaba River (South America) or Fly River (New Guinea). For the Piracicaba River, river flow data and drainage areas were taken from the online database provided by the Department of Water and Electric Energy of São Paolo (http://www.daee.sp.gov.br) and data for the Ok Tedi and Upper Fly River watersheds was taken from [Ferguson, 2007].
 Collectively, on the basis of global area and river flow estimates from Dai and Trenberth , the watersheds encompass 5% of global non-ice, nondesert area, and collective annual flow represents 972 km3 or 2.6% of global freshwater discharge to the oceans. Using annual volumetric flow data and gross watershed areas, R was calculated as the mean rate of annual runoff for each watershed and the 95% confidence intervals were reported as estimates of uncertainty. In actuality, the proportion of gross watershed area that contributes to annual river flow on an annual basis probably fluctuates due to intermittent or restricted hydrologic connectivity and this would affect definition of R, but for consistency, gross watershed areas were utilized throughout this study. This may affect the magnitudes of the different water balance components, but not their relative proportions. Corresponding river flow data for the period 1961–1990 were used to match R with P. Where not available, data for the existing durations were used to define R. The records for most rivers were between 20–50 years and annual flow volumes for each river were relatively consistent (i.e., no statistically significant change at the 0.05 level). Only the river flow for the Bani River exhibited a statistically significant decline at the 0.05 level over the period 1960–1988 and therefore the annual river flow for the period 1922–1994 was used. The reason for the use of multiyear precipitation and river flow data is that for relatively long time intervals, hydrologic steady state can be assumed (Figure 7). For the watersheds in New Guinea, P and R were based upon precipitation and streamflow data for stations located within the Ok Tedi and Upper Fly River watersheds for the period 2004–2005. Due to the intensity of daily rainfall in this region and the short-term sensitivity of river flow to rainfall in upstream regions (Figure 8), the assumption of negligible changes in water storage was considered valid despite the short duration of the study. The relatively consistent monthly ratio of R to P for the Ok Tedi and Upper Fly River watersheds (Figure 9) further supports the validity of this assumption [Ferguson, 2007].
Figure 7. Conventional annual water balances for the selected watersheds. ET often exceeds R, implying that more water may be transferred to atmosphere each year as water vapor than is transferred to the oceans via river water. 1, North Saskatchewan River; 2, South Saskatchewan River; 3, Ottawa River; 4, St. Lawrence River; 5, Mississippi River; 6, Bani River; 7, Upper Niger River; 8, Black Volta River; 9, White Volta River; 10, Oti River; 11, Nyong River; 12, Piracicaba River; 15, Murray-Darling River.
Download figure to PowerPoint
Figure 8. Rainfall and river runoff for (a) the Upper Fly River and (b) the Ok Tedi. Rainfall exceeds runoff, implying some transfer of water vapor via evaporation and transpiration (see Figure 11 for the locations of rainfall stations and streamflow gauges).
Download figure to PowerPoint
Figure 9. Mean monthly precipitation (P) and river runoff (R) for the (a) Ok Tedi and (b) Upper Fly River watersheds of central New Guinea, 2004–2005. (P − R), an approximation of total annual evaporation (ET), is relatively constant with respect to P and represents 30 – 35% of P on an annual basis for these watersheds.
Download figure to PowerPoint
2.2. Methodology: Partitioning Evaporation and Transpiration Water Vapor Fluxes
 Apportionment of ET from equation (1) into Ed and T is difficult without additional information and for this reason, Ed is often assumed to be negligible in densely vegetated regions by necessity [Fleischbein et al., 2006], implying that the entirety of the terrestrial water vapor flux (less rainfall interception) can be attributed to T. Conversely, in sparsely vegetated regions [Choudhury, 2000] or those covered to a large extent by open water [Machavaram and Krishnamurthy, 1995], Ed is often considered more significant than T. Intuitively, it should be apparent that neither Ed nor T are likely to be completely negligible in most hydrologic systems, yet a quantitative description of their relative contributions to ET often requires the use of nonconventional hydrological techniques, principally the use of the stable isotopes of water as tracers of the hydrologic cycle.
 Isotope-based techniques for tracing water movement vary from approaches which entail integration of stable isotope data with computer simulations of the water cycle [Henderson-Sellers et al., 2004; McGuffie and Henderson-Sellers, 2004; Miller et al., 2005; Henderson-Sellers, 2006; Fekete et al., 2006] to field-oriented studies which rely more directly on measurements of ambient moisture [Moreira et al., 1997] or employ hydrologic and isotope mass balance equations and the isotope separation between water input by precipitation and eventual outflow in order to infer the relative contributions of evaporation and transpiration [Gibson et al., 1993; Gibson and Edwards, 2002; Lee and Veizer, 2003; Welp et al., 2005; Mayr et al., 2006; Gammons et al., 2006; Twining et al., 2006; Wolfe et al., 2006; Ferguson et al., 2007; Freitag et al., 2007; Karim et al., 2007]. Each technique is, however, fundamentally based upon the principle that the mass-dependent isotope fractionation that accompanies the phase transitions of water (i.e., evaporation, sublimation, and condensation) causes the predictable apportionment of the heavier, rarer isotopes of water (oxygen-18, 18O; deuterium, 2H) within the principal isotopologues of the water molecule (1H216O, 1H2H16O, 1H218O). By convention, the 18O and 2H content of water samples are hereafter expressed as δ values, representing deviation in parts per thousand (‰) from Vienna-Standard Mean Ocean Water (VSMOW), such that δ2H or δ18O = [Rsample/Rstandard) − 1] × 1000 where R represents 18O/16O and 2H/1H, respectively.
 As a consequence of mass-dependent isotope fractionation, δ18O and δ2H values of global precipitation exhibit systematic spatial and temporal variations and collectively define a regression line that is commonly referred to as the global meteoric water line (GMWL, Figure 10) [Craig, 1961; Craig and Gordon, 1965; Dansgaard, 1964; Rozanski et al., 1993]. At a regional scale, mass-dependent isotope fractionation leads to the existence of two linear trends on a conventional δ18O-δ2H diagram, one trend defined by waters which retain their original isotope composition derived from precipitation and a second trend defined by waters which, to some extent, have undergone heavy-isotope enrichment due to evaporation (i.e., surface waters) (Figure 10) [Froehlich et al., 2002; Gibson et al., 2005]. The former trend defines a local meteoric water line (LMWL) with a slope and δ2H-intercept value governed by the prevailing meteorological conditions of the moisture source area, rainout and the trajectory of the air mass, and second-order kinetic effects such as those associated with snow formation and evaporation from raindrops [Clark and Fritz, 1997; Araguás-Araguás et al., 1998]. The second linear trend, commonly referred to as a local evaporation line (LEL), is defined by δ18O and δ2H values of river water collected from the point of outflow from a watershed.
Figure 10. A conceptual representation of the relationship between a local evaporation line (LEL) defined by δ18O and δ2H values of river water and the global meteoric water line (GMWL). GMWL is based on monthly δ18O and δ2H values of precipitation from each GNIP station used in this study. δ18O and δ2H values of precipitation for Cayenne (French Guyana) and Calgary (Canada) are shown to illustrate differences in the isotope composition of precipitation in the tropics versus higher latitudes. For a given region or watershed, δI and δS represent the mean annual δ18O and δ2H values of water input by precipitation and eventual outflow via river water, respectively (see section 2.2).
Download figure to PowerPoint
 Analogous to the isotope composition of precipitation at a specified location, which is determined by the initial isotope composition of water evaporated at the moisture source and the path of moisture transport to the location of eventual condensation [Jouzel, 2006], river water at a given location within a watershed is primarily determined by the initial isotope composition of water input and the cumulative influence of evaporation upstream, as mass-dependent isotope fractionation causes an evaporating moisture flux to become preferentially depleted in 18O and 2H and leaves the residual liquid in the watershed enriched in these heavier isotopes [Gibson and Edwards, 2002]. Consequently, δ18O and δ2H values of river water collected from the point of outflow from a watershed can be interpreted as a proxy for the residual liquid that remains after evaporation has occurred. In watersheds where direct evaporation from soils and water bodies is volumetrically significant relative to annual water input, δ18O and δ2H values of river water define a LEL with a shallower slope than the LMWL, such that the offset between δ18O and δ2H values of individual samples from the LMWL increases in approximate proportion to the cumulative fraction of water transferred to the atmosphere by evaporation in upstream areas.
 According to Gibson and Edwards , the intersection of the LEL and LMWL provides an amount-weighted estimate of the δ18O and δ2H values of water entering a hydrologic system and thus is unaffected by heavy-isotope enrichment due to evaporation within the watershed. Inferences regarding the influence of evaporation within a hydrologic system are usually based upon the relationship between the LMWL and LEL, as these linear features are considered more representative of prevailing hydrologic conditions than individual δ18O and δ2H values. Fundamentally, studies that purport to separate fractionating and nonfractionating water vapor fluxes also rely upon the assumptions that evaporation is incomplete (i.e., a residual, 18O-2H-enriched liquid remains), the isotope composition of soil water is relatively unaffected by the process of water uptake by plants [Moreira et al., 1997; Twining et al., 2006], and that a relatively small number of δ18O and δ2H values are representative of processes occurring over large areas [Gibson et al., 1993, 2005; Fekete et al., 2006]. If these assumptions are deemed valid and only direct evaporation from soils and water bodies affects the isotope composition of water during its movement through a watershed, the proportion of fractionating water vapor fluxes to water input can be resolved by examining the isotope separation between δ18O and δ2H values of water input (i.e., precipitation) and output (i.e., river outflow) from a particular geographical region [Gonfiantini, 1986; Gibson et al., 1993; Lee and Veizer, 2003; Welp et al., 2005; Gammons et al., 2006]. Rivers are often the focus of these studies because their watersheds usually define a closed hydrologic system and the isotope composition of river water reflects, to an extent, the influence of evaporative water vapor fluxes in regions that contribute to river flow.
 On the basis of the magnitude of the isotope separation between water input by precipitation (δI) and river outflow (δS), a variety of isotope mass balance equations are proposed to quantify the ratio of evaporation to water input under steady state [Gonfiantini, 1986; Gat and Bowser, 1991; Gibson et al., 1993] and nonsteady state hydrologic conditions [Gibson, 2002]. In situations where volumetric changes are negligible and the volume of water in the hydrologic system can be considered constant (i.e., P = R + ET from equation (1)), a steady state isotope mass balance equation enables quantification of the proportion of evaporation (Ev) with respect to water input (I). Assuming conditions of hydrologic and isotope steady state, [Gonfiantini, 1986] derived an isotope mass balance equation that enables an approximation of the Ev relative to annual water input (I),
where h is the ambient humidity normalized to saturation vapor pressure, α is the equilibrium fractionation factor for oxygen (lnα = 1137T−2 − 0.4156T−1 − 0.00207) and hydrogen (lnα = 24844T−2 − 76.248T−1 + 0.05261) isotopes during evaporation [Friedman and O'Neil, 1977], Δɛ is the kinetic enrichment factor for oxygen [14.2(1 − h)] and hydrogen isotopes [12.5(1 − h)], ɛ = α − 1, and, δI, δA and δS are the mean δ18O (or δ2H) values of precipitation, ambient moisture, and outflow, respectively (see Ferguson et al.  for a more detailed description of equation (2)). The range of δ18O and δ2H values defined by the intersections of the 95% confidence intervals of the LEL with the LMWL was considered an approximation of the uncertainty related to δI, which enabled information on the proportion of variance described by the LEL for each watershed to be incorporated. δS represents the flow-weighted mean δ18O (or δ2H) value of river water at the point of outflow from the watershed (i.e., the river mouth). The 95% confidence intervals for δI and δS were used to approximate the error associated with the Ev/I value from equation (2). Mean annual values for temperature and humidity from [New et al., 2000] were substituted into equation (2) and were used to calculate the isotope fractionation factors for this equation. In order to approximate the isotope composition of atmospheric moisture, isotope equilibrium between atmospheric moisture and mean annual precipitation was assumed such that δA = δI − ɛ* [Gat and Matsui, 1991; Gammons et al., 2006].
 The Ev/I values from equation (2) enabled the separation of ET from equation (1) into fractionating water vapor fluxes (i.e., direct evaporation from water bodies and soils, Ed) and nonfractionating water vapor fluxes (i.e., canopy evaporation, In, plus plant transpiration, T). In represents the proportion of precipitation that is caught or intercepted by vegetated surfaces and subsequently evaporated before infiltration into the soil zone can occur. By definition, In represents complete evaporation from plant surfaces (i.e., no residual liquid remains) and thus does not appreciably alter the isotope composition of water entering the soil zone. Consequently, In must be accounted for in order to approximate the proportion of ET attributable to T. In order to calculate In, vegetation within each of the fifteen watersheds was classified on the basis of the Global Land Cover (GLC) 2000 data sets for North America [Latifovic et al., 2002], Africa [Mayaux et al., 2004], Southeast Asia [Stibig et al., 2003] and South America [Eva et al., 2003] and this classification was, by necessity, used to calculate the net water input to each watershed (i.e., P − In). In order to calculate an area-weighted estimate of In for each watershed, the corresponding GLC2000 land-cover class for each grid point from New et al.  was determined using ArcGIS™ software and literature In values were assigned to each point. An area-weighted estimate of In for each watershed was calculated as the sum of In values for each data point completely contained by the watershed boundaries.
 Ed was then calculated as the product of the Ev/I value from equation (2) and (P − In), which represents the net water input to the soil zone. T was considered the residual amount of water required to balance equation (1) and the error assigned to T was the propagated errors of P, R, In, and Ed calculated by a sum of squares.
2.3. Sources of Isotope Data
 In North America, the selected rivers included the North and South Saskatchewan Rivers (the principal tributaries of the Nelson River, which flows into Hudson Bay), the Mississippi River, the Ottawa River (a tributary of the St. Lawrence River), and the St. Lawrence River (Figure 6). The isotope compositions of river water collected near the mouths of the North Saskatchewan River at Prince Albert and South Saskatchewan River near Saskatoon (Figure 11a) were available from Ferguson et al. . The isotope compositions of precipitation at Calgary (51.02°N, 114.02°W) and Edmonton (53.57°N, 113.52°W) are from the Global Network of Isotopes in Precipitation (GNIP) network (IAEA/WMO, 2004, http://isohis.iaea.org) and additional isotope data for precipitation at Saskatoon were available from [Ferguson et al., 2007] (Figure 12). For the St. Lawrence River, the isotope composition of river water collected near Cornwall (51.02°N, 114.02°W) was taken from Barth and Veizer  and represents outflow from the Great Lakes region of North America. The isotope composition of precipitation in this region was characterized by data from GNIP stations located at Ottawa (45.32°N, 75.67°W), Atikokan (48.75°N, 91.62°W), Simcoe (42.85°N, 80.27°W), and Chicago (41.78°N, 87.75°W). For the Ottawa River, isotope data for river water were from Carillon, a station located near the river mouth [Telmer and Veizer, 2000; Myre and Hillaire-Marcel, 2004], and isotope data for precipitation from the GNIP station at Ottawa. The isotope composition of river water collected near the mouth of the Mississippi River at Melville was taken from Lee and Veizer  and the isotope composition of precipitation for the Mississippi River watershed was from GNIP stations located at Coshocton (40.36°N, 81.80°W), Chicago (41.78°N, 87.75°W), Calgary (51.02°N, 114.02°W), Hatteras (35.27°N, 75.55°W), and Denver (39.77°N, 104.88°W).
Figure 11. (a) North and South Saskatchewan River watersheds in North America, (b) Volta River watershed in West Africa, and (c) sampling locations for river water at Konkonda (Ok Tedi), Kiunga (Upper Fly River), and Nukumba (Middle Fly River) and rainfall at Tabubil.
Download figure to PowerPoint
Figure 12. The δ18O and δ2H values of (a) rainfall and (b) river water at monitoring stations in the North and South Saskatchewan River watersheds (95% confidence intervals are in parentheses). In most of the “water-limited” watersheds, δ18O and δ2H values of river water defined a LEL with a shallower slope than the corresponding LMWL (Table 2), implying that some evaporative enrichment of river water occurs on an annual basis.
Download figure to PowerPoint
 In South America, the Piracicaba River watershed is located in the state of São Paolo in southern Brazil and flows into a tributary of the Paraná River. The isotope composition of river water collected near the river mouth at Artemis (22.67°S, 47.77°W) was available from Martinelli et al.  and the GNIP database (IAEA/WMO, 2004, http://isohis.iaea.org). For precipitation, data from the GNIP station located near Rio de Janeiro (22.90°S, 43.17°W) were used.
 In Africa, the Black Volta River, the White Volta River, and the Oti River are the principal tributaries of the Volta River, which flows into Lake Volta in southern Ghana before eventual outflow occurs into the Atlantic Ocean (Figure 11b). The Upper Niger River and the Bani River are headwater tributaries of the Niger River. The Nyong River (Cameroon) flows into the Atlantic Ocean. The isotope composition of river water collected near the mouths of the Black Volta River, White Volta River, and Oti River are from Bamboi, Nawuni, and Sabari, respectively [Freitag et al., 2007] (Figure 13). The isotope composition of river water collected from the Upper Niger River at Banankoro and the Bani River at Douna were available from the GNIR database. Douna is located immediately before the confluence of the Bani River and Upper Niger River whereas the station at Banankoro is located approximately 230 km upstream of this confluence. In the Nyong River watershed, located between the Niger River and Congo River watersheds in Cameroon, the isotope composition of river water collected near the river mouth at Dehane was available from F. Brunet (personal communication, 2006). The isotope composition of precipitation in West Africa was from the GNIP stations at Kano (12.05°N, 8.52°E), Bamako (12.32°N, 7.57°W), and Niamey (13.52°N, 2.08°E), five stations within the Volta River watershed [Freitag et al., 2007], and five stations located within the Nyong River watershed.
 The Ok Tedi and Upper Fly River watersheds are located along the southern flank of the Central Cordillera of New Guinea and are tributaries of the Fly River (Figure 11c). The isotope compositions of river water at Konkonda (6.00°N, 141.15°E) and Kiunga (6.12°N, 141.30°E) were used to characterize outflow from the Ok Tedi and Upper Fly River watershed, respectively, and the isotope composition of precipitation at Tabubil (5.26°S, 141.21°E) was considered representative of the headlands region of these rivers [Ferguson, 2007] (Figures 11c and 14).
Figure 14. The δ18O and δ2H values of (a) rainfall and (b) river water at monitoring stations in the Fly River watershed, New Guinea (95% confidence intervals are in parentheses). The similarity of LEL and LMWL in watersheds covered primarily by tropical rain forest suggests that the volume of water “lost” via direct evaporation from water bodies and soils (Ed) may be much less than the nonfractionating water vapor fluxes in these regions (i.e., T plus In).
Download figure to PowerPoint
 In Australia, outflow from the Murray-Darling River watershed was characterized by samples collected from several stations located below Rufus Junction [Simpson and Herczeg, 1991]. The isotope composition of precipitation was characterized by data from GNIP stations located near Brisbane (27.43°S, 153.07°E), Adelaide (34.92°S, 138.57°E) and Melbourne (37.82°S, 144.97°E).
 All discussed isotope data for precipitation and river waters are included as auxiliary material (see Tables S1 and S2).