4.1. Monthly and Synoptic Variability
 All of the vapor isotope measurements from each site are plotted in δ18O-δD space in Figure 1. In each case, the δ18O and δD data cluster around the GMWL, with higher summer values showing a tighter fit to the GMWL than lower winter values. The large scatter around the GMWL at Rosemount may be because this was the only site that did not use a buffer volume on the ambient intake sampling line to dampen short-term variability.
Figure 1. Vapor measurements of δD and δ18O at each site, over the entire period of measurement (light gray crosses) and for June–August (black circles). For reference, the GMWL is plotted with each site (dark gray line). Deuterium-excess (d) is a measure of the deviation from the GMWL in δD (‰).
Download figure to PowerPoint
 The New Haven and Beijing sites had the longest duration of continuous measurements of about a year, making it possible to examine the seasonal variability of d in vapor and precipitation. Figure 2 shows the monthly arithmetic mean dvand the amount-weighted precipitationd. At both sites, dv was usually greater than precipitation d. In general, both vapor and precipitation d were lowest during summer months. This is in contrast to the seasonal cycles in δ18O and δD, which have summer maximums. Summer minima in dv was previously observed by Jacob and Sonntag  over eight years in Heidelberg, Germany and by Angert et al.  over nine years in Rehovot, Israel. The mean seasonal cycle of d in northern hemisphere precipitation from the IAEA GNIP database is also at a minimum in the summer months [Araguas-Araguas et al., 2000]. Angert et al.  describe how the seasonal variability in dv in Israel can be explained by changes in the initial vapor from the Mediterranean Sea superimposed on Rayleigh isotopic distillation. Initial vapor is modified by monthly variability in temperature and relative humidity near the surface of the Mediterranean Sea and varying rates of moisture entrainment from the lower troposphere into the PBL.
Figure 2. Monthly arithmetic mean of vapor d(solid lines) and amount-weighted precipitationd (dashed lines) for the New Haven (black) and Beijing (gray), January 2007 through May 2008.
Download figure to PowerPoint
 In Beijing, monthly variability in dv has a negative linear correlation with temperature (p = 0.001, slope = −0.46‰ per °C), water vapor mixing ratio (p = 0.002, slope = −0.52‰ per mmol mol−1), but not well correlated with precipitation amount (p = 0.10) or relative humidity (p = 0.24). In New Haven, monthly variability in dv is not well correlated with any of these environmental variables (p > 0.20 in all cases).
 Mean summer dv at the North American sites were higher than at the Chinese sites by approximately 7‰ (Table S1 in the auxiliary material) even though they are all at nearly the same latitude. This general pattern is predicted from a model interpolation of the IAEA GNIP database, showing high values of d of precipitation downwind of the Great Lakes [Bowen and Revenaugh, 2003]. Evaporation from the Great Lakes is expected to have higher d values than the air masses transported to this region. Gat et al.  used high d in precipitation from the Northeastern U.S., presumably resulting from dv anomalies, to estimate that recycled continental moisture from evaporation off the Great Lakes contributed 4–16% of precipitation downwind of the lakes in the summer.
 We also observed differences in the synoptic variability in the concurrent measurements at New Haven and Beijing (Figure 3). The 10-day running means were more variable at New Haven than at Beijing. New Haven is located in the northeast of U.S. where all the major storm tracks converge and experiences weather cycles every 7–10 days throughout the year [Zielinski and Keim, 2003] bringing with them moisture from different source regions. In comparison, frontal activities are less frequent in Northeast China in the warm season. Wen et al.  proposed that the Asian monsoon suppresses variability in dv in the summer by maintaining a relatively consistent moisture source region.
4.2. Day-to-Day Variability
 If dv is a conservative tracer of conditions in the moisture source region, we would not expect it to vary with local h unless there is a local source of moisture to the atmosphere. Afternoon averages (12:00–18:00 local standard time) of dv were correlated with h at New Haven (dv/h = −0.36‰/%, r = −0.74, p < 0.001) and Borden Forest (dv/h = −0.22‰/%, r = −0.57, p < 0.001) sites during the summer months (June–August), suggesting that local contributions of moisture are relatively high (Figure 4 and Table 2). The afternoon period was chosen because correlations during this time were the higher than any other time of day or the daily mean. This is also the time of the day when the surface measurements are most representative of regional conditions due to increased atmospheric turbulence. These two particular sites are located near large bodies of water, which may contribute significant amounts of moisture to the sites. New Haven sits on Long Island Sound adjacent to the Atlantic Ocean. Located between Lake Huron and Lake Ontario, Borden Forest is approximately 20 km southeast of the Georgian Bay. Although we did not observe significant linear correlations between dv and h afternoon averages at Beijing and Luancheng during the summer, Wen et al.  reports weak nonlinear correlations between hourly dv and local h at these sites outside periods of peak summer monsoon activity from the same data set. That Borden and New Haven show significant correlations, whereas the more inland sites do not, suggests that a stronger correlation may be indicative of a higher contribution of local water sources to the atmospheric moisture. The footprint of the source influence appears to have extended tens of kilometers beyond the local site scale, even though the measurements were made near the ground.
Figure 4. Afternoon (12:00–18:00 local time) mean dv versus afternoon mean h in the summer (June–August). Only New Haven and Borden sites have significant negative correlations. Changing the averaging time window within the daytime period does not affect the correlation pattern across sites. Fit statistics are summarized in Table 2.
Download figure to PowerPoint
Table 2. Linear Least Squares Fits Between Afternoon Means (12:00–18:00) During June–August, Correlations and Statisticsa
|dv (‰) Versus h (%)|
|dv (‰) Versus Temperature (°C)|
|dv (‰) Versus H2O (mmol/mol)|
|dv (‰) Versus PBL Height (km)|
|h (%) Versus PBL Height (km)|
|δ18O (‰) Versus PBL Height (km)|
|δD (‰) Versus PBL Height (km)|
 The variability in dv/hrelation in marine-type settings in the literature ranges from −0.31 to −0.61‰ per % for vapor evaporated from the Mediterranean Sea and southern ocean [Gat et al., 2003; Pfahl and Wernli, 2008; Uemura et al., 2008]. The New Haven slope (−0.36) is within the range of ocean/sea vapor observations from the literature, but the Borden Forest slope (−0.22) is lower. To our knowledge, this is the first terrestrial observation showing a relation between dv and h. A possible explanation for the lower observed dv/h slope at Borden Forest, compared to literature values, is that they calculate h relative to the sea surface temperature at or near the site of evaporation, whereas our comparison used h calculated relative to air temperature at the vapor measurement site downstream of the suspected moisture source. However, our attempt to correct measured h relative to NOAA buoy water temperature or air temperature just above the water surface of the assumed moisture source regions (Long Island Sound in the case of New Haven and Lake Ontario in the case of Borden Forest) revealed lower dv/h slopes (Table S2 in the auxiliary material).
 It is well known that water vapor mixing ratio controls much of the variability in δ18O and δD [Lee et al., 2006; Wen et al., 2010] due to Rayleigh distillation processes. It is not expected that this would extend to dv as well because Rayleigh distillation should not greatly affect dv, with the exception of the nonlinearity effects described in the Methods section. We found that afternoon averages of dv were strongly negatively correlated with water vapor mixing ratio at all sites except Beijing (Table 2 and Figure S1 in the auxiliary material, p < 0.001). Similar to the case of the correlation with h, changes in water vapor mixing ratio appear to have a larger effect on dv (i.e., higher slopes) at the North American sites than at the Chinese sites. It could be that water vapor mixing ratio is also a good predictor of dv at least at some sites, but the mechanism causing such a relationship is difficult to identify at this point. None of the sites showed a significant correlation between dv and local air temperature (Table 2).
4.3. Diurnal Variability
 Figure 5a shows the mean diurnal cycles of dvfrom each site during the summer months (June–August). This three-month period was selected because measurements were made at all sites during these months, and they represent the growing season at midlatitudes in the northern hemisphere. We found thatdv varied diurnally, showing a clear, robust pattern of maximum dv during the afternoon at all sites. This extends the finding of Wen et al. , which previously published the results from the Beijing site, and it is consistent with the observations of Lai and Ehleringer  over a few days in the Pacific Northwest. The study presented here is the most extensive set of continuous measurements of diurnal variability in dv to date and shows that daytime dv increase is not a pattern unique to any one location or vegetation type. The nighttime ‘baseline’ values of the Chinese sites are approximately 6‰ lower than the North American sites, similar to the offset in summer mean values already discussed.
Figure 5. Mean diurnal cycles from June–August of (a) d of vapor, (b) δ18O of vapor, (c) δD of vapor, (d) h relative to local air temperature, (e) water vapor mixing ratio, and (f) planetary boundary layer (PBL) height shown for each site.
Download figure to PowerPoint
 The phases are remarkably similar despite widely varying local vegetation types. The site that is most different from the others in terms of amplitude and phase is Beijing. It is also the most densely populated urban environment with large amounts of impermeable surfaces, and the influence of transpiration should be negligible and evaporation should be small at this site. In contrast, New Haven, the other urban site, has a large fraction of tree vegetation cover (49%, http://nrs.fs.fed.us/data/urban/) and is located in close proximity to the Atlantic Ocean, resulting in higher transpiration and evaporation influence.
 Although the phases are similar at all sites, the peak-to-trough amplitudes vary greatly: 3.5‰ at Beijing, 7.7‰ at Luancheng, 9.8‰ at New Haven, 10.0‰ at Borden Forest, 13.5‰ at Duolun, and 17.1‰ at Rosemount (Figure 5a). Note that the Rosemount amplitude may be overestimated in this analysis because of the large amount of noise in the data. Clearly within the diurnal time scale, dv is not a conserved tracer of humidity conditions at the marine moisture source region. For example, a diurnal amplitude of 9.8‰ at New Haven would imply an uncertainty of 27% in the inferred relative humidity at the source.
 We find that the mid-day decrease inδ18O of water vapor, combined with little to no change in δD, leads to the variability in dv (Figures 5b and 5c). Several previous studies have examined controls of the diurnal variability of δ18O of water vapor, and it is likely that some of the same processes may explain the variability in dv. The morning decrease in δ18O has been attributed to the rapid increase of air entrained into the boundary layer during convective mixing [Lai et al., 2006; Welp et al., 2008; Zhang et al., 2011]. Later in the day, the increase in δ18O may be driven by surface evapotranspiration, because evapotranspiration acts to enrich the surface vapor in 18O [Lee et al., 2012]. Lai and Ehleringer  is one of the few studies that investigated the diurnal variability in dv, and they also conclude that entrainment and evapotranspiration play dominate roles in dv variability in the early morning and late afternoon, respectively.
 The nonlinearity of the delta-notation definition may contribute to the increaseddvduring mid-day (see comment on the approximation of the delta-notation in the Methods section). This effect was taken into account recently in the seasonal cycle ofdv in Angert et al. . We estimate that the nonlinearity of the delta-notation contributes to ∼10% or less of the diurnal amplitude indv. Because the daily minimum of δ18O and δD are not at the same time, the nonlinearity errors are either positive or negative at different sites, ranging from −1.6 to 0.4‰ (Table S3 in the auxiliary material). We conclude that this effect is not a significant contributor to the diurnal cycle of dv at the sites we observed.
 To examine entrainment, which has been identified as a major contributor to the diurnal cycle of vapor isotopes from other studies, we compared afternoon mean vapor isotopes at each site with modeled estimates of the PBL height (Figure 6 and Table 2). Although the PBL height is controlled by several factors, it has been shown that the entrainment of free tropospheric air is the leading order control providing the energy to grow the PBL [Medeiros et al., 2005], and we use it here as a rough proxy for entrainment. Afternoon dv was significantly correlated with PBL height at New Haven (slope = 8.5‰/km, p < 0.001) and Borden Forest (slope = 8.7‰/km, p < 0.001). On days with higher PBL heights, dv was higher at these sites. Rosemount also showed a similar slope, but with less statistical significance (slope = 8.6‰/km, p = 0.021). This site also has a higher degree of noise in the data, which could reduce the correlation significance. Correlations were not significant at the Chinese sites (p > 0.13).
Figure 6. Afternoon (12:00–18:00 local time) mean dv versus afternoon mean PBL height in the summer (June–August). Fit statistics are summarized in Table 2.
Download figure to PowerPoint
 Interestingly, afternoon PBL height correlations with δ18O were the largest and most significant at the North American sites (slopes range from −2.9 to −4.9‰ km−1, p < 0.003) as were the correlations with afternoon δD (slopes range from −14.5 to −30.6‰ km−1, p < 0.03) (Table 2). These results suggest that the isotope ratios of water vapor could vary more with height in the atmosphere in some areas like northeastern North America, than others, like near Beijing, China.
 An increase in dv with height in the atmosphere could explain our observations at the North American sites. To our knowledge, there is limited evidence of dv increasing with height in the lower troposphere. Griffis et al. [2011a] showed that dv measured in Rosemount was, on average, 5.1‰ greater at 200 m than at 3 m, from measurements made including the cold season, starting April 2010 through June 2011, and that the difference approaches zero during periods of strong daytime mixing. Measurements made near New Haven by He and Smith  showed that the free atmosphere dv was ∼8‰ higher than the atmospheric boundary layer value. There is also evidence of extremely large dv values in the upper troposphere and lower stratosphere of the tropics and subtropics [Webster and Heymsfield, 2003]. While our analysis suggests that entrained air has higher dv than the surface, Lai and Ehleringer  predicted that air entrainment exerted a negative isotope forcing in the early morning on their three study days in the Pacific Northwest and switched to a positive isotope forcing on the afternoons of the last two days. It should be noted that they used measurements of dv at 60 m to estimate the isotopic composition of the entrained air. While the addition of surface moisture via evapotranspiration may be able to explain vertical profiles of δ18O and δD, it has the wrong sign in the case of dv, adding higher values of dv near the surface.
 In the interpretation of Figure 6, we assume that a higher PBL height in mid-afternoon is associated with faster PBL growth and therefore a stronger entrainment rate. If there is a vertical gradient indv increasing with height, entrainment could easily explain the diurnal cycle in dv. To support this argument, we parameterized the ISOLES model with typical midlatitude summer conditions. In the model domain, the evolution of the PBL was forced by a time-varying solar radiation, a prescribed initial specific humidity profile in the early morning [Lee et al., 2012] and initial profiles of δ18O and δD according to the observed relationships with specific humidity at Beijing [Wen et al., 2010]. To remove the influence of evapotranspiration on dv, the d of the surface water vapor flux was held at a constant equal to the initial surface layer dv value. The results demonstrate that mixing and entrainment could produce increases in dvfrom early morning to mid-afternoon in magnitude similar to that of the observed changes (Figure S2 in theauxiliary materials).
 We present two testable hypotheses for vertical profiles of dv from the surface to the lower troposphere to explain our observations. (1) An increase in dv with altitude could be unique to eastern North America caused by evaporation off the Great Lakes resulting in high dvat altitude downwind, similar to Gat's lee-side convective model [Gat et al., 2003]. (2) Partial re-evaporation of raindrops below the cloud base increasesd of the surrounding vapor [Lee and Fung, 2007]. This process could be common to most regions, but during the summer months, the Chinese monsoon has erased such a gradient in dv because of enhanced vertical mixing and large downpours with little chance for raindrop evaporation during decent. We examined the dvoutside of the summer monsoon season to test this hypothesis and found no correlation with PBL height at either New Haven or Beijing; however, greater noise in the data during the drier non-summer months may hide any such signal.
 Turning our attention to plant transpiration, a common assumption has been that isotope ratios of plant transpiration equal those of xylem water. This is true for δ18O and δD when averaged over days, or during short time-scales under steady state conditions, but not under nonsteady state conditions [e.g.,Farquhar and Cernusak, 2005; Lai et al., 2006; Welp et al., 2008; Xiao et al., 2010]. The isotopic land surface model SiLSM [Xiao et al., 2010] provides support for the role of nonsteady state plant transpiration in the daytime increase in dv. This model consists of a standard land surface sub-model for water and carbon dioxide exchange, a nonsteady state theory of leaf water isotopic composition, and a formulation of the kinetic fractionation at the canopy scale. The model has been tested against field observations of theδ18O composition of evapotranspiration at Rosemount during a soybean year [Xiao et al., 2010] and at the Luancheng cropland [Xiao et al., 2012]. Here we have extended the model to the δD tracer with the inclusion of the appropriate fractionation factors. Model simulations for the Luancheng cropland were parameterized using field measurements of plant xylem water isotopes, averaging −6.3‰, −53.5‰ and −3.3‰ for δ18O, δD and d respectively during the wheat phase and −7.0‰, −59.5‰ and −3.5‰ respectively during the corn phase. Since corn and wheat have different leaf physiology and stomatal control, they have been parameterized differently in the model. The model results reveal a strong diurnal cycle in the d of transpiration (dT) for both the corn and wheat growing periods that are more or less in phase with the diurnal composition of dv observed at this site (Figure 7). Averaged over the entire growing season, the dTwas negative at night (−200 to −150‰), increased rapidly after sunrise and peaked at 20 to 50‰ in mid-morning hours and remained high through the mid-afternoon. The model results are in agreement with recent observations in a forest showing thedof evapotranspiration peaked in mid-day [Wen et al., 2012]. The SiLSM-modeleddT results show plant transpiration is a likely contributor to the diurnal variability in dv.
Figure 7. SiLSM model results of the mean diurnal cycle of d of transpiration (dT) at Luancheng for the wheat (black solid) and corn (gray solid) phases during the 2008 growing season. In the mid-afternoon,dT approaches the average d value of xylem water in the model, approximately −3.5‰ for both the wheat and corn phases. Also shown are the results of a sensitivity test where canopy resistance (rc) was doubled (dotted lines). Higher canopy resistance results in a departure from steady state transpiration and larger diurnal amplitude in the isotopic forcing of transpiration.
Download figure to PowerPoint
 We do not have complete data on the forcing variables to run SiLSM for other sites. Sensitivity analysis suggests that ecosystems with a lower stomatal resistance should have more diurnally varying dT (Figure 7). Perhaps this explains why the diurnal amplitude at the irrigated Luancheng crop site was lower than at the rainfed maize site in Rosemount and at the two natural ecosystems (Borden and Duolun).
 Finally, we discuss the potential role of surface evaporation from soils. It is possible that the daytime increase in dv is partially driven by the diurnal pattern in the soil evaporation, a small flux of water, but which tends to have large dv values. Soil evaporation is expected to be a minor component of evapotranspiration at all sites except Beijing during the summer. Field observations show that at times when LAI is greater than 2, evaporation is generally less than 10% of the total evapotranspiration flux to the atmosphere, based on lysimeters measuring soil evaporation and eddy covariance measuring total evapotranspiration flux [Lee et al., 2007b; Griffis et al., 2010]. Soil evaporation typically follows a diurnal pattern nearly identical to transpiration that peaks in mid-afternoon following incoming solar radiation making it difficult to separate from plant transpiration. The evaporation flux is also enhanced by the lowerh during the daytime (Figure 5d). Using the observed isotope compositions of soil water at Luancheng, the dof soil evaporation varied in the range of 160 to 210‰ through the diurnal cycle according to the Craig-Gordon model prediction in SiLSM. These larged values could be enough to contribute to dv variability even by a relatively small portion of total evapotranspiration.