With the help of a Lagrangian moisture source diagnostic, linkages between stable isotope measurements in water vapor in Rehovot (Israel), with typical sampling times of 8 hours, and the meteorological conditions in the evaporation regions are established. These linkages can be formulated in quantitative terms, and are also quantitatively comparable with other data from isotope measurements over the ocean and with simple theoretical calculations. On the one hand, a strong negative correlation (r = −0.82) between relative humidity with respect to sea surface temperature in the source regions and measured deuterium excess (d) is found, corroborating results from isotope global circulation model simulations. This relationship can also be applied to model d in a larger region, as shown for a sample case. On the other hand, sea surface temperature in the evaporation regions does not correlate well (r = −0.21) with measured d. This finding contradicts results from other models. Although requiring confirmation by isotope data from different regions, this weak correlation is potentially of major importance for the interpretation of deuterium excess measured in ice cores, which has been used to reconstruct moisture source temperatures for past climates.
 Stable hydrogen and oxygen isotopes in atmospheric waters are useful in various ways: they can be applied as tracers in order to improve our understanding of the hydrological cycle [e.g., Gat, 1996; Henderson-Sellers et al., 2004], and their concentration in ice cores and other paleo-archives is used to reconstruct past climate variability [e.g., Dansgaard et al., 1993; Petit et al., 1999; North Greenland Ice Core Project members, 2004]. Essential for all these applications are fractionation effects which change the concentration of heavy isotopes with respect to their lighter counterparts at phase transitions of water. Two types of fractionation can be distinguished: Equilibrium effects occur at all phase transitions owing to the different bonding energies of molecules made of different isotopes in the condensed phase, leading to differing water vapor saturation pressures. These processes, according to their quantum mechanical origin, are controlled by temperature [Gat, 1996]. In addition, the slower diffusion velocities of heavy molecules can lead to fractionation during transport, called nonequilibrium (or kinetic) fractionation. In the atmosphere, the latter occurs mainly during evaporation of water [Craig and Gordon, 1965; Merlivat and Jouzel, 1979], reevaporation of rain drops under the cloud base in unsaturated air [Steward, 1975; Gedzelman and Arnold, 1994] and the formation of ice clouds, when supersaturation is possible [Jouzel and Merlivat, 1984].
 Whereas equilibrium fractionation is well understood theoretically [Bigeleisen, 1961] and in laboratory experiments [Majoube, 1971; Horita and Wesolowski, 1994], the situation for nonequilibrium effects is less clear. In particular, there are several models describing evaporation of water from the sea [Craig and Gordon, 1965; Merlivat and Jouzel, 1979], based upon the theory of one-dimensional advective-diffusive systems [He and Smith, 1999], but there is only very little isotope data available which can be used to validate these models. Hence, it is not clear if some of the assumptions are oversimplified and if all critical physical processes are properly represented. For instance, newer lab measurements by Cappa et al.  showed that surface cooling due to evaporation, which is not included in the standard models, might play an important role. Under realistic atmospheric conditions, additional factors like sea spray evaporation in the case of strong winds come into play [Gat et al., 2003], and feedbacks between the isotopic composition of the evaporation flux and the ambient air have to be taken into account [Gat and Bowser, 1991]. Difficulties in quantifying these various processes arise not only from the paucity of isotope measurements on short timescales (<1 day), which are required for establishing direct linkages to the involved mechanisms, but also from the complexity of the interpretation of these measurements.
1.1. Deuterium Excess
 The deuterium excess (hereafter called d excess or briefly d, measured in ‰) is defined as d = δ2H − 8δ18O, whereas δ2H and δ18O are given relative to Vienna standard mean ocean water (VSMOW) [Dansgaard, 1964]. For evaporation under equilibrium conditions from ocean water with isotopic composition equal to that of VSMOW, the ratio δ2H/δ18O is approximately 8, thus the d excess is a measure for nonequilibrium fractionation. In addition, the equilibrium ratio for the liquid-vapor transition slightly depends on temperature, hence equilibrium fractionation can also lead to small changes in d. (However, at the transition from vapor to ice, even large changes in d can occur, also under equilibrium conditions.) The global average value of d in meteoric waters is 10‰ [Craig, 1961], indicating that atmosphere and ocean typically are out of thermodynamic equilibrium.
 The dominating effect that determines the d excess on a global scale is nonequilibrium fractionation during water evaporation from the ocean [Craig and Gordon, 1965]. Other nonequilibrium effects mentioned above are thought to be important on local to regional scales [Jouzel et al., 2007]. From theoretical evaporation models, it can be deduced that d is primarily determined by relative humidity, measured with respect to saturation at the sea surface temperature (RHSST), temperature, isotopic composition of the ocean and, to a lesser degree, the wind regime at the evaporation site [Merlivat and Jouzel, 1979; Johnsen et al., 1989] (the global closure assumption made in these studies is not crucial for this general conclusion, but can affect the quantitative relationships [cf. Jouzel and Koster, 1996; Armengaud et al., 1998]). Simulations with isotope global circulation models (GCMs), in which parametrization of the evaporation flux is based on the theoretical model described by Merlivat and Jouzel  without global closure, show the same governing factors for d of water vapor over the sea [Jouzel and Koster, 1996]. Furthermore, simple Rayleigh-type isotope models indicate that the information about the evaporation conditions is at least partially conserved along the subsequent trajectory of the vapor until rainout [Johnsen et al., 1989; Petit et al., 1991; Ciais et al., 1995], as supported again by results from GCM simulations [Armengaud et al., 1998]. On the basis of these findings, d excess records from ice cores have widely been used to reconstruct source region conditions of precipitation in past climates, thus providing information that cannot be derived from δ2H or δ18O alone [e.g., Vimeux et al., 1999, 2001, 2002; Uemura et al., 2004; Masson-Delmotte et al., 2005; Jouzel et al., 2007]. However, these studies restrict their analysis on the reconstruction of source temperature, not taking into account the dependence on relative humidity. The authors justify this focus on temperature with the decreasing influence of source RH moving inland in Antarctica as seen in simple Rayleigh-type models, the weak glacial-interglacial changes of RH in GCM simulations and, as the main argument, the linear relationship between sea surface temperature (SST) and RH in the overlying air observed in GCMs [Vimeux et al., 1999, 2001, 2002; Armengaud et al., 1998]. This relationship is used to include the dependence of d from RH in the coefficient of the temperature relation.
1.2. Interpretation of Atmospheric Stable Isotope Data
 There is a large number of observations of stable isotopes in precipitation with a monthly time resolution all over the globe. The largest data collection, the Global Network for Isotopes in Precipitation, is operated by the International Atomic Energy Agency and the World Meteorological Organization [see, e.g., Araguás-Araguás et al., 2000]. The climatological isotope patterns obtained from these measurements are usually interpreted in terms of different isotope effects [Dansgaard, 1964], for example, the latitude and altitude effects. These are not related to single physical processes, but reflect a complex sequence of fractionation during the phase transitions of water from evaporation to precipitation. For instance, isotope distillation via rainout, the temperature dependence of equilibrium fractionation factors and the increasing fraction of solid precipitation with latitude contribute to the poleward decrease of the concentration of heavy isotopes. Since this effect is particularly important for the calibration of the isotope-temperature relationship used for temperature reconstruction from ice cores, a better physical understanding of the individual processes is required to overcome uncertainties resulting from variations of the relation in space and time [Jouzel et al., 1997, 2000; Noone and Simmonds, 2002; Helsen et al., 2007; Sodemann et al., 2008b].
 Water evaporation from the ocean is particularly important in this context. Nonequilibrium fractionation in the evaporation regions can be best explored with the help of isotope measurements in water vapor, because the additional isotopic signal due to condensation can be excluded more easily compared to measurements in precipitation. Another advantage of measuring isotopes in vapor is that it is possible throughout the year, independent of the synoptic situation and not limited to rain days. On the other hand, there are also potential problems related to the measurement of isotopes in water vapor, because technically, the sampling procedure is more difficult than for liquid water. For future applications, new approaches to determine isotope ratios in vapor from remote sensing might be used in addition to surface in situ measurements [see, e.g., Worden et al., 2006; Payne et al., 2007].
 GCMs fitted with water isotope processes have been used to model stable isotope patterns on the global scale [e.g., Joussaume et al., 1984; Hoffmann et al., 1998; Noone and Simmonds, 2002]. These models are able to reproduce climatological patterns of δ2H and δ18O in precipitation, however their representation of d excess is not yet satisfying [Werner et al., 2001; Jouzel et al., 2007]. Also, there are first hints that they do not capture the annual cycle of isotopes in water vapor on local scales [Angert et al., 2008]. In spite of their high suitability to explore stable isotopes on climatological timescales, it is more difficult to use them for simulating single meteorological events (e.g., to compare with measurements from a specific storm), not least because of their coarse spatial resolution.
 Recently, several studies have been conducted that modeled stable isotope ratios in high-latitude precipitation using a combined approach [Helsen et al., 2004, 2006, 2007; Sodemann et al., 2008b]. As a first step, backward trajectories were calculated from the precipitating air mass with the help of (re-)analysis data, representing realistic water transport paths. Parameters like temperature, pressure and specific humidity were extracted from the analysis data along the trajectories and used as input for a Rayleigh-type isotope model, which calculated fractionation along the (in work by Sodemann et al. [2008b] idealized) transport paths. In these studies, no event-based comparison with measured isotope ratios has been performed (except for by Helsen et al. ), owing to the focus on climatological interpretations. Helsen et al. [2004, 2006, 2007] initialized the Rayleigh-type model with three-dimensional, monthly averaged isotope fields from a GCM without explicitly considering moisture source regions. Sodemann et al. [2008b] extracted evaporation regions from the backward trajectories quantitatively. In order to initialize the isotope model, which had the diagnosed moisture source regions as the starting points, GCM isotope fields at the lowest model level have been applied.
 In this study, a backward trajectory diagnostic based on the algorithm of Sodemann et al. [2008a] is used to detect moisture sources for a set of isotope measurements in water vapor in Rehovot (Israel). Unlike Sodemann et al. [2008b], we do not use these diagnosed regions to initialize an isotope model, but try to establish a direct, quantitative link between the measured isotope ratios and meteorological quantities in the source regions. A special focus is laid on the d excess as a measure for nonequilibrium fractionation.
2. Data and Method
2.1. Measurements of Isotopes in Water Vapor
 From 1998 till 2006, measurements of stable isotopes in water vapor were performed at the Weizmann Institute of Science in Rehovot, Israel (31.9°N, 34.8°E, 76 m above sea level). The sampling site was on a rooftop, about 6 m above the ground. Samples were taken roughly twice a week, with some bigger gaps in between. Altogether, 433 measurements were conducted up to July 2006. In this study we only use measurements since October 2000 (265 measurements), because not all ECMWF data necessary for the analysis are available previous to this date (see section 2.2). The water vapor collection time usually was around 8 hours, but there are some exceptional cases with times between 3 and 24 hours. After collection, the water was analyzed with the help of a Finnigan MAT 250 mass spectrometer. The precision of the analysis was 0.1‰ for 18O and 1‰ for 2H. Technical details about the analysis procedure are described by Angert et al. .
 Stable isotopes in precipitation have been studied to a large extent in the entire eastern Mediterranean region [cf. International Atomic Energy Agency, 2005; Lykoudis and Argiriou, 2007] and particularly in Israel [e.g., Rindsberger et al., 1990; Gat et al., 1994]. Compared to other coastal regions, the isotopic composition of rainwater in the Mediterranean, due to the geographical location, is particularly strongly influenced by the interaction of continental air with the sea surface, leading to rather high values of and large variations in d excess [Gat and Carmi, 1970]. Hence, the measurement site is not representative for coastal conditions worldwide, but provides the possibility to explore the diversity of meteorological conditions that lead to a large spectrum of d excess values.
2.2. Moisture Source Diagnostic
 For each day with isotope measurements in water vapor at Rehovot, moisture source regions have been identified. In order to do this, we have used a Lagrangian diagnostic, introduced by Sodemann et al. [2008a], with some minor modifications. The diagnostic is based on the calculation of air parcel trajectories using analysis data. Trajectories have been started at Rehovot and calculated 10 days backward in time. Several meteorological variables have been traced. In order to concentrate on evaporation processes and transport in cloud-free conditions, trajectories have been clipped to exclude clouds or rain from above. The change in specific humidity along the trajectories has been used to detect moisture uptake. In particular, uptake in the oceanic boundary layer can be directly related to evaporation from the sea. Only those measurement days have been taken into account for which sources can be attributed to a substantial fraction of the measured vapor with our method. Finally, a weight has been assigned to each uptake according to its contribution to the final humidity, and different meteorological quantities have been averaged over the uptake regions of a specific date, weighted with the respective contributions. From this, it has been possible to explore statistical relationships between measured isotope ratios and average moisture source conditions. The following paragraphs describe each technical step of the analysis procedure in full detail.
2.2.1. Analysis Data and Principles of Trajectory Calculation
 Three-dimensional, kinematic backward trajectories [Wernli and Davies, 1997] have been calculated, using wind fields and other meteorological variables from operational analysis data of the European Centre for Medium-Range Weather Forecasts (ECMWF). These data are available every 6 hours, with a spectral resolution of T511 and 60 vertical levels until February 2006. Thereafter, the model resolution corresponds to T799 and 91 vertical levels. The data has been interpolated on a horizontal grid with a spacing of 0.75°. Along the trajectories, several variables have been recorded, including specific humidity, temperature, cloud liquid water and ice content, sea surface temperature (SST) at the respective horizontal position and boundary layer height (which is calculated by the boundary layer parametrization scheme of the ECMWF model according to Troen and Mahrt ). In time, linear interpolation between the analysis steps has been applied. The spatial interpolation used to obtain all the variables at a specific trajectory position has been linear in the vertical and bilinear in the horizontal. SST is not defined over the continents, hence there are points close to the shore where a bilinear interpolation is not possible. Thus, the following scheme has been applied at these locations: In the presence of one continental point at one of the four adjacent grid points, the value has been estimated with linear interpolation using the three oceanic points. If there were two continental points, linear interpolation has been applied ignoring the variation of the SST field perpendicular to the direction given by the two remaining oceanic points. Finally, if three of the adjacent grid points have been located over land, the SST at the trajectory location has been set to the value of the fourth oceanic point. This special interpolation scheme is necessary, because otherwise SST cannot be determined for a lot of humidity uptake points close to the continental shore. Furthermore, precipitation has been estimated along the trajectories by averaging the 6-hourly accumulated prognostic ECMWF field (sum of large-scale and convective precipitation) over the respective trajectory section. To this end, forecast steps between 6 and 18 hours have been used from forecasts started at 0000 and 1200 UTC. Because of the relatively low temporal resolution, this averaging provides only approximate values. Also, precipitation estimates from model forecasts have to be handled with care as they are affected by spin-up and forecast errors. Finally, because the vertical position of the trajectories is calculated in pressure coordinates, the pressure at the top of the boundary layer (BLP) has to be calculated. Therefore, the ECMWF boundary layer height (in meters) has been multiplied with a factor of 1.5 and then converted to pressure by linear interpolation in logarithmic pressure. The factor 1.5 has been introduced because of the uncertainties of the boundary layer parametrization and in order to address the fact that a lot of moisture uptake occurs close to the boundary layer top [see Sodemann et al., 2008a].
2.2.2. Specific Trajectory Setup
 Backward trajectories have been started every hour during a measurement period. As horizontal positions of the starting points, the coordinates of Rehovot and four additional points, displaced by 0.25° in longitude or latitude, have been chosen. In the vertical, starting coordinates have been taken as the pressure values of the uneven model levels 1,3,…,15, if they were located within the diagnosed boundary layer at the respective horizontal position and point of time (linear interpolation in time has been applied). The trajectories have been calculated 10 days backward, about the maximum timescale on which integrity of the traced air parcel can be assumed. Calculations have been performed for every measurement date between October 2000 and July 2006, excluding days when rain was detected at the measurement site. These days have been excluded because reevaporation of rain drops might bring moisture into the boundary layer at Rehovot that cannot be diagnosed by our method. (Since 2004, no appropriate precipitation data have been available at Rehovot during the measurement periods. Hence, an explicit exclusion of rain days has only been performed up to December 2003. For later times, rain days have been excluded implicitly, see below.) Altogether, about 50,000 trajectories have been calculated for 238 measurement days with a total measurement duration of circa 1900 hours. As an example, all trajectories for a specific date, the 17 May 2001, are shown in Figure 1a. The air that reached Rehovot at this date originated from Northern Europe. In a northerly flow, very dry air was transported toward the Eastern Mediterranean. After a relatively fast descent to levels primarily below 900 hPa, humidity was taken up over the eastern Mediterranean Sea during the last 2 days of transport.
 Precipitation and cloud formation during transport alter the isotopic composition of the moisture along the trajectories. Because in this study the focus is on relationships between measured isotope values and the moisture source conditions, trajectory segments where these additional processes occur should be excluded. This has been done by detecting, for each trajectory, the first point (starting from Rehovot and going backward in time) where either interpolated precipitation exceeds 1 mm (per 6-hour time step) or the sum of cloud liquid water and ice is larger than 0.1 g/kg. Only the trajectory section in between this point and Rehovot has been kept for further analysis (see Figure 1b for an example). In section 3.3, the influence of this criterion on the results will be explored.
2.2.3. Moisture Uptake
 The key parameter used to diagnose moisture sources along a trajectory is the change of specific humidity q within one trajectory time step,
where (t) denotes the position of the air parcel at time t [Sodemann et al., 2008a]. When this parameter is positive, humidity is taken up by the air parcel. Different types of uptakes can be distinguished; especially important for this study is humidity uptake within the oceanic boundary layer (OBL). This is associated with mixing of the advected air with moisture that has locally evaporated from the sea at the respective position. Turbulent fluxes, not explicitly resolved by the coarse model, are responsible for this mixing. In order to detect such OBL uptake, average values of BLP and air parcel pressure for the respective uptake have been calculated, i.e., averages of the quantities at t and t-6 h. If this averaged air parcel pressure is larger than the respective BLP, the uptake is supposed to be located in the boundary layer. The second criterion for OBL uptake is that at least one of the trajectory positions at t and t-6 h is located over the sea. If both conditions have been fulfilled, the average trajectory position during the 6-hour interval has been considered as a moisture source location. All other uptakes that occur over the continents or above the boundary layer cannot be attributed to a specific oceanic source. The former is related to water recycling, i.e. evaporation from soils and transpiration from plants, the latter can occur, for example, owing to processes like convection, small-scale turbulence and reevaporation of precipitation, which are not resolved by the model, as well as owing to numerical errors and inconsistencies in the analysis data set.
 After detecting all moisture sources along a trajectory, the relative contribution of each source to the final humidity of the air parcel has been determined. A weight has been assigned to each uptake according to this relative contribution. However, the weight of a specific moisture source has been reduced if humidity has been lost later along the trajectory (for details, see again Sodemann et al. [2008a]). A measure for the total amount of water vapor gathered at Rehovot during a measurement period (qfinal) has been estimated by averaging first for each hour the specific humidities at all trajectory starting points and then the hourly values over the entire measurement period (this two-step process is necessary because the boundary layer height and thereby the number of starting points vary in time). The total weighted humidity for a specific date for which uptakes can be identified in the oceanic boundary layer, qobl, has been obtained following the same averaging process. The ratio of these quantities, denoted Ra, corresponds to the fraction of measured water vapor for which sources can be attributed with our method. Figure 1c shows the weighted source regions for the example trajectories shown in Figure 1b. Colors indicate the percentage of final humidity per square kilometer which is taken up at the respective position. For this example, two major uptake spots can be identified, one close to the Turkish coast, the other just before the arrival in Israel.
2.3. Statistical Analysis
 In the following, only those measurement dates are kept for further analysis for which sources have been attributed for at least sixty percent of the water vapor collected at Rehovot, i.e.,
This criterion is fulfilled for only 45 of the 238 days. The main reason for this low percentage exceeding the Ra threshold is the strict exclusion of cloudy trajectory sections. The exclusion of the majority of measurements seems necessary, because reliable statements about the relationship between isotope ratios and moisture source conditions are solely possible if sources can be detected for the greater part of the measured water vapor. In section 3.3 the sensitivity of the results with respect to this exclusion criterion will be investigated. The moisture source regions for all measurement days with Ra > 0.6 are shown in Figure 1d. Humidity uptake takes place almost exclusively over the eastern Mediterranean Sea, corresponding to a transport duration of 2 days or less from the evaporation to the measurement site. A clear domination of a northwesterly vapor transport path, associated with uptake close to the Turkish coast, is also visible from Figure 1d. It should be kept in mind that these uptake regions do not necessarily represent the climatological average of water vapor sources for Israel, but are surely biased by our approach (e.g., the trajectory clipping to avoid clouds and the selection of trajectory starting times solely during measurement periods).
 For the remaining 45 dates, in addition to the variables mentioned above, the following quantities have been determined at the moisture source regions: (1) relative humidity with respect to the temperature at the air parcel position, RH() = q()/qsat(T()), and with respect to saturation at the sea surface, RHSST() = q()/qsat(SST(x,y)) (where T denotes temperature and = (x,y,z) position in space); (2) wind velocity at the air parcel position; (3) evaporation flux (as parameterized by the ECMWF model); (4) 2-m temperature and dew point as obtained from the ECMWF boundary layer parametrization, and derived relative humidity RH2M and relative humidity with respect to sea surface RH2MSST (defined analogously to RHSST); and (5) wind velocity at 10 m height (from the ECMWF boundary layer parametrization). All these variables have been interpolated linearly to the uptake position; that is, they have been averaged between two trajectory time steps. If only one of the trajectory positions at t and t-6 h is located over the sea, the variables have been analyzed at this position (for air parcels that move from land to sea or vice versa within one time step, the uptake over the ocean is assumed to be dominating). Finally, all available variables have been averaged over the humidity sources, weighted with the contribution of the specific uptake to the final humidity at Rehovot (see section 2.2.3). Mean values for each measurement day have been obtained by averaging over the trajectories as described for water vapor in the previous section.
 Correlation analysis and linear regression have been used to explore the statistical relationships between measured isotope values and diagnosed water vapor source conditions. Assuming Gaussian distributions for all variables, we have used the Pearson correlation coefficient (denoted with r) and a least square method to calculate regression lines. Additionally, Spearman rank correlation coefficients, independent of the underlying distribution, have been calculated to ensure universality of the results (in a statistical sense) [see, e.g., Wackerly et al., 2002]. In the following, only Pearson coefficients are specified, because the differences between the two methods appeared to be small. As a second step, multiple linear regression (with two independent variables) has been applied to investigate connections between different combinations of source condition parameters and the isotope measurements.
3. Results and Interpretation
 In the following section, the results of our correlation analysis are described. In the first subsection, correlations of d excess measured at Rehovot with various meteorological parameters determined at the moisture source regions are shown. Additionally, the influence of surface wind velocity is examined, and a definition of an adjusted deuterium excess is presented, which describes nonequilibrium during water evaporation more precisely. In the second subsection, results of the correlation analysis of δ2H and δ18O are shown. Finally, in the third subsection we explore the sensitivity of the results to parameter settings in the analysis.
3.1. Correlation Analysis of d Excess With Meteorological Parameters
3.1.1. Correlation With RH2MSST
Figure 2a shows the measured d excess, plotted against averaged relative humidity at 2 m height above ground, calculated with respect to saturation at the respective sea surface temperature (RH2MSST). The two parameters show a strong negative correlation (r = −0.82, as indicated in Figure 2a). A linear regression leads to the following equation: d = 48.3‰ − 0.53‰/% · RH2MSST. The standard errors for intercept and slope are 3‰ and 0.06‰/%, respectively. This substantial correlation indicates that information from the source region, particularly about relative humidity, is preserved along the water vapor trajectories and can be measured in terms of d.
 A similarly negative, but slightly less significant correlation (r = −0.74, not shown) has been found between d excess and relative humidity with respect to sea surface temperature at the trajectory level (RHSST). The height of the trajectories at humidity uptake, averaged for all uptakes during a measurement day, varied between circa 250 and 1000 m above ground. Hence it is reasonable that the correlation with RHSST is lower than with RH2MSST, because the conditions close to the surface are most important for the isotopic composition of the evaporation flux. Nevertheless, the relation with RHSST is still highly significant, showing that the humidity deficit, which determines d, is a property of the whole air column in the boundary layer and not just the surface layer. According to this result, it is evident that the value of d in the evaporation flux is not only determined by surface quantities, but might also crucially depend on properties of advected air within the boundary layer (as incorporated in simple theoretical models in terms of a constant upper boundary condition, [cf. He and Smith, 1999]). It is important to note that the calculation of relative humidity with respect to saturation at the sea surface is essential for our results. Calculation of RH relative to the local temperature at the trajectory level does not lead to a significant correlation (r = 0.05), showing that the gradient of relative humidity within the air column at the evaporation site is indeed the critical quantity.
3.1.2. Correlation With SST
 In Figure 2b, measured d excess is plotted against averaged sea surface temperature (SST) at humidity uptake. The plot shows that the correlation between these variables is rather weak (r = −0.21). With a multiple statistical model with SST and RH2MSST as independent variables, 70% of the variance in d can be explained, that is only 2% more than with the model based on RH2MSST only. Thus, SST at the humidity source is not the appropriate quantity to explain the variability in measured d excess in this data set. Implications of these results, i.e., the low correlation of d with SST and the strong negative correlation with RH2MSST, are discussed in section 4.
3.1.3. Correlation With Other Quantities
 Correlations with d have also been calculated for all other parameters listed in the previous section and for some derived quantities (see Table 1). The derived quantities include the time that an air parcel was situated within the boundary layer after humidity uptake. The latter potentially has an influence on the isotopic composition of the air parcel by mixing with surrounding water vapor without net moisture uptake. However, this potential influence does not lead to a substantial correlation coefficient. A larger negative correlation is obtained for the average evaporation flux at the uptakes (r = −0.69). This is in agreement with our previous findings, because the flux, as parameterized in the ECMWF model, directly depends on the vertical humidity gradient above sea surface.
Table 1. Correlation Coefficients of Different Diagnosed Moisture Source Parameters With Measured Deuterium Excessa
VEL denotes wind velocity, VEL10M wind velocity at 10 m height. Δp indicates the difference between surface pressure and pressure at the trajectory level. Δq/q is the uptake of specific humidity, divided by the total humidity in the air parcel, and OBL denotes the oceanic boundary layer.
T2M – SST
T – SST
Time in OBL after uptake
 Another quantity that shows a strong negative correlation with d excess is the difference between temperature at a height of 2 m (T2M) and SST (i.e., the temperature gradient in the surface layer) at humidity uptake, as shown in Figure 3a (r = −0.82). The main reason for this correlation is the clear functional relation between this temperature difference and RH2MSST, as indicated in Figure 3b. Physically, this functional relation indicates that humidity uptake generally occurs in relatively cool air compared to SST, for example, in continental air moving over a warm sea surface. This results in a negative temperature difference as shown in Figure 3b on the horizontal axis. (Note as an aside that positive temperature differences at the humidity uptakes do not occur, since they would imply an inversion layer just above the surface, which hampers mixing and humidity uptake by overlying air.) The functional relation between T2M-SST and RH2MSST occurs because RH2M is limited by the saturation vapor pressure at T2M, and therefore, a large difference between T2M and SST implies a low value of RH2MSST, i.e., of RH2M with respect to the warmer sea surface. In contrast, if the temperature of the overlying air is not much lower than the SST, its water vapor capacity is larger, leading on average to higher values of RH2MSST. Besides this physical mechanism, the observed functional relation might partly be generated by the boundary layer parametrization scheme of the ECMWF model (note that both T2M and RH2M are not explicitly calculated model variables). However, the functional dependence shown in Figure 3b is still present if using temperature and RH at the trajectory level instead of the 2 m quantities, corroborating the physical explanation given above. Additionally, the correlation between temperature difference and d (Figure 3a) also occurs if temperature is considered at the trajectory level (r = −0.77). Also T and T2M are both negatively correlated with d, but the absolute values of the coefficients are distinctly smaller than for the temperature differences (T-SST and T2M-SST). These correlations are mainly due to the same physical effect as described above and the fact that SST varies much slower in time than air temperature. This implies that the high-frequency part of the correlations of the temperature differences with d is generated by variations in T2M and T. Nevertheless, the absolute values of the correlation coefficients show that the temperature gradient and by this the humidity gradient in the surface layer is the most important quantity for understanding variations in d.
 In addition, multiple linear statistical models have been specified with d as dependent and all pairwise combinations of the source parameters as two independent variables. The highest values for the explained variance appear in models including RH2MSST, but none of these is substantially larger than 70%; that is, no significant increase in explained variance can be obtained compared to the model based on RH2MSST alone.
3.1.4. Influence of Surface Wind Velocity
 In their model, Merlivat and Jouzel  found that isotopic fractionation during evaporation from the sea depends on the roughness regime of the sea surface, which is in turn related to the mean wind speed at a height of 10 m. They defined a smooth regime for wind speeds below and a rough regime for velocities above 7 m/s. In order to compare this finding with our data, linear models for the dependence of d on RH2MSST have been defined separately for the two regimes, using the diagnosed ECMWF 10-m wind velocity, averaged over all uptake points, as indicator for surface roughness. The slopes of the two regression lines differ significantly (−0.5‰/% for low, −0.73‰/% for high wind speeds, not shown), confirming the dependence of the relation on surface roughness. However, this conclusion is preliminary and might be misleading owing to several reasons: First, by averaging the wind velocity and applying the threshold for the roughness regime afterward, an intrinsically nonlinear process is treated as linear, what might introduce distinct errors. Secondly, there are only 11 measurement days with diagnosed mean wind speeds larger than 7 m/s, and for none of these days RH2MSST is much larger than 60%; hence, this regression is not based on a profound statistic.
3.1.5. Adjusted d Excess
 As already mentioned in section 1.1, d is not an exact measure of nonequilibrium fractionation, because the equilibrium ratio δ2Heq/δ18Oeq slightly depends on temperature. An adjusted deuterium excess dadj, which takes this temperature dependence into account, can be calculated as follows:
Here, T denotes the temperature at the evaporation site and has been taken as the average value of T2M derived from our moisture source diagnostic. Assuming no further fractionation during water vapor transport, dadj measures nonequilibrium fractionation at evaporation exactly. Equilibrium values δ2Heq(T2M) and δ18Oeq(T2M) have been calculated from the fractionation factors given by Majoube , assuming Vienna standard mean ocean water. From this equilibrium values and the measured isotope ratios δ2H and δ18O, dadj has been computed according to (3). Figure 4 shows dadj plotted against mean RH2MSST. The correlation (r = −0.88) is even more significant than for standard d excess (r = −0.82), indicating that the adjusted excess is more directly related to nonequilibrium fractionation, and thereby to RH2MSST, than d. Overall, almost 80% of the variance in dadj can be explained with the variation of mean RH2MSST in the moisture source regions.
3.2. Correlation Analysis of Oxygen and Deuterium Isotopes
 Additionally to the results for d excess shown in the previous section, correlations between the directly measured isotope ratios δ2H and δ18O and the diagnosed moisture source conditions have been calculated. Figure 5 shows these isotope ratios, together with source SST (horizontal axis) and RH2MSST (color). Also shown are linear regression lines for all points and for points with RH2M > 62%, as well as theoretical equilibrium lines, calculated on the basis of equilibrium fractionation factors given by Majoube . The correlation coefficients for δ2H versus SST and RH2MSST are r = 0.47 and r = 0.43, respectively. Also with respect to all other source variables, no correlations with absolute values larger than 0.5 are obtained for δ2H, and the multiple statistical models using two independent variables do not lead to a substantial increase in explained variance. These rather low correlations indicate that the imprint of source conditions on δ2H directly is lower than for d. Equilibrium fractionation during transport, not materially affecting d, can distinctly change the isotope ratio itself. Such fractionation, as mentioned above, can occur, for example, during subgrid-scale cloud formation. Apart from these transport effects, the correlation between source SST and δ2H, which would be one in idealized thermodynamic equilibrium, is diminished by nonequilibrium effects during evaporation. This becomes clear when looking only at points with relatively high source RH2MSST (larger than 62%). For these points, nonequilibrium fractionation is less pronounced (as indicated by the results in section 3.1.1 and discussed further in section 4), and the correlation coefficient between SST and δ2H increases to 0.57. Using d as a more explicit measure for nonequilibrium, the increment in correlation is even a bit larger (r = 0.60 for all points with d < 16‰). However, these increases are statistically not very robust because of the low number of data points in the reduced cases. Despite the relatively low correlation coefficients, the correspondence between both regression lines shown in Figure 5a and the theoretical equilibrium line is surprisingly good. The deviations of points to both sides of this equilibrium line indicate that transport effects dominate in the generation of these deviations (contrasting the increase in correlation described above), because nonequilibrium effects during evaporation solely cause larger depletion than in equilibrium. It is not clear if the balance of these deviations, determining the correspondence of regression and equilibrium lines, is just a coincidence in this data set.
 The correlation coefficient between source SST and δ18O is even a bit lower than for δ2H (r = 0.44), whereas the correlation to RH2MSST is very high (r = 0.81). All other correlations are smaller, and the multiple models do not lead to substantial gains, as described for d in section 3.1. The similarly large correlation of δ18O and d with RH2MSST implies a correspondence between these variables, which is due to the fact that the relative importance of nonequilibrium effects is much larger for δ18O than for δ2H [see Jouzel et al., 2007]. This is manifested in the negative deviation of all points in Figure 5b from the equilibrium line and the decrease of δ18O with increasing RH2MSST, as also indicated by the positioning of the regression lines. Owing to the bad correlation, these regressions can certainly be regarded as rough estimates only. For δ18O, omitting points with low RH2MSST does not lead to a better correlation with SST (r = 0.33 for points with RH2MSST > 62%). Most probable, this is again caused by the strong influence of nonequilibrium fractionation, also for high RH2MSST.
3.3. Sensitivity of the Results to Parameter Settings in the Analysis
 The most important result from the previous paragraphs is the strong correlation between d and RH2MSST. This result is based on moisture uptake regions diagnosed from trajectories that have been cut at the first point, going backward in time, where rain from above or cloud water larger than a given threshold have been detected. The reason for this trajectory clipping is to exclude, as strictly as possible, other fractionation processes apart from evaporation. Now, the influence of this clipping on our correlation result is explored. Figure 6 shows d excess plotted against RH2MSST, identified from trajectories for which this clipping criterion has been moderated (Figure 6a) or totally omitted (Figure 6b). For Figure 6a, clipping has been performed only at rain points; that is, condensation and clouds have been allowed along the trajectories. In this case, there are of course more measurement days for which Ra exceeds 60% than with the full clipping (59 instead of 45). The correlation coefficient remains the same, indicating that cloud formation during transport from the evaporation site to Rehovot has no substantial influence on d. The reason for this is most probably that in the region and altitudes considered here, almost no ice clouds occur and warm clouds are assumed to condense in thermodynamic equilibrium, which does not materially affect d. Figure 6b shows that, if rain is also allowed along the trajectory (116 days with Ra > 0.6), the correlation decreases only slightly. Rain from above can influence d of the trajectory's water vapor in two ways: First, through equilibration the vapor adapts to the isotopic signature of the water that falls into the air parcel from higher levels, which in general has different uptake regions and a different d excess. Secondly, the evaporation of raindrops in unsaturated air itself is a nonequilibrium process, changing d of the vapor. The small reduction of r in Figure 6b, compared to Figures 2a and 6a, shows that the influence of these processes is small, at least for the present data set in the Eastern Mediterranean. Indeed, this reduction might also be caused by the fact that the source regions in the no clipping case are, on average, at a larger distance from Rehovot (not shown), possibly introducing larger uncertainties of the trajectory approach, for example, owing to subscale mixing. Hence, it is not possible to estimate more quantitatively the impact of rain reevaporation in this study.
 Another parameter, in addition to the thresholds for clouds and rain, which has been chosen subjectively is the threshold for Ra, the fraction of collected water vapor of which sources can be attributed. In Figure 7, d is plotted against RH2MSST for the original threshold value of 0.6 and for reduced values, 0.4 and 0.2. The absolute value of the correlation coefficient decreases slightly when adding measurements with lower Ra, but unexpectedly remains large (r = −0.77). Even most of the points with Ra < 0.4 can be found in the same range as the original points with Ra > 0.6. Several reasons for this high correlation of d with source RH2MSST at days with low fractions of attributable water vapor are conceivable. First, our source diagnostic might underestimate the contribution of local moisture sources close to the Israeli coast. Since the algorithm detects local sources also in cases with low Ra, these sources might explain a larger fraction of measured isotope ratios than estimated from the humidity increase along the trajectories. Secondly, the meteorological conditions in the diagnosed uptake regions are possibly representative for larger areas at the respective time instant. Thus, the uptake conditions (e.g., the relative humidity) might be similar at locations where uptake takes place in reality, but which are not identified by our method. All together, the results shown in Figure 7 indicate that the method introduced in this study might also yield useful results in cases with low values of Ra.
 In addition to the correlation coefficients, also the linear regression lines shown in Figures 6a, 6b and 7 do not differ significantly from the original results given in section 3.1. Consequently, the relationship between d and RH2MSST is very robust; it does not depend quantitatively on the clipping criterion and the threshold for Ra.
 Finally, the sensitivity of the correlation results with respect to the degree of sophistication of the source region attribution has been explored. Therefore, as a simpler alternative to using the relative humidity at the exact moisture uptake regions determined with the Lagrangian method described in section 2.2, we averaged RH2MSST over the whole Eastern Mediterranean Sea (defined as all Mediterranean ocean grid points in the ECMWF analysis data east of 24°E and south of 40°N). This region has been assumed to be the main water source region for Israel. For each measurement day, the temporal average of this domain-mean relative humidity has been calculated over 48 hours prior to 0600 UTC at the respective day and correlated with the measured deuterium excess. The correlation coefficient obtained from this analysis is r = −0.69. This correlation is surprisingly high, showing that the distribution of RH2MSST in the Eastern Mediterranean region is rather representative for the respective conditions at the actual evaporation sites of the water transported to Rehovot. For days included in the original analysis (i.e., days with Ra > 0.6), the reason for this is that almost all detected source regions are located over the Eastern Mediterranean Sea (see Figure 1d). In spite of the good results that can be obtained with such a simple (Eulerian) approach, which is based upon averages over a large region that is considered as representative for the actual water origins, a detailed (Lagrangian) analysis of moisture source regions is still meaningful and rewarding owing to the following reasons: (1) The correlation coefficient is substantially larger with the Lagrangian method, showing that detected source conditions represent actual evaporation conditions in a better way. (2) Without a detailed knowledge of source regions, one can hardly be sure whether a subjectively chosen region of average origin adequately represents reality. This might be relatively simple for Israel and the Mediterranean, but for other locations an a priori assumption is less straightforward. (3) Detailed source regions can be used as starting points for Lagrangian modeling in future research, as pointed out in section 5.
 In section 3, we have demonstrated a strong negative correlation between deuterium excess in water vapor measured at Rehovot, Israel and the relative humidity in the water evaporation regions, defined with respect to saturation at the sea surface. This result is physically very plausible, because the importance of nonequilibrium effects during evaporation is basically determined by the humidity gradient above the ocean surface. The strong influence of relative humidity on nonequilibrium fractionation and, by this, d excess is documented in several studies [e.g., Craig and Gordon, 1965; Merlivat and Jouzel, 1979; Gat et al., 2003]. Furthermore, the preservation of source region information along water trajectories in the atmosphere and its measurability in terms of d has also been deduced from Rayleigh type models and isotope GCMs (see section 1.1).
 In the following paragraphs, we discuss some aspects of the results presented in the previous section in more detail. First, we compare our data with other data sets from the literature, addressing the quantitative commensurability of the relationships deduced with the help of our method. Second, some implications of the low correlation of d with source SST found in this study for the interpretation of d as temperature proxy are discussed. Finally, possible error sources of our method are addressed, and an example is given illustrating the possibility of modeling d excess as an application of our results.
4.1. Quantitative Comparison With Other d Excess Data
 Unfortunately, measurements of stable isotopes in water vapor over the sea that can be used to directly validate our results are very rare. Gat et al.  performed such measurements for vapor collected during a cruise in the Mediterranean Sea in winter 1995. In Figure 8, d excess from these measurements is plotted against locally measured RHSST (not taking into account the advection of moisture from other sources in this case) together with the Rehovot data already shown in Figure 2a. Only measurements are included for which no rain occurred in the vicinity of the ship during sampling. It should be noted that the relative humidity applied for the cruise data is based on measurements at mast height, i.e., 27.9 m above the waterline, in contrast to the diagnosed values of RH at 2 m altitude used for the Rehovot data. The red dotted line shown in the plot is a linear regression line for the ship data. Overall, the agreement between the two data sets is surprisingly high. All points lie in the same range, and the regression lines do not differ significantly. Hence, the values of RH2MSST during evaporation, obtained from reanalysis data with the help of our moisture source diagnostic, are quantitatively commensurable to measured RHSST over the ocean. Furthermore, this agreement confirms that the relationships between measured isotope values and diagnosed source parameters can also be interpreted in quantitative terms. The correlation coefficient for the ship data is r = −0.73, i.e., lower than for the Rehovot data set. The reason for this is that for the latter multiple moisture sources are considered which contribute to the local signal by mixing of advected air, whereas for the ship data only the vapor that has evaporated locally, under the measured RHSST conditions, is taken into account. This local evaporate dominates water vapor at a height of circa 20 m over the ocean, but advection of moisture obviously cannot be neglected in this situation.
 The green, dashed line in Figure 8 has been taken from Jouzel and Koster . It shows results from an isotope GCM. For all pure open ocean grid points of the model, average isotope values of water vapor and RH near the surface for the months June, July and August were derived for two years of a last glacial maximum simulation. The line in Figure 8 is the result of a linear regression based on these data. In spite of the different timescale and climate conditions, the modeled relationship is in line with the measurement results discussed before. This might lead to the conclusion that the processes which relate RH and the isotopic ratio of the evaporation flux are properly represented in the GCM. In addition, it gives a first hint that the results found in this study for the Mediterranean might be generalized to global scales. However, these conclusions are very preliminary due to the different timescales used for the analysis of the measured and GCM data (synoptic versus climatological scales).
 Finally, the blue dash-dotted line is the result of a theoretical evaporation model, derived from a global closure assumption (i.e., assuming the identity of the isotopic composition of evaporation flux and water vapor on a global scale) and for a temperature of 25°C. It shows data from Table 1 of Merlivat and Jouzel . A comparison with our data is not totally fair, because assumptions on temperature and surface wind speed were made in this paper, which do not necessarily match local conditions in the Eastern Mediterranean. Furthermore, Merlivat and Jouzel  defined the d excess in a somewhat different way (with slopes in the δ18O−δ2H space slightly departing from 8). However, an offset of these results from the other data in the plot is obvious, corroborating that the global closure assumption can lead to wrong results on the local scale, as already pointed out by Jouzel and Koster .
4.2. Implications for the Interpretation of d as Temperature Proxy
 A positive correlation between d excess and SST observed in GCM simulations is the major fundament of source temperature reconstructions from ice cores. (In fact, source and site temperatures are often estimated in parallel using multivariate statistical models based on d and either δ18O or δ2H; nevertheless, d is assumed to describe the major part of source temperature variability in these models.) However, there is little to no confirmation of such a correlation from measurements. In this context, it might be important that the analysis performed in this study and the ship measurements by Gat et al.  disagree with the GCM simulations at this point: there is a negative and weak correlation between d and SST in our data (see Figure 2b) and a stronger negative correlation in the ship data (r = −0.53, not shown). Also, T and T2M are negatively correlated with d in data presented here. The positive correlation between d and SST observed in GCMs is related to the fact that RH is itself negatively correlated with SST in these models [Jouzel and Koster, 1996; Armengaud et al., 1998]. This correlation in turn is positive in our data (r = 0.42 for RH2MSST against SST), as it is in the ship data (r = 0.31).
 At this point, it is not clear if the parametrization of isotope fractionation or the simulation of mean humidity and temperature fields in the GCMs is responsible for the observed discrepancies. This might be a starting point for future research, comparing correlations between modeled humidity and temperature with observations. Nevertheless, the deviations might also be due to the different timescales and the restriction on moisture uptake points in the Mediterranean in this study and in the study of Gat et al. .
 Of course, the results presented here alone are not sufficient to put the validity of source temperature reconstructions with the help of the d excess into doubt. The main reason for this is the local character of our data set and the different timescales involved. Ice core data are interpreted on scales from decades to glacial-interglacial cycles, while our analysis operates on diurnal scales. The fact that correlations are absent or present on these short timescales does not automatically imply the same for longer scales. There is a need for clarification of these issues within future research, and the present study can serve as a starting point for such an undertaking.
4.3. Modeling d Excess
 Relying on the linear relationship between d and RH2MSST deduced above and the moisture source diagnostic, it is possible to model three-dimensional fields of deuterium excess, also at locations where no measurements are available. This is demonstrated in an example for the Mediterranean region on 18 November 2001, the day with the highest measured d excess at Rehovot for which Ra is larger than 0.6. At 1200 UTC on this day, backward trajectories have been started from grid points of the four lowest model levels of the ECMWF analysis data set in the region from 10°W–40°E, 30°N–50°N with a horizontal grid spacing of 0.75°. For all these horizontal grid points, the moisture source diagnostic has been performed as described in section 2.2, for each point based on the four trajectories from slightly different vertical levels. Average RH2MSST, derived from the moisture sources of each grid point separately, has then been used to calculate d via the linear relationship given in section 3.1. Figure 9 shows the resulting spatial pattern of d on the lowest ECMWF model levels (i.e., at about 980 hPa). Values are only shown for points where Ra is larger than 0.6. A very strong gradient in d is visible from the very east of the Mediterranean (with high values, as measured at Rehovot at this date) to its western part. Such a gradient, albeit less pronounced, can also be found in the climatological pattern of d in precipitation [Lykoudis and Argiriou, 2007]. At the western shore of the Mediterranean sea, close to the Strait of Gibraltar, d increases again to relatively high values (see Figure 9). It is also apparent from Figure 9 that the method can be applied more easily to locations over or close to the sea. For these points, Ra in general is higher. Reducing the threshold for Ra or omitting the trajectory clipping can improve the applicability over continental areas, and both does, according to section 3.3, not substantially lower the accuracy of the method.
4.4. Methodical Error Sources
 Certainly, there are also some drawbacks related to our moisture source diagnostic that might influence the deduced correlations between source parameters and measured isotopic values [cf. Sodemann et al., 2008a]. The Lagrangian scheme applied here assumes coherence of an air parcel on timescales up to 10 days, in accordance with the results of Stohl and Seibert . But of course, the timescale of this coherence depends on various parameters and varies from case to case. Besides, there are many small-scale processes, like convection and turbulence, which are not resolved by the model and might lead to undetected moisture transport. The moisture content of air parcels along a trajectory in the model world is additionally altered by nonphysical factors like numerical diffusion. However, these effects obviously do not affect the source diagnostic strongly enough to prevent the detection of physically reasonable and quantitatively meaningful correlations between source parameters and isotope measurements.
 More specifically, one can ask the question if shortcomings of the method are the reason that potential correlations between d excess and SST, which might exist in nature, are not observed here. However, we think that this is very unlikely because (1) for RH2MSST, a strong correlation has been detected, and (2) the SST field shows less spatial variation than RH. In order to give a more quantitative estimate of the difference in spatiotemporal variability of these fields, for each measurement day the range of source parameter variation that accounts for 90% of the humidity uptake has been calculated. For RH2MSST, the mean value of this daily source variability is 22%, for SST it is 4 K. Relative to the entire range of source parameter variation (i.e., the difference between maximum and minimum of all daily ranges), this average daily variability is 31% for RH2MSST, but only 18% for SST. Hence, it should in fact be easier to detect source SST properly with our method, and errors due to averaging in space and wrong attribution of source regions should be smaller.
5. Conclusions and Outlook
 In order to link isotope measurements at Rehovot with the transport history and particularly the meteorological conditions at the source regions of the respective water vapor, a Lagrangian moisture source diagnostic has been applied. With the help of this diagnostic, it has been possible to establish quantitative relationships between the source conditions and the measured isotope ratios. The significance of these relations has been analyzed by calculating correlation coefficients. Establishing such linkages enhances our understanding of stable isotopes in the atmospheric water cycle in several ways. First, it is a first step toward a quantitative interpretation of isotope measurements on short timescales in terms of their very complex transport and fractionation history. So far, this history was comprised qualitatively in such an interpretation by several studies [Rindsberger et al., 1983; Gat et al., 2003; Anker et al., 2007; Strong et al., 2007]. However, an estimation of the effects of many different processes during water transport requires their quantification. In this context, our work can be seen as a starting point, which can be expanded by integrating processes like cloud formation more directly and exploring other measurement data sets, specifically measurements of isotopes in precipitation. In addition, it might be very interesting to include measurements of δ17O in the analysis, because recent studies showed that the 17O excess (defined with respect to 18O) is also related to moisture evaporation conditions [Barkan and Luz, 2007; Landais et al., 2008]. Secondly, the methodology applied here can be used in the context of Lagrangian modeling of water isotopes. For example, the moisture source diagnostic, together with a parametrization of the isotopic composition of evaporation fluxes, might be used to test predictions of the Craig-Gordon model [Craig and Gordon, 1965]. Thirdly, the diagnosed relationships can, of course, be used for further applications, for example, the interpretation of ice core data or modeling of d excess fields in time and space, as shown in section 4.3. Another interesting application of the relationship might be the determination of water source regions from isotope measurements in terms of footprint analyses. Yamanaka et al.  performed such an analysis for Japan, based on the theoretical relation deduced by Merlivat and Jouzel  with the help of a global closure assumption.
 In principle, connections between the isotopic composition of the evaporation flux and the local meteorological conditions can be measured directly over the ocean. Our method shall not replace these direct measurements, but complements and extends results from such measurements by incorporating transport effects. Furthermore, the paucity of isotope measurements over the sea requires to implement more advanced techniques to investigate the local isotopic composition of the atmosphere there.
 In section 3.1 it has been shown that a major part of the variance of measured deuterium excess can be explained with varying relative humidity at a height of 2 m at the moisture uptake, calculated with respect to saturation at the sea surface. The correlation coefficient between the two parameters is −0.82 and becomes even more significant taking into account the temperature dependence of the equilibrium ratio δ2H/δ18O. The deduced relationship is very robust when varying different thresholds of the diagnostic. Besides, a comparison with similar relations from ship measurements and GCM data shows a very good correspondence. Finally, it has been demonstrated that the relationship can be used to model d in water vapor for regions where no measurements are available with high spatial and temporal resolution.
 In contrast to results from isotope GCMs, no significant correlation between source SST and d excess could be detected in this study. A possibly spurious correlation of these quantities in GCMs might be induced by a likewise questionable anticorrelation of SST and RH in the models. Owing to the fact that source temperature reconstruction from ice cores is basically founded on such correlations observed in GCMs, the work presented here might also have important consequences in this context. However, in this study we have just looked at an isotope data set from a single location, and the global applicability of the results still has to be examined. Furthermore, the timescale of our data set differs substantially from the scales important for ice core interpretation.
 Correlations between meteorological source conditions and the isotope ratios δ2H and δ18O have also been explored. Whereas δ18O has been shown to be mainly determined by nonequilibrium fractionation during evaporation, as it is the case for d, δ2H depends more strongly on equilibrium effects during vapor transport. These transport effects have not explicitly been considered in this study. As mentioned above, this might be another starting point for future research.
 This study was partly funded by the Landesexzellenscluster “GEOcycles” (GEOcycles publication 421). We thank Dan Yakir, Ruth Ben-Mair, and Joel Gat from the Weizmann Institute of Science for providing the isotope data. In particular, we are very grateful to Joel Gat for initiating this study and for his continuous input. Furthermore, we thank Harald Sodemann for many helpful comments on the manuscript. We are grateful to three anonymous reviewers for various constructive comments, improving this manuscript significantly. The German Weather Service DWD is acknowledged for providing access to ECMWF analysis data. Some of the analysis and graphics were produced using the software package R.