Environmental drivers of evapotranspiration in a shrub wetland and an upland forest in northern Wisconsin

Authors


Abstract

[1] To improve our predictive understanding of daily total evapotranspiration (ET), we quantified the differential impact of environmental drivers, radiation (Q), and vapor pressure deficit (D) in a wetland and upland forest. Latent heat fluxes were measured using eddy covariance techniques, and data from four growing seasons were used to test for (1) environmental drivers of ET between the sites, (2) interannual differences in ET responses to environmental drivers, and (3) changes in ET responses to environmental drivers between the leaf expansion period and midsummer. Two simple ET models derived from coupling theory, one radiation-based model, and another using mass transfer were used to examine the mechanisms underlying the drivers of ET. During summer months, ET from the wetland was driven primarily by Q, whereas it was driven by D in the upland. During the leaf expansion period in the upland forest the dominant driver was Q. ET from the wetland was linearly related to net radiation using coupling coefficients ranging from a low of 0.3–0.6 to a high of 1.0 between early May and midsummer. Interannually, ET from the upland forest exhibited near linear responses to D, with an effective reference canopy stomatal conductance varying from 1 to 5 mm s−1. The results show that ET predictions in northern Wisconsin and other mixed wetland-upland forests need to consider both wetland and upland forest processes. Furthermore, leaf phenology effects on ET represent a knowledge gap in our understanding of seasonal environmental drivers.

1. Introduction

[2] In mesic forested regions, evapotranspiration during the growing season can represent a large fraction of the precipitation received during the same period. As such, it is a crucial mechanism for both water and energy cycling, and a significant source of uncertainty in making predictions in watersheds lacking instrumentation [Sivapalan et al., 2003]. Evapotranspiration (ET) has two primary environmental drivers, radiation (Q) and atmospheric vapor pressure deficit (D). Q and D are inputs to the Penman-Monteith (P-M) combination equation [Monteith, 1965], which is routinely incorporated into large-scale models [Aber and Federer, 1992; Band et al., 1993; Famiglietti and Wood, 1994; Wigmosta et al., 1994; Vertessy et al., 1996; Sellers et al., 1997; Foley et al., 2000]. The relative importance of Q and D as drivers of ET in landscapes containing both upland forests and bottomland wetlands has generally not been considered, and yet these represent potentially large sources of nonlinearity in the emergent landscape ET. Models that are sensitive to these nonlinearities can potentially be used to fill in data gaps or as substitutes for lack of data in remote regions and will begin to provide predictive understanding of the interaction between ET and climate change. In this paper we analyze the role of Q and D controls over ET from multiple growing seasons of eddy covariance and micrometeorological data from a wetland and an upland forest.

[3] ET from vegetated land surfaces can be predicted from environmental drivers using the P-M combination equation:

equation image

where s is the rate of change of saturation vapor pressure with temperature [kPa °C−1], Rn is net absorbed radiation [W m−2], G is ground heat flux [W m−2], ρa is air density [kg m−3], cp is the specific heat canopy of air [J kg−1 °C−1], Ga [ms−1] is aerodynamic conductance, D is vapor pressure deficit [kPa], ρw is density of water [kg m−3], λ is latent heat of vaporization [J kg−1], γ is the psychrometric constant [kPa °C−1], and Gv is a combination of leaf boundary layer (Gb) and canopy stomatal conductance (Gc):

equation image

where Gc = GS * L, L is leaf area [m2 m−2], and GS is canopy average leaf level stomatal conductance. The response of GS to governing variables has previously been quantified by Jarvis [1976]:

equation image

where GSmax is maximum GS, Qo is photosynthetic photon flux density, TA is air temperature, and θ is soil moisture. Formulations such as equation (3) are not strictly mechanistic, but they highlight potentially important drivers of GS and consequently ET. Equations (1) and (3) use Rn and D as direct drivers of ET and as modifiers of stomatal control. TA affects GS through its role on leaf photosynthetic activity. θ affects root water availability and is thus a proxy variable for detecting water stress. Other factors could be included, including soil temperature, which influences root activity and consequently plant water uptake, especially at low temperatures or when soils are freezing.

[4] A more mechanistic view of controls on GS follows from plant hydraulic theory. As a result of atmospheric dryness and/or high photosynthetic rates, woody plants experience water stress at high transpiration rates due to hydraulic limitations to water transport from the roots to the leaves [Tyree and Sperry, 1988; Sperry et al., 1998]. Transpiration rates that exceed the plant's ability to transport water to the leaves cause leaf water content or potentials (ΨL) to decrease to a point where stomates close to prevent runaway cavitation. Although transpiration rates respond to D, stomata respond more directly to ΨL or transpiration rate rather than D [Mott and Parkhurst, 1991]. Available evidence suggests that plants regulate transpiration via changes in ΨL resulting from whole plant water status [Meinzer and Grantz, 1991; Saliendra et al., 1995; Cochard et al., 1996; Nardini et al., 1996; Salleo et al., 2000; Ewers et al., 2000; Franks, 2004]. Oren et al. [1999] showed that

equation image

where GSref is reference canopy stomatal conductance defined at D = 1 kPa, and −m is the rate of change of GS with respect to ln(D). The ratio of −m to GSref has been shown using plant hydraulic theory to be between 0.54 and 0.6 when plants are isohydric and regulating minimum leaf water potential to prevent excessive cavitation [Oren et al., 1999; Ewers et al., 2005]. GSref incorporates GSmax and all other limits to stomatal conductance other than D. Consequently, an analysis of drivers of ET should consider all these variables.

[5] Previous researchers have examined a variety of models for estimating ET in wetlands [Drexler et al., 2004; Rosenberry et al., 2004] and have found no ideal model. However, radiation-based methods outperformed mass transfer methods when the mass transfer component was relatively small and thus had amplified uncertainty associated with it [Rosenberry et al., 2004]. Assuming G is negligible with respect to daily total energy flux [Amiro and Wuschke, 1987], and following the 1983 work of McNaughton and Jarvis (as discussed by Monteith and Unsworth [1990]), equation (1) can be rewritten in the form

equation image

where Ω is a coupling coefficient, defined as equation image , which varies between 0 and 1. The adiabatic term in equation (5) accounts for a lack of equilibrium between the state of the atmosphere at a reference height and the state of the evaporating surface through D. When the atmosphere and evaporating surface are in equilibrium then D approaches 0, and the adiabatic term becomes negligible. For a sufficiently large wetland surface equation (5) can be simplified to

equation image

Note that Ω varies with the ratio of aerodynamic to canopy conductance, and so it retains the physical meaning of these conductances. A low-stature shrub wetland might be considered decoupled or only weakly coupled, in which case Ω tends toward 1 [Monteith and Unsworth, 1990]. Since Ω declines with canopy conductance it also retains the physical meaning of Gv, or some counterpart such as soil vapor conductance.

[6] For tall stature, closed forest canopies with a high Ga and low Ω, we assume that the mass transfer contribution to equation (5) is relatively large in comparison to the radiation contribution [Jarvis and McNaughton, 1986]. If this is the case for our upland forest we can simplify equation (5) to an upland evapotranspiration model (ETU):

equation image

in which ETU is proportional to D and scaled by Gv. When plants are regulating minimum leaf water potential to prevent excessive cavitation (isohydric plants), the relationship between ETU and D will saturate or even decline at higher D [Jarvis, 1980; Pataki et al., 2000; Ewers et al., 2005].

[7] Boreal forests, mixed forests bordering the south shores of the Great Lakes and along the Appalachian Mountains in North America, and similar forests of this climatic regime consist of a continuum of moisture regimes from lowland forested and shrub wetland to mesic, upland forests [Baldocchi et al., 1997; Fassnacht and Gower, 1997]. Northern Wisconsin forests lie at the boundary between northern temperate and boreal systems. Their moisture regimes are largely determined by glacial deposits that produced a topography consisting of wetlands surrounding low-rising upland forests at 10–20 m elevation above the wetland areas. This subtle topographic relief has produced, in a first-order analysis, a bimodal pattern consisting of wetland and upland forests, with a deterministic length scale [Seyfried and Wilcox, 1995] on the order of 102 m [Burrows et al., 2002]. Sparsely vegetated wetlands represent one end-member in which Q is potentially the key driver of ET, while for closed canopy upland forests that are aerodynamically rough the more important driver may be D.

[8] Given the above theory we tested four hypotheses related to daily ET. (1) ET from the wetland site is driven by variations in solar radiation (i.e., ET = ETW), while ET responds primarily to D in the upland hardwood stand (i.e., ET = ETU). (2) The primary environmental driver of ET at either site is invariant among years. (3) Because of phenological changes in leaf area, evapotranspiration in the upland varies between radiation driven and vapor pressure deficit driven. (4) The primary environmental driver of ET does not change in the wetland during the growing season. To test these hypotheses, we present four growing seasons of eddy covariance latent heat flux data collected on towers situated in a wetland and an upland forest in northern Wisconsin. We then determine the key driver, Q or D, of flux data for each tower year, and on the basis of this analysis we apply the appropriate model to help explain observed intersite, interannual, and interseasonal variations in ET.

2. Methods and Materials

2.1. Site Descriptions

[9] We made eddy covariance and micrometeorological measurements in two ecosystem types within the Chequamegon-Nicolet National Forest in northern Wisconsin. The area is situated in the Northern Highlands geographic province of Wisconsin. The bedrock is composed of Precambrian metamorphic and igneous rock and overlain by 8 to 90 m of glacial and glaciofluvial material deposited approximately 10 to 12 kyr before present. Topography is slightly rolling with a range of 45 m within the defined study area. Outwash, pitted outwash, and moraines are the dominant geomorphic landforms. Annual precipitation (1971–2000 average) is about 800 mm, with mean January and July temperatures of −12°C and 19°C, respectively [Burrows et al., 2002]. There is no marked dry season, which ensures that even the upland systems in the region are well watered [Ewers et al., 2002; Cook et al., 2004].

[10] The Lost Creek (hereafter called wetland) AmeriFlux site is located at 46°4.9′N, 89°58.7′W. Vegetation is a shrub height of 1–2 m, consisting of an overstory of speckled alder (Alnus regosa) and willow (Salix spp), and an understory of sedge (Carex, spp.). Soils consist of poorly drained Totagatic-Bowstring-Ausable complex and Seelyeville and Markey mucks formed on outwash sand, and are composed primarily of sapric material about 0.5 m thick [Natural Resources Conservation Service, 2006]. The Willow Creek (hereafter called upland) AmeriFlux site is a mature, second-growth hardwood forest about 70 years old, which is located at 45°48.47′N, 90°04.72′W. Dominant overstory species at this site are sugar maple (Acer saccharum Marsh), basswood (Tilia Americana L.), and green ash (Fraxinus pennsylvanica Marsh), with an average canopy height of approximately 24 m. Leaf area index of 5.3 m2 m−2 was measured [Desai et al., 2005] at the site during the period of flux measurement reported here. Soils consist of sandy loam overlying coarse glacial till. A detailed description of the site is given by Cook et al. [2004].

2.2. Flux and Environmental Measurements

[11] Three-axis sonic anemometers (Campbell Scientific Inc., Logan, Utah, Model CSAT) and closed path infrared gas analyzers (Li-Cor Inc., Lincoln, Nebraska, model LI-6262) were deployed above canopy at a height of 10.2 m in the wetland and 29.6 m in the upland forest. Continuous measurements have been made since mid-1999 and late 2000 in the upland and wetland, respectively. Latent heat fluxes were calculated using established methods [Berger et al., 2001; Cook et al., 2004]. A detailed discussion of these calculations, spectral corrections, storage fluxes, screening for instrument error and low friction velocity, and quality control per an AmeriFlux relocatable reference system for the upland site are given by Cook et al. [2004]. The same methods were applied at the wetland site.

[12] Basic micrometeorological measurements, including air temperature (Ta) and relative humidity, precipitation, irradiance, Rn, and surface soil temperature (TS) were made at each tower [Cook et al., 2004]. TS was taken at the soil surface in all years except in 2004 when the wetland peat subsided by approximately 20 cm. Continuous measurements of water table height (ZW) were made in the wetland using a submerged pressure transducer (Omega Engineering, Stamford, CT, model PX242-005G). Measurements of soil moisture (θ) at 5 cm below the soil surface were made in the upland with a horizontally installed water content reflectometer probe (Campbell Scientific, Logan Utah, model CS615). For the purpose of our analysis we focused on the period from 1 May to 10 September. Data for the upland forest was available for years 2000 to 2003, and for the wetland data was available for 2001 to 2004. Growing season precipitation for all years was within one standard deviation of the 30 year average (48 ± 12 cm) at the National Climatic Data Center station in Minocqua, WI. Among the years in this study, 2000 and 2002 were relatively wet (53.8 cm and 53.0 cm, respectively), 2001 was near average (46.6 cm), and 2003 and 2004 were relatively dry (38.0 and 36.0 cm, respectively). Table 1 summarizes the precipitation and temperatures for May and for June–August for each site.

Table 1. Measured Precipitation and Temperatures for the Study Sitesa
YearWetland Precipitation, mmForest Precipitation, mmTemperature, °C
MayJJAMayJJAMayJJA
  • a

    JJA is sum of June, July, and August precipitation or average of June, July, and August temperatures.

2000  5427912.817.0
2001882189821012.318.6
2002167260983237.817.7
200314418510619511.217.7
200474170832189.215.5

[13] Our criteria for selecting days for analysis were as follows. Days in which rainfall exceeding 5 mm was recorded between 6 pm of the previous day and 6 pm of the current day were not considered. When a day was missing more than one consecutive midday (8 am to 4 pm) half-hourly flux measurement it was not used. We did not adjust this range to account for changes in day length, as even midsummer before 8 a.m. and after 4 p.m. latent heat fluxes were relatively small in comparison to midday fluxes. More than one consecutive missing observation was accepted before 8 a.m. and after 4 p.m. when latent heat fluxes were on average less than 15% of the fluxes during the midday period and therefore not expected to contribute a large amount of error to the daily sum of evapotranspiration. Single missing observations during midday were corrected using mean diurnal variations [Falge et al., 2001] when multiple days with similar light and vapor pressure deficit (VPD) conditions were available. When data from similar light and VPD conditions were not available to fill in gaps we replaced the missing observation with the average of the fluxes from one observation prior to and one observation following the missing observation. This approach was used only when increases or decreases in light levels or VPD over the averaging time period were less than 20%.

[14] Half-hourly latent heat flux (LE; W m−2) was converted to water depth equivalent (mm) flux footprint evapotranspiration as follows:

equation image

where λ is the latent heat of vaporization calculated as a function of air temperature at the respective tower measurement heights on a half-hourly basis. Daily total tower evapotranspiration fluxes, ED, were aggregated from half-hourly E obtained over daylight hours as follows:

equation image

where b and e respectively refer to time at the beginning and end of day, adjusting for day length changes from early May through mid-September. We adjusted b and e so that they delimited a daylight period. For the remainder of the analysis ED will refer just to these aggregated measurements of ET.

[15] Rn was measured at the top of the upland and wetland towers using CNR1 radiometers (Kipp and Zonan Inc.), and Q was measured at the top of the upland and wetland towers using a CNR1 radiometer and silicon pyranometer (Li-Cor, Lincoln, NE, model LI-200X), respectively. D was calculated from relative humidity and air temperature measurements [Goff and Gratch, 1946] obtained just below the top of the canopy in the upland forest and about 8 m above the vegetation in the wetland. Where necessary we gap-filled values of D to ensure a more complete data set. However, the use of gap-filled values of D did not influence gap filling of E, since in the cases where D was gap-filled we relied only upon observations of Q and not D at the respective towers to guide gap-filling of E. It should be noted that a large number of such gap-filled values can potentially bias the analysis of E in response to D. However, the number of such gaps was small and values of D among sites were very similar [Mackay et al., 2002]. Gaps were filled using linear fits to the AmeriFlux WLEF tower (30 m above ground) [Davis et al., 2003], to the upland tower in the case of the wetland, to the wetland tower for the upland, to four micrometeorological stations located in red pine, alder, mixed species, and aspen stands (at 1.5 m above ground) near the WLEF tower [Cook et al., 2004; Mackay et al., 2002], or to diurnal average values obtained at each respective tower. To determine daily mean D we retained only days in which either the maximum recorded half-hourly D exceeded 0.6 KPa [Ewers and Oren, 2000], or the daily average D (DD) was at least 0.1 kPa [Phillips and Oren, 1998]. This screening process generally eliminated only days that were immediately preceded by nighttime rainfall, and it was applied at both the wetland and upland sites using the same thresholds of D. This conservative approach ensured that our analysis was not based on days with very low D which tend to correspond with erroneous flux measurements [Ewers and Oren, 2000]. DD was determined as an average of the half-hourly D values and daily Q (QD) as the sum of half-hourly radiation values (W m−2 30 min−1) between times b and e. In addition, we determined daily average TA, TS, and ZW in the wetland or θ in the upland. The daily variables were used for statistical analyses, but all calculations using equations (1)(5) were made half-hourly and summed to daily.

2.3. Seasonal Definitions

[16] For each site year we divided the flux values into two groups. The first group (spring) spans a period from preleaf out (1 May) to about mid-June. By mid-June full leaf expansion has generally taken place in northern Wisconsin. The end date was determined partly by breaks in the flux data, with the constraint that the same data was used for each year for a flux tower. For Willow Creek we used 9 June as the leaf out period end date, while for Lost Creek we used 14 June. The second group (summer) extends from the end of spring to about 10 September in any given year.

2.4. Statistical Analysis

[17] Statistical analyses were performed using SAS (version 9.1, SAS Institute, Cary, NC, USA); Proc Reg was used for stepwise multivariate regression. Linear and nonlinear curve fits were performed in SigmaPlot (version 9.01, Systat Software Inc., Richmond, CA, USA). Curve fits were performed on individual groups of flux measurements and then Student's T was used to test for differences in slopes and intercepts among groups.

[18] From equations (2), (4), and (7) it is clear there is nonlinear response of ETU to D, which can be closely approximated by an exponential rise to a maximum (i.e., ETUa · (1 − exp(−bD), where a and b are fitting parameters [Ewers et al., 2005]). We anticipated that a number of other factors may preclude detecting a nonlinear response of E to D when measured from eddy covariance data. A relatively large, free, unrestricted evaporation source in the flux footprint would demonstrate a linear response of evaporation to D, which could mask or even hide the hydraulically limited signal (equation (4)) of the trees. Also, a set of observations made over a narrow and low range of D can produce a near-linear response of ET to D because ΨL will not be low enough to trigger stomatal closure.

2.5. Modeling Analysis

[19] As further evaluation of the environmental controls on evapotranspiration, we applied equation (6) to the 1 May to 10 September periods for each of the four years of Lost Creek data. Ω was adjusted weekly or when there were gaps in ETW to minimize bias in ETW versus ED. We chose not to adjust Ω at shorter intervals to reduce the amount by which we were subjecting the fitting procedure to noise in the micrometeorological and flux data.

[20] To evaluate the drivers for the upland forest we employed equation (7). We calculated Gb = 0.025 ms−1 assuming an average leaf width of 0.06 m and mean sunlit hours wind speeds [Campbell and Norman, 1998]. Although boundary conductance varies with wind speed and leaf display this variation contributes little to total conductance in comparison to variations in stomatal conductance. We adjusted GSref (reference canopy stomatal conductance; equation (4)) among weekly intervals or where there were extended breaks in ED. It should be noted that this adjustment of GSref also accounts for actual changes in L, which would occur through leaf phenology as well as interannual changes in leaf area. The variability of Gv partly reflects changes in both L and GSref, but there was insufficient leaf area data to adequately separate the effects of both variables and so we adjusted only GSref. Among years we adjusted the value of m (sensitivity of stomata to the rate of water loss; equation (4)), which has the effect of adjusting the curvature of the relationship between ETU and D. Since equation (7) assumes a fully coupled canopy (i.e., Ω = 0) we tested the validity of this assumption by inverting the full ETPM formulation (equation (1)) to estimate GV and then solved for Ω in equation (5) at Ga = 0.2 m s−1. To avoid declines in Ω due to water stress, which would falsely imply stronger coupling [Monteith and Unsworth, 1990], we limited this analysis to well water conditions between 1 May and 31 July. We also compared GSref values derived using equation (1) to the values derived using equation (7) for both spring and summer periods.

3. Results

3.1. Hypothesis 1: Environmental Drivers

[21] Overall energy balance was 72% at both the upland [Cook et al., 2004] and wetland sites. Some researchers suggest that flux calculations should be corrected on the basis of energy budget errors [Twine et al., 2000]. Such correction was not attempted as it was difficult to confirm that the energy imbalances were not partially due to errors in estimating available energy [Wilson et al., 2002; Mahrt, 1998; Cook et al., 2004]. Moreover, there was no guarantee that the latent and sensible heat fluxes had the same fetch. There was no indication at either site that energy closure was correlated with meteorological conditions, and so we suppose that the relationships between fluxes and drivers would only change in flux amplitude, not shape.

[22] Table 2 summarizes the number of flux days used for the statistical analysis. A small number of data gaps were due to the instruments being off-line. These gaps ranged from 2–4 weeks in length, depending upon when technicians could visit these relatively remote sites (5 hour drive from Minneapolis, Minnesota) to make instrument repairs. Shorter gaps of typically 1–5 days were due to our criteria for selecting days as outlined in section 2.2. With the exception of instrument failure there was generally a balanced sampling of days within and among years at both sites.

Table 2. Number of Days, by Year and Period, When Flux Data Were Used for the Analysesa
 YearMayJuneJulyAugust/September
  • a

    Footnotes b–f indicate where gaps in the data are due to instrument failure. The remaining missing days are due to a relatively short period when meteorological conditions precluded using the flux data.

  • b

    No data available 21 June to 19 July.

  • c

    No data available 21 July to 10 August.

  • d

    No data available 8–30 July.

  • e

    No data available 18 June to 4 July.

  • f

    No data available 13–29 July.

Wetland20012016b533
Wetland2002222217c13
Wetland200316231825
Wetland200419192321
Upland200015132024
Upland200113126d17
Upland2002010e1437
Upland200317176f13

[23] The results of a stepwise multiple regression using QD, DD, TA, TS, (ZW or θ), QD * DD, and DD * DD as predictors of ED is summarized in Table 3. An additional variable, Julian day (Jday), was included in the multivariate analysis to rule out the possibility that additional changes in the system were occurring during the analysis periods. Since we have a comprehensive set of environmental variables covered already, Jday can be thought as a proxy for changes in leaf area or at least effective leaf area over time [Samanta et al., 2007].

Table 3. Variance Explained in ED by Environmental Drivers, Incoming Solar Radiation (QD), and Vapor Pressure Deficit (DD)a
YearLost CreekWillow Creek
SpringSummerSpringSummer
QDDDQDDDQDDDQDDD
  • a

    Numbers in bold indicate the most significant driver of ED for the respective year, site, and season. A dash indicates no data.

  • b

    Variable was not significant (P > 0.1).

2000----0.510.21b0.780.87
20010.860.790.790.720.530.26b0.43b0.81
20020.750.750.840.76--0.660.76
20030.37b0.610.690.670.38b0.40b0.340.79
20040.600.600.890.65----

[24] In all cases the quadratic terms either did not significantly (P > 0.10) explain variance in ED or the variance explained was at most 1%. During the summer periods in all years in the wetland the most significant driver of the variance in ED was QD. Correlation with DD was indirectly related to correlation with QD, with DD explaining only an additional 2–7% of variance in ED. During 2003, TA and TS explained 9 and 6%, respectively, of the variance in ED (P < 0.0001). However, for the spring period there was no consistent most significant predictor of ED, with QD dominating in 2001 and 2004 and DD dominating in 2002 and 2003.

[25] For the upland forest most of the summer ED was best explained as a response to DD, with less than 6% of the variance explained by adding in QD. TS, TA, and θ were significant (P < 0.08) in 2002, but each contributed less than 2% of the variance. QD was the dominant driver of spring ED in years 2000 and 2001. We could not completely rule out TS and θ during the spring period in the upland site. TS was significant (P < 0.0001) in 2003 and contributed 51% of the variance, and θ was significant (P < 0.07) in 2001 and explained 9% of the variance in ED during the spring period.

3.2. Hypothesis 2: Interannual Variability

[26] Linear fits for the most significant drivers of ED in the wetland are shown in Figure 1. The apparent interannual variability in ED in response to QD was negligible among years 2001, 2002, and 2003 (P > 0.2 in all combinations). However, the response in 2004 was significantly different from the responses in the other years (P < 0.1).

Figure 1.

Response of ED measured from eddy covariance in the wetland to daily above-canopy radiation (QD) and vapor pressure deficit (DD) for years (a, b) 2001, (c, d) 2002, (e, f) 2003, and (g, h) 2004. All regressions are linear fits.

[27] Figure 2 shows linear fits of ED to QD and saturating nonlinear fits to DD for the upland forest site. Although saturating curves explained slightly more of the variation in ED than linear fits, this difference amounted to at most 2% of the total variation. However, an examination of the residuals among the linear and nonlinear fits showed a better fit with the saturating fits, which had both small mean residuals and more constant variance. Among linear fits of ED versus DD significant interannual differences were found between 2003 and other years (P < 0.03).

Figure 2.

Response of ED measured from eddy covariance in the upland forest to above-canopy radiation (QD) and vapor pressure deficit (DD) for years (a, b) 2000, (c, d) 2001, (e, f) 2002, and (g, h) 2003. Curves in Figures 2a, 2c, 2e, and 2g are linear fits. Curves in Figures 2b, 2d, 2f, and 2h are exponential saturation curves of the form, Y = a(1−exp(−bX)).

3.3. Hypotheses 3 and 4: Seasonal Variability

[28] As hypothesized, the driver of upland ED changed from QD in the spring to DD in the summer (Table 3). To compare curves we tested for significant differences in slopes among the same environmental drivers. With respect to responses to QD there were significant seasonal differences in years 2000 (P < 0.001) and 2001 (P < 0.025), but not in year 2003 (P > 0.2). With respect to DD there were significant (P < 0.001) seasonal differences in slopes for all three years.

[29] The significant driver of spring wetland ED changed among years, with QD being most important in 2001 and 2004, and DD driving the flux in 2002 and 2003. There were also no consistent patterns in terms of the absolute values of explanatory variables during the spring period, but TS explained 9% and 13%, respectively, of the variance in ED during the 2003 and 2004 (P = 0.002) spring period. We also could not rule out Jday, our proxy for phenology, which was significant (P = 0.0009) in 2001 and 2003, and explained 15% of the variance in 2003. ZW was generally not significant, except in 2002 (P = 0.03) when it explained 3% of the variance in ED.

3.4. Modeling Evaluation of Environmental Drivers

[30] Comparisons of modeled ETW versus measured ED are shown in Figure 3. The predicted evapotranspiration closely matched the observations in terms of high degree of fit and low bias, although there was slight overestimation of low fluxes and underestimation of high fluxes in 2001 and 2002. Figures 4 and 5show the values for the coupling coefficient, Ω, with Ta and ZW, respectively. Midsummer values of Ω generally varied from 0.8 to 1.0, but were lower in the spring and late summer. The lower spring values (0.3 to 0.6) closely followed air temperature, with the lowest values occurring in 2004 during an unusually cool May with mean daily temperatures only slightly above freezing. Values for Ω decreased during periods of water table drawdown (Figure 5), especially in the late summer periods of 2003 and 2004.

Figure 3.

ETW modeled on the basis of equation (3) versus ED from eddy covariance in the wetland. Shown are linear regressions with dashed lines representing the 95% confidence intervals. Solid lines are one-to-one relationships.

Figure 4.

Adjustments to the coupling coefficient over time for each season at the wetland site. Also shown is air temperature at the site.

Figure 5.

Adjustments to the coupling coefficient over time for each season at the wetland site. Also shown is water table height at the site.

[31] The results for the upland forest are shown in Figure 6. Good fits were achieved between the modeled and measured ETU for each year, but there was a significant bias at low flux in 2002. The sensitivity of stomatal conductance to the rate of water loss (m) was 0.6·GSref in 2000 and 2001, 0.5·GSref in 2002, and 0.54·GSref in 2003. These values are within the range reported for a variety of woody species including northern hardwoods [Oren et al., 1999; Ewers et al., 2001; Wullschleger et al., 2002; Addington et al., 2004; Ewers et al., 2005, 2007b]. Figures 7 and 8show how GSref varies in comparison to Ta and θ, respectively. GSref was lowest during May and generally peaked in July. This trend was consistent with leaf phenology during May and early June, during which time the increasing GSref reflected increases in L as well as increases in stomatal conductance. The continued increase into July could not be explained from the data at the Willow Creek site. GSref had a lower peak in 2003, which correlated with a steady decline in surface soil moisture (Figure 8). When we inverted equation (1) to obtain GV we obtained a mean Ω = 0.14 with 90% of the values falling between 0.01 and 0.27. GSref derived using equation (7) was 1.6 times as large as GSref derived using equation (1) during summer months, and 2.4 times as large during spring.

Figure 6.

ETH modeled on the basis of equation (4) versus ED from eddy covariance in the upland forest. Shown are linear regressions, with dashed lines representing the 95% confidence intervals. Also shown are one-to-one lines.

Figure 7.

Variability in parameterized GSref over time for each season at the upland site. Also shown is air temperature at the site.

Figure 8.

Variability in parameterized GSref over time for each season at the upland site. Also shown is soil moisture in the top 5 cm of the soil at the site.

4. Discussion

4.1. Hypothesis 1: Environmental Drivers of ET

[32] Wetlands and upland forests represent end-members along edaphic gradients in northern Wisconsin, with transpiration dominating the evaporative signal in upland forests and soil evaporation dominating in wetlands. Hydrologic models for these types of systems may utilize just a single ET formulation, a single environmental driver (e.g., radiation, vapor pressure deficit, temperature), or a full combination method to estimate ET without knowing which environmental driver is dominant. We used four hypotheses to better understand the nonlinearities associated with these different environmental drivers in northern Wisconsin. Our first hypothesis, ET in the wetland and upland forests is driven by variations in Q and D, respectively, was not rejected. The statistical analysis and simulations with equation (6) support the claim that the wetland ET is driven primarily by Q, as has been demonstrated in other studies [Drexler et al., 2004; Rosenberry et al., 2004].

[33] The statistical analysis and simulations with equation (7) support the claim that upland ET is driven primarily by D. The mean value of Ω(=0.14) supports the assumption of strong coupling for the upland stand [Jarvis and McNaughton, 1986]. However, by assuming fully coupled (Ω = 0) instead of fully P-M conditions, GSref was forced to compensate by increasing by a factor of 1.6 during the summer months and a factor of 2.4 during spring. This is due to the weaker ET response to changes in GS at higher values of Ω [McNaughton and Jarvis, 1991]. When GSref is determined using ETPM then the values we obtain are similar to values reported for other sugar maple stands in northern Wisconsin and the upper peninsula of Michigan [Mackay et al., 2003; Ewers et al., 2007a, 2007b]. We note that an analysis using just equation (1) would mask the effects of drivers seen here. Moreover, an analysis with equation (5) requires simultaneous adjustments to partially correlated parameters, Ω and GSref, which potentially masks the relationships found here.

4.2. Hypothesis 2: Interannual Variability

[34] Our second hypothesis, that the primary driver at each site is invariant among years, was also supported by the available data. The Ω values for the wetland generally varied from about 0.6 to 1.0 in response to variations in soil surface temperature, water table depth, and potentially leaf phenology. This variability is consistent with changes in surface conductance [Monteith and Unsworth, 1990; L'Homme, 1997]. At peak water table levels (positive ZW in Figure 5) or even shallow depths to the water table the surface soil may be considered above field capacity. One issue to consider further is that our wetland site is not a true well watered bare soil, an irrigated crop or grassland, a bog, or a true forest. With its shrub height vegetation this system is perhaps more analogous to a rice paddy than these other systems. Gao et al. [2003] showed that rice growth affected aerodynamic roughness properties, but not energy partitioning patterns. Although our system is not a bog it shares some of the vascular plant characteristics reported by Lafleur et al. [2005], who show weak relationships between ET and water table height in a system with a wider range of water table fluctuations than observed here. In our shrub wetland we could be seeing differences in the timing and rates of sedge, willow and alder leaf expansion among years, which would impact both aerodynamic and stomatal conductances. Further research, especially seasonal dynamics of leaf area and physiology, is needed to determine how leaf phenology may contribute to differences in Ω in such shrub wetlands.

[35] It is encouraging that with the exception of the dry 2003 the maximum GSref among years varied by only about 1 mm s−1, suggesting that at least under optimal conditions this parameter is robust. Maximum GSref among years appears to be related to surface soil moisture, with the drier conditions in 2003 having the lowest peak GSref. Ewers et al. [2001] showed that GSref also declined with decreasing soil moisture in Pinus taeda while retaining the same ratio between −m and GSref. Another possible interpretation of the relationship between GSref and soil moisture is that soil evaporation is contributing significantly to the midsummer peak flux, although it is generally relatively small under closed forest canopies [Baldocchi et al., 2000] as has been shown in other sugar maple systems in northern Wisconsin [Mackay et al., 2002]. We cannot rule out the possibility of additional sources of moisture. In particular, there is a wetland that is sometimes within the fetch of the Willow Creek tower depending on wind direction [Cook et al., 2004; Desai et al., 2005]. As such, it is possible that the GSref values here reflect the flux contributions from the adjacent wetland. Further analysis of the flux footprint is needed to test this hypothesis.

[36] The peak ET in 2001 was smaller than that for 2000 and 2002. During June 2001, a widespread outbreak of tent caterpillars defoliated aspen and partially defoliated some other upland stands, including the Willow Creek stand. Reduction in GSref in 2001 is at least partly attributable to a reduction in L associated with the defoliation without a compensating effect from remaining foliage [Pataki et al., 1998; Ewers et al., 2007b]. In addition, volumetric soil moisture in the lower half of the rooting zone declined nearly monotonically from 0.30 m2 m−2 in early July to 0.24 m2 m−2 by late August [Cook et al., 2004], whereas soil moisture remained above 0.30 m2 m−2 for the whole 2000 growing season. The decline in rooting zone soil moisture during 2001 likely contributed to a small decline in GSref. Carryover effects of the defoliation and subsequent use of resources to grow new leaves in the same summer on reduced radial growth increment both during summer 2001 and early spring growth in 2002 could explain why total reference canopy stomatal conductance in 2002 did not recover to year 2000 levels.

[37] Year 2003 was the driest summer of the study period, and volumetric soil moisture in the top 100 cm declined steadily from an average of 0.30 m2 m−2 in early July to 0.20 m2 m−2 by late August. This drop in soil moisture can represent a 0.1 to 0.2 MPa [Clapp and Hornberger, 1978] decline in soil water potential, which may partly account for the lower ETU response to D in comparison to the other years (Figure 2). This result is consistent with Ewers et al. [2007a] who showed a 25% decline in 2003 sugar maple GSref in comparison to 2002 GSref derived from inverting a canopy model driven by sap flux data inputs, although they found soil moisture to be a significant but small factor in this decline.

[38] The variability in m from 0.5 to 0.6 times GSref is well within that expected given the relatively short range of D over which most of the flux data values are distributed in some years, and is not likely attributed to changes in plant function. Even under a host of conditions affecting the interannual variability of GSref for a variety of northern Wisconsin species m was approximately equal to 0.6·GSref [Ewers et al., 2007b]. One possible source of variability in m is that nonstomatal sources of water are included within the flux signal. Although the near linear response of ED to DD (Figure 2) gives the appearance of nonstomatal sources of water, such linearity is also not found in species regulating leaf water potential. Ewers et al. [2005] found that old black spruce exhibited linear flux responses to D and Ogle and Reynolds [2002] found similar results in desert shrubs due to a lack of minimum leaf water potential regulation. If one adds to this open water sources and bryophytes without stomata, then systems such as the northern Wisconsin forests are especially complicated. A way around this problem is to distinguish these linearly responding systems from the nonlinear ones, by making a more thorough mapping of the component fluxes along moisture gradients.

4.3. Hypotheses 3 and 4: Seasonal Variability

[39] Our third hypothesis, that upland forest ET is driven by Q during spring and D in the summer due to phenological changes was not rejected. During the spring phenology period upland ET was explained by radiation. Three potential factors during phenological changes could explain the response at the upland site. First, with a more open canopy prior to leaf expansion a greater proportion of the total flux is expected to occur from below the canopy as a response to a relatively larger penetration of radiation to the forest floor [Baldocchi et al., 2000]. Second, leaf budburst follows shortly after snowmelt, and so surface soil moisture content is high. During summer months soil evaporation in the hardwood stands comprises a relatively small (<10%) proportion of total evapotranspiration within the Chequamegon forest [Mackay et al., 2002] and in other studies [Moore et al., 1996; Kelliher et al., 1995; Wilson et al., 2000]. A third potential factor is that GSref during the period of leaf expansion may have been limited by low daytime air temperatures (Figure 7), low soil temperatures, nighttime freezing, or by limited development of gas exchange and photosynthetic capacity [Gratani and Ghia, 2002].

[40] To achieve a good fit between simulated and observed ET at Willow Creek required us to make adjustments to GSref. Seasonal variations in GSref generally followed leaf phenology for the region rather than tracking air temperature during the May–June period. An improved understanding of leaf phenology for the region would therefore reduce the variability of GSref. However, an explanation for why the increasing trend continued well into middle to late summer remains elusive. Ewers et al. [2007b] and Mackay et al. [2003] found a similar unexplained trend in sap flux data at another sugar maple stand in northern Wisconsin.

[41] Our fourth hypothesis, the primary environmental driver of ET at the wetland site is always Q, was rejected because ET was better explained by D in two of the springtime periods. However, there was insufficient data or evidence of changes in environmental conditions to fully explain why D was the dominant spring flux driver in 2002 and 2003. DD may be statistically the more significant driver during these periods due to the relatively small sample sizes, 22 and 16 days, respectively. Moreover, evaporation was equally correlated to Q and D in 2002. This underscores the importance of recognizing that if DD is the dominant driver, QD may also be strongly correlated. It also shows a need for information on how evapotranspiration from contrasting wetland and upland sites responds to changing environmental conditions during the spring-to-summer transition period.

[42] It is clear that the upland flux during May is lower than the JJA fluxes and seasonal wetland fluxes, suggesting that an improved understanding of leaf phenology and its effects on evaporative response to environmental drivers in the upland forests is needed. While measurements of eddy covariance have shown that water loss and carbon uptake at the stand level are correlated with leaf phenology [Goulden et al., 1996; Granier et al., 2000], these types of correlations have not been extensively tested and mechanistic connections have not been established [Turner et al., 2003]. There is physiological evidence that such correlations between water and carbon fluxes and leaf phenology are not robust. For instance, Gratani and Ghia [2002] found increases in stomatal conductance and photosynthesis through the leaf expansion period. Furthermore, the susceptibility of larger xylem conduits to cavitation from freezing is well known [Sperry, 1995], but the repair of freezing induced cavitation is poorly understood [Hacke and Sperry, 2001]. Given the importance of phenology in regional and global-scale modeling [Myneni et al., 1997; Schwartz, 1998; Menzel and Fabian, 1999; Schwartz and Reiter, 2000], and the limited ability of current models to predict phenology [Botta et al., 2000], this represents an important area of ecohydrologic research.

5. Conclusions

[43] Our results show that estimates of forest evapotranspiration in northern glaciated and similar systems should be made using an understanding of both wetland and upland processes, and their responses to environmental conditions. Evapotranspiration fluxes from our two end-member sites representing upland closed forest and short stature vegetation-dominated wetland were more sensitive to vapor pressure deficit and radiation, respectively, during summer months. Our analyses also show that these results are primarily dependent on season, and the key drivers can change between the leaf expansion period and summer. Evapotranspiration fluxes from our upland forest responded to radiation during the leaf expansion periods, whereas the key driver for the wetland site varied among years during the same periods. The results of this study show that when trying to determine evapotranspiration across a wetland-upland mosaic landscape, it is important to select models that are sensitive to the key drivers of evapotranspiration across the span of environmental conditions from upland to wetland sites.

Acknowledgments

[44] The authors are grateful to three anonymous reviewers for their input on this manuscript. This research was partially supported by the NSF Hydrologic Sciences Program through grants EAR-0405306 to D.S.M. and EAR-0405381 to B.E.E.. Flux tower research was funded in part by the National Institute for Global Environmental Change through the U.S. Department of Energy (DOE). Any opinions, findings, and conclusions or recommendations herein are those of the authors and do not necessarily reflect the view of DOE.

Ancillary