Environmental drivers of spatial variation in whole-tree transpiration in an aspen-dominated upland-to-wetland forest gradient



[1] Assumed representative center-of-stand measurements are typical inputs to models that scale forest transpiration to stand and regional extents. These inputs do not consider gradients in transpiration at stand boundaries or along moisture gradients and therefore potentially bias the large-scale estimates. We measured half-hourly sap flux (JS) for 173 trees in a spatially explicit cyclic sampling design across a topographically controlled gradient between a forested wetland and upland forest in northern Wisconsin. Our analyses focused on three dominant species in the site: quaking aspen (Populus tremuloides Michx), speckled alder (Alnus incana (DuRoi) Spreng), and white cedar (Thuja occidentalis L.). Sapwood area (AS) was used to scale JS to whole tree transpiration (EC). Because spatial patterns imply underlying processes, geostatistical analyses were employed to quantify patterns of spatial autocorrelation across the site. A simple Jarvis type model parameterized using a Monte Carlo sampling approach was used to simulate EC (EC−SIM). EC−SIM was compared with observed EC(EC−OBS) and found to reproduce both the temporal trends and spatial variance of canopy transpiration. EC−SIM was then used to examine spatial autocorrelation as a function of environmental drivers. We found no spatial autocorrelation in JS across the gradient from forested wetland to forested upland. EC was spatially autocorrelated and this was attributed to spatial variation in AS which suggests species spatial patterns are important for understanding spatial estimates of transpiration. However, the range of autocorrelation in EC−SIM decreased linearly with increasing vapor pressure deficit, implying that consideration of spatial variation in the sensitivity of canopy stomatal conductance to D is also key to accurately scaling up transpiration in space.

1. Introduction

[2] Hydrologic studies utilizing estimates of forest canopy transpiration or evapotranspiration typically use a mean observed species sap flux value and an estimate of sapwood area per unit ground area to scale to the stand level [Cermak et al., 1995; Ewers et al., 2002; Hatton et al., 1995; Oren et al., 1998; Santiago et al., 2000]. Such an approach assumes that transpiration per tree is spatially well mixed at the stand level, an assumption that has not been explicitly tested. Hydrologic models operating at catchment, regional, and global scales all typically incorporate transpiration estimates [Band, 1993; Band and Moore, 1995; Famiglietti and Wood, 1994; Foley et al., 1996, 2000; Gedney et al., 2006; Running and Coughlan, 1988; Sellers et al., 1997; Wigmosta et al., 1994], and the implications of spatially varying transpiration are particularly pertinent for catchment scale models [Seyfried and Wilcox, 1995]. Advances in model sophistication and incorporation of spatial heterogeneity into hydrologic models has resulted in a demand for complementary spatial data [Grayson et al., 2002].

[3] Advances in statistical methods designed for dealing with spatial data have drawn increased attention to the role of spatial heterogeneity in ecosystem function [Legendre, 1993]. Spatial autocorrelation in areas considered to be environmental gradients are particularly important in this context. Recognizing this functional heterogeneity inherently suggests that it should be included in ecosystem and hydrologic models [Legendre, 1993; Legendre and Legendre, 1998]. In the present study we characterize spatial variation in transpiration per tree in forest stands with the ultimate goal of identifying the underlying mechanisms and incorporating this into a hydrologic model used to estimate transpiration.

[4] Transpiration per unit xylem area (JS) and transpiration per tree (EC), are regulated by a number of biological and environmental variables. It is reasonable to expect some if not all of these variables to vary in space at the stand level. As such, JS, and EC may also be expected to vary spatially as well. A number of studies have reported an observed decrease in JS, and/or EC as a result of declines in soil moisture for a variety of species [Gazal et al., 2006; Lagergren and Lindroth, 2002; Oren and Pataki, 2001; Pataki et al., 2000]. Although these observations deal primarily with temporal variation in soil moisture, JS and EC spatial heterogeneity could be expected as well [Katul et al., 1997], particularly during periods of drought [Granier et al., 2000]. Additional heterogeneity in soil moisture attributable to topography, soil composition, and soil depth has also been shown to cause spatial variation in EC particularly during transitions between wet and dry periods [Granier et al., 2000; Tromp-van Meerveld and McDonnell, 2006]. Biological differences within and between species can also underlie variations in JS, and EC. Changes in JS with stem diameter have been reported in the literature [Oren et al., 1999]. Declines of JS with age, sapwood area (AS) and height attributed to lower hydraulic conductance (KS) have been observed in several species [Alsheimer et al., 1998; Hubbard et al., 1999; Lundblad and Lindroth, 2002; Schafer et al., 2000; Wullschleger and King, 2000]. In addition to spatial variation, short-term (i.e., diurnal) temporal variation in JS, and EC may be expected because of the relationship between stomatal conductance (GS) and sensitivity to vapor pressure deficit (D) [Ewers et al., 2007b].

[5] Our aim in this study is to explore spatial variation in JS and EC across a moisture gradient between a forested upland and a forested wetland using a combination of field and simulation techniques. To achieve this we employed approximately 173 heat dissipation sap flux sensors [Granier, 1987] in northern Wisconsin and used observed values for model parameterization. Geostatistical analysis techniques have become increasingly prevalent in hydrology and ecosystem studies [Gallardo and Covelo, 2005; Grayson et al., 2002; Western et al., 2004] and we have employed these tools to characterize spatial patterns across the site. Simulated EC values were used to analyze temporal trends in spatial autocorrelation as gaps in the observed data prevented robust spatial analyses. We tested the following four hypotheses using both observed and simulated data where applicable: (1) JS is spatially autocorrelated across a moisture gradient from a forested wetland to forested upland; (2) EC is spatially autocorrelated across a moisture gradient from a forested wetland to forested upland; (3) patterns of spatial autocorrelation in EC are driven by environmental variables and so they can be simulated using a simple Jarvis model (equation (3), see below); and (4) patterns of spatial autocorrelation in a) JS and b) EC vary temporally with D.

2. Methods

2.1. Site Description

[6] The study was conducted in northern Wisconsin, near Park Falls (45.9458°N, 90.2723°W). The study site was situated less than 1 km southeast of the WLEF very tall (447 m) tower instrumented to measure fluxes of carbon, water, and energy between the land surface and the atmosphere [Bakwin et al., 1998]. The site and the tower are located within the Chequamegon-Nicolet National Forest, and are both associated with the Chequamegon Ecosystem Atmosphere Study (ChEAS) [Davis et al., 2003]. This area is located within the Northern Highlands physiographic province, which is a southern extension of the Canadian Shield. The bedrock is composed of Precambrian metamorphic and igneous rock, overlain by 8 to 90 m of glacial and glaciofluvial material. Topography of the area is slightly rolling, with an approximate variation in elevation of 3 m across most of the study site. Outwash, pitted outwash, and moraines are the dominant geomorphic features. The growing season is short and the winters are long and cold where mean July and January temperatures are 19°C and −12°C respectively [Fassnacht and Gower, 1997].

[7] The study site is a 120 m 120 m area that captures the transition between a forested upland and forest wetland (Figure 1). The forest is in secondary succession, regenerating from a timber harvest approximately 25 years ago. The upland is dominated by quaking aspen (Populus tremuloides Michx) with balsam fir (Abies balsamea (L.) Mill) also present in the overstory. Speckled alder (Alnus incana (DuRoi) Spreng) and white cedar (Thuja occidentalis L.) dominate the wetland. Aspen, alder, and cedar were identified as the dominant species for sap flux sampling (Table 1).

Figure 1.

Map of the study site showing cyclic sampling design and species sampled for each plot.

Table 1. Number of Sampled (n), and Mean, Total, and Percent Sapwood Area, and Mean Daily Sap Flux per Unit Sapwood Area by Species for Alder, Aspen, and Cedar
Speciesn*Mean AS, cm2Total AS, cm2Mean JS, g cm−2 day−1
  • a

    Values in parentheses are 1 standard error of the mean.

  • *

    n = Total number of individuals instrumented for sap flux sampling.

Alder4122.0 (1.5)880117.0 (8.8)
Aspen7936.8 (1.2)2763122.2 (5.7)
Cedar9123.0 (13.8)1230101.7 (11.0)

2.2. Data Collection

[8] Cyclic sampling was employed to establish plots for sap flux instrumentation [Burrows et al., 2002]. We used a 3/7 cyclic sampling design (Figure 1), where three plots in seven are sampled, and this cycle is repeated. By using such an approach we were able to maximize the spatial information from our site while simultaneously minimizing the number of samples required to generate robust semivariograms. A total of 144 circular plots with a 5 m diameter were established, and the species and diameter at breast height of each tree in every plot was recorded. At least one tree per plot was selected for sap flux sampling. Trees were selected based on species with priority given to aspen, alder, and cedar, respectively. Where more than one member of a species was present the largest tree based upon diameter at breast height (DBH) was chosen. For each tree in each plot we recorded DBH, distance from the tree to the plot center, and the tree's direction from plot center. In plots with multiple species present, trees from each species were sampled with sap flux sensors where resources allowed. A total of 173 trees were instrumented in the site (Table 1). Trees were instrumented for sap flux measurements from 28 July to 6 August 2004 using Granier-type sensors [Granier, 1987], and data from 1, 3–5 August were used for analyses.

[9] Sapwood area (AS) for each tree was calculated from DBH using species-specific allometric relationships established from a study within 10 km of the site [Ewers et al., 2002]. Ewers' study reports a positive correlation between DBH and sapwood depth for all species except cedar, which exhibits a constant sapwood depth. In addition cedar was the only species to display circumferential differences in JS, with slightly higher fluxes observed on the south side of stems. A radial decline in JS was observed between 0–20 mm and 20–40 mm sapwood depth for aspen, and this relationship was independent of DBH and height. No trends are reported for Alder, however the relationship between DBH and sapwood depth differs from all other species that were sampled because there was no heartwood formation. From the results Ewers and colleagues were able to derive relationships between DBH and AS capable of accurately scaling point measurements of JS to EC. We calculated observed transpiration per tree (EC−OBS) with estimates of AS based on the findings of Ewers et al. [2002] using the following equation:

equation image

where JS is transpiration per unit xylem area (g m−2 sec−1) measured within the active sapwood zone, and AS is sapwood area (m2). Note that EC−OBS is transpiration for an individual tree, and not transpiration per unit ground area (ECG) which is often reported. Scaling JS to ECG is accomplished by substituting the ratio of sapwood area to ground area for a plot (AS: AG) for AS in equation (1) [Oren et al., 1998]. To examine ECG would be useful for determining whether net transpirative fluxes per unit ground area vary spatially near stand boundaries. However, it would be difficult to assess whether such variation was attributable to physiological differences between individual plants or rather to stem size and/or densities across a stand. As such we chose to examine transpiration per tree in order to highlight differences in transpiration attributable to physiological variation between individuals. Looking at individuals is crucial for identifying the mechanisms underlying any spatial autocorrelation.

[10] Temperature and relative humidity measurements (Vaisala HMP 45C, Vaisala Oyj, Helsinki, Finland) were made at two thirds canopy height, ∼7 m. Sap flux and environmental measurements were recorded every 30 s (CR10X, Campbell Scientific, Logan, UT, USA) and aggregated to 30 m values. Volumetric surface soil moisture (0–6 cm) estimates for each plot were obtained on July 28th, 30th, and August 5th by averaging three measurements taken at random locations within each plot (Theta Probe, Delta-T, Cambridge, UK). Additional measurements including wind speed, photosynthetically active radiation (Q0), and precipitation were obtained from the nearby Lost Creek eddy flux tower [Cook et al., 2004].

2.3. Model Description

[11] The Terrestrial Regional Ecosystem Exchange Simulator (TREES) [Ewers et al., 2007a, 2008; Mackay et al., 2003a; Samanta et al., 2007] was used to simulate transpiration per tree (EC−SIM). Gaps in EC−OBS because of isolated power and sensor failures necessitated simulating EC so that a data set robust enough to conduct geostatistical analyses at finer temporal scales (i.e., binned by time or hourly D) could be achieved. Total above canopy radiation was partitioned into sun and shade canopy elements with beam, scattered and diffuse radiation components [Spitters et al., 1986], using light extinction methods described by Campbell and Norman [1998]. Beam and diffuse radiation within the canopy were further partitioned into photosynthetically active radiation (Q0) and near infrared radiation for each canopy element so that Q0 could be used in simulating stomatal conductance. Simulated transpiration was calculating in each element with the Penman-Monteith [Monteith, 1965] combination equation to calculate EC−SIM. The canopy is treated as two parallel “big leaves” representing the sun and shade elements, with separate element level stomatal conductances and EC values calculated in each element. EC values for each element were summed to obtain whole canopy transpiration per tree.

[12] Water flow through woody plants is driven by an increasingly negative water potential gradient between the soil and atmosphere at the leaf surface. Water stress at high rates of transpiration in woody plants is typically attributed to hydraulic stress induced by high atmospheric D [Sperry et al., 1998; Tyree and Sperry, 1989]. Stomata close to prevent hydraulic failure at high D in response to increased leaf water potential that is caused by high transpiration rates although the signal is still unknown [Franks, 2004; Mott and Parkhurst, 1991]. Models exist that describe (GS) as a function of environmental factors [Jarvis, 1976], or as a function of photosynthetic carbon uptake [Ball et al., 1987]. GS can also be defined in terms of Darcy's Law [Whitehead and Jarvis, 1981; Whitehead et al., 1984]:

equation image

where GS is average canopy stomatal conductance, KS is whole tree hydraulic conductance, AS is sapwood area, AL is leaf area, ψS is soil water potential, ψL is leaf water potential, h is water column height, ρw is the water density, and g is acceleration due to gravity. Empirical forms of equation (2) are commonly used in conjunction with the Penman-Monteith [Monteith, 1965] equation to estimate transpiration [Bosveld and Bouten, 2001; Ewers et al., 2008; Mackay et al., 2003a, 2003b; Scanlon and Albertson, 2003; Van Wijk et al., 2000]. In TREES each canopy element conductance is treated as a function of turbulent transport defined by combining boundary layer conductance and vapor conductance of the canopy surface in series [Campbell and Norman, 1998], using a Jarvis equation [Jarvis, 1976] to model GS. The Jarvis equation uses a series of multiplicative functions to constrain a theoretical maximum stomatal conductance (Gsmax), and has the following form in TREES:

equation image
equation image
equation image

where Q0 is photosynthetically active radiation (μmol m−2 sec−1). The parameters Gsmax, δ, and α are theoretical maximum stomatal conductance, sensitivity of stomata to D, and absolute sensitivity to Q0 respectively. The EC−SIM values were simulated using a mean species leaf area (AL) as reported by Ewers et al. [2002] for these stand types and described by Oren et al. [1999].

[13] A Monte Carlo sampling approach was used to generate parameter values from uninformed distributions. A comparison between EC−SIM and EC−OBS was assessed using a linear least squares analysis and evaluated using the slope of the best fit line and the index of agreement (IOA) [Willmott, 1982]. For our analyses we first sorted the simulation results by slope, and then simulations within the top 1% according to slope were sorted by IOA to find the best fit model.

equation image

2.4. Data

[14] Individual trees identified as outliers [Adelman et al., 2008] were removed from the data set using the following logic. For each species diurnal values of EC−OBS for each tree were plotted to identify erratic behavior, and/or extreme values that indicated potential sensor issues, or physiological problems (i.e., defoliation or mortality). In addition, the slope and IOA for the corresponding models were examined. Trees were excluded when both slope < 0.80 and IOA < 0.80. These thresholds were chosen because trees with known sensor issues resulting in erroneous data values typically exhibited slope and IOA < 0.80. In addition, one aspen tree on the edge of the site was a remnant from before the site was clear-cut. This individual had a DBH of 31.9 cm, which was large in comparison to the mean DBH of 9.5 cm for all other aspen sampled for sap flux. While sap flux per unit sapwood area for this individual was within the range of values for the site, whole tree sap flux was well above the site average and had a profound affect on the structure of the semivariograms because differences are squared (see below). This individual was ultimately excluded from the analysis on the grounds that it was not a true member of the regenerating stand.

[15] Half-hourly data values of JS, and EC−SIM were sorted according to D, and then aggregated into bins of 0.2 kPa in order to examine trends in spatial variability with this environmental driver. We found this approach to be the more effective than binning by time or Q0 for elucidating temporal trends in spatial patterns. This is likely due to the known relationship between D and GS described by Oren et al. [1999] and general lack of one-to-one relationship between D and time.

2.5. Statistical Analysis

[16] We employed geostatistics to examine our fluxes spatially, using the semivariogram. Semivariance was calculated using the following equation:

equation image

where γ(h) is the semivariance between two points separated by a lag distance (h), (v) is the difference in values between pairs of points (i, j) separated by h, and N(h) = the total number of point pairs separated by h. A semivariogram is created by plotting h on the abscissa, and γ(h) on the ordinate.

[17] Geostatistical analyses were performed with GS+ (version 7, Gamma Design Software, Plainwell, MI, USA). Common semivariogram models assume normally distributed data [Cressie, 1993; Diggle et al., 1998], and so for each variable analyzed using geostatistics we applied one of two available transformations, log or square root, if they resulted in a distribution closer to normal than the untransformed data. Skewness of the distribution was used to gauge normalcy and determine whether transformation yielded an improvement. Spherical models were fit manually to minimize the residual sum of squares, and maximize R2. Differences in the absolute semivariance (i.e., the sill) result from different transformations. Relativized semivariograms were created by relativizing equation (8) to a sill of 1 [Isaaks and Srivasta, 1989] for each semivariogram model. The following equation was used to calculate 95 percent confidence intervals for the semivariograms:

equation image

where γ is the semivariance of a lag class of points with separation distance h, and N is the number of point pairs within the lag class. We analyzed drift to detect spatial trends in mean and variance that would violate the assumption of second order stationarity associated with geostatistics [Legendre and Legendre, 1998]. Where trends were identified appropriate de-trending measures were taken.

[18] Semivariograms are useful for quantifying spatial autocorrelation. However, in order to relate autocorrelation to geographic locations additional tools must be employed. Interpolated maps are typically used for this purpose, and kriging is a method exclusive to geostatistics that is designed to minimize error variance. Kriging produces estimates based on a weighted linear combination of the covariance of nearby sample points, accounting for statistical rather than geographic distance between pairs of points. We use the point kriging method of interpolation to create maps displaying spatial patterns in transpiration. Our maps are intended to be visual aids, and not predictions. All transformed data were back transformed for interpolation.

[19] Regression analysis was performed in Sigmaplot (version 9.01 Systat Software, CA, USA).

3. Results

[20] Soil moisture across the site exhibited strong spatial autocorrelation (Figures 2 and 3) . The soil moisture semivariogram had a range of 90 m (r2 = 0.99) and the structural variance represented 88 percent of the total variance. The semivariograms of mean daily JS revealed no spatial autocorrelation between values across the site (Figure 3). Examination of the semivariograms for AS, mean daily EC−OBS, and mean daily EC−SIM revealed drift, which violated the assumption of second-order stationarity. For data used in these semivariograms the variance increased with increasing separation distance indicating a bimodal distribution. This trend was the result of the differences in mean values of AS between alder, and aspen and cedar, and so we conducted a parallel set of analyses using a detrended data set consisting of only aspen and cedar. Hereafter all analyses conducted using alder, aspen, and cedar are denoted as 3-species, and those using only aspen and cedar are denoted as 2-species.

Figure 2.

Semivariogram of soil moisture (r2 = 0.99) for the study site (a) and corresponding kriged map (b), and a kriged map of AS (c) for the study site. Note that kriged maps in this study are intended to aid interpretation of spatial patterns described by semivariograms, and not as predictions.

Figure 3.

Semivariograms for 2-species sap flux (JS) (a), 2-species sapwood area (AS) (b), 2-species observed transpiration (EC−OBS) (c), 2-species simulated transpiration (EC−SIM) (d), 3-species sap flux (JS) (e), 3-species sapwood area (AS) (f), 3-species observed transpiration (EC−OBS) (g), and 3-species simulated transpiration (EC−SIM) (h). All semivariograms shown are plotted with 95% confidence intervals (equation (8)).

[21] As shown by both 2-species and 3-species semivariograms of AS (Figure 3), there was spatial autocorrelation among the samples. Ranges of 73 m (r2 = 0.73) and 89 m (r2 = 0.85), and proportions of structural variance of 0.50 and 0.49 were observed for 2-species AS and 3-species AS, respectively. Similar results were obtained for both 2-species and 3-species mean daily EC−OBS, and 2-species and 3-species mean daily EC−SIM. For 2-species and 3-species mean daily EC−OBS, respectively, ranges of 72 m (r2 = 0.47) and 118 m (r2 = 0.93) and proportions of structural variance of 0.30 and 0.50 were observed. Semivariograms of 2-species and 3-species mean daily EC−SIM respectively exhibited ranges of 70 m (r2 = 0.59) and 89 m (r2 = 0.81), and proportions of structural variance of 0.37 and 0.43.

[22] Figure 4 shows average EC−OBS and EC−SIM for each species along with D and Q0 for four days in August 2004. EC−SIM matched EC−OBS values reasonably well (Table 2). Mean slope and IOA values for alder, aspen, and cedar were 0.99 and 0.88, 0.98 and 0.90, and 0.97 and 0.89 respectively. Diurnal patterns of EC−OBS and EC−SIM responded to D as expected. A 2 h lag in the response of D to changes in Q0 was observed as well (Figure 4).

Figure 4.

Time series of observed transpiration (EC−OBS) and modeled transpiration (EC−SIM) for alder (a), aspen (b), and cedar (c), and (d) vapor pressure deficit (D) and photosynthetically active radiation (Q0) for 1, 3–5 August 2004.

Table 2. Mean Slope and IOA for EC-SIM by Species
SpeciesMean SlopeMean IOA
  1. a

    Values in parentheses are 1 standard error of the mean. Sample sizes are given in Table 1.

Alder0.99 (0.006)0.88 (0.03)
Aspen0.98 (0.011)0.90 (0.01)
Cedar0.97 (0.007)0.89 (0.01)

[23] Analyses of binned mean hourly EC−SIM exhibited little spatial autocorrelation, and no discernable changes in spatial autocorrelation with time were observed (data not shown). Analyses of EC−SIM binned by Q0 yielded similar results (data not shown). Subsequent analyses of EC−SIM binned by D revealed differences in spatial autocorrelation across bins. Figure 5 shows the actual semivariograms, relativized semivariograms, and corresponding kriged maps for bins of 2-species EC−SIM where mean D = 0.30 kPa, D = 0.92 kPa, and D = 1.51 kPa with ranges of 87 m (r2 = 0.57), 69 m (r2 = 0.53), and 48 m (r2 = 0.44) respectively. Figure 6 shows the actual semivariograms, relativized semivariograms, and corresponding kriged maps for bins of 3-species EC−SIM where mean D = 0.30 kPa, D = 0.92 kPa, and D = 1.51 kPa with ranges of 101 m (r2 = 0.84), 100 m (r2 = 0.88), and 82 m (r2 = 0.81), respectively. The range of EC−SIM spatial autocorrelation was negatively correlated with D for all data (Figure 7). Slopes of −25.7 (r2 = 0.89; p < 0.001), and −24.1 (r2 = 0.78; p < 0.001) were observed for 2-species and 3-species EC−SIM respectively.

Figure 5.

Semivariograms for 2-species EC-SIM binned by (a) D = 0.3 (r2 = 0.57), (b) 0.9 (r2 = 0.53), and (c) 1.5 (r2 = 0.50), and (d) relativized by sill semivariograms for all three bins. Corresponding kriged maps of 2-species EC-SIM where D = 0.3 kPa (e), D = 0.9 kPa (f), D = 1.5 kPa (g). Actual semivariograms shown are plotted with 95% confidence intervals (equation (8)). Note that kriged maps in this study are intended to aid interpretation of spatial patterns described by semivariograms, and not as predictions.

Figure 6.

Actual semivariograms for 3-species EC-SIM binned by (a) D = 0.3 (r2 = 0.84), (b) 0.9 (r2 = 0.88), and (c) 1.5 (r2 = 0.88) and (d) relativized by sill semivariograms for all three bins. Corresponding kriged maps of 3-species EC-SIM where D = 0.3 (e), D = 0.9 (f), D = 1.5 (g). Actual variograms shown are plotted with 95% confidence intervals (equation (8)). Note that kriged maps in this study are intended to aid interpretation of spatial patterns described by semivariograms, and not as predictions.

Figure 7.

Plot of D versus Range (equation (12)) of spatial autocorrelation for 2- and 3-species EC−SIM with regression lines.

4. Discussion

[24] Our results show that spatial autocorrelation exists in forest canopy transpiration across an aspen-dominated upland-to-wetland transition. We rejected our first hypothesis due to the absence of spatial autocorrelation in the semivariograms for 2-species and 3-species JS (Figure 3). Conversely, given the spatial autocorrelation exhibited in semivariograms for 2-species and 3-species EC−OBS and EC−SIM (Figure 3) meant that we failed to reject our second and third hypotheses. Our results indicate that spatial autocorrelation in EC (EC−OBS and EC−SIM) is an effect of scaling JS by AS. Our model was able to effectively simulate transpiration for each species we measured, and replicate the patterns of spatial autocorrelation present in the data. Semivariograms of 2-species and 3-species EC−OBS and EC−SIM binned by D indicated a dynamic pattern of spatial autocorrelation, and therefore we failed to reject hypothesis 4.b. (Hypothesis 4.a was precluded from analysis and therefore rejected because the first hypothesis was rejected.) Interestingly, we found the range of spatial autocorrelation to be negatively correlated with D, indicating that AS is not solely responsible for spatial variation in EC so that scaling in space requires knowledge of responses of key attributes, namely GS, to variation in temporal drivers such as D.

4.1. Spatial Patterns of Js

[25] We expected JS fluxes to vary, both within and between species along the observed gradient in soil moisture across the site. Possibly, this failed to occur because the upland plots, although much drier then the wetland plots, had sufficient moisture so that soil moisture was not limiting transpiration. It should be noted that our soil moisture measurements characterized the upper 6 cm of the soil, which may not be indicative of moisture availability in the entire root zone and probably underestimated it [Wilson et al., 2003]. However, soil moisture rarely limits transpiration in these forests [Desai et al., 2005; Ewers et al., 2002, 2007a; Mackay et al., 2002, 2007].

[26] Characteristics of the species observed in this study likely contributed to the lack of structure in the spatial variation of JS as well. In the wetland, alder and cedar were not water limited as the water table was typically within 1m of the surface as evidenced by piezometer observations across a wide range of environmental conditions in subsequent field seasons. The relative lack of aspen in the wetland (Figure 1) was consistent with its comparative intolerance of excess water [Landhausser et al., 2003]. On the other hand, minimal variation in JS among the upland aspen is consistent with reports that the species is relatively drought tolerant and has exhibited relatively little sensitivity to moisture limitation even in xeric environments [Hogg et al., 2000; Pataki et al., 2000].

4.2. Spatial Patterns of Ec

[27] Following equation (4) and the semivariograms for JS, AS, EC−OBS, and EC−SIM for both 2-species and 3-species (Figure 3) it is clear that spatial patterns of AS explained the spatial patterns of EC−OBS and EC−SIM. A positive correlation between AS and EC may also be inferred from a visual comparison of their respective kriged maps (Figure 2). Our results suggest that traditional estimates of stand transpiration derived using AS from trees sampled within a representative plot to scale mean JS to EC [Oren et al., 1998; Ewers et al., 1999, 2002, 2008; Hatton et al., 1995] should consider the location of the plot. In order for this approach to yield an accurate estimate of stand EC the representative plot would have to be located in an area representative of mean spatial whole-tree EC, i.e., near a stand boundary in our case. Situating a plot near the center of a stand may preclude trees from the lower end of mean EC distribution, particularly where plot sizes are small, and/or the stand is substantially larger than the typical plot size (∼30 m). Although the assumption of spatially well-mixed transpiration per tree does not hold for this site, this scaling method still applies because the spatial pattern of transpiration can be attributed to spatial patterns of AS, which can be inexpensively measured spatially through its allometric relationship with DBH. In water-limited ecosystems where spatial patterns of EC may be driven by spatially variable JS, up-scaling may be more complicated, requiring spatially explicit sampling of JS.

[28] Our results emphasize the importance of considering ecological boundaries. A biological perspective may be appropriate for exploring the spatial pattern of transpiration at this site in the sense that it is the physical size of individuals that dictates EC. In order to understand this spatial pattern it may be necessary to consider environmental constraints exerted by biophysical characteristics throughout the site such as competition between species, and in the case of aspen the possible growth limiting effects of excessive soil moisture in the wetland. Kriged maps in Figure 2 suggest that AS is negatively correlated with soil moisture at our site. Differences in decomposition rates and subsequent nutrient turnover, or simply anoxia alone could cause such differences in growth rates which lead to differences in AS between the upland and wetland. Accounting for boundary gradients characterized by spatial patterns of dominant species may improve regional EC estimates as well, especially as forests such as this one are increasingly fragmented by natural and anthropogenic disturbances [Vitousek et al., 1997].

4.3. Intraspecific Versus Interspecific Species Effects

[29] We found that comparing EC between species violated the assumption of stationarity associated with geostatistical analyses. To overcome this we chose to examine only species with similar EC values. The presence of a trend in the 3-species data set exposes several issues that need to be considered in the context of studies examining spatial transpiration. Understanding changes in processes such as transpiration per tree across ecosystem boundaries is important for scaling up and understanding ecosystem processes. However, species differences that are likely to be of interest may exacerbate the magnitude of variation across such transitions. This is particularly important to consider in the context of the techniques employed here where species differences created additional spatial autocorrelation on a per tree basis. This suggests that a more suitable approach may be to preclude trends by initially examining one species at a time. This is reflected in the focus on aspen in the interpretation of our results. Aspen comprises nearly 90 percent of the 2-species data set, and is a more important species in the region from a hydrological perspective in terms of its abundance and high rates of transpiration [Ewers et al., 2002, 2007a; Mackay et al., 2002]. The spatial extent and sample size for each species prohibited geostatistical analysis for individual species. We retained results from analyses of the 3-species data as support for the 2-species results and to illustrate the effects of sample size on our semivariance confidence intervals.

4.4. Drivers of Spatial Ec

[30] On average, simulations of EC fit the observed data reasonably well. Noticeable deviations occurred during midday periods of high D. Here some spikes in EC−OBS were not reproduced by EC−SIM. It is possible that these spikes were not in response to D or to Q0, but rather a result of some other physiological or physical process such as a delayed stomatal response to D, or canopy self shading [Beadle et al., 1985; Ewers et al., 2007b; Zweifel et al., 2002]. Semivariograms of EC−SIM were smoother than those for EC−OBS for both 2-species and 3-species. EC−OBS data contains gaps, while EC−SIM is continuous throughout the time period and this likely contributes to the smoothing effect exhibited in the analyses. Additionally, the model is driven by both D and Q0, and as such estimates exhibit a strong response to these variables. Our model performs well at times when Q0 typically limits EC during periods of low D in early morning and late afternoon so we can exclude it as an explanation for discrepancies between EC−SIM and EC−OBS. Our results indicate that soil moisture is not likely to offer an explanation. Temperature can be eliminated as an explanation as well because our data was collected on warm summer days with midday temperatures in excess of 20°C on all days used for analyses. Currently no other variables that affect EC and that would allow it to respond to other drivers are incorporated into the model. Additional variables may be required to improve model accuracy. Temporal variation in the proportion of sunlit versus shaded leaves and Q0 due to inter and intracanopy interactions is a possible inclusion that may improve the model.

4.5. Dependence of Spatial Patterns of Ec on D

[31] The general response of EC to D observed in our results agreed with common findings from the literature [Oren et al., 2001]. However, it is unlikely that changes in EC spatial heterogeneity were caused by differences in D among plots. Aspen canopies are typically well coupled with the atmosphere due to their rotating petiole [Hogg and Hurdle, 1997], and atmospheric D should be relatively homogenous above and within the canopy at the scale of this study [Jarvis and McNaughton, 1986] which has been verified in these forests by comparing D between a forested wetland, an aspen stand and above canopy [Ewers et al., 2008]. A plausible explanation is physiological differences manifested in GS. There is tree-to-tree variation in JS, but no autocorrelation could be detected with variogram analysis. However, individual tree responses to increasing rates of EC require that KS drops in response to lower water potentials. This reduces the effective sapwood area while having no effect on JS. Physically this can be described in terms of the response of KS to changes in ψL as EC increases. Potential determinants of sensitivity to water loss rate may include soil textural effects on KS, [Hacke et al., 2000], nutrient impacts [Ewers et al., 2008], competition for light effects on photosynthetic limitations to GS, and topographic position [Franks, 2004; Mott and Parkhurst, 1991; Sperry et al., 1998; Tyree and Sperry, 1989]. Future research is aimed at testing this hypothesis with more detailed measurements and TREES.

[32] Another possible cause for the changes in the range of EC semivariograms are variations in shading within the canopy caused by competition for light between individuals which is probably manifest in the increased variability in EC−OBS during high fluxes (Figure 4). Values of D typically follow a diurnal pattern that peaks after mid day and lower D values typically occur closer to sunrise and sunset when zenith angles are lowest. As such the proportion of sunlit and shaded leaves may exhibit temporal variation accordingly. Concurrently, spatial variation in the proportion of sunlit versus shade leaves as a result of competitive shading likely exists as well. This could be due to either tree height heterogeneity, and/or the presence of remnants from previous stands interspersed throughout the site. Although our model incorporates sunlit and shaded canopy elements, spatially accounting for temporal variation in the proportion of these elements could explain some of the variation of EC range with D [Boulain et al., 2007].

[33] In light of this result, variation in spatial patterns of EC with D over time becomes very important in the context of scaling up to landscape and regional scales. Simply considering spatial patterns may not be sufficient to calculate accurately scaled estimates of EC. A more sophisticated scaling approach that involves a temporal consideration of environmental drivers and subsequent effects on spatial patterns of EC may be required. The ability of the model to capture such temporal variation in spatial patterns (Figures 5 and 6) suggests that we are close to accomplishing this.

4.6. Implications for Landscape Scale Ec

[34] Our results suggest that it is necessary to consider landscape heterogeneity in relation to EC. It may no longer be sufficient to rely on general vegetation classification and a few key environmental variables to create adequate estimates of EC or evapotranspiration (ET). By incorporating atmospheric and edaphic drivers in up-scaling efforts, and doing so in a spatial context, more accurate estimates will be achieved. Ecological gradients and stand boundaries are becoming more easily identifiable with increasingly available high resolution remote sensing data. General spatial patterns could be captured by including edaphically controlled drivers such as soil texture, moisture, or nutrients. Spatially explicit estimates could be further refined by including atmospheric drivers such as D that effect spatial patterns temporally. Ideally, further investigation will elucidate links between EC and these spatial drivers, with the ultimate goal of improving predictive understanding by discovering the mechanisms behind spatially varying transpiration.

5. Conclusions

[35] Using geostatistics we have shown that transpiration per tree varies spatially across an aspen dominated upland-to-wetland transition in northern Wisconsin. Spatial variation was not exhibited in transpiration per unit xylem area, but was observed in sapwood area. As such it appears that spatial patterns for individual species are the key to obtaining more accurate estimates of stand transpiration, and that future efforts to characterize such patterns may rely on simple species relationships with environmental drivers rather than geospatial techniques. This suggests that scaling to the stand level may be accomplished using traditional methods if the representative plot is located in an area that captures mean spatial EC, despite the apparent invalidity of the assumption that transpiration is homogenous at the stand level. We have shown that a simple model is capable of replicating species specific spatial patterns observed in field data. Spatial patterns of whole-tree transpiration changed temporally with D, indicating that scaling efforts should consider spatial and temporal heterogeneity in relation to D. This demonstrates that models relying on soil moisture gradients as primary limiting factors of spatial transpiration are insufficient at this particular site, underscoring the importance of seeking alternative ecohydrological controls on water fluxes and not just soil water in forest ecosystems, which in many regions are not typically water-limited. On the contrary, we suggest that excessively high soil moisture near and within the wetland had a long-term effect of limiting tree growth and thus only in the long-term contributed to spatial autocorrelation in EC.


[36] Funding for this study was from the National Science Foundation, Hydrologic Sciences Program (EAR-0405306 to D.S.M., EAR-0405381 to B.E.E., and EAR-0405318 to E.L.K.). In addition, M.L. was supported by the NSF IGERT program. The statements made in this manuscript reflect the views of the authors and do not necessary reflect the views of NSF. We would also like to thank personnel at Kemp Natural Resource Station in Woodruff, Wisconsin for logistical support.