Separating the effects of climate and vegetation on evapotranspiration along a successional chronosequence in the southeastern U.S. Authors: Stoy, Paul C., Gabriel G. Katul, Mario B. S. Siqueira, Jehn-Yih Juang, Kimberly A. Novick, Heather R. McCarthy, A. Christopher Oishi, Joshua M. Uebelherr, Hyun-Seok Kim, and Ram Oren This is the peer reviewed version of the following article: see citation below, which has been published in final form at https://doi.org/10.1111/j.1365-2486.2006.01244.x. This article may be used for non-commercial purposes in accordance with Wiley Terms and Conditions for Self- Archiving. Stoy, Paul C., Gabriel G. Katul, Mario B. S. Siqueira, Jehn-Yih Juang, Kimberly A. Novick, Heather R. McCarthy, A. Christopher Oishi, Joshua M. Uebelherr, Hyun-Seok Kim, and Ram Oren (2006) Separating the effects of climate and vegetation on evapotranspiration along a successional chronosequence in the southeastern U.S. Global Change Biology 12: 2115-2135. DOI: 10.1111/j.1365-2486.2006.01244.x. Made available through Montana State University’s ScholarWorks scholarworks.montana.edu Abstract We combined Eddy-covariance measurements with a linear perturbation analysis to isolate the relative contribution of physical and biological drivers on evapotranspiration (ET) in three ecosystems representing two end-members and an intermediate stage of a successional gradient in the southeastern US (SE). The study ecosystems, an abandoned agricultural field [old field (OF)], an early successional planted pine forest (PP), and a late-successional hardwood forest (HW), exhibited differential sensitivity to the wide range of climatic and hydrologic conditions encountered over the 4-year measurement period, which included mild and severe droughts and an ice storm. ET and modeled transpiration differed by as much as 190 and 270mmyr1, respectively, between years for a given ecosystem. Soil water supply, rather than atmospheric demand, was the principal external driver of interannual ET differences. ET at OF was sensitive to climatic variability, and results showed that decreased leaf area index (L) under mild and severe drought conditions reduced growing season (GS) ET (ETGS) by ca. 80mm compared with a year with normal precipitation. Under wet conditions, higher intrinsic stomatal conductance (gs) increased ETGS by 50mm. ET at PP was generally larger than the other ecosystems and was highly sensitive to climate; a 50mm decrease in ETGS due to the loss of L from an ice storm equaled the increase in ET from high precipitation during a wet year. In contrast, ET at HW was relatively insensitive to climatic variability. Results suggest that recent management trends toward increasing the land-cover area of PP-type ecosystems in the SE may increase the sensitivity of ET to climatic variability. Keywords: Eddy covariance, evapotranspiration, oak-hickory forest, old field, Pinus taeda Received 28 November 2005; revised version received 5 June 2006 and accepted 11 June 2006 cially as predictions of vegetative and climatic changes, including extreme events, become more spatially and temporally refined (Houghton et al., 2001; Wear & Greis, 2002; Katz et al., 2005). For example, predictive skill will increase with improved understanding of the mechan- isms responsible for hurricanes (Xie et al., 2002, 2005) and ice storms (Ramos da Silva et al., 2006), and the El Nin˜o events (Hoerling & Kumar, 2003) that are asso- ciated with summer drought in the southeastern US (SE, Peters et al., 2003). Studies are now exploring the implications of these exogenous events on the differen- tial loss of productivity within various SE ecosystems (McNulty, 2002; McCarthy et al., 2006); however, their impact on the differential changes in water cycling in general and ET in particular has received much less attention, the subject of this work. Separating the effects of climate and vegetation on evapotranspiration along a successional chronosequence in the southeastern US PAUL C . S TOY, GA B R I E L G . KATUL , MAR I O B . S . S I QU E I R A , J E HN - Y I H J UANG , K IMB E R LY A . NOV I CK , HEATH E R R . M C CARTHY, A . CHR I S TO PHE R O I SH I , J O SHUA M . UE B E LHER R , HYUN - S EO K K I M and RAM OREN Nicholas School of the Environment and Earth Sciences, Duke University, Box 90328, Durham, NC 27707–0328, USA Introduction Evapotranspiration (ET) is controlled by external (i.e. physical, climactic) and internal (i.e. biological) drivers, both of which are projected to change on multiple spatial and temporal scales due to the coupled effects of climate change and anthropogenic ecosystem man- agement (Pielke et al., 1998; Houghton et al., 2001; Wear & Greis, 2002; Foley et al., 2003). Thus, understanding the relative roles of climate vs. vegetation on ET is critical for predicting how water cycling will respond to future physical and biological perturbations, espe- 0.22million km2, with concomitant declines in upland hardwood forested area from 0.27 to 0.23million km2 and agricultural land from 0.40 to 0.26million km2 (Wear & Greis, 2002). If ET in these three vegetation types shows distinctly different rates and responses to climatic forcing, such a dramatic change in vegetation cover could substantially impact surface fluxes and water cycling across the SE, including the triggers of convective rainfall (Juang et al., 2006). In a first-order analysis, ET might be considered a conservative quantity (Roberts, 1983; Gholz and Clark, 2002), especially in wet temperate climates if it is limited by energy availability irrespective of ecosystem type. For example, Gholz and Clark (2002) found that ET was controlled by climate rather than ecosystem type along a chronosequence of slash pine (P. elliottii Englm.); net radiation (Rn) explained more than 80% of the variability in LE. However, one might also expect to find some modulations in ET and especially its compo- nent fluxes – transpiration (T) and evaporation (E) – due to soil-plant hydraulics. In particular, T should vary among ecosystems comprised of different species with Fig. 1 A color-inverted infrared aerial photograph of the three study ecosystems in the Blackwood Divison of Duke Forest near Durham, NC. Measurement towers are separated by less than 800m. The clearcut to the south of HW lies on private land. Fluxes dominated by the signature of the clearcut are excluded from this analysis. The objective of this study is to isolate the contribu- tion of vegetation from that of climate and soils on controlling ET at three adjacent SE Piedmont ecosys- tems. We measured ET using the Eddy-covariance (EC) technique for 4 years at old field (OF), early succes- sional planted pine forest (PP), and late-successional oak-hickory forest (HW) ecosystems (Fig. 1). These ecosystems represent a typical postagricultural succes- sionary sequence in the SE, are adjacent with towers sufficiently separated such that their flux footprints rarely overlap, and experience similar climatic and edaphic conditions. Thus, any interecosystem differ- ences in ET in response to common external drivers can be attributed to the effects of vegetation rather than climate and soils. Furthermore, a common shallow rooting depth of ca. 35 cm ensures that ET is controlled by the recent precipitation (P) signal without confound- ing effects from groundwater. The study ecosystem types are common on the land- scape, but land cover is rapidly changing. For example, over the next 40 years, the area of land in pine planta- tion in the SE is projected to increase from 0.13 to differences in drought sensitivitiy and canopy morphol- ogy (Lai & Katul, 2000; Oren & Pataki, 2001; Pataki & Oren, 2003) in response to available water, radiation, and vapor pressure deficit (i.e. atmospheric coupling, Jarvis & McNaughton, 1986, but see Roberts, 1983). However, the magnitudes of these responses over long terms are unknown. We address the study objective in two ways. First, we quantify the influence of vegetation on the hydrologic and energy budgets, focusing on net differences in ET in response to mild and severe droughts, a wet year following an ice storm and a year with normal precipi- tation. Next, we analyze interecosystem sensitivity to the specific mechanisms responsible for variability in ET at the growing season (GS) time scale (ETGS, see Table 1 for abbreviations). We show that ETGS is a linear function of the product of the physical drivers P and vapor pressure deficit (D), and the biological drivers leaf area index (L), and intrinsic stomatal conductance (gs). Thus, we can assess the sensitivity of the ecosys- tems to these drivers through a linear perturbation analysis, which we discuss after presenting the experi- mental setup and GS and annual sums of ET and its components. We finish by discussing some broader implications of the experimental findings on water resources in the SE. Materials and methods Site description EC measurements of ET and associated environmental drivers were collected from 2001 through 2004 at OF, PP and HW ecosystems in the Blackwood Division of the Duke Forest near Durham, NC (35158041.43000N, 79105039.08700W, 163m a.s.l. – see Fig. 1). The three study ecosystems model a typical SE ecological succession from OF to PP to HW (Oosting, 1942), and represent dominant ecosystem types in the SE. The ecosystems lie adjacent to one another on Enon silt loam, a low fertility Hapludalf typical of the SE US Piedmont, with a transition to Iredell gravelly loam toward HW and the northern part of OF (Pataki & Oren, 2003). EC measure- ment towers lie within 800m of each other (Fig. 1). An impervious clay pan underlies the research sites at ca. 35 cm belowground (Oren et al., 1998; Lai & Katul, 2000) thereby imposing similar constraints on root-water access for all three ecosystems. Long-term mean annual Ta and P are 15.5 1C and 1145mm, respectively. Detailed characteristics of the ecosystems can be found elsewhere (Ellsworth et al., 1995; Oren et al., 1998; Lai & Katul, 2000; Pataki & Oren, 2003; Novick et al., 2004; Stoy et al., 2005). We briefly describe each for completeness. OF is approximately 480m 305m and was established after a burn in 1979. It is mowed at least once annually during the summer for hay and to check woody encroachment (Novick et al., 2004). The vegeta- tion is dominated by the C3 grass Festuca arundinaria Shreb., with minor forb and other C3 and C4 grass species including Lespedeza cuneata (Dum. Cours.) G. Don, Andropogon virginicus L., and Sorghum halepense (L.) Pers. Canopy height (h) is spatially and temporally variable and averages ca. 0.1–1m depending on harvest and GS. EC instrumentation is at 2.8m on a 6m tall tower (Table 2). PP was established in 1983 following a clear cut and a burn. Pinus taeda L. (loblolly pine) seedlings were planted at 2.4m 2.4m spacing and ecosystem devel- opment has not been managed after planting. h in- creased from 16m in 2001 to 18m in 2004. The canopy is comprised primarily of P. taeda with some emergent Liquidambar styraciflua L. and a diverse and growing understory with 26 different woody species of diameter breast height 42.5 cm. The flux tower lies upwind of the CO2-enriched components of the free atmosphere carbon enrichment (FACE) facility (Hendrey et al., 1999) located in the same pine forest. EC instrumentation is at 20.2m (Table 2) on a 22m tower. HW is classified as an uneven-aged (80–100-year old) oak (Quercus) – hickory (Carya) forest with L. styraciflua and Liriodendron tulipifera L. also contributing to the canopy and a diverse understory with similar species as PP. The ecosystem has not been managed after estab- lishment. h averaged 25m with some emergent treetops reaching over 35m, and the canopy has large and frequent gaps. EC instrumentation is at 39.8m on a 42m tall tower (Table 2). There was an 11.9 ha clearcut in HW 200m south of the measurement tower on private land adjacent to the Duke Forest in November 2002 (Fig. 1). We minimize the effects of this disturbance as described in the ‘‘Data Filtering’’ section below. Measurements We measured and modeled the components of the energy balance at half-hourly time scales for 4 years at the three study ecosystems: Rn  LE H  G  M ¼ I; ð1Þ with a focus on latent heat flux (LE), where H is sensible heat exchange, G is energy storage below the canopy (predominantly in soil) and is near 0 when averaged for 24 h periods, and M accounts for photosynthesis and plant metabolism and is assumed to be trivial, ca. 0.1–0.4% under field conditions (Odum, 1971). The imbalance (I) often arises because of scale issues in measurements, advective energy transport, and the fact that the finite sampling duration and frequency in EC Table 1 List of abbreviations with units and definitions Abbreviation Units Definition a ms2 Gravitational acceleration An mmolm 2 s1 Net photosynthesis b Scale parameter of the Laplacian distribution Ca ppm Atmospheric CO2 concentration Cd Canopy drag coefficient Ci ppm Leaf-internal CO2 cp Jmol 1 C1 Specific heat of air at constant pressure d m Characteristic length scale D0 m Zero-plane displacement D kPa Vapor pressure deficit E mm Evaporation EC Eddy covariance EF Modeling efficiency ET mm Evapotranspiration ga molm 2 s1 Atmospheric conductance gb molm 2 s1 Leaf boundary layer conductance gc molm 2 s1 Canopy conductance gH molm 2 s1 Conductance to sensible heat gs molm 2 s1 Stomatal conductance gv molm 2 s1 Conductance to water vapor G Wm2 Soil heat flux GS April–September growing season Gr Grashof number h m Canopy height H Wm2 Sensible heat flux HW Hardwood forest ecosystem I Wm2 Radiation closure imbalance Km m 2 s1 Turbulent diffusion coefficient l m Mixing length L m2leafm 2 ground Leaf area index LE Wm2 Latent heat flux OF Old field (abandoned agricultural) ecosystem pa kPa Atmospheric pressure P mm Precipitation PAD m2plantm 2 ground Plant area density PAI m2plantm 2 ground Plant area index PAR mmol photonsm2 s1 Photosynthetically active radiation PM Penman–Monteith equation PP Planted pine ecosystem Re Reynolds number Ri, s Wm 2 Incident shortwave radiation Rn Wm 2 Net radiation S C1 Slope of the saturation mole fraction function SE Southeastern United States Ta C Air temperature Tc C Canopy temperature u ms1 Wind speed uc m s 1 Mean within-canopy wind speed u * ms1 Friction velocity u0w0 m2 s2 Momentum flux Vcmax mmolm 2 s1 Maximum carboxylation efficiency x Horizontal : vertical leaf projected area z m Height zH m Heat roughness length zm m Momentum roughness length measurements do not resolve the entire spectrum of eddies at both the low- and high-frequency ends (Wil- son & Baldocchi, 2000; Wilson et al., 2002) as discussed in more detail below. We note that LE (Wm2) is commonly used in energy flux studies and ET (mm 12 h1) is used in the hydrologic studies. These two terms are used interchangeably, depending on context. EC measurements do not independently resolve T or E, but estimates are desirable to elucidate physical and biological controls on ET. The model used to partition ET into T and E is discussed in Appendix A. Latent and sensible heat flux measurements LE and H were measured using EC systems comprised of triaxial sonic anemometers (CSAT3, Campbell Scien- tific, Logan, UT, USA) and in the case of LE, open-path infrared gas analyzers (IRGA, LI-7500, Li-Cor, Lincoln, NE, USA). Measurements of vertical wind velocity, temperature, and scalar concentrations of H2O were collected at 10Hz and flow statistics were processed in real time using 23X data loggers (Campbell Scienti- fic). The Webb–Pearman–Leuning correction (Webb et al., 1980) for the effects of air density fluctuations on flux measurements was applied to scalar fluxes mea- sured with the open-path LI-7500. A closed-path gas analyzer (LI-6262, Li-Cor) was used at PP before May 1, 2001 and 5Hz measurements were postprocessed using procedures described elsewhere (Katul et al., 1997). Topographic variations are minor (o5%) and influence flux measurements negligibly (Kaimal & Finnigan, 1994). To quantify energy partitioning differences among ecosystems and canopy coupling to the atmosphere, two dimensionless parameters are often used. The Bowen ratio (b) is the ratio of H to LE and the decou- pling coefficient (O, Jarvis & McNaughton, 1986) is a metric for stomatal control of transpiration: O ¼ 1þ Slc 1 p 1þ Slc1p þ gvag1vc ; ð2Þ where S is the slope of the saturation mole fraction and is a function of air temperature (Ta), cp is the specific heat of air and gvc is surface conductance in the original formulation and modeled gc here (Appendix B). Environmental measurements Daily P was measured with a rain gauge near the NOAA meteorological station located at OF, and half- hourly P was measured using a tipping bucket (TI, Texas Instruments, Austin, TX, USA) at PP. PAR, Rn, Ta, and relative humidity (RH) were sampled every second and averaged for half-hour periods at all three sites. PAR was measured using LI-190SA quantum sensors (Li-Cor). Rn measurements were made with Fritschen-type net radiometers (Q7, REBS, Seattle, WA, USA) through 2003 and with CNR1 net radio- meters (Kipp & Zonen, Delft, the Netherlands) in 2004. The Q7 and CNR1 showed good agreement for Table 1. (Contd.) Abbreviation Units Definition Zr m Rooting depth aleaf Leaf absorptivity to radiation b Bowen ratio dETGS mm Change in growing season evapotranspiration e Emissivity l Jmol1 Latent heat of vaporization r Soil reflectivity (i.e. albedo) s mm or Wm2 here Standard deviation sSB Wm 2 K4 Stephen–Boltzmann constant y m3waterm 3 soil Volumetric soil moisture content u m2 s1 Kinematic viscosity c degrees Zenith angle C Atmospheric stability O Decoupling coefficient Oc Leaf clumping factor Table 2 Characteristics of the old field (OF), pine plantation (PP) and hardwood forest (HW) ecosystems and associated eddy covariance (EC) and net radiation (Rn) measurements OF PP HW Mean canopy height (m) o1 16 (2001)–18 (2004) 25 L (m2m2) 0.1–4 1–5.5 0.1–7 Basal area (m2 ha1) 0 26.4 28.4 EC measurement height (m) 2.8 20.2 39.8 Rn measurement height (m) 4.8 22.2 41.8 ometer, respectively. (3) Flux data were removed if the inverse tangent of the ratio between u and 12 hourly mean vertical wind velocity (w) was greater than 151 to ensure that the mean streamlines remain almost parallel to the ground (i.e. o5% of measured flux as in Detto et al., 2006), and HW flux data were removed if mean wind direction was less than 401 or exceeded 3501 when distortion from the tower and a nearby canopy gap was largest. (4) To ensure that data represent a turbulent flux that originates from the ecosystem of interest in a probabilistic sense, we used the rigorous atmospheric stability (C) filter of Novick et al. (2004), which requires that night-time data (defined here as periods for which solar zenith angle 4901) are measured under near- neutral C, and that the peak of the source-weight function (using the footprint model of Hsieh et al., 2000) does not exceed ecosystem dimensions. In this way measured fluxes for which the peak of the source weight function originates outside the ecosystem of interest (e.g. from the clear-cut at HW) were excluded from this analysis. The percentages of EC data remain- ing after progressively applying each filter are listed in Table 3. Gapfilling EC time series incorporate many missing data ‘gaps’, but continuous time series are preferred for ET estima- Table 3 The percentage of flux data remaining after progres- sively employing four data filters Ecosystem Filter Daytime Total OF Logical 85 83 Instrument 83 81 Wind Directional 79 77 Atmospheric Stability 79 41 PP Logical 83 83 Instrument 77 76 Wind Directional 74 73 Atmospheric Stability 65 35 HW Logical 93 92 Instrument 89 89 Wind Directional 76 75 Atmospheric Stability 61 32 The ‘logical’ filter removes unrealistic data points. The instru- ment filter ensures proper sonic anemometer and IRGA func- tion. The wind directional filter removes fluxes potentially contaminated by advection. The atmospheric stability filter ensures both that measurements are taken under conditions of sufficient turbulence and that the bulk of the flux footprint Hsieh et al. (2000); Detto et al. (2006) arises from the ecosystem of interest. OF, old field; PP, planted pine forest; HW, hardwood forest. all ecosystems (data not shown). The CNR1 measures incoming and outgoing solar and far-infrared radiation separately using a coupled pyranometer/pyrgeometer design, enabling surface albedo (i.e. reflectivity, r) mea- surements. PAR and Rn sensors were 2 m above EC instrument height at each ecosystem (Table 2). Ta and RH were measured with HMP35C Ta/RH probes (Campbell Scientific) positioned at 2 m at OF and at two-thirds canopy height at PP and HW. At PP, inte- grated 0–30 cm soil moisture (y) measurements were made at 12 locations using CS615 y sensors (Campbell Scientific), and the mean of these measurements was taken to be site-wide y. At OF and HW, y was measured using Type ML1 ThetaProbe soil moisture sensors (Delta-T Devices, Cambridge, UK). Six sensors were positioned at 10 cm depth and two sensors at 25 cm depth near the respective tower. Average y was com- puted for the 10 and 25 cm depths and the mean of the two depths was taken to obtain a single value for y that is comparable with PP. Leaf area index measurements In 2001, L at OF was estimated by measuring PAR transmission with a series of 80 quantum sensors (AccuPAR model PAR-80 Ceptometer, Decagon Instru- ments, Pullman, WA, USA) to calculate gap fractions, which were inverted to calculate L after Norman & Campbell (1989; see Novick et al. 2004). After 2001, L at OF was estimated using a combination of litter data and LAI-2000 plant-canopy analyzer (Li-Cor) measure- ments. At PP, the contribution to L from P. taeda trees was calculated using needle elongation and litterfall measurements, and L of understory hardwood species was calculated using degree-day sums and litterfall measurements (McCarthy et al., 2006). At HW, L was determined using LAI-2000 measurements adjusted to match litterfall data (Palmroth et al., 2005). Plant area density (PAD) was determined using profile LAI-2000 measurements at all ecosystems. Data filtering Care must be taken to ensure that EC measurements represent a turbulent flux from the land surface, and in our case must be rigorously filtered to ensure that the flux footprint lies within ecosystem dimensions (Fig. 1). Flux data were filtered using four criteria. (1) Data were removed if they exceeded logical maxima and minima, determined to be 800 and 100 Wm2, respectively, for LE and H (i.e. ‘Logical’ filter, Table 3 below). (2) The effects of excessive sensor noise were filtered by remov- ing points for which scalar variance exceeded 4 gH2Om 3 and 4 1C for the IRGA and sonic anem- tion on seasonal or annual time scales (Falge et al., 2001). When the measurement of any meteorological variable was unavailable due to equipment failure or other error, a continuous record was obtained by fitting a linear regression between measurements from the sensor of interest and a nearby sensor of the same type using a windowing function that modeled temporally local data appropriate for the size of the gap [i.e. at least 25% of the size of the gap, with a coefficient of deter- mination (r2) of at least 0.9]. Missing H data were also gapfilled in this way using linear temporally local relationships with Rn. When sonic anemometer measurements of u were not available, we approximated u using cup anemometer measurements made at PP and maintained by Brookha- ven National Laboratories. We note that cup anem- ometer measured mean wind speed is ‘contaminated’ by the turbulent kinetic energy, but cup and sonic anemometer mean velocity measurements were strongly related (r25 0.75). Missing LE data was gapfilled using the Penman Monteith (PM) equation and tested against a simpler gc model (Jarvis, 1976; Oren et al., 1999) as described in Appendix B. Canopy wind speed (uc) is required to model conductances to water vapor and sensible heat (gv and gH, respectively) for the PM equation and was modeled by a first-order canopy turbulence model also described in Appendix B. Canopy conductance para- meters for both models were fit for monthly periods via least squares regression using the Gauss–Newton algo- rithm standard in MATLAB (Mathworks Inc., Boston, MA, USA). Error estimation Error in annual ET estimates was calculated following a standard approach (Goulden et al., 1996; Moncrieff et al., 1996) that combines error due to gapfilling – also called ‘sampling uncertainty’ (SU) – in flux estimates with the ‘uniform systematic error’ (USE) of each EC system. SU was determined by simulating the impact of the stan- dard error of the fitted monthly canopy conductance parameters in the PM equation (Appendix B) on annual ET estimates using Monte Carlo simulations with 100 realizations. The variance from the resulting distribu- tion of annual flux estimates represents an estimate of the error due to SU. In the original formulation, USE was estimated for each EC system for each year by sampling night-time LE, which should approximate 0. Rn is negative at night, thus night-time LE measurements may deviate from 0 (e.g. due to condensation). Also, for high u * , the Eddy- diffusivity near the canopy top is large; hence, any small gradients in mean water vapor concentration may lead to a finite flux not associated with water vapor production from foliage or forest floor, but rather stored water vapor concentration leaving the canopy air space. Finite LE may also occur when nocturnal radia- tive perturbations associated with the passage of clouds induce instabilities and local production of turbulence within the canopy (Cava et al., 2004). Again, this flux is better approximated as a storage flux rather than a biosphere–atmosphere flux. Thus, we excluded conditions for which u * 40.2m s1, the absolute value ofC iso0.1, or if the standard deviation of Rn is greater than 5Wm2, a surrogate for potential passage of clouds. Deviance from 0 flux in the remaining data is assumed to be ‘inherent’ error in the flux measure- ments. These inherent errors followed a Laplacian dis- tribution (Hollinger & Richardson, 2005; Richardson et al., 2006), the standard deviation of which is s ¼ ffiffiffi2p b and the unbiased estimator for the scale para- meter b is PN i¼1 xi  xj j=N where N is the length of the data series and x is the mean night-time LE measure- ment. This standard deviation of night-time fluxes was divided by average daytime LE and multiplied by the annual or seasonal flux estimate to calculate USE (Goul- den et al., 1996). Total error was calculated by combin- ing variances due to SU and USE, and is reported as a 1s interval about estimated annual ET. Results We begin by briefly discussing climatic variability across the measurement period and analyzing the Rn balance closure at each site to assess how much of the differences in observed ET are due to vegetation. We then discuss the partitioning of Rn, focusing on annual and ETGS and its relationship to external and internal drivers and its components, T and E, with an analysis of techniques for gapfilling and error estimation. Climatic variability GS (April–September) precipitation (PGS) was 529mm in 2001 (hereafter ‘mild drought’), 371mm in 2002 (‘severe drought’), 790mm in 2003 (‘wet’), and 661mm in 2004 (‘average’), about 1s below, 2 s below, 1s above and near the long-term (111 year) mean PGS of 632  130mm, respectively (Fig. 2a). The sum of PARGS was nearly equal in 2001 and 2002, but ca. 10% less than 2001–2002 levels in 2003 and 2004, due to the cloudier conditions (Fig. 2b). Mean Ta,GS differed by more than 1.5 1C among years; 2001 and 2003 were relatively cool- er and 2002 and 2004 were relatively warmer (Fig. 2c). P (Fig. 2a), PAR (Fig. 2b), and Ta at 2/3 canopy height in the forested ecosystems and at 2m at OF (Fig. 2c) did not differ appreciably among the adjacent ecosystems, with the latter influenced slightly by the state of the canopy. Parts of OF are shaded at sunup and sundown due to nearby forest edges. These conditions occur during low PAR periods and are considered negligible for long-term ET estimates. Any differences among ecosystems in micrometeor- ological and edaphic variables such as D (Fig. 3a) and y (Fig. 3b) can be attributed to the influence of vegetation to a first order (Palmroth et al., 2005). For example, y was higher at HW than OF and PP during winter when leaf area is absent and transpiration ceases (Fig. 3b). As expected, the rougher canopy (HW) also experienced the highest momentum sink (¼ u2) during maximum foliage (Fig. 3c), and in the winter, the u * for PP and HW were comparable despite significant differences in win- tertime L (Fig. 3e). At OF, u * did not exhibit marked seasonal trends despite strong L variability. Radiation balance closure LE and H dominated the radiation balance and com- prised 66%, 75%, and 66% of Rn at OF, PP, and HW, respectively, on the 12 hourly basis. The greater contribu- tion of LE and H at PP is logical given its higher relative leaf area in winter and lower G. We measured G and below canopy Rn intermittently and found this term to be small relative to LE and H, but it could exceed 100Wm2 when LAI was low at OF and HW, and when direct solar radiation penetrated a gap at PP 0 50 100 150 200 250 Pr ec ip ita tio n (m m mo nt h− 1 ) 947 1092 1346 982 529 371 790 661 Ann. Σ (mm) G.S. Σ (mm) 500 750 1000 1250 100% 93% 88% 91% 100% 98% 89% 92% Rel. Ann. Rel. G.S. PA R (m ol ph ot on s m − 2 m o n th − 1 ) 2001 2001.5 2002 2002.5 2003 2003.5 2004 2004.5 2005 0 5 10 15 20 25 T a (°C ) Year 14.5 15.1 14.3 14.8 20.3 21.9 20.5 21.1 Ann. mean (C) G.S. mean (C) OF PP HW (a) (b) (c) Fig. 2 Precipitation (P, a), photosyntheically active radiation (PAR, b) and air temperature (Ta, c) did not differ appreciably among the adjacent ecosystems. 0 0.5 1 1.5 2 D (kP a) 0 0.1 0.3 0.5  (m 3 m − 3 ) 0 0.1 0.3 0.5 u * (m s− 1 ) 0 0.1 0.2 0.3 0.4 u c (m s− 1 ) 2001 2001.5 2002 2002.5 2003 2003.5 2004 2004.5 2005 0 2 4 6 LA I (m 3 m − 3 ) Year (a) (b) (c) (d) (e) Fig. 3 Vapor pressure deficit (D, a), soil moisture (y, b), friction velocity (u * , c), modeled canopy wind speed (uc, d) and leaf area index (L, e), differed among ecosystems. and HW. To minimize the effects of G on the energy balance closure, we computed radiation balance closure at the daily time step, which was 72% at OF, 77% at PP, and 65% at HW. This lack of closure is greater than reported in the synthesis of Wilson et al. (2002) who found a mean radiation balance closure of 80% across a wide range of FLUXNET sites. We attribute the ob- served I to low-frequency losses and entrainment dur- ing convective conditions rather than other common explanations for I (e.g. Wilson et al., 2002) as discussed in Appendix C; these events are likely to impact H more than LE. Appendix C (see Fig. 10) suggests that the lack of energy closure strongly varies with atmospheric stability – with near-convective conditions experiencing the largest I. Models for gapfilling missing data Compared with the Jarvis-type model, the PM model for bimonthly periods with full accounting of atmo- spheric, leaf boundary layer and stomatal conductances (Appendix B) resulted in the best fit with measured LE for all ecosystems (Table 4). The slope between model and measurements was closest to unity, the intercept was closest to 0, and the parameter estimates consis- tently converged. Hence, it was the logical choice for gapfilling missing ET data. Interestingly, model-fitting statistics were not com- promised by dramatically simplifying the gapfilling model. Replacing PM with the Jarvis type gc model [i.e. Eqns (B2) and (B3) alone], with proper unit correc- tion, resulted in comparable root mean-squared error (RMSE) and modeling efficiency (EF, Loague & Green, 1991; Meyer & Butler, 1993) despite the fact that it is a model for T, not ET (Table 4). In this way, accurate simplifications to ET models may be achieved with knowledge of only PAR, D, L, and, during drought periods, of y (or as a surrogate, P) as well. This finding already suggests that much of the variability in ET is driven by T. Radiation balance partitioning Rn (Fig. 4a) followed the seasonal pattern at all ecosys- tems. H (Fig. 4b) values were comparable at OF and PP during all periods except after the ice storm. LE was lower at OF than the forested ecosystems during most GS periods, and LE at PP was generally larger than or comparable to HWwith the exception of the 2002 severe drought-GS (Fig. 4c). HW had the largest wintertime Rn; low LE was compensated by larger H fluxes. Low LE and high H during severe drought at OF and PP resulted in a dramatic increase in mean daytime b (Fig. 5a); the summertime b at HWwas consistently low due to high LE and low H fluxes regardless of drought conditions, and the wintertime b was consistently high- est due to very low LE. The O showed the expected response under all conditions except severe drought (Fig. 5b). High O values at OF were indicative of Rn limitation, and the O was lowest on average at the needle-leafed PP because gc was well-coupled to atmo- spheric demand. Interestingly, the O at OF approached values typical of PP during the peak of the drought in 2002, indicating that gc at these two different canopies was similarly limited. Water vapor fluxes at the annual time scale Interecosystem differences in Rn balance partitioning (Fig. 4a–c) resulted in annual and GS sums of ET, T, and E that varied among ecosystems and across years and Table 4 Model-fitting statistics for ET for the full Penman– Monteith model (PM, Eqn 2), and a Jarvis-type model after Oren et al. (1999) (Eqn B3) Ecosystem Model Slope Intercept (Wm2) RMSE (Wm2) EF OF PM 0.87 16 62 0.56 Jarvis 0.70 26 59 0.60 PP PM 0.82 22 45 0.81 Jarvis 0.82 20 46 0.80 HW PM 0.77 27 62 0.73 Jarvis 0.75 29 61 0.73 RMSE is the root mean square error and EF is modeling efficiency. OF, old field; PP, planted pine forest; HW, hardwood forest. 0 50 100 150 200 R n (W m − 2 ) 0 25 50 75 100 H (W m − 2 ) OFPP HW 2001 2001.5 2002 2002.5 2003 2003.5 2004 2004.5 2005 0 25 50 75 100 LE (W m − 2 ) Year (a) (b) (c) Fig. 4 Monthly average net radiation (Rn a), latent heat (LE, b), and sensible heat (H, c) fluxes for the three study ecosystems. gapfilling are sufficiently accurate for estimating seaso- nal or annual ET sums, and also that efforts to reduce error in EC systems should focus on systematic error. Modeled T (Table 5, Fig. 6b) decreased from 290 to 160mm at OF as the drought progressed, then increased dramatically to 430mm during the wet year and de- creased by 70mm between the wet and normal years. 0 1 2 3 4 5 (a) (b)  OF PP HW 2001 2001.5 2002 2002.5 2003 2003.5 2004 2004.5 2005 0 0.1 0.2 0.3 0.4 0.5 Ω Year Fig. 5 Mean Bowen ratio (b, a), simplified here as the monthly average of daytime H/LE, and the decoupling coefficient (O, b). 0 200 400 600 800 (a) (b) (c) ET (m m) 0 200 400 600 T (m m) 2001 2001.5 2002 2002.5 2003 2003.5 2004 2004.5 2005 0 200 400 600 Year E (m m) OF PP HW Fig. 6 Cumulative evapotranspiration (ET) for old field (OF), pine plantation (PP) and oak-hickory hardwood forest (HW) ecosystems for 2001–2004 with flux error (1s) estimated after (Goulden et al., 1996). Growing seasons are indicated by hor- izontal bars. (b) Same as (a) but for cumulative transpiration (T) estimates. (c) Same as (a) but for cumulative evaporation (E) estimates. Table 5 Annual and growing season (GS) evapotranspiration (ET), transpiration (T), and evaporation (E) estimates with associated error about annual ET estimates as estimated after Goulden et al. (1996) Ecosystem Year ET T E ETGS TGS EGS USE SU Total Error OF 2001 560 290 260 420 250 170 59 4 59 2002 460 160 300 320 120 200 68 3 68 2003 650 430 220 480 350 130 79 2 79 2004 580 360 220 430 290 140 65 2 65 PP 2001 660 500 160 510 420 90 53 7 53 2002 560 400 160 410 320 90 51 10 52 2003 670 500 180 510 410 100 44 7 45 2004 740 560 180 550 460 100 51 3 51 HW 2001 610 440 170 490 410 80 63 10 64 2002 580 410 160 480 390 90 59 9 60 2003 640 460 180 510 420 90 62 7 62 2004 640 460 180 500 420 80 58 4 58 USE is ‘uniform systematic error’ and SU is ‘sampling uncertainty’. All units are in mm. Component fluxes may not equal ET due to rounding. OF, old field; PP, planted pine forest; HW, hardwood forest. GSs (Fig. 6a, Table 5). Annual ET was characteristically lower at OF than the forested ecosystems except during the wet 2003, when it was similar to ET at HW. Annual ET at PP was in general greater than at HW, but usually within the range of estimated error. ET at HW slightly exceeded PP during the severe drought (2002) and exhibited relatively low interannual variability. Total error in annual ET varied between 7% and 14%, consistent with other studies (Meyers, 2001; Wilson & Meyers, 2001). Total error was dominated by the USE (i.e. systematic error) component rather than the SU (gapfilling) component (Table 5), suggesting that error due to gapfilling is small and that models used for At PP, T decreased by 100mm between mild and severe droughts, then returned to 2001 levels during the wet year (2003) after the ice storm. T increased to 560mm in 2004 despite less P. T was nearly invariant at HW and changed by a maximum of ca. 50mm between severe drought and wet conditions. Modeled E (Fig. 6c) was larger at OF than the other ecosystems, increased by 40mm from mild to severe drought, and was similar (220mm) in 2003 and 2004 with different seasonal patterns. E at PP was similar for all years and increased by only 20mm during the wet year when L was lower due to ice-storm impacts. Like- wise, E at HW was consistent between years and increased by only 20mm between severe drought and wet years. Relationships between ETGS and environmental drivers ETGS at OF and PP was sensitive to P, less so at HW (Table 5, Fig. 7a). The large reduction in LGS at PP after the December 2002 ice storm (McCarthy et al., 2006) caused the relationship between ETGS and PGS to fall below the linear response observed during other years (Fig. 7a). This effect was not observed in the other ecosystems. Consequently, average PGS in 2004 de- creased ETGS from 2003 levels by over 10% at OF, but increased ETGS by nearly 8% at PP as the canopy recovered from ice-storm damage (Table 5, Fig. 7a). In contrast, ETGS at HW changed by 6% or less between consecutive years (Table 5). The resulting relationship between PGS and ETGS was nearly linear for all ecosys- tems, especially HW (Fig. 7a). ETGS decreased with increasing GS mean D (DGS) at all ecosystems (Fig. 7b), suggesting that plant controls on transpiration, not increased E, dominated the ET signal under high D at the GS time scale. It is interesting to note that OF increased growing season L (LGS) and thus ET in response to optimal growth conditions dur- ing 2003 and 2004 (Fig. 7c), but the forested ecosystems increased mean growing season gs (gs, GS), as estimated from Eqn (B3), rather than L (Fig. 7d). Thus, ETGS was primarily driven by changes in P, D, L, and gs. The response of ETGS to the product of these variables is approximately linear across all ecosystems (Fig. 8): ETGS  aPGSDGSLGSgs;GS þ b: ð3Þ We use this simple model to isolate the relative contribution of the physical (P,D) and biological factors (L, gs) that give rise to changes in ETGS through a linear perturbation analysis similar to Wilson and Baldocchi (2000). Briefly, if we consider changes in ETGS (dETGS), the total derivative of Eqn (3) is represented by a 300 350 400 450 500 550 600 650 700 750 800 850 300 350 400 450 500 550 600 650 (a) (b) (c) (d) 2002 2001 2004 2003 OF PP HW 1 1.2 1.4 300 400 500 600 Σ ET G S (m m) Σ ET G S (m m) Σ PGS (mm) 0 0.5 1 1.5 2 0.015 0.025 0.035 0.045 Mean gs, GS (mol m−2 s−1)Mean LGSMean DGS (kPa) Fig. 7 The sum of April–September growing season evapotranspiration (ET) vs. precipitation (P, a), mean vapor pressure deficit (D, b), mean leaf area index (L, c), and stomatal conductance (gs, d). multivariate Taylor’s expansion: dETGS  @ETGS @PGS dPGS þ @ETGS @DGS dDGS  þ @ETGS @LGS dLGSþ @ETGS @gs;GS dgs;GS  þ 1 2! @2ETGS @2PGS ðdPGSÞ2 þ    ; ð4Þ where the higher-order terms are neglected in this first- order analysis. Using Eqn (3), @ETGS @PGS ¼ aDGSLGSgs;GS; @ETGS @DGS ¼ aPGSLGSgs;GS; @ETGS @LGS ¼ aPGSDGSgs;GS; @ETGS @gs;GS ¼ aPGSDGSLGS: ð5Þ It is important to note that gs is expected to change with time owing to both external and internal factors. For example, the Fick’s law relationship for net photo- synthesis An coupled with T, which is a function of gs [Eqn (B2)], can be expressed as T ¼ ð1:6AnDÞ=ðCa  CiÞ (Katul et al., 2003) where Ca is atmospheric CO2, and Ci is leaf-internal CO2. Thus, gs is expected to change with variables that influence leaf An including, for example, maximum carboxylation efficiency Vcmax, which is vari- able in time in these ecosystems due to temperature acclimation and leaf nitrogen variations (Ellsworth, 2000; Wilson et al., 2001; Juang et al., 2006). The model matches measurements well despite the linearity assumption (Figs 8 and 9). ETGS at OF and PP were the most sensitive to the combination of external and internal drivers (i.e. they had the highest slope in Fig. 8), and HW was least sensitive. Internal (i.e. biolo- gical) adjustment dominated dETGS at OF over the measurement period (Fig. 9a) and dominated dETGS at PP during the drought years (Fig. 9b). The impact of biological and climatic drivers opposed each other at PP during the wet year after the ice storm. dETGS at HW was small, and the contribution of physical and biolo- gical factors were approximately equal (Fig. 9c). To ensure that these findings are not overly sensitive to the choice of the model, we repeated the entire analysis using the full PM model [Eqn (B1)] in Appendix D. From Appendix D, we found that the relative role of internal vs. external drivers is the same as Fig. 9. Individual aspects of the linear perturbation analysis (Fig. 9d–f) are discussed in more detail below. Discussion We begin by noting some aspects of the environmental measurements that are of particular importance to water cycling, namely y, canopy wind speed and mo- mentum flux, and Rn. Next, we discuss the outcome of the linear perturbation analysis within the context of this experiment and contrast our findings with other studies, focusing on longer-term dynamics. The broader implications of these findings on water resources in the SE are also discussed. Site characteristics Interestingly, ETGS exceeded PGS in 2002 by 30 and 100mm at PP and HW, respectively, suggesting that the GS root-zone storage must have been depleted by these amounts. If we consider the soil water balance nZrðdy=dtÞ (Porporato et al., 2002) where n is porosity (0.54m3m3, Oren et al. (1998) and dy=dt is the change in soil moisture during the 2002 GS (0.19m3m3 at PP and 0.27m3m3 at HW, Fig. 3b), then rooting depth (Zr) is estimated to be shallow, ca. 30–50 cm at both ecosys- tems. A clay pan was observed at 35 cm, which is also the depth at which water uptake balanced transpiration at PP (Oren et al., 1998). Trenching experiments and tip- ups in the forested ecosystems further suggest that this approximation is robust. At OF, roots did not develop below 45 cm (Lai & Katul, 2000), and observed ET was well described using a root water capture model that was active to 35 cm depth. Coupled with existing stu- dies, the findings here are consistent regarding the 10 20 30 40 50 60 70 80 90 100 110 300 350 400 450 500 550 600 PGS × DGS × LGS × gs, GS (mmol H2O kPa m−1 s−1) ET G S (m m) OF PP HW Fig. 8 The differential sensitivity of growing season evaportan- spiration (ETGS) of the study ecosystems to the combination of precipitation (P), vapor pressure deficit (D), and leaf area index (L) and stomatal conductance (gs). assumption that rooting depth is approximately equal among ecosystems and is on the order of 35 cm, but we cannot exclude the possibility that parts of HW may experience deep rooting. With respect to the momentum flux (u0w0, i.e. u2), Fig. 3(c and e) suggests that much of its seasonal varia- bility is primarily driven by the alterations in roughness density that vary with the drag coefficient and its dependence on the local Reynolds number (Poggi et al., 2004a, b), and the leaf area density distribution, rather than the variation in L (Poggi et al., 2004a, b). In contrast, canopy wind speed (uc, Fig. 3d) differed markedly among ecosystems and across seasons because of its strong dependence on L (Table 2, Fig. 3e). Despite large changes in surface albedo r (data not shown) and surface temperature among ecosystems and across seasons, interecosystem differences in Rn were typically smaller than differences in H and LE (Fig. 4a–c). Thus, if we separate the Rn into its short- wave (1 rRi;s) and longwave (sSB½eaT4a  esT4s ) com- ponents, where Ri,s is incident shortwave radiation, sSB is the Stephen–Boltzmann constant, and ea and es are air and surface emissivity, respectively, we find that changes in Rn were dominated by changes in Ri, s, which was identical among ecosystems (Fig. 2b). Physical controls on ET The analysis of Eqn (4) in Fig. 9d–f reveals the relative importance of PGS, DGS, LGS, and gs,GS on dETGS at the GS time scale. All ecosystems were sensitive to changes in P, as expected from Fig. 7a. This response was particularly strong at PP (Fig. 9e) where approximately 40 and 70mm of the observed dETGS during 2001 and 2002, respectively, was due to low PGS. Interestingly, lower PGS in 2001 and 2002 decreased ETGS by approximately equal amounts – between 10 and 20mm – at OF and HW (Fig. 9d and f). The increase in ETGS due to increased PGS during 2003 was ca. 30mm at OF and PP, and 10mm at HW. Changes in ETGS due to the direct impacts of DGS were smaller in magnitude and of opposite sign than changes due to PGS for all ecosystems and GSs, and were never larger than 25mm (Fig. 9d–f). Thus, avail- able water, rather than atmospheric demand, was the primary external control on dETGS in the study ecosys- tems, although we note that the two are not indepen- dent; PGS was lower and DGS higher during drought. dETGS due to DGS was largest at PP for all GSs, as expected for the canopy with lowest mean O (Fig. 5b) and, thus, the strongest coupling of canopy response to atmospheric demand for water vapor. −150 −100 −50 0 50 100 OF −150 −100 −50 0 50  ET G S fro m 2 00 4 (m m) PP 2001 2002 2003 −150 −100 −50 0 50 Growing season (April–September) HW Measured Physical Biological −150 −100 −50 0 50 100(a) (d) (e) (f) (b) (c) OF −150 −100 −50 0 50 PP 2001 2002 2003 −150 −100 −50 0 50 Growing season (April–September) HW I (P ) II (D ) III (L ) IV (gs) Fig. 9 (a–c) The contribution of physical [i.e. precipitation (P) and vapor pressure deficit (D)] and biological [i.e. leaf area index (L) and stomatal conductance (gs)] drivers to measured changes in April–September growing season evapotranspiration (dETGS) at old field (OF) planted pine (PP) and hardwood forest (HW) ecosystems in the Duke Forest, NC. (d–f) Same as (a–c) but for all physical and biological drivers separately. 2003). The effect of the ice storm on LGS (McCarthy et al., 2006) decreased ETGS to the same degree that high PGS increased ETGS during the wet year. The biological response of gs,GS closely matched the changes in PGS for all years. In contrast, ETGS at HW was relatively insensitive to the wide range of climatic and hydrologic conditions experienced during the measurement period. Comparison with other studies The finding that ET varied due to both biological and climatic responses among ecosystems and across years contains similarities and differences to the results of Gholz & Clark (2002), who found that annual ET across a chronosequence of clear-cut, mid-rotation and full- rotation of slash pine plantations in Florida (FL) was more sensitive to climate than management activities (Table 6). Annual ET in the FL ecosystems was largely independent of plantation age, and the ratio ET/P was between 80% and 86% for most ecosystems and years except for a drought year in the mid-rotation stand, when it approached unity. This interaction follows logically from deep-rooted trees in a sandier soil; shal- low-rooted post-clearcut vegetation could not access groundwater in the FL case. ET averaged between 870 and 1170mm in the FL study (Table 6), but only between 460 and 740mm here despite the fact that P was similar for the study periods (930–1350mm in NC and 880–1390mm in FL). Part of the ET differences can be explained by the lower temperatures (including more instances of subfreezing temperature), shorter GS, and lower available energy in NC. For example, given the relationships between LE and Rn reported by Gholz & Clark (2002) and had their ecosystems experienced the same Rn as measured at PP, ET over the 4-year measurement period would have been 2680mm at their mid-rotation plantation, almost exactly the same as what was measured at PP (2650mm, Table 5). Along these lines, ET at HW varied from 580 to 640mm (Table 5), very similar to 3 years of ET estimates at a hardwood forest (HW) at Oak Ridge, TN (with approximately the same latitude as our study sites) that also experienced drought (Table 6, 537–611mm, Wilson & Baldocchi, 2000). Interannual ET changes in the TN ecosystem were shown to be dominated by changes in canopy conductance. Hence, it is clear that ET in the SE is primarily driven by available energy, which is then modulated by other climatic and hydrologic constraints in which the differential sensitivities of the vegetation to climatic variability become important. Estimated ET at OF for the period between April 11, 2001 and April 11, 2002 was 544mm, similar to the value estimated by Novick et al. (2004), who reported 568mm (Table 6), noting that different gapfilling mod- Biological controls on ET The ecosystems differed strongly in their biological response to prevailing physical conditions. Dramatic reductions in LGS at OF in response to drought reduced ET from 2004 levels by ca. 80 mm in 2001 and 2002. In contrast, differences in LGS between 2004 and 2002 at PP, attributable to both drought and lagged recovery from ice storm damage in 2004 (McCarthy et al., 2006), changed ETGS by less than 10 mm, but reductions in LGS due to ice storm damage in 2003 decreased ETGS by 50 mm. This decrease was nearly as strong as the reduction of ETGS due to decreased PGS during drought. The compensatory effect of increased gs during low L conditions during 2003 is consistent with progressive defoliation studies on P. taeda. Pataki et al. (1998b) used sapflux and porometry measurements to find that in- creased gs compensated completely for the physical removal of ca. 50% of L. LGS varied little at HW among GSs and minimally impacted changes in ETGS. At PP, declines in ETGS due to gs, GS were slightly greater than those due to PGS in both 2001 and 2002. In contrast, changes in gs, GS dominated the biological response of HW and was closely correlated to, but not as strong as, changes in ETGS due to PGS. If we take P to be a surrogate of y, it is clear that PP is more sensitive to drought than HW. This result agrees with those from sap-flux scaled transpiration and mean canopy conduc- tance at the same and similar ecosystems (Oren et al., 1998; Oren & Pataki, 2001; Pataki & Oren, 2003). Nonlinear responses of ETGS to PGS at OF can be attributed in part to vegetation death during the severe drought and a combination of mowing and cloudy conditions early in the 2004 GS. Maximum L at OF was not attained until later in the GS in 2004 in contrast to 2003, when OF was mowed later and was not accompanied by an uncharacteristic decrease in ETGS. Mowing was found to affect the carbon balance and L dynamics for a period of about 1 week at OF in 2001 (Novick et al., 2004), but it impacted L and, thus, ETGS for longer time scales in 2004. Late growing season ETGS was similar at OF in 2003 and 2004, but the combination of mowing and cloudiness resulted in lower ETGS in the early 2004 GS (Figs 3a, 5c and 6a). To summarize results from the linear perturbation analysis, low LGS at OF decreased ETGS during drought and high gs, GS increased ETGS during wet conditions. The effect of the decrease in LGS on ETGS was greater than the decrease due to PGS, indicating that the biolo- gical response to drought ‘overcompensated’ for the physical signal due to the drought sensitivity of the vegetation. PP was most sensitive to the external drivers PGS and DGS due to a combination of drought sensitivity and low O (Fig. 5b; Oren et al., 1998; Pataki & Oren, els were used. Novick et al. (2004) developed a model for gc similar to the Jarvis-type model described in Appendix B, then removed y limitations on gc to esti- mate the magnitude of ET that would have occurred in the absence of drought. This resulted in 738mm of ET under drought-free conditions, over 13% more than reported here for the wet year (650mm, Table 5). The modeling approach and measurement period in Novick et al. (2004) required the assumption that physiological and L dynamics remained unaltered, but L changed over the full 4-year measurement period in a way that reduced ET below the expectations of the big leaf model that was parameterized for the 2001 mild-drought con- ditions. These results highlight the importance of stu- dies designed to elucidate the long-term dynamics of compositionally and structurally dynamic ecosystems, but also suggest that reasonable predictions of ET can be obtained using simple modeling approaches with short-term data sets. Comparative sapflux analyses on species that dom- inate the canopies at PP and HWagree with the findings here that PP is more coupled to atmospheric demand and is more drought sensitive. Estimates of T at PP agree well with sapflux based estimates at the same forest for the period 1998–2000 (520–530mm, Scha¨fer et al., 2002), and range from 400 to 560mm except in the case of severe drought (Tables 5 and 6). In a chamber study of species common to PP and HW, Tof the species that dominate the canopy at PP, P. taeda and the minor component L. styraciflua, showed greater sensitivity to increasing D than Quercus phellos, a species common at HW (Pataki et al., 1998a). Of the three species, P. taeda had the largest cumulative soil water uptake, indicating the greatest potential for T under well-watered condi- tions, and Q. phellos had the lowest T, similar to results here at the ecosystem scale. At HW, sapflux in oak (Quercus alba and Q. rubra), hickory (Carya tomentosa), ash (Fraxinus americana), and sweetgum (L. styraciflua) species was insensitive to y depletion during a mild drought, only tulip poplar (L. tulipifera) showed a decrease in Tunder water-limited conditions (Pataki & Oren, 2003). Species sensitivity to hydrologic variability was similar between HW and structurally and compositionally similar stands else- where in Duke Forest (Oren & Pataki, 2001), suggesting that results here may hold at larger spatial scales. Potential impacts of land cover change on water resources Agricultural and OF-type ecosystems have been con- tinuously converted to PP-type ecosystems throughout the SE through ecosystem succession and anthropo- Table 6 ET and T estimates from the study ecosystems, and ET estimates from Eddy-covariance measurements in other south- eastern US (SE) ecosystems Study Ecosystem Time Period ET ETGS T TGS Novick et al. (2004) OF April 11, 2001 to April 11, 2002 568 Scha¨fer et al. (2002) PP 1998 537 519 1999 575 533 2000 614 519 Pataki & Oren (2003) HW 1997 264 Emanuel et al. (2006, personal communication) Blandy Exp. Farm, VA 2001 572 445 2002 522 Gholz & Clark (2002) Clearcut P. elliottii, FL 1998 1048 1999 869 Mid-rotation P. elliottii, FL 1998 1014 1999 887 Full rotation P. elliottii, FL 1996 1001 1997 1171 Wilson & Baldocchi (2000) Deciduous forest, Oak Ridge, TN 1995 537 1996 554 1997 611 Hanson et al. (2004) Deciduous forest, Oak Ridge, TN 1995 515 1996 584 1997 624 1998 583 1999 658 2000 644 All flux units are in mm. to PP-type ecosystems (Wear & Greis, 2002) may induce subtle influences on the summer-time convective pre- cipitation. The comprehensive effects of plantation forestry on regional water cycling should be examined (Jackson et al., 2005), particularly in areas such as the SE where their contribution to land cover is large. Acknowledgments This research was supported by the Office of Science (BER), US Department of Energy, Grant No. DE-FG02-00ER63015, by the National Institute of Global Environmental Change (NIGEC) through the Southeast Regional Center at the University of Alabama, Tuscaloosa (DOE cooperative agreement DE-FC030- 90ER61010) and by the SERC-NIGEC RCIAP Research Program. Cup anemometer data was obtained under the US Department of Energy Contract No. DE-AC02-98CH10886 with Brookhaven National Laboratory. We would like to thank A. Porporato and R. Jackson for helpful comments on the manuscript, K. Novick, Y. Parashkevov, B. Poulter, K. Lewin, R. Nettles, G. Hon, and R. LaMorte for assistance with data collection in the Duke Forest, Judd Edeburn and the Office of the Duke Forest for logistical support, Y. Parashkevov and R. M. Broadwell for assistance in creating Fig. 1, and R. Emanuel, H. Epstein, C. Williams and J. Albertson for providing ET data from the Blandy Experimental Farm, VA. References Baldocchi D, Meyers T (1998) On using eco-physiological, micrometeorological and biogeochemical theory to evaluate carbon dioxide, water vapor and trace gas fluxes over vegetation: a perspective. Agricultural and Forest Meteorology, 90, 1–25. Campbell GS, Norman JM (1998) An Introduction to Environmental Biophysics, 2nd edn. Springer, New York. Cava D, Giostra U, Siqueira MBS et al. (2004) Organized motion and radiative perturbations in the nocturnal canopy sublayer above an even-aged pine forest. Boundary Layer Meteorology, 112, 129–157. Detto M, Montaldo N, Albertson JD et al. (2006) Soil moisture and vegetation controls on evapotranspiration in a heteroge- neous mediterranean ecosystem on Sardinia, Italy. Water Resources Research, 42, doi: 10.1029/2005WR004693/. Ellsworth DS (2000) Seasonal CO2 assimilation and stomatal limitations in a Pinus taeda canopy. Tree Physiology, 20, 435–445. Ellsworth DS, Oren R, Huang C et al. (1995) Leaf and canopy responses to elevated CO2 in a pine forest under free-air CO2 enrichment. Oecologia, 104, 139–146. Emanuel RE, Albertson JD, Epstein HE et al. (2006) Carbon dioxide exchange and early old-field succession. Journal of Geophysical Research, 111, G01011. Engel V, Jobbagy EG, Stieglitz M et al. (2005) Hydrological consequences of eucalyptus afforestation in the Argentine pampas. Water Resources Research, 41, W10409. Falge E, Baldocchi D, Olson R et al. (2001) Gap filling strategies for long-term energy flux data sets. Agricultural and Forest Meteorology, 107, 71–77. 0 0.1 0.3 0.5 0.7 0.9 10 0 10 20 30 40 50 60 〈R n − H − LE 〉 (W m − 2 ) 〈u * 〉 (m s−1) 15 10 5 0 〈Ψ〉 OF PP HW (a) (b) −−− Fig. 10 Mean daily (h  i) radiation balance closure for different hu*i (a) and atmospheric stability (hCi, b) bins. genic management since the reconstruction period following America’s Civil War (Oosting, 1942; Wear & Greis, 2002). Recent regional assessments have predicted that the landcover area of pine plantation ecosystems will continue to increase at the expense of agricultural and HW ecosystems (Wear & Greis, 2002). ET at PP was greater than or equal to ET at OF and HW during all periods except severe drought (Table 5, Fig. 6) although some biases due to I cannot be ignored (Ap- pendix C). These results are consistent with the recent finding that plantation ecosystems often use more water than the ecosystems that they replace (Engel et al., 2005; Farley et al., 2005), but also suggest that efforts to reduce I, particularly the role of convection in creating I (Fig. 10b), should be further explored. The abandoned agricultural field and pine plantation ‘early successional’ ecosystems showed remarkably similar H dynamics for all periods except during the winter and GS following the ice storm. The two cano- pies also showed similar behavior under severe drought; metrics of dryness through b and O were nearly identical despite dramatically different canopy morphologies. The magnitude of H dominates the de- velopment of the convective boundary layer with its critical influence on convective storm triggers. If H is similar in these two ecosystem types for normal to dry precipitation conditions, this change in land cover may minimally impact some aspects of micro and mesoscale meteorology, primarily the statistics of con- vective precipitation (Juang et al., 2006), but higher LE at PP may tend to increase triggers of P because the increase in water vapor concentration reduces the height of the lifting condensation level. The H and b dynamics of HW contrast those of both OF and PP (Figs 4 and 5), thus, the widespread conversion of HW Farley KA, Jobbagy EG, Jackson RB (2005) Effects of afforestation on water yield: a global synthesis with implications for policy. Global Change Biology, 11, 1565–1576. Foley J, Costa M, Delire C et al. (2003) Green surprise? How terrestrial ecosystems could affect earth’s climate. Frontiers in Ecology and the Environment, 1, 38–44. Gholz HL, Clark KL (2002) Energy exchange across a chronose- quence of slash pine forests in Florida. Agricultural and Forest Meteorology, 112, 87–102. Goulden ML, Munger JW, Fan S et al. (1996) Measurements of carbon sequestration by long-term eddy covariance: methods and a critical evaluation of accuracy. Global Change Biology, 2, 169–182. Hanson PJ, Amthor JS, Wullschleger SD et al. (2004) Oak forest carbon and water simulations: model intercomparisons and evaluations against independent data. Ecological Monographs, 74, 443–489. Hendrey G, Ellsworth D, Lewin K et al. (1999) A free-air enrich- ment system for exposing tall forest vegetation to elevated atmospheric CO2. Global Change Biology, 5, 293–309. Hoerling M, Kumar A (2003) The perfect ocean for drought. Science, 299, 691–694. Hollinger DY, Richardson AD (2005) Uncertainty in eddy covar- iance measurements and its application to physiological models. Tree Physiology, 25, 873–885. Houghton JT, Ding Y, Griggs DJ et al. (2001) Climate Change 2001: The Scientific Basis. THIRD Assessment Report of Working Group. Cambridge University Press, Cambridge. Hsieh CI, Katul G, Chi T (2000) An approximate analytical model for footprint estimation of scalar fluxes in thermally stratified atmospheric flows. Advances in Water Resources, 23, 765–772. Jackson RB, Jobba´gy EG, Avissar R et al. (2005) Trading water for carbon with biological carbon sequestration. Science, 310, 1944–1947. Jarvis PG (1976) The interpretation of the variations in leaf water potential and stomatal conductance found in canopies in the field. Philosophical Transactions of the Royal Society of London B, 273, 593–610. Jarvis PG, McNaughton KG (1986) Stomatal control of transpira- tion: scaling up from leaf to region. Advances in Ecological Research, 15, 1–49. Juang J-Y, Porporato A, Stoy PC et al. (2006) Hydrological and atmospheric controls on the statistics of convective precipita- tion events. Water Resources Research, in press. Kaimal JC, Finnigan JJ (1994) Atmospheric Boundary Layer Flows: Their Structure and Measurements. Oxford University Press, Oxford. Katul GG, Hsieh CI, Kuhn G et al. (1997) Turbulent eddy motion at the forest-atmosphere interface. Journal of Geophysical Research-Atmospheres, 102, 13409–13421. Katul G, Leuning R, Oren R (2003) Relationship between plant hydraulic and biochemical properties derived from a steady- state coupled water and carbon transport model. Plant, Cell and Environment, 26, 339–350. Katul GG, Mahrt L, Poggi D et al. (2004) One- and two-equation models for canopy turbulence. Boundary Layer Meteorology, 113, 81–109. Katz RW, Brush GS, Parlange MB (2005) Statistics of extremes: modeling ecological disturbances. Ecology, 86, 1124–1134. Lai C-T, Katul GG (2000) The dynamic role of root-water uptake in coupling potential to actual transpiration. Advances in Water Resources, 23, 427–439. Loague K, Green RE (1991) Statistical and graphical methods for evaluating solute transport models: overview and applica- tions. Journal of Contaminant Hydrology, 7, 51–73. Luo YQ, Medlyn BE, Hui DF et al. (2001) Gross primary pro- ductivity in Duke Forest: modeling synthesis of CO2 experi- ment and eddy-flux data. Ecological Applications, 11, 239–252. McCarthy HR, Oren R, Johnsen KH et al. (2006) Ice storms and management practices interact to affect current carbon seques- tration in forests with potential mitigation under future CO2 atmosphere. Journal of Geophysical Research-Atmospheres, 111, D15103, doi: 10.1029/2005JD006428/. McNulty SG (2002) Hurricane impacts on US forest carbon sequestration. Environmental Pollution, 116, S17–S24. Meyer DG, Butler DG (1993) Statistical validation. Ecological Modelling, 68, 21–32. Meyers TP (2001) A comparison of summertime water and CO2 fluxes over rangeland for well watered and drought condi- tions. Agricultural and Forest Meteorology, 106, 205–214. Moncrieff JB, Mahli Y, Leuning R (1996) The propagation of errors in long-term measurements of land-atmosphere fluxes of carbon and water. Global Change Biology, 2, 231–240. Norman JM, Campbell GS (1989) Canopy structure. In: Plant Physiological Ecology Field Methods and Instrumentation (eds Pearcy RW, Ehleringer JR, Mooney HA, Rundel PW), pp. 301–325. Chapman & Hall, New York. Novick KA, Stoy PC, Katul GG et al. (2004) Carbon dioxide and water vapor exchange in a warm temperate grassland. Oeco- logia, 138, 259–274. Odum EP (1971) Fundamentals of Ecology. W. B. Saunders, Phila- delphia. Oosting HJ (1942) An ecological analysis of the plant com- munities of Piedmont, North Carolina. American Midland Naturalist, 28, 1–126. Oren R, Ewers BE, Todd P et al. (1998) Water balance delineates the soil layer in which moisture affects canopy conductance. Ecological Applications, 8, 990–1002. Oren R, Pataki DE (2001) Transpiration in response to variation in microclimate and soil moisture in southeastern deciduous forests. Oecologia, 127, 549–559. Oren R, Sperry JS, Katul G et al. (1999) Survey and synthesis of intra- and interspecific variation in stomatal sensitivity to vapor pressure deficit. Plant, Cell and Environment, 22, 1515–1526. Palmroth S, Maier CA, McCarthy HR et al. (2005) Contrasting responses to drought of the forest floor CO2 efflux in a loblolly pine plantation and a nearby oak-hickory forest. Global Change Biology, 11, 421–434. Pataki DE, Oren R (2003) Species differences in stomatal control of water loss at the canopy scale in a mature bottomland deciduous forest. Advances in Water Resources, 26, 1267–1278. Pataki DE, Oren R, Katul GG et al. (1998a) Canopy conductance of Pinus taeda, Liquidambar styraciflua and Quercus phellos under stomatal conductance of Pinus taeda L. Trees to stepwise reductions in leaf area. Journal of Experimental Botany, 49, 871–878. Peters AJ, Ji L, Walter-Shea E (2003) Southeastern US vegetation response to ENSO events (1989–1999). Climatic Change, 60, 175–188. Pielke RA, Avissar R, Raupach M et al. (1998) Interactions between the atmosphere and terrestrial ecosystems: influence on weather and climate. Global Change Biology, 4, 461–475. Poggi D, Katul GG, Albertson JD (2004a) Momentum transfer and turbulent kinetic energy budgets within a dense model canopy. Boundary Layer Meteorology, 111, 589–614. Poggi D, Porporato A, Ridolfi L et al. (2004b) The effect of vegetation density on canopy sub-layer turbulence. Boundary Layer Meteorology, 111, 565–587. Porporato A, D’Odorico P, Laio F et al. (2002) Ecohydrology of water-controlled ecosystems. Advances in Water Resources, 25, 1335–1348. Ramos da Silva R, Bohrer G, Werth D et al. (2006) Sensitivity of ice storms in the southeast US to Atlantic SST – insights from a case study of the December 2002 storm. Monthly Weather Review, 134, 1454–1464. Richardson AD, Hollinger DY, Burba GG et al. (2006) A multi-site analysis of uncertainty in tower-based measurements of carbon and energy fluxes. Agricultural and Forest Meteorology, 136, 1–18. Roberts J (1983) Forest transpiration: a conservative hydrological process? Journal of Hydrology, 66, 133–141. Scha¨fer KVR, Oren R, Lai CT et al. (2002) Hydrologic balance in an intact temperate forest ecosystem under ambient and elevated atmospheric CO2 concentration. Global Change Biology, 8, 895–911. Stoy PC, Katul GG, Siqueira MBS et al. (2005) Variability in net ecosystem exchange from hourly to inter-annual time scales at adjacent pine and hardwood forests: a wavelet analysis. Tree Physiology, 25, 887–902. Wear DN, Greis JG (2002) The southern forest resource assess- ment summary report. In: Southern Forest Resource Assessment. Gen. Tech. Rep. SRS-53. (eds Wear DN, Greis JG), US Depart- ment of Agriculture, Forest Service, Southern Research Station, Asheville, NC. Webb EK, Pearman GI, Leuning R (1980) Correction of flux measurements for density effects due to heat and water vapour transfer. Quarterly Journal of the Royal Meteorological Society, 106, 85–100. Wilson KB, Baldocchi DD (2000) Seasonal and interannual varia- bility of energy fluxes over a broadleaved temperate decid- uous forest in North America. Agricultural and Forest Meteorology, 100, 1–18. Wilson KB, Baldocchi DD, Hanson PJ (2001) Leaf age affects the seasonal pattern of photosynthetic capacity and net ecosystem exchange of carbon in a deciduous forest. Plant, Cell and Environment, 24, 571–583. Wilson KB, Goldstein A, Falge E et al. (2002) Energy balance closure at FLUXNET sites. Agricultural and Forest Meteorology, 113, 223–243. Wilson KB, Meyers TP (2001) The spatial variability of energy and carbon dioxide fluxes at the floor of a deciduous forest. Boundary-Layer Meteorology, 98, 443–473. Xie L, Pietrafesa LJ, Wu KJ (2002) Interannual and decadal variability of landfalling tropical cyclones in the southeast coastal states of the United States. Advances in Atmospheric Sciences, 19, 677–686. Xie L, Yan TZ, Pietrafesa L (2005) The effect of Atlantic sea surface temperature dipole mode on hurricanes: implications for the 2004 Atlantic hurricane season. Geophysical Research Letters, 32, L03701, doi: 10.1029/2004GL021702/. Appendix A: separating evaporation from transpiration using simplified models and EC measurements Transpiration is not measured directly by the EC sys- tem, which measures entire ecosystem fluxes, but esti- mates are desirable for ecosystem intercomparisons and to elucidate canopy physiology. We used a simple radiation transfer model (Campbell & Norman, 1998) to separate E from T and, thus, gc, and estimated an average gs for all leaves in the canopy by dividing gc by L [Eqn (B2)]. The model uses an exponential radiation extinction model after Beer’s Law tðcÞ ¼ exp½ ffiffiffiffiffiffiffiffialeafp KðcÞPAIOc; ðA1Þ where c is zenith angle, aleaf is leaf absorptivity taken to average 0.5 across the photosynthetically active and near infrared bands (Campbell & Norman, 1998), PAI is plant area index (i.e. the sum of LAI and stem area) and Oc is a leaf clumping factor taken to be 1 at OF (i.e. leaves are randomly distributed), 0.6 at PP, and 0.84 at HW after HW measurements at Oak Ridge, TN (Bal- docchi & Meyers, 1998). K is the extinction coefficient for an ellipsoidal leaf distribution KðcÞ ¼ ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi x2 þ tan2 c p x þ 1:774ðx þ 1:182Þ0:733 ; ðA2Þ where x is the ratio of horizontal vs. vertical canopy projected area, taken to be 0.7 at OF (Novick et al., 2004), 1.64 at PP (Luo et al., 2001), and 1 at HW. The relation- ship between radiation that penetrates that canopy when the canopy is inactive (i.e. when L is near 0 at OF and HW, and when Ta is less than 10 1C at PP) and measured LE is used to model E. T was determined by subtracting modeled E from measured ET. We note that such an analysis is not intended to replace rigorous analyses of water balance at the plot level, rather it is intended to provide a method for acquiring variables of ecological interest from EC-measured ET. varying atmospheric and soil water conditions. Tree Physiology, 18, 307–315. Pataki DE, Oren R, Phillips N (1998b) Responses of sap flux and Appendix B: PM equation and models for water vapor and heat conductance The PM equation, after taking into account I (Wilson & Baldocchi, 2000) is LE ¼ lgvS Rn  G  I  lgvD=palgvS þ cpgH   þ lgvD pa ; ðB1Þ gv and gH require modeling. gc can be defined as gc ¼ gs; L; ðB2Þ where gs is stomatal conductance modeled after Jarvis (1976) and modified by Oren et al. (1999) for stomatal sensitivity to D gs ¼ ðaPPARþ bPÞð1 m log½DÞ; ðB3Þ m takes a theoretical value of 0.5–0.6 (Oren et al., 1999). Conductance at the leaf boundary layer (gb) is a combination of forced convection and free convection owing to the temperature difference between leaf and surrounding air. Forced convection can be considered dominant if the ratio between the Grashof number (Gr, the ratio of buoyant and inertial forces to squared viscous force, Gr ¼ ad3ðTc  TaÞ=Tku2, where a is accel- eration due to gravity, Tk is Ta in degrees Kelvin, u is the kinematic viscosity and d is the characteristic length scale of the leaves) and squared Reynolds number (Re, the ratio between inertial and viscous forces, Re ¼ ucd=u) is small. d for typical P. taeda needles was estimated using leaf width measurements to be 0.75mm (Campbell & Norman, 1998). Using leaf temperature measurements from an infrared temperature sensor (Model 4000, Everest Interscience, Tuscon, AZ, USA) during the day during a typical summer period (May 2004), Gr/Re averaged less than 0.01. Thus, we can simplify gb to both sensible heat (gHb) and latent heat (gvb) by considering only the forced convection case which, when using typical values, the thermal diffusiv- ity and density of air (Campbell & Norman, 1998) equal gHb ¼ 0:135 ffiffiffiffiffi uc d r L; gvb ¼ 0:147 ffiffiffiffiffi uc d r L; ðB4Þ where gHb is boundary layer conductance to sensible heat, gvb is boundary layer conductance to water vapor, and uc is canopy wind speed, the estimation of which is described below. Atmospheric conductance to sensible heat (gHa) and water vapor (gva) in the turbulent surface layer are equal gHa ¼ gva ¼ kr^u ln zdzH   þCH h i ; ðB5Þ where k is the von Karman constant (5 0.4) and zH is roughness lengths for heat (Campbell & Norman, 1998). The nonlinear atmospheric stability term (CH) tends to 0 for neutral conditions. gc [i.e. gsL, Eqn (B2)] is approximately an order of magnitude less than gva and gvb for the forested eco- systems and twice as large as gva and gvb at OF. However, gHb and gHa are roughly equal and depend on two different terms related to the wind speed, uc and u * , respectively. Thus, for the purposes of PM model differentiation (Appendix D), gv was simplified as gc, but for flux gapfilling the series conductance gv ¼ gvagvbgvc=ðgvagvb þ gvagvc þ gvbgvcÞ; was used. gH was taken to be the series combination of gHa and gHb for both flux gapfilling and PM model differentiation, gH ¼ gHagHb=ðgHa þ gHbÞ. Canopy mean wind speed model uc was modeled for different canopies using first- order closure principles assuming a constant mixing length (l) inside the canopy as described in (Katul et al., 2004), who found that first-order closure models match measured values as well as higher-order models if l is a priori specified. Using first-order closure princi- ples, the turbulent diffusion coefficient (Km) is modeled as Km ¼ l2 @ u @z  ; ðB6Þ the momentum flux is modeled as u0w0 ¼ Km @ u @z ; ðB7Þ and the mean momentum budget is given by @u0w0 @z ¼ CdPADðzÞu2; ðB8Þ where Cd is the drag coefficient assumed constant at 0.2, u0w0 is the momentum flux (equal to u2 at the top of the canopy) and measured u is a specified upper boundary condition. l is specified as 0.2h, z is height. The system of Eqns (B6)–(B8) has three unknowns (u0w0, Kt and u) and can be readily solved using standard numerical proce- dures. An appropriately weighted uc is obtained by multiplying the u profile obtained from (B7)–(B8) by normalized PAD. Appendix C: an analysis of radiation balance closure Plausible reasons for the observed lack of energy bal- ance closure at FLUXNET sites (e.g. Wilson et al., 2002) include footprint differences between radiometers and EC-measured fluxes, instrument bias, neglected storage sinks, high frequency losses, and advection. These explanations can be sequentially negated, but the latter is more complicated. The differing foot- r 2006 The Authors Journal compilation r 2006 Blackwell Publishing Ltd, Global Change Biology, 12, 2115–2135 We find a relationship between I and u * (Fig. 10a), as found in other studies (Wilson et al., 2002) suggesting that either a lack of turbulent transport or diminished role of turbulent transport (or both) are responsible for the magnitude of I. At the diurnal time step, I was clearly related to C and increased during unstable conditions when convection dominates turbulent trans- port (Fig. 10b). If convective cells form, EC measure- ments may effectively sample both the biosphere– atmosphere interface, as well as entrainment from the top of the atmospheric boundary layer. These mesoscale atmospheric events act on a time scale longer than the typical 12 hour EC averaging period, consistent with the notion that low-frequency events are the largest con- tributors to I. The top of the boundary layer will be relatively dry at times (which would decrease measured LE), but it will almost certainly be colder than the surface layer. Ac- cordingly, one might expect EC-measured H to depart more from the true surface flux than LE if convection explains the bulk of I as hypothesized. Also, H is an active scalar that impacts the buoyant production of TKE and its estimation may be more impacted by convection. However, there is no consistent way to test for the specific effects of convective transport given our measurements and there is no agreement within the FLUXNET community regarding how or if corrections should be made to account for I. Future research should investigate any relationships between the state of the entrainment zone and I. When interpreting our results, the true surface flux of H1LE may be 20–30% higher than EC measurements (the magnitude of I), but it is likely that the under- estimation in LE is less than 1/2 of I and lower than 10–15% if the top of the boundary layer may be rela- tively wet or dry but is consistently cold. (Additional support for this argument is the good agreement be- tween sapflux and EC measurements during dry peri- ods as found at PP.) Also, the length scale of the convective cells is on the order of the boundary layer height (here commonly 1000m; Juang et al., 2006) and any effects would impact our measurements equally because towers are separated by only 750m. Thus, when comparing ecosystems the difference – not the magnitude – of I among sites may be a better indicator of potential bias. This difference does not exceed 13%, so any comparative bias in ET is likely less than 6.5%. Bias in b, a minor component of this study, may be higher. Appendix D: perturbation analysis using the PM equation The total derivative of the PM Eqn (B1) is: dET ¼ @ET @Rn dRn I þ @ET @D dD II þ @ET @S dS III þ @ET @gH dgH IV þ @ET @L dL V þ @ET @gs dgs VI ; ðD1Þ and the corresponding partial derivatives are: @ET @Rn ¼ lgsLS cpgH þ lgsLS ; ðD2Þ @ET @D ¼ lgsL pa ; ðD3Þ @ET @S ¼  lcpgHgsLðlgsL þ pa½G  I  RnÞ paðcpgH þ lgsLSÞ2 ; ðD4Þ @ET @gH ¼ lcpgsLSðlgsL þ pa½G  I  RnÞ paðcpgH þ lgsLSÞ2 ; ðD5Þ −200 −150 −100 −50 0 50 (a) (b) (c) OF −200 −150 −100 −50 0 PP 2001 2002 2003 −200 −150 −100 −50 0 Growing season (April–September) HWδ ET G S fro m 2 00 4 (m m) R D S g L g Fig. 11 Same as Fig. 9d–f but for the change in growing season ET (dETGS) attributable to each term of the Penman–Monteith Eqn (B1) for old field (OF, a), pine plantation (PP, b) and hard- wood forest (HW, c) ecosystems. prints will not produce a consistent sign in I. We found good agreement between Q7 and Kipp and Zonen radiometers (biases rarely exceeded 5%) and the EC systems at PP and HW matched results from the Ameriflux roving system. Storage fluxes average nearly 0 at the daily time step. Flux-transporting eddy sizes are comparable with h in the forested ecosystems (far exceeding instrument separation), thus minimizing high-frequency losses. The night-time C filter is meant to reduce potential advection, but does not ensure that the role of advection-type events are isolated. @ET @L ¼ ðlgs½cpgHSðpa½G  I  Rn þ 2gsL½D  1Þ þ l 2g2sL 2S2½D  1 þ c2pg2HDÞ paðcpgH þ lgsLSÞ2 ; ðD6Þ @ET @gs ¼ ðlL½cpgHSðpa½G  I  Rn þ 2gsL½D  1Þ þ l 2g2sL 2S2½D  1 þ c2pg2HDÞ paðcpgH þ lgsLSÞ2 : ðD7Þ The analysis of (D1)–(D7) at the GS time scale gave similar results to the linear model with some differences due to averaging terms of the PM [Eqn (B1)] at the GS time scale (Figs 9 and 11). Changes in Rn, S (related to Ta) and gH between GSs changed ETGS by less than D, L, and gs when considering all ecosystems and all years (Fig. 11). However, changes in Rn, S and gH changed ET to a similar degree as D at OF in years without severe drought. These drivers were more im- portant (but minor) contributors to dETGS than LGS at HW, because LGS changed minimally among years. Thus, to a first order, the linear perturbation analysis captured the dominant drivers of ET change at the GS time scale. Undertaking a more rigorous analysis with the PM model did not alter results.