Scaling gross ecosystem production at Harvard Forest with remote sensing: a comparison of estimates from a constrai ned quantum-use efficiency model and eddy correlation

Two independent methods of estimating gross ecosystem production (GEP) were compared over a period of2 years at monthly integrals for a mixed forest of conifers and deciduous hardwoods at Harvard Forest in central Massachusetts. Continuous eddy flux measurements of net ecosystem exchange (NEE) provided one estimate of GEP by taking day to night temperature differences into account to estimate autotrophic and heterotrophic respi ration. GEP was also estimated with a quantum etliciency model based on measurements of maximum quantum effi ciency (Qmax>· seasonal variation in canopy phenology and chlorophyll content, incident PAR, and the constraints of freezing temperatures and vapour pressure deficits on stomata! conductance. Quantum efficiency model esti mates of GEP and those derived from eddy flux measure ments compared welJ at monthly integrals over two con secutive years (R 2 = 0·98). Analysis of video data collected from the ultralight aircraft indicated that the fraction of conifer cover varied from < 7% near the instrument tower to about 25% for a larger sized area. Al 25% conifer cover, the quantum efficiency model predicted an increase in the estimate of annual GEP of < 5% because unfavourable environmental conditions limited conifer photosynthesis in much of the non-growing season when hardwoods lacked leaves.


A B STRACT
Two independent methods of estimating gross ecosystem production (GEP) were compared over a period of2 years at monthly integrals for a mixed forest of conifers and deciduous hardwoods at Harvard Forest in central Massachusetts. Continuous eddy flux measurements of net ecosystem exchange (NEE) provided one estimate of GEP by taking day to night temperature differences into account to estimate autotrophic and heterotrophic respiration. GEP was also estimated with a quantum etliciency model based on measurements of maximum quantum efficiency (Qmax>· seasonal variation in canopy phenology and chlorophyll content, incident PAR, and the constraints of freezing temperatures and vapour pressure deficits on stomata! condu ctance. Quantum efficiency model estimates of G E P and those d erived from eddy flux measurements compared welJ at monthly integrals over two consecutive years (R 2 = 0·98).
Remotely sensed data were acquired seasonaUy with an ultralight aircraft to provide a means of scaling the leaf area and leaf pigmentation cha nges that affected the light absorption of photosynthetically active radiation to larger areas. A linear correlation between chlorophyll concentrations in the upper canopy leaves of four hardwood species a nd their quantum efficiencies (R 2 = 0·99) suggested that seasonal changes in quantum efficiency for the entire canopy can be quantified with remotely sensed indices of chlorophyll. An alysis of video data collected from the ultra light aircraft indicated that the fraction of conifer cover varied from < 7 % near the instrument tower to a bout 25 % for a larger sized area. Al 25 % conifer cover, the quantum effi ciency model predicted a n increase in the estimate of annual GEP of < 5 % because unfavourable environmental conditions limited conifer photosynthesis in much of the non-growing season when hardwoods lacked leaves.

INTRODUCTION
In most ecosystem models the general assumption is made that photosynthesis by an entire canopy i usually lightlimited and. as a result. that the assimilation rate can be assumed to be near linearly related to the absorbed flux density of photosynthetically active radiation (Jarvis & Leverenz 1983). This assumption has been widely accepted. variation in the efficiency by which PAR is absorbed under diffuse and direct-beam solar radiation at varying solar zenith angles being taken into account (Charles-Edwards 1982; Wang. Mc Murtrie & Landsberg 1992). If solar radiation data over periods of months or longer arc integrated. the linearity of canopy photosynthetic response to absorbed photosynthetically active radiation increases as differences in the beam fraction and solar zenith angle of incident PAR are balanced out (Wang & Polglase 1995). To date, suppo rt for a simple canopy light-absorption model has been provided mainly from ' >tudies on homogeneou!. canopies of single-species composition (Baker. Hesketh & Duncan 1972: McCree 1984Caldwell et al. 1986).
Natural vegetation is often a mixture of species or varies in structure. in contrac;t 10 a homogeneous monoculture of crops or tree plantations. h is desirable therefore to test how well the linear light absorption model applies to vegetation composed of a· mixture of species. We had such an opportunity at the Harvard Forest which is situated in the temperate forest region of the nonheastem USA. There, long-term studies of physiology and ecology have been combined with continuous eddy flux data since 1990 which provide measures of whole-system net ecosystem exchange (NEE) and gross ecosystem production (GEP) (Wofsy et al. 1993).
Calculations of net ecosystem production showed that Harvard Forest was accumulating 2-4 Mg C ha 1 annually in biomass and soil organic mauer (Wofsy et al. 1993). an interpretation that may apply to a wide range of temperate forests (Tans, Fung & Takahashi 1990). To date, few ecosystem models have been scaled regionally to evaluate these predictions. Most models are not dynamic, in the sense that vegetation boundaries, phenology, and weather conditions are assumed to be stable from year to year (Melillo et al. 1993;Potter et al. 1994). To make regional or global scale models more dynamic, there has been an effort to integrate satellite-derived data in a form useful for parametrizing and driving simplified ecosystem models. Current global satellite coverage can provide remotely sensed estimates of solar radiation, ambient air temperature, atmospheric humidity deficits and surface moisture conditions for areas of the order of 25-100 km 2 or larger (Dye & Goward 1993;Waring et al. 1993;Goward et ai. l 994a;Prince & Goward 1995). ln addition, spectral vegetation indices, calculated from the red (R) and near-infrared (NIR) portions of the spectrum, provide indicators of seasonal changes in the amount of green leaf cover (Kumar & Monteith 1982). Studies have shown that the normalized difference vegetation index [NOVI: (NIR-R)l(NIR + R)] is a linear to near-linear function of the fraction of photosynthetically active radiation (PAR) intercepted by many types of green vegetation (Asrar et al. 1984;Law & Waring J 994a,b;Goward er al. 1994a). Monthly integration of data is usually required to reduce errors to less than 5% in estimating solar radiation and to obtain c loud-free estimates of canopy properties a nd surface moisture conditio ns by remote sensing (Goward et al. l 994a). In combination, these remotely sensed variables are employed to provide time-integrated estimates of photosynthesis across landscapes for broadly similar types of vegetation, given knowledge of the apparent quantum effic iency of canopies and the extent to which environmental factors constrain stomata! conductance (Running et al. 1989;Sellers et al. 1992;Nemani, Pierce & Running 1993;Vermaetal. 1993).
Al Harvard Forest, all the requisite data required to drive the photosynthetic component of a whole-system model were available. A tower equipped with eddy-flux instrumentation provided an estimate of gross ecosystem production (GEP) that could be compared to independent estimates of GEP derived from a model based on measurements of maximum quantum efficiency COmux>· seasonal variation in canopy leaf area index, leaf chlorophyll content. incident PAR. and restriction on C0 2 diffusion through stomata imposed by freezing temperatures, vapour pressure deficits and other environmenta l factors.
From an Oregon State University ultralight aircraft, video and spectro radiometer data were acquired seasonally to provide the necessary spatial resolution to c haracterize variation in conifer and deciduous hardwood cover and to document pigment and leaf area changes in the hardwood component of the canopy. With remotely sensed data obtained from the aircraft we gained an independent measure of phenology to compare with that obtained from ground observations. In this paper we compare predictions of monthly gross ecosystem production derived from the quantum efficiency model with those determined by edd~ correlation and suggest how remotely sensed data can ai( in scaling stand-level observations to larger areas.

Site description
The study site was located in the Prospect Hill tract o Harvard Forest, north-central Massachusetts (42.54° N 72. 18° W; elevation 340 m). The forest. composed of 50 to 70-year-old trees, regenerated following a hurricane ir 1938 and has reached a height of 20-24 m and supports a leaf area index of between 3.5 and 4.0 during the growing season. At present, dominant species are red oak (Quercu~ rubra L.) and red maple (Acer rubrum L.), which together make up 80% of the total 30 m 2 ha-1 of basal area. Some black birch (Betula lenta L.), white birch (Betula papyrifera Marsh.), yellow birch (Betula alleghaniensis Britton) . black cherry (Prunus serotina Ehrh.), white pine (Pinus strobus L.). and hemlock (Tsuga canadensis (L.) Carr.) are also present. but together represent less than 20% of the total basal area and less than I 0% of the 1300 stems ha-1 • The c limate is cool temperate (January mean weekly temperature -6 °C , July mean 20 °C). and humid, with precipitation distributed evenly throughout the year (annual mean 1100 mm). Summer drought rarely occurs. T he terrain is moderately hilly. about 95% forested, with the nearest paved roads> I km away and smaU towns >10 km away. Soils are mainly of the Gloucester series (fine loamy, mixed mesic Typic Dystrochrept) with a surface pH of 3.8 and a subsurface pH of 4.8 (Pererjohn et al. 1994 ).

Meteorological data
Meteorological sensors were installed at the top of a 30 m tower (Rohn 25G), 6-10 m above the top of the forest canopy (Wofsy et al. 1993 ). Data were logged on a computer situated in a climate-controlled shack 20 m from the base of the tower. Throughout the investigation, the system logged 60 min mean air temperature measured with an aspirated thermistor, relative humidity determined with an electrical-capacitance sensor, and horizontal photosynthetically active photon flux density obtained with a pair of silicon quantum sensors one above the canopy and the other below at a height of 8 m. The soil-surface temperature was a lso recorded using an array of potted thermistors. An integrated monthly measure of the fraction of light intercepted by the canopy (f 1 PAR) was subsequently calculated as Ithe fraction of inc ide nt PAR that was transmitted through the canopy.
The observations were interrupted by occasional extended gaps as a result of equipment failure, and by more frequent gaps associated with calibration, maintenance or data transfer. The meteorological data for 199 I were comparatively patchy, with 20% missing, while 10% were missing for 1992. In the current analysis we replaced miss-ing meteorological data with the mean for that hour plus and minus 2 d.

Eddy flux measurements
Eddy-correlation measurements of net ecosystem C0 2 exchange were initiated at Harvard Forest in April 1990 on the same 30 m tower where meteorological instruments were installed. A three-axis sonic anemometer (Applied Technologies) and a closed-path infrared gas analyser (IRGA, LiCor 6262 or 6251) were used to sample simultaneously the venical wind speed and the C0 2 mixing concentration at 30 m. The raw signals were digitized at 4 Hz and these data logged to disk for subsequent analysis. Air was drawn to the instrument shelter down a 50 m tube and through the IRGA at 6 to 8 dm:i min-1 , introducing a lag of several seconds which we adjusted for in the flux calculation (Wofsy et al. 1993: Fan et al. 1995. In computing the flux we accounted for the orientation of the streamlines by rotation to the plane where the mean venical wind speed was zero (McMillen 1988).
A eries of simulations, laboratory tests. and spectral analyses indicated a small underestimation of flux caused by the reliance on a closed path sensor for the C0 2 measurement (overall 90'k respon e faster than 1 s). We corrected for this error by scaling the measured C0 2 flux by the ratio of the mea ured raw sensible heat flux to the sensible heat flux calculated after filtering the temperature signal to simulate an instrument with limited high-frequency response (Leuning & King 1992: Wofsy et al. 1993). This correction was generally small (< 10%) because most flux is carried by eddies with a frequency of O· l to 0·002 Hz. Uncertainty associated with calibration was less than 10%. Additionally, we anticipate <IOo/r underestimation of flux caused by finite sensor path length and spatial separation (Baldocchi & Meyers 1991: Lee & Black 1994. Data gaps occurred occasionally. particularly following lightning strikes. For these gaps we substituted values from hourly averages determined from the closest 5 d set of data. Change'> in atmo'>pheric stability may cau. e shifts in the quantity of C0 2 stored in the air space below the sensor array. decoupling the flux through the eddy plane from biotic activity. We accounted for storage by monitoring the change in C0 2 in the air <,pace beneath 30 m with a second IRGA to sample '>equentially inlets at 29, 24, 18. 12. 5. 3. I and 0.5 m. We then applied the rate of C0 2 storage change in combination with the eddy flux measurement to calculate net ecosystem exchange (NEE) (Wofsy et al. 1993).
In the convention of eddy-flux measurements. photosynthesis is a negative flux into the system and respiration is a positive flux out. Net ecosystem exchange me~sured during daylight hours include!. gross photosynthesis (P 8 ) photorespiration (Rp). maintenance respiration (R.,,) and growth respiration (Ri;) of autotrophic plants as well as all sources of heterotrophic respiration (Rh). DayNEE=Ps+Rp+R.,,+Rg+Rh.

Scaling gross ecosystem production 1203
At night the photosynthetic terms, P & and RP, are absent.

Night NEE= Rm+ Rg +Rh = Re'
(2) where Re is the total ecosystem respiration. Rm and Rh are largely controlled by temperature, allowing us to approximate Re during daylight periods as a function of temperature. The relation between Night NEE during well mixed periods (Fan et al. 1995) and soil temperature (T_) was Gross ecosystem production (GEP) is then calculated as Day Re is typically 20-30% of Day NEE. Consequently. large uncertainties in Day Re have a relatively small effect on the accuracy of GEP.

Vegetation measurements
The basal area of tree species was determined by measur- to e-,timate the leaf area index (LAI) of deciduous hard-wood<, and the relative contribution of each species. Fewer than 10 white pine tree tern<; were present within the o;ampling area and were excluded from the analysis. Independent estimates of maximum and minimum LAI, which did include the conifers. were made in the same area using a LiCor LAI 2000 Plant Canopy Analyzer (Welles & Norman 1991 ). Phenological observations were provided by John O'Kcefe at Harvard Forest, who recorded the dates of initial bud break. when leaves were at 75 and 95% of full expansion, when autumn colour was first observed, and when I 0-25 and 98% of leaves had been shed. These data were available for all major tree species but we referenced whole-canopy phenology to only the two dominant species, red oak and red maple. Apparent maximum quantum efficiency was determined in AuguM when leaves were fully expanded on four hardwood '>pecie~ from samples of representative leaves growing near the top of the canopy. A Li-C?r 6200. gas exchange '>ystem (Lincoln. NE. USA) equipped with a 0·25 dm 1 cuveue and a PAR !>ensor was used to obtain measurement'> of dark respiration, net photosynthesis and photon nux den'>ity. Neutral shade cloth incrementally reduced the light level incident on the leaf surface. allowing for repeated measurements of the same leaves at different light levels. Leaves were allowed to acclimate to the light levels for at least 5 to 8 min before instantaneous ga!. exchange measurements were taken. For red oak and red maple, measurements were made on I 0 leaves located in the upper ca:nopy and an additional 10 in the lower canopy strata (five leaves at both heights on two trees of each species). For other specie!>, measurements were made on five leaves from each strata (one tree per species). Apparent maximum quantum efficiency and maximum photosynthetic rates were determined from least-squares fitting of the basic photosynthesis, light-saturation equation described by Ogren & Evans ( 1993): where P is the photosynthetic rate, PAR is the inc ident PAR irradiance, Qmax is the apparent maximum quantum efficiency, Pm is the maximum photosynthetic rate at light saturation, and C is the curvature of the light saturation curve at ambient C0 2 and 0 2 partial pressure observed in our field analysis. The apparent maximum quantum effic iency is determined from the linear portion of the C0 2 uptake respo nse as incident PAR increases from zero. Gross photosynthesis (Pg), leaf maintenance respiration (R 01 ), and photorespiration (Rp) are incorporated such that where PAR 1 represents incident PAR as it approaches 0 and PAR 2 refers to the case when PAR is 0. Because the cuvette temperature was constant. we assume that Rm 1 =

Rm2·
Because our method for determining quantum efficiency in the field illuminated leaf surfaces predominantly from only the upper side. the Qmax values should be about half those determined o n leaves receiving illumination from all sides within an integrating sphere (Leverenz & Oquist 1987: Ogren & Evans 1993. Chlorophyll analyses were performed on deciduous leaves collected in August 1991 when leaves were fully green and in September 1991 when autumn colour was first noted using N, N-Di-methylformamide extraction (Moran & Porath 1980;Moran 1982;Inskeep & Bloom 1985). Leaf mass per unit area (LMA) was also determined to allow expression of results in mg chlorophyll per gram of tissue and mg chlorophyll per m 2 of leaf area. For red oak and red maple, 10 leaves were sampled and analysed from the upper and lower canopy strata. For othe r species, five leaves were collected from each stratum for analysis.

Quantum-use efficiency model
The quantum-use efficiency model assumes that photosynthetically active radiation (PAR) absorbed by the vegetation is converted into photosynthate based on the quantum efficiency of all leaves (Qmux• g carbon fi xed per mol photo n). The maximum quantum efficiency (Qmax) for the canopy was determined from photosynthesi!> measurements on upper canopy foliage of key species near the tower. weighted by the fraction of the total basal area that each species represented within the 50 x 50 m vegetation plo t. We constrained the estimates of photosynthesis in this study to take into account partial or complete closure of stomata attributed to environme ntal factors , notably below-freezing conditions and extremes in vapour pressure deficits (VPD) (Law & Waring l 994a). The model predicts gross ecosystem production (GEP) over a given time integral: (7) where PAR is incident photosynthetically active radiation and ftPAR is the fraction of incident PAR intercepted b) vegetation. Qc is the constrained quantum-use efficienc) derived from multiplying Qmax by values ranging from 0 (complete constraint on photosynthesis) to l ·O (no environmental constraints), a convention which follows earlier work in Oregon on a variety of tree-and shrub-dominated landscapes where drought was an additional constraint on photosynthesis (Law & Waring I 994a;Runyon et al. 1994).
The constrained quantum-use efficiency model wa!> driven with hourly meteorological data from the tower in Harvard Forest and parameterized separately for conifers and hardwoods, although the same Qmax was applied to conifers for lack of other data. For the dec iduous hardwoods, we spec ified the timing of leaf-off and assumed that no PAR was intercepted during that period. Phenological data from Harvard Forest indicated that 10-25% leaf-off occurred on 21 October 1991 and on 23 October 1992, and bud break occurred approximately on 6 May 1991 and o n 13 May 1992. From the onset of deciduous leaf colour in autumn to the onset of leaf fall, we incrementally decreased Qc from a maximum to zero to account for progressive reductions in chlorophyll during this transition period.
When temperatures were less than or equal to -2°C, we assumed no light was utilized for photosynthesis for 24 h so that Qc= 0. Each additional hour below this limit reduced Qc to zero for another hour. The fractionaJ reductio n in Q . .. caused by high VPD was applied hourly. We assumed no reduction for VPD less than l ·5 kPa and a linear reduction in the scalar between 1 a nd 0 for the fractional constraint between l ·5 and 2·5 kPa (Runyon er al. 1994;Law & Waring 1994a). The Qc was determined hourly with a statistical programming package (SAS Institute, Inc .. Raleigh, NC, USA) that incorporated meteorological data and the physiological constraints (Law & Waring 1994a).
Mean monthlyf 1 PAR• calculated from mean dailyf 1 PAR between 1000 a nd 1400 h, was combined with incident PAR to estimate intercepted PAR (!PAR= fiPAR . PAR). Because the forest was a mixture of hardwoods and conifers, we determined the fraction of totalflPAR for both components from the classification of cover fractions in video images obtained with the aircraft from the 1000 x 1000 m area around the tower (see next section for details).
Finally, GEP was calculated from Eqn 7 and summed momhly. These constrained quantum-use efficiency estimates of monthly GEP were regressed against those derived from eddy flux data. We also ran the quantum-use efficiency model with VPD constraints from eddy flux measurements made at Harvard Forest and compared the resulting monthly estimates of gross ecosystem production with the VPD response described earlier.

Remotely sensed measurements
From a Quicksilver Model GT-500 ultraJight aircraft we collected remotely sensed data from aJtitudes of 300-400 and IOOO m and at an air speed of 55 km h-1 over a 1000 x IOOO m area centred around the eddy flux tower. Flights were made between 0800 and 1200 on October IO and April 10 199 1. and on May 19 and July 6 1992. Video images were acquired continuously with a Sony Model TR5 8mm camera (230 000 images on a 2 h tape). Selected images were transferred to computer with a Matrox graphics board, and analysed with Resource Imaging Graphics System software (version 3·24, 1989; Decision Images. Inc. , Princeton, NJ. USA). The spatial coverage of video images corresponds to a square with sides approximately equal to the altirude. Spatial resolution from an altitude of IOOO m altitude was 4 m 2 . In April, when hardwoods were leafless. a supervi ed classification was performed on a total of I 0 reference . ites where patches of hardwoods and conifers were clearly evident. Image analyses to separate these two types were made over an increasing area ( 125 m x 125 m, 250 mx250 m. 500 m x500 mand IOOOm x IOOOm) centred on the eddy flux tower. Analyses were repeated three time to obtain standard errors of estimates.
Spectral reflectance measurements were collected at lower altitudes between 300 and 400 m to resolve differences between areas with conifer and hardwood cover. A Spectron Engineering SE-590 spectroradiometer (Denver. CO) was interfaced with a data logger. The SE-590 provides continuous spectra between 380 and I I 00 nm ai nominal 10 nm spectral resolution. In natural sunlight. instrument sensitivity is limited below 400 nm and above 900 nm (Goward, Huemmrich & Waring 1994b). We therefore limiced the processing of data to the 400-900 nm range for all analysis reported in this paper. Equipped with a I 0 lens, the circular resolution (R) in m 2 was determined from the formula: where 8 i!. the view angle in degrees and h is the altitude in m. At 300 m, the spatial resolution within a circular area was 2 1 ·5 m 2 with a diameter of 5·2 m. With a flight speed of 55 km h 1 and an instrument scan time of I s. a rectangular ground area of 5.2 m x 15·3 m wa' captured. During all but che last sampling period we did not have a global po:o,itioning system (GPS) on board. This made it necessary to cr,ss-link time codes on the video and specrroradiometer d~ta to identify areas ampled.
Spectroradiometer data were edited to extract average reflectance vaJues in the red (655-665 nm) and nearinfrared (785-795 nm) acquired over specific types of cover identified from the time-linked video images. Software developed by Moon Kim (Goddard Space Flight Center, MD) was utilized in processing the reflectance data. The normalized difference vegetation index was calculated separately for areas dominated by conifer and by hardwood cover. Beneath these two distinct types of canopy, the fraction of intercepted light <tiPAR) was mea-Scaling gross ecosystem production 1205 sured under clear sky conditions in mid-afternoon following morning flights on three out of four dates: when deciduous trees were leafless in April, when they had begun leaf ~xpansion in May. and when leaves were fu lly developed m July. No f IPAR data were gathered in October when leaves were in full autumn colour.
To compare remotely sensed estimates of canopy leaf area andJiPAR• eparate estimates of fiPAR for hardwoods and conifers were obtained with a single quantum sensor (Li-Cor. Lincoln, NE, USA) under clear sky conditions over an area of 0·5 ha. Seasonally. sample sizes ranged from 410 to 765 under the deciduous canopy and from J 14 to 237 under the more unifonn conifers. Light measurements were not made in October when canopy leaf area was still high but chlorophyll pigments had decreased. For reference. incident PAR was measured in the open before and after completion of mea urements along the transects. The fraction of transmitted PAR was determined after log transformation to normalize for direct light penetration through canopy gaps (Lang & Yueqin 1986). The mean transformed value was back-transformed and subtracted from I to provide the mean fraction of intercepted PAR if IPAR). which was regressed against respective NOVI values.
The seasonal variation inf IPAR determined from a pair of stationary quantum sensors appeared similar for both 1991 and 1992 (Fig. I ). Higher values off IPAR recorded in January than in April reflect differences in midday solar zenith angle. with more light interception by tree boles in January. The full canopy intercepted >80% of incident PAR from May to September. The measured light intercepted during the middle of the growing season was equivalent to that expected for a hardwood forest with a LAI of 3.2, applying the Beer-Lambcn Law with an assumed extinction coefficient of 0.5 . The values were similar to estimatei. of LA I made with a LiCor 2000 Plant Canopy Analyzer. The Plant Canopy Analyzer samples a hemi-sphericaJ area. and may include the contribution of some scattered conifers in the estimates of projected LAI. Litter trap~ provided a range of LAI estimates between 2.7 and 4.1 for hardwoods.
The NOVI values for areas with pure hardwoods decreased from a maximum of 0.85 on 6 July 1992 to 0.6 1 on I 0 October 199 1 before a major decrease in ftPA R was recorded near the tower ( Table I   beneath the two types of vegetation is presented in Fig. 2. We emphasize that the regression equation for Fig. 2 was developed for green foliage only and is not appropriate for application during leaf senescence because of the influence of chlorophyll on red reflectance. These restricted c learweather measurements were not appropriate for comparison with daily mean values of transmitted PAR acquired from sensors near the tower. The maximum quantum efficiency determined in midsummer varied nearly 2-fold, from a low value of 0·026 mol C mol-1 photon for Acer rubrum to a high of 0·047 mo! C mo1-1 photon for Berula alleghaniensis (Table 2). When weighted by the representative basal area. which was 50% Quercus rubra and 30% for Acer rubrum, the average quantum efficiency at mid-summer was 0·0298 mo! C mo1-1 photon or 0·36 g C mol-1 photon C Omax>· We observed on I 0 October 199 1 that the NOVI value~ acquired from the ultralight aircraft would suggest that < fully green canopy had been reduced in leaf area since Jul) by -50% (from Fig. 2: NOV I = 0·61 is equivalent to ar f iPAR vaJue of0·63, which corresponds to an LAI -2.0).
To account for seasonal differences in incident solat radiation, we calculated the ratio of the reflectance value~ recorded in the red (655-665 nm) and near-infrarec (785-795 nm) bands against lhe total reflectance summed for all bands (400-900 nm). This normalizing procedure showed that reflectance in the red band increased from 3 tc 6% between July and October, indicating a comparable reduction in chlorophyll absorption. Over the same period. near-infrared reflectance decreased by less than 30%, from 35 to 25·5% (R. W. Mccreight, unpublished results). Phenological observations confirmed that only 10-25% of the leaves had fallen by 3 October 1991. The video image acquired on I 0 October also confirmed lhat a dense canopy of hardwoods leaves was still present but in nearly full autumn colour (Fig. 3). We therefore interpreted the decrease in the hardwood NOVI in October (Table I) to reflect mainly a reduction in chlorophyll absorption, which was confirmed by mid-September measurements that were nearly 50% below those measured in mid-August (Table  3).
We had hoped to identify species composition from video images acquired in October when hardwoods were in full colour. but were unsuccessful because colour differences varied with the precise stage in leaf senescence. An  (Fig. I). indicating that NDV I integrates chlorophy ll content as well as leaf area index (Table 3). alternative means of assessing changes in the maximum quantum efficiency of the entire canopy with remote sensing was available through use of the recently established near-linear relations chat exist between red reflectance (and NDVI) and chlorophyll concentrations in upper canopy leaves and between chlorophyll concentrations and photosynthetic capacity. These relationships have been confirmed for a variety of hardwood and conifer species with various spectral reflectance indices (Yoder & Waring 1994;Curran 1995: Gamon er al. 1995. To confirm the first relation we compared the average chlorophyll concentrations measured on upper canopy leaves collected from the four species of hardwoods in August (Table 3) with their corresponding values of maximum quantum efficiency ( Scaling gross ecosystem production 1207 canopy leaves was linear with an R 2 of 0·99 (Fig. 4). These results suggest that seasonal estimates of NDVJ or related reflectance indices may provide an integrated estimate of Qmax for forests of mixed composition. The constrained modelled quantum efficiency {Qc) for conifers shown in Fig. 5. remained above 0·3 g C mo1-1 photon from April to October in both 1991 and 1992. The YPD proved to be only a minor constraint. reducing monthly net photosynthesis by a maximum of 8% in May 1991and6% in June 1992. Freezing temperatures reached a maximum constraint of 82% for conifers in January 1991 and 86% in February 1992. The combined climatic constraints resulted in estimates of Qc for conifers which ranged from 0·07 g C mo1-1 photon in January to a maximum of O· 36 g C mor 1 photon in September 199 1, and from 0·05 g C mo1-1 photon in February to 0·36 g C mol 1 photon in September 1992 (Fig. 5). The derived Qc values were the same for hardwoods when green leaves were on the irees as for conifers, because the same quantitative environmental constraincs were applied. The quantum-use efficiency model estimates of GEP by hardwoods and conifers within the 125 x 125 m area around the tower during l 991 ranged from 0·45 g m-2 month-' in January to 351 g m-2 month-' in June (Table 4). For 1992, the range was from 0.55 g m 2 month-' in February to 304 g m-2 month-1 in July. The annual sums were 1467 g m-2 in 1991 and 1305 g m-2 in 1992. The hardwoods accounted for 1328 g m-2 in 1991. and 1177 g m 2 in 1992 (90% of total).
Comparison of monthly GEP calculated from the quantum-use efficiency model using the standard VPD constraints versus those calculated from the flux estimates of VPD constraints showed no significant difference between SD Chti Table 3 . Seasonal changet. in chlorophyll concentrations and specific leaf area (mg cm-2 ) (mgm-2 ) (SLA) with standard deviations (SD) detcnnined on 5-10 leaves collected from the top and lower canopy in August and 0·96   estimates. Monthly GEP estimates from the quantum-use efficiency model and from the flux residual approach were also similar throughout the year ( Figure 6, Table 4). A linear regression of monthJy GEP estimated using the two separate approaches showed a high correlation ( Fig. 7; 1991 R 2 = 0·97; 1992 R 2 = 0-99). The slope was slightly, but significantly, different from unity in 1992 (P = 0·01), but not in L 991 (P = 0· 11 ). The intercept was essentially zero in both years (P = 0·52 for both years).
When conifer cover was increased from 7% for the 125 x 125 m area to 25% representative of larger areas up to 1000 x I 000 m, the model predicted a maximum monthly increase in photosynthesis during the months of April and October of 14 to 18 g C m-2 when hardwood trees were leafless (data not shown). At these times temperatures were still above freezing and PAR was well above the light compensation point for conifers (Fig. l). Model estimates of annual GEP increased by onJy 4-5%, to 1525 g m-2 year-1 in 1991 and to 1369 g m-2 year-1 in 1992 by an increase in conifer cover from 7 to 25%.  (Table 3) were linearly related to the maximum quantum efficiency (Qm•~> determined on the same leaves in August (Table 2). y = -0·044 + 0·0002r; R 2 = 0·99. Sample size varied from 5 10 I 0 leaves with the higher numbcri. collected from the dominant oaks and maples.

DISCUSSION
The quantum-use efficiency model requires an accurate estimate of the apparent maximum quantum efficiency derived from cuvette gas-exchange measurements in the upper. most active canopy strata. We were fonunate to have cower access and to sample on days with high humidity so chat VPD had lictle effect on stomata and thus Qmax· In more arid environments. cuvettes with environmental controls would be required 10 obtain maximum values of quantum efficiency. T he weighting by species fraction of the total site basal area appears to be a reasonable approach Scaling gross ecosystem production 1209 to esti mating the average Q max· The oak and maple leaves collected in 1 itter traps for both 199 1 and 1992 represented 75-90% of the total leaf area collected whereas their basal area was 80% of the total (conifer basal area was -I 0% but is not listed in Table 3).
In more open canopies, or canopies with lower quantum efficiencies, the assumption of linearity between light interception and photosynthetic activity may not be as valid as a c urvilinear light-satu racion model (McCree . of quantum-use efficiency (QJ for conifer:. ranging from 0.07 g C mo1-1 pho1on in January 10 a maximum of0.36 g C mol 1 photon in September 1991. and from 0.05 g C mo1· 1 photon in February to 0.36 g C mol 'photon in September 19920.36 g C mol 'photon in September . 1992

400
.c 1984; Holl inger er al. 1994). With a monthly resolution. it seems appropriate to treat the canopy as a single unit, particularly when/ IPAR values are acquired continuously and include overcast and clear sky conditions. Other studies have demonstrated that in moderately dense deciduous hardwood canopies, near-linear relations exist between light conditions and leaf mass m-2 foliage, nitrogen content m-2 of foliage. and photosynthetic capacity (Ellsworth & Reich 1993). Similar reports have also been published for evergreen hardwoods (Hollinger 1989) and for conifers (Oren er al. J 986). suggesting that estimates of quantum efficiency may be similar for leaves throughout closed canopies (Wang & Polglase 1995).
We were concerned that high photosynthetic rates on one day might cause accumulations. of starch that would lead to photosynthetic inhibition on proceeding days (Luxmoore 199 1 ). From inspection of data, however, we found little indication that previous high rates of photosyn-thesis in the preceding 1 to 24 h affected observed rates (M. L. Goulden, unpublished results).
In applying the quantum efficiency model. we assumed that common physiological thresholds apply to all species. We also ignored the possibility of delayed response to subfreezing or abnormally high temperatures (Leverenz & -Oquist 1987;Bassow, McConnaugbay & Bazzaz 1994). We recognize that species differ in the rates in which their stomata respond to light fluctuations. YPD and other environmental factors (Schulze et al. 1994 ). Because hourly data are integrated over monthly periods, however, errors in determining physiological thresholds are less critical than for models that predict in shorter time steps.
ln this study, the major e nvironmental constraint on canopy photosynthesis was IPAR. Subfreezing tempera- tures had a significant effect only on the conifers during mid-winter at the time deciduous trees were leafless. Vapour pressure deficits were not extreme during the summer and therefore had JjttJe effect on Qc. VPD and subfreezing temperatures restrict photo ynthesis by 50-75% in shrub and evergreen forest vegetation in temperate regions such as the Pacific Northwest of the USA (Runyon et al. 1994;Law & Waring 1994a) and to a somewhat lesser extent on the South Island of New Zealand (Hollinger et al. 1994). Good agreement between the constrained quantum-use efficiency estimates of GEP and those determined from two years of eddy flux measurements further suppom the value of integrating at monthly time steps. It is not reasonable to expect model predictions to agree well with hourly or daily measurements of photosynthesis because varying quantities of direct and diffuse radiation influence photosynthetic efficiency differently. Diffuse light casts less shadow and therefore penetrates more efficiently through canopies. As a result, photosynthetic rates may not decrease as much as expected (Sheehy & Chapas 1976;Hollinger et al. 1994;Fan er al. 1995). Our approach may have partly solved this problem by obtaining .f tPAR values under a range of cloud conditions each month and by integrating model estimates of photosynthesis at monthly intervals.
We recogni.ced that flux measurements taken from a single stationary tower may not represent gas exchange from a large forest (Hollinger et al. 1994). When conifers are clumped in their distribution, the wind direction may have a major influence on their contribution to area-wide fluxes. Similarly, restricted area sampling off tPAR measurement<, can bias estimate!. of photosynthe<.is if the sampling b not representative of the larger area. Both of these problem!. can be solved in part by measuring fluxes and remotely sensing canopy properties from aircraft (Dejardins er al. : Spanner et al. 1994. ScaJing of the quantum-use efficiency model to larger areas ma) be accomplished through satellite or airborne eMimates off IPAR and Qm.,,· Satellite estimates off tPAR require atmospheric correction. in particular for visible wavelengths. Low-altitude airborne sensors provide an advantage by acquiring reflectance data with linle atmo-.,phere between the ground and the sensor. Data acquired from aircraft can also validate atmospheric corrections modelled for satellite data (Goward et al. 1994a). In regions where snow cover is likely, NOV I estimate!. would be misrepresentative of canopy properties. We recommend obtai ning minimum estimates off IPAR for the year by concentrating observations on clear days in the spring or autumn when deciduous trees are no1 in leaf to minimi1e effects of snow cover. Where evergreens dominate, :.easonal variation in leaf area is often less than 30% (Spanner et al. 1994 ).
The potemial exists for scaling the quanrum-use efficiency model 10 biomes and the globe, as suggested by Running & Hunt ( 1993). A physiologicaJ model (BIOME-BGC) is available to estimate Qc from climatic data and © 1995 Blackwell Science Ltd. Plant. Cell andEnvironment. 18. 1201-1213 Scaling gross ecosystem production 1211 certain biome-specific assumptions about nitrogen. carbon, and water movement through the system (Running et al. 1994). MeteorologicaJ data are also necessary as driving variables in this approach, and have been successfully estimated across mountainous topography where temporary weather stations were installed for vaJidation (Glassy & Running 1994). An aJternative would be to apply the simplified constraints of freezing 1emperature, drought. and high VPD to Qmax• using remote sensing to estimate the meteorological variables :is well as f IPAR and incident PAR (Goward et al. l 994a: Prince & Goward 1995).
We recommend that monthly resolution models of net photosynthesis and other ecosystem properties hould be tested where continuous flux measurements arc being made throughout the year. Harvard Forest is one of the first sites to a11emp1 these long-term measurements (Wofsy er al. 1993). To scale predictions to larger areas, remote sen:.ing of red and near-infrared reflectance. coupled with video analysis to re olve variation in deciduous and evergreen cover. showed promise as a viable approach at Harvard Forest.