Sie sind auf Seite 1von 11

Global Change Biology (1996) 2, 231-240

The propagation of errors in long-term measurements of land-atmosphere fluxes of carbon and water
J.B.MONCRrEFF*, Y. MALHI* andR. LEUNING +
*Jnstitute of Ecology and Resource Management, University of Edinburgh, Edinburgh EH9 3JU, UK. fComTnonwealth Scientific and Industrial Research Organization, Centre for Environmental Mechanics, PO Box 821, Canberra ACT 2601, Australia

Abstract For surface Ouxes of carbon dioxide, the net daily flux is the sum of daytime and nighttime fluxes of approximately the same magnitude and opposite direction. The net flux is therefore significantly smaller than the individual flux measurements and error assessment is critical in determining whether a surface is a net source or sink of carbon dioxide. For carbon dioxide flux measurements, it is an occasional misconception that the net flux is measured as the difference between the net upward and downward fluxes <i.e. a small difference between large terms). This is not the case. The net flux is the sum of individual (half-hourly or hourly) flux measurements, each with an associated error term. The question of errors and uncertainties in long-term flux measurements of carbon and water is addressed by first considering the potential for errors in flux measuring systems in general and thus errors which are relevant to a wide range of timescales of measurement. We also focus exclusively on flux measurements made by the micrometeorological method of eddy covariance. Errors can loosely be divided into random errors and systematic errors, although in reality any particular error may be a combination of both types. Systematic errors can be fully systematic errors (errors that apply on all of the daily cycle) or selectively systematic errors (errors that apply to only part of the daily cycle), which have very different effects. Random errors may also be full or selective, but these do not differ substantially in their properties. We describe an error analysis in which these three different types of error are applied to a long-term dataset to discover how errors may propagate through long-term data and which can be used to estimate the range of uncertainty in the reported sink strength of the particular ecosystem studied. Keywords: carbon balance, eddy covariance, enrors, long-term measurements, micrometeorology.
Received 18 August 1995; revision accepted 6 January 1996

1 Introduction
All long-term fluxes are estimated from a series of measurements made at shorter intervals. Even within a campaign of long-term flux measurements for carbon dioxide and water vapour there will be a range of scientific objectives relevant to differing periods over which answers are sought. At short timescales (hours to days) interest may focus on the control of surface fluxes by the biology, soil and atmosphere. At longer timescales (weeks to years) we may seek to determine whether the ecosystem is a net source or sink for carbon, and to describe how evaporation from the Correspondence: Dr J. B. Moncrieff, tel + 44-(0)l31-650-5402, fax + 44-(0)l31-662-0478, e-mail J.Moncrieff@ed.ac.uk 1996 Blackwell Sdence Ltd. system responds to seasonal fluctuations in soil water supply and atmospheric demand. Errors occur at each stage of measurement and the question of error propagation from short- to long-term needs to be addressed. In this paper we briefly examine errors assodated with the measurement of fluxes using the eddy covariance technique at short timescales, and discuss the consequences of these errors in long-term measurements of CO2 and HjO exchange. The error analysis is then used to assess the confidence with which a mean source/sink strength can be stated for a particular ecosystem given the nature of the instrumentation, and underlying assumptions in the eddy covaritince technique.

232

J.B.MONCRIEEE etai where r^s is the correlation coefficient and c^^ and the standard deviations of w and s, respectively.
are

2 The eddy covariance technique


2.1 Basic assumptions The eddy covariance technique measures the flux of a scalar (heat, mass) or momentum at a point centred on instruments placed at some height above the surface. For these measurements to be identical to the flux at the underlying surface, the instruments must be located in the internal boundary layer where the flux is constant with height. A clear understanding of the conditions required for a constant flux layer is obtained by examining the conservation equation for a scalar quantity in a control volume placed over the surface:
as

2.2 Errors associated with the eddy covariance method Figure 1 is a schematic diagram which summarizes the factors micrometeorologists consider when assessing the accuracy of their field measurements (Businger 1986). This shows that errors in the eddy covariance technique are assodated with instrumentation as well as a failure to meet the assumptions required for Eq. (3). We first discuss errors inttoduced by non-steady state conditions, advection and complex terrain, and then sources of instrumentation error. This is followed by an examination of the impact of random and systematic errors on longterm flux measurements. Under non-steady state but horizontally homogeneous conditions, Eq. (1) may be integrated with respect to height to give
f.-F.= ds

-D
dt

si. S
dx]

In this equation s is the density of the scalar, t is time, u, is the wind velocity component in the orthogonal directions Xi (i = 1,2,3), D is the molecular diffusivity for quantity s in air, S is the volumetric source/sink strength, and the overbar represents a time average. The first term on the left is the mean rate of change of mass per volume, while the second and third terms are the flux divergence arising from net advection and molecular diffusion through the sides of the control volume. A constant flux layer requires stationarity (ds/dt = 0), the absence of sources or sinks between the surface and the instruments (S = 0), and a horizontally homogeneous surface for a sufficiently long upwind area, so that d(u^)/dx, = 0 and Dd^/dx^ = 0(i = 1, 2). When these assumptions are met, we obtain

dz.
dt

(5)

The term on the right represents the change in storage of s in the air mass between the surface and the height z. Baldocchi et ai (1988) showed that these errors are generally small during the day but can be significant at dusk, overnight and at dawn when turbulent mixing is low. This is particularly relevant to fiux measurements over forests, as discussed later. Ideally the measurement site should be extensive, flat and horizontally uniform to ensure that fluxes are measured in the constant flux layer. Under stationary but non-homogeneous conditions, error in the flux measurement at height z due to advection is given by j- dus i dx 'r ds i dx
dz.

dz

322

= 0,

(2) (6)

where w = ujis the vertical windspeed, and z = X3 is the vertical coordinate. Turbulence is suppressed near the surface while turbulent transport is many orders of magnitude greater than molecular diffusion at height z above the surface (Businger 1986). Thus integrating this equation and using the Reynolds convention that w = w' +w and s = s' + s, yields
fn=-D
dz
0

To a first-order approximation, the error Fj - FQ, following a change in surface with a flux of f, to one with flux FQ may be estimated using = 10 z ln(z/2o)
(7)

X ln(X/10zo)

(3)

where FQ is the flux resulting from molecular diffusion at the underlying soil and leaf surfaces, and F^ is the turbulent eddy flux at height 2. Note that the eddy flux can also be written as
* ?

(4)

where X is the upwind distance from the change in surface and ZQ is the roughness length of the downwind surface. To derive this equation from Eq. (6), we have assumed that the internal boundary layer height, h O.IX (see Kaimal and Finnigan 1994; p. 115), and that wind profiles are logarithmic, i.e. neutral stability. It is clear from this expression that the error in the flux measurement fj - FQ, is proportional to the difference in fluxes 19% Blackwell Sdence Ltd., Global Change Biology. 2. 231-240

E R R O R S IN L O N G - T E R M F L U X M E A S U R E M E N T S ATMOSPHERE - stationarity INSTRUMENTATION - one point sampling - time response - sensor separation - tube losses - digital Filtept - suppqptirig structi

233

flux footprint

Mean horizontaJ wind

Fig. 1. Schematic of the main processes and terms which have to be considered when assessing the validity of flux measurements made by an eddy covariance system.

SURFACE - heterogeneity - complex tertain

katabatic flow path-lengths of sonic anemometers or by air flow through trace gas analysers, also reduces the measured flux through underestimates of the standard deviations in eqn (4). Corrections to the measured flux may be estimated provided spectral transfer functions can be defined for any particular set of instrumentation, including hardware and processing software (Moore 1986; Moncrieff et al. 1996). Such correction terms of course, have uncertainties assodated with them as they are usually based either on engineering formulae which are partly empirical or rely on similarity theory which is itself dependent on atmospheric and surface properties which may be difficult to spedfy exactly. Nonetheless, such transfer functions do permit the approximate correction of flux estimates for a wide variety of error terms assodated with using the inevitable compromises assodated with field instrumentation. However, even when seemingly identical eddy covariance sensors are intercompared at an ideal site, random errors tire present and this generally sets a lower limit of several percent on the overall accuracy of the measurements (Moncrieff et al. 1992). Most micrometeorologists accept that variability in the natural environment (atmosphere and surface) and instrumentation errors restricts the accuracy of an individual turbulent flux measurement to between 10 and 20% (Wesely and Hart 1985). Baldocchi et al. (1988) point out that, rather than having fixed guidelines about where and for how long measurements should be made, there has grown up a number of rules-of-thumb. One of the most widely quoted of these guidelines is concerned with the degree of departure possible from level terrain. In sloping ground, an apparent mean vertical velodty may be recorded by the sonic anemometer unless the coordinate system is rotated to ensure the turbulent flux is calculated with respect to the perpendicular to the local streamlines (Dyer et al. 1982; McMillen 1988). It is generally accepted that once coordinate rotation of the axes is performed, it is possible to make reliable measurements on slopes of up to about 15%. Dyer et al. (1982) pointed

of the upwind and downwind surfaces, F| - FQ. The error also diminishes as z / X decreases, and is < 5% of F-[ - FQ for z/X < 0.01, which conforms with the micrometeorological rule-of-thumb that the measurement height should be < X/100 to ensure accurate measurements. This ratio needs to be increased for stable conditions and may be relaxed under unstable stratification (Kaimal and Finnigan 1994). The flux footprint schemes of Schmid and Oke (1990) and Schuepp et al. (1990) and the work of Mulheam (1977) provide a more detailed analysis of the sensitivity of flux measurements to advective conditions. Instrumentation must have a suffidently high frequency response to measure all the turbulent fluctuations of w and s that contribute to the flux (Eq. 3), and the product of w and s must be formed over a suitable length of time to capture the low frequency parts of the spectrum. Errors assodated with the instrumentation have been widely studied as they are more easily quantified and amenable to correction. Errors associated with the three main compranents of an eddy covariance system, namely sonic anemometer, gas analyser, and software have been published frequently, usually as a description of a particular system. Examples for sonic anemometers include the papers by Kaimal and Gaynor (1991) and de Bruin et a!. (1993); Businger and Delany (1990) present nomograms for the chemical resolution required of a sensor to achieve a certain level of accuracy when used in an eddy covariance system; for software see Shuttleworth (1988). A discussion of errors associated with open- and closedpath sensors for measuring trace gas concentrations is presented by Leuning and Judd (1996, this volume, p.241). It is implidt in Eq. (3) that fluctuations in the vertical wind and scalar are measured at the same point in space. This is generally impossible physically and often leads to underestimates in the measured flux because the correlation coeffident r^s decreases with increasing instrument separation (Eq. 4). Inadequate instrument frequency response, caused by averaging along finite 1996 Blackwell Sdence Ltd., Global Change Biology, 2, 231-240

234

j . B . M O N C R I E F F etai

Chamber Micro mel

a o

Fig. 2. Cumulative carbon uptake by scaled-up chamber measurements and by eddy covariance (labelled micromet in the figure) at an old black spruce stand during BOREAS (rc-drawn from Jarvis et al. 1995).

out that bit errors for momentum and scalar Suxes are different, i.e. momentum flux would be in error by about 14% per degree of tilt but only about 3% per degree of tilt for scalar fluxes. When tissessing the representativeness of point measurements of fluxes in the surface boundary layer it can be difficult to differentiate between an uncertainty caused by the variability of natural surfaces and uncertainties caused when there is no independent check of the measurements available, as is often the case. For the sum of fluxes of sensible and latent heat it is possible to attempt to close the energy balance over a daily cycle or over several days, by comparing eddy covariance measurements against independent measurements of net radiation. Satisfactory closure can give confidence in the overall measurement accuracy of the whole system (Lloyd et al. 1984). This may not always work, however, in terrain which is extremely variable in surface cover as the sources and sinks for heat, water and energy may interact in a non-linear fashion (Lloyd 1995). For fluxes such as carbon dioxide it is difficult to conceive of any sort of closure statistic which could be used in a similar fashion to energy-balance closure. One possible way to close the carbon balance of a stand would be to measure all the separate components contributing to the net flux, i.e. leaf photosynthesis, soil and stem respiration and then scale them up. Figure 2 shows a comparison between the net CO2 flux as measured by an eddy covariance system over an old black spruce stand during the BOREAS exfjeriment of 1994, and the sum of the simultaneous scaled-up chamber measurements made on the soil, leaves and stems (Jarvis et al. 1995). The agreement is good but, presumably, the errors which occur in the scaling up of chamber measurements will account for a large part of the discrepancy.

^^"

true flux measured flux

Random
e-g- stochastic nature of one-point sampling varying flux footprint

si

^A ^ V

T'

V
Time

Fully Systematic
e.g. consistently missing high and/or low frequency components of cospeclnim

o u
Time

SelectiveSystematic
e.g. under-reading of nighl-Eime fluxes because of drainage flows, different turbulence spectra

Time

Fig. 3. An illustration of the main types of error: random, fully and selective systematic errors with examples of the processes which might bring them about. 1996 Blackwell Sdence Ltd., Global Change Biology, 2, 231-240

ERRORS IN LONG-TERM FLUX MEASUREMENTS


Table 1 Examples of random and systematic errors in an eddy covariance system Random one-point sampling of turbulence; varying size of flux footprint and surface heterogeneity; inadequate length of sampling interval; random noise in the signal; Systematic consistent over- or underreading of fluxes by calibration error; use of incorrect spectral forms in calculation of transfer functions; under reading of flux at night because of katabatic flow below sensor; incorrect application of the corrections due to simultaneous fluxes of heat and water vapour {the Webb, Pearman and Leuning (1980) correction); inadequate sensor response or fiow distortion; inadequate height above the surface;
E 5 --

235

22

24

inadequate sensor response; limited fetch and nonstationarity;

3 Random and systematic error


For our purposes, errors can loosely be divided into
random errors and systematic errors, and the essential

differences as they impact on flux studies is shown schematically in Fig. 3 with examples of the type of mechanism which brings about each type of error. It is accepted that in reality, however, any particular error may be a combination of both types. Systematic errors can be fully systematic errors (errors that apply on all the
diumal cycle) or selectively systematic errors (errors that

Fig, 4. The mean daily flux density of carbon dioxide (a) and water vapour (b) obtained from 44 days of data in Rondonia as described by Grace ek ai (1995). The dashed lines are upper and lower quartiles. The x-ajris is local time.

apply to only part of the daily cycle), which, as we shall demonstrate, can have very different effects. Random errors may also be full or selective, but these do not differ substantially in their properties. Table 1 lists the error terms which were illustrated in Fig. 1 as either random or systematic. When errors are random, errors in estimated means and variances diminish with increasing size of data set according to 1 / VW, where N is the number of data points (Barlow 1989). Thus random errors can be detected, estimated, and minimized by examining the convergence of calculations of the net flux with increasing size of data set (as long as the data set is not so enlarged that, for example, seasonal trends become important). In contrast, systematic errors are not affected by increasing data set size ('multiplying your mistakes produces no reward'). This is because, being persistent offsets or multipliers to the data, they simply add in a linear fashion. These errors can be very difficult to detect. For the daily cycles illustrated in Fig. 4, we can apply

a random error, p^, to each half-hour datum. The random error in the mean flux for one mean daily cycle is

= Pr
N

where N is the number of flux measurements In one day, and F, is the ith flux measurement of the daily cycle. For a data set consisting of several days of data, the random error is further reduced, such that E,(Nd) = Er{l)/VN2 where Nij is the number of days. If instead we apply a systematic error, p,, to each halfhour datum, the systematic error on one mean diumal cycle is simply the sum of the individual systematic errors, such that

This systematic error is identical to the overall systematic error on the data, Es(N^), irrespective of the

1996 Blackwell Sdence Ltd., Global Change Biology, 2, 231-240

236 1.0

J.B.MONCRIEFF

etai.

r 0.5
en

\1

\ ,
5%

C-0.5
O
Mean Daily Rux

10% 20% 50%

...--""

I -.-

I...

-2.0 0.1 1 10 Number of Days' Data, W^ size of flie data set. In the case of a selective systematic error, eqn (9) should only be applied to the relevant portion of the daily cycle. We have only considered fractional systematic errors here, as these are the type most likely to be encountered in field measurements (e.g. persistent non-measurement of a certain fraction of the flux). However, it is conceivable that some systematic errors may be fixed rather than percentage offsets, and a similar analysis ctin be applied in that situation. As part of a review of measurements of dry deposition, Businger (1986) presented a table of the magnitude of typical random and systematic errors assodated with six eddy covariance systems reported in the literature. He identified ID sources of error, most of which have been discussed in relation to Fig. 1. Although there was quite considerable variability in whether or not some error terms had been accounted for in different systems, in general systematic errors on the order of 30% or more and random errors of the same magnitude were found.
1.0

100

1000

Fig. 5. The effect of applying random errors of varying magnitudes on the mean daily CO^ flux for the Rondonia data set. The solid, horizontal line is the mean, net daily CC)2 flux of -0.92 ^ ^

0.5 -

.,.

- .1.

- ^ J

0.0
I I Mean Daily Flui

E -0.5 - - -t.O -1.5 - -2.0 - -2.5 -3.0 -3.5 - -4.0


T
., J ,

..-A-.-i

f. i - t

.^

1,_ _ ..^

j_

-100 -80 -60 -40 -20 0 20 40 60 Percentage Systematic Error, p, Fig. 6. As for Fig. 5 but for fully systematic errors.

80 100

4 Case study
How do errors in flux estimates for individual half-hour periods translate into those for a week, a month or a year? To examine this question we shall take some data from a field campaign in the Brazilian Amazon reported by Grace et al. (1995, 1996), where details of the measurements can be found. Our task is to establish the level of confidence in the net flux and to decide whether the reported long-term sink strength was real or whether it could have come about through some combination of uncertainties assodated with the methodology. This case study will allow similar assessment of results from other field studies. The data used here were collected during

April and May 1995 over tropical rain forest in Rondonia, Brazil. Half-hour averaged values of turbulent fluxes of heat, water vapour and carbon dioxide were measured continuously over a period of 44 days using the University of Edinburgh's EdiSol eddy covariance system (Moncrieff et al. 1996). The measurements indicated that over this period, there was a net flux of carbon dioxide into the forest of about 0.92 ^unol COj m ' V (= 0.95 g C m'^ d"'), suggesting that the forest was a substantial sink of carbon. We will draw attention in particular to the effects of errors that are important during one phase of the daily cycle (e.g. at night). To examine these and other types of 1996 Blackwell Sdence Ltd., Global Change Biology, 2, 231-240

ERRORS IN L O N G - T E R M FLUX M E A S U R E M E N T S error, a useful technique is to examine the properties of the mean daily cycle, constructed by taking the mean according to time of day of the entire data set. The net flux over the mean daily cycle is directly proportional to the overall net flux over the measurement period, and by introducing 'errors' at specific times of day one can investigate the effects of these errors on the overall data set. This approach is valid as long as the basic character of the daily cycle does not vary significantly from day to day. An alternative approach which has been used on a much longer dataset, is described by Goulden et ai (1996). 4.1 Diumal cycles of fluxes of carbon dioxide and water vapour far the Ronddnia 1994 data set The mean daily cycle for carbon dioxide fluxes is shown in Fig. 4a. The upper and lower quartiles of the data are also plotted, giving an indication of the day-to-day variability in each phase of the daily cycle. The flux is negative (i.e. into the canopy) during daytime, when photosynthetic uptake dominates over respiration, and positive at night, when respiration, particularly from the soil, is the only significant process. At night the turbulent flux is partially inhibited by the stable stratification of the atmospheric surface layer above the canopy; thus there is an overnight accumulation of carbon dioxide within the canopy, a large proportion of which is 'flushed out' after sunrise with the onset of convection, resulting in the prominent 'spike' in the CO2 flux between 07.00 and 09.00 hours. The magnitude of the spike varies significantly from day to day, being dependent on the amount of within-canopy CO2 accumulation, and thus on the overnight wind speed and stratification. The mean daily turbulent flux is the average of daytime and night-time fluxes of approximately the same magnitude and opposite direction. In this case the mean flux is -0.92 fimol COT m " V , significantly smaller than most of the individual flux measurements. It is an occasional misconception that each value of the CO; flux is measured as the difference between the net upward and downward fluxes (i.e. a small difference between large terms), which would have drastic consequences for sensitivity to errors. While this is true in terms of the competing biological processes (photosynthesis and respiration) generating the net flux, this is not true for any actual measurement of net flux. The mean daily flux is the sum of individual (e.g. half-hourly or hourly) flux measurements, each with an associated error term. The mean daily cycle for fluxes of water vapour is illustrated in Fig. 4b. As before, the upper and lower quartiles of the data are also plotted. The flux is large and positive during daytime, when evapotranspiration is driven by solar heating of the canopy, and much smaller but still positive at night, when there is no solar
1996 Blackwell Sdence Ltd., Global Change Biology, 1, 231-240 -4.0 -100 -80 -60 -40 -20 0 20 40 60 80

237

Daytime Error Only Night-time Error Only

100

Percentage Systematic Error, p.

Fig. 7. The effect of applying selective systematic errors (by day and by night) of varying magnitudes on the mean daily CO; flux for the Rondonia data set. The solid, horizontal line is the mean net daily CO2 flux of -0.92 |imol m"^ s^'.

heating and, moreover, turbulent fluxes are inhibited by stable stratification. The mean daily water vapour flux out of the canopy is 1.64 mmol H2O m"^ s"', corresponding to 2.55 kg m"^ d"^ Thus, for water vapour the net daily flux is the sum of large daytime and small night-time fluxes and is effectively unidirectional. The mean flux is therefore approximately half the magnitude of the daytime fluxes. 4.2 Consequences of each type of error
4.2.1 Carbon dioxide fluxes. Figure 5 shows the results of

applying various values of percentage random error, p,, to each half-hour CO2 flux of the Rondonia data set. The solid horizontal line represents the net daily carbon dioxide flux measured at the site, -0.92 ^mol CO2 m~^ s'^ The symmetrical sets of lines above and below this line represent the overall random error on this net flux, dependent on the magnitude of p^ and the number of days data collected. The magnitude of the random errors decrease with increasing data set size, but at an everdecreasing rate (Eq. 8). This graph allows an estimation of how many days' data are required to ensure that a measured net CO2 balance is not an artifact of random errors. For example, assuming 20% random errors on each half-hour data point, we can see that 1 day's data would not be sufficient to resolve a net sink of 0.92 |imol CO2 m"^s"V On the other hand, 10 days' data would resolve a net sink of about 0.4 jimol CO2 m"^"V and 100

238

J.B.MONCRIEPF etai 3.0


2.5-., 2.0 - =
I

\ \

\
U.

Pr 5% 10% 20%

u.

1.5 --

r'

-y/

50%

0.5 H
0.0 0.1

//
1 10 Number of Days' Data. N^ 100

1000

Fig. 8. The effect of applynng random errors of varying magnitudes on the mean daily HjO flux for the Ronddnia data set. The solid, horizontal line is the mean, net daily H2O flux of 1.64 mmol m~^ s"^

days' data would resolve a net sink of 0.2 ^mol CO2 m"^"'. In the particular case of the Ronddnia data set, with 44 days' data and an assumption (possibly overpessimistic) of 20% random errors per half-hour value, the overall random error on the net carbon dioxide flux is 0.40 ^mol CO2 m"V^ (i.e. 44% random error). Figure 6 shows the consequences of applying various magnitudes of percentage systematic error, p^, to the same data set. As before, the solid horizontal line represents the measured mean CO2 flux. The dashed line represents the actual net CO2 flux (vertical axis), for a given percentage error, p^ (horizontal axis), on each measured half-hour flux. For example, if there is a tendency for each halfhour flux measurement to underestimate the actual flux by 20% (i.e. p^ = -20%), and the measured mean flux is -0.92 ^imol CO2 m-2 s"', then the actual net flux is -1.23 ^mol CO2 m~^s"^ (i.e. there is a net systematic error of -26%). Note that, in the case of CO2 fluxes, full systematic errors are not as severe in their consequences as might be expected. This is because the CO2 flux has opposite signs by day and night; an over- or underestimation of daytime flux is to some extent compensated for by an over- or underestimation of night-time flux. The shape of the plot in Fig. 6 is very sensitive to the asymmetry between day and night of the diumal cycle shown in Fig. 4; if the cycle were totally symmetrical, the net systematic error would always be zero. Another feature to note is that a percentage full systematic error on its own cannot force the net CO2 flux to change sign (e.g. it cannot make an actual CO2 source appear as an apparent CO2 sink). While partial synnmetry reduces the impact of full systematic errors, this Is not the case for selective systematic errors that exist for only part of the daily cycle.

Figure 7 shows the consequences of applying various magnitudes of selective percentage systematic error, p^ to the data shown in Fig. 4. The thin, solid line indicates the actual net flux if a given percentage systematic error applies only to night-time fluxes, and the dotted line indicates the same if the error applies only to daytime fluxes. The slopes of the lines are greater than in Fig. 6, indicating the greater importance of selective systematic errors. Another notable feature is that the lines cross the zero-axis; thus a selective systematic error can, for example, convert an actual net CO2 source into an apparent net CO2 sink. As another example, if the actual CO2 flux balance of a surface is zero, the apparent measured CO2 sink of 0.92 ^mol CO2 m"^ s"' could be generated by underestimation of night-time flux by about 40% (by, for example, the presence of lateral within-canopy flow), or by overestimation of daytime fluxes by 60% (by, for example, the presence of an anomalous CO2 source upwind). An analysis such as this gives an estimate of the magnitude of the possible error factors that need to be searched for. 4.2.2 Water vapour fluxes. Figure 8 illustrates the influence of random errors on the measurement and calculation of net water vapour flux. Because the random errors on each half-hour flux measurement are generally smaller than the mean water vapour flux, the overall random errors are much smaller than for CO2. In the particular case of the Rondonia data set, with 44 days' data and the assumption of 20% random errors per half-hour value, the overall random error on the net water vapour flux is 0.21 mmol H2O m"^ s~^ (i.e. 13% random error, a significantly smaller percentage than for the CO2 flux). The daily cycle of the water vapour flux is unidirec1996 Blackwell Sdence Ltd., Global Change Biology, 2. 231-240

ERRORS IN L O N G - T E R M FLUX M E A S U R E M E N T S tional and highly asymmetrical between day and night. Systematic errors can be important, and are simply the same percentage of the net flux as p^, the percentage error on each individual half-hour measurement. Thus, for example, a systematic underestimate of the flux by 10% in each half-hour flux measurement (due to, say, nonmeasurement of the high-frequency component of the flux) results in a systematic error of -10% on the value of the flux. As the night-time flux of water vapour is negligible compared to the daytime flux, any selective systematic error occurring only at night (e.g. the presence of drainage flows) would have little effect on the net water vapour flux. For the same reason, any selective systematic error biased towards daytime will have exactly the same influence as a full systematic error.

239

5 Discussion
In any long-term flux measurement, a sensitivity analysis of the effects of errors on an averaged day can be a key tool in assessing the likelihood of certain errors producing spurious results. Thus, for the Rondonia data set, if we allow for a 20% random error on each half-hour data value, the total random error on the net flux is 0.40 pjnol CO2 m"^ s"^ (i.e. 44% of the net flux). If we also allow for systematic errors of 10% and that they may be selective by day or night, we see from Fig. 7 that the total systematic error on the net flux is 0.25 jimoi CO2 m'^ s"' (i.e. 27%). Adding these two errors in quadrature, the total uncertainty of the net flux is 0.47 ^mol CO2 m^^s"' (i.e. 53% of the total net flux). Thus we can conclude that the observation that the forest appeared to be a net sink of CO2 over the experimental period was valid. Using the error assessments shown in Figs 4-8, it is possible to establish (i) how many days' data are needed before random errors are likely to be significantly smaller than the mean measured net flux; (ii) the magnitude of nocturnal under-measurement of flux that could lead to a spurious apparent sink term and whether this magnitude of underestimation is likely; and (iii) whether errors in high-frequency flux corrections have a significant impact on the overall result. The relationship between the errors on individual halfhour flux measurements and the net flux to or from a surface depend strongly on the shape of the daily cycle of the flux, in particular on the uru- or bidirectionalty of the flux and on the degree of symmetry between the daytime and night-time fluxes. Therefore, this relationship is very different for CO2 and water vapour fluxes. For daily variation in fluxes that are approximately symmetrical about zero (as for CO2), full systematic errors partially cancel. Therefore, they are not as important as might be expected. This is not the case for heat or water
1996 Blackwell Sdence Ltd., Global Change Biology, 2, 231-240

vapour fluxes where fluxes are generally much larger during the day than at night. Partial systematic errors (e.g. under-reading of night-time fluxes because of drainage flows, or different turbulence spectra) do not cancel for either quantity. Therefore, they are potentially the most important type of error and should be considered with the greatest attention when planning any field experiment or analysing any field data. For carbon dioxide fluxes, tihe net daily flux is the difference between daytime and night-time fiuxes in opposite directions. Therefore, it is essential to calculate error estimates when presenting any values for the net carbon balance of a surface. It is a common convention in physical statistics to quote values as X a b, where X is the mean value of the result, a is the estimate of stochastic (random) uncertainty, andfeis an assessment of systematic uncertainty because of possible systematic errors in the system. This might be a good convention to adopt in quoting results from studies employing eddy covariance systems. References
Baldocchi DD, Hicks BB, Meyers TP (1988) Measuring biosphereatmosphere exchanges of biologically related gases with micrometeorological methods. Ecology, 69, 1331-1340. Barlow RJ (1989J Statistics. A guide to the use of statistical methods in the physical sciences. Wiley, Chichester, 204 pp. Businger JA (1986) Evaluation of the accuracy with which dry deposition can be measured with current micrometeorological techniques, journal of Climate and Applied Meteorology, 25, 1100-1124. Businger JA, Delany AC (1990) Chemical sensor resolution required for measuring surface fluxes by three common micrometeorological techniques. Journal of Atmospheric Chemistry, 10, 399-410. de Bruin HAR, Kohsiek W, van der Hurk BJJM (1993) A verification of some methods to determine the fluxes of momentum, sensible heat and water vapour using standard deviation and structure parameter of meteorological quantities. Boundary-Layer Meteorology, 63, 231-257. Dyer AJ. Garratt JR, Francey, R.J, et ai (1982) An international turbulence comparison experiment (ITCE 1976) Boundaryiayer Meteorology, 24, 181-209. Goulden ML, Munger JW, Fan S-M, Daube BC, Wofsy SC (1996) Measurements of carbon sequestration by long-term eddy covariance: methods and a critical evaluation of accuracy. Global Change Biology, 2, 169-182. Grace J, Lloyd J, Mclntyre J, Miranda A, Meir P, Miranda H, Moncrieff JB, Massheder JM, Wright I, Gash J (1995) Fluxes of carbon dioxide and water vapour over an undisturbed tropical forest in South-West Amazonia. Global Change Biology, 1, 1-12. Grace J, Malhi Y, Lloyd J, Mclntyre J, Miranda AC, Meir T, Miranda HS (1996) The use of eddy covariance to infer the net carbon dioxide uptake of Brazilian rain forest. Global Change Biology, 2, 209-217.

240

J . B . M O N C R I E F F etai

Jarvis PG, Moncrieff JB, Massheder JM, Rayment MB, Hale SE, sensible heat, water vapour and carbon dioxide. Journal of Scott SL (1995) Carbon dioxide exchange of boreal forest in Hydrology (in press). BOREAS. Final Report to NERC TIGER Committee on GST/ Moore CJ (1986) Frequency response corrections for eddy 02/610. correlation systems. Boundary-Layer Meteorology, 37, 17-35. Kaimal JC, Finnigan JJ (1994) Atmospheric Boundary Layer Flows. Mulhearn, P.J (1977) Relations between surface fluxes and mean Oxford University Press, New York, 289 pp. profiles of velodty, temperature and concentration, Kaimal JC, Gaynor JE (1991) Another look at sonic thennometry. downwind of a change in surface roughness. Quarterly Journal Boundary-Layer Meteorology, 56, 401-410. of the Royal Meteorological Society, 103, 785-802. Leuning R, Judd MJ (1996) The relative merits of open- and Schmid HP, Oke, T R (1990) A mode! to estimate the source area closed-path analysers for measurement of eddy fluxes. Global contributing to turbulent exchange in the surface layer over Change Biologj^ 2, 241-253. patchy terrain. Quarterly Journal of the Royal Meteorological Lloyd CR, Shuttleworth WJ, Gash JHC, Turner M (1984) A Society, 116, 96S-988. microprocessor system for eddy correlation. Agricultural and Schuepp PH, Leclerc MY, MacPherson JI, Desjardins RL (1990) Forest Meteorology, 33, 67-80. Footpint prediction of scalar fluxes from analytical solutions Lloyd CR (1995) The effect of heterogeneous terrain on of the diffusion equation. Boundary-Layer Meteorology, 50, micrometeorological measurements: a case study from 355-373. HAPEX-SAHEL. Agricultural and Forest Meteorology, 73, 209- Shuttleworth WJ (1988) Corrections for the effect of background 216. concentration change and sensor drift in real-time correlation McMillen RT (1988) An eddy correlation technique with extended systems. Boundary-Layer Meteorology, 42, 167-180. applicability to non-simple terrain. Boundary-Layer Webb EK, Pearman GI, Leuning R (1980) Correction of flux Meteorology, 43, 231-245. measurements for density effects due to heat and water Moncrieff JB, Verma SB, Cook DR (1992) Intercomparsion of eddy vapour transfer. Quarterly Journal of the Royal Meteorological correlation sensors during FIFE 1989. Journal of Geophysical Society, 106, 85-100. Research, 97, 18,725-18,730. Wesely ML, Hart RL (1985) Variability of short term eddyMoncrieff JB, Massheder JM, De Bruin H, Elbers J, Friborg T, correlation estimates of mass exchange. In: The Forest- ' Heusinkveld B. Kabat P, Scott S, Soegaard H, Verhoef A Atmosphere Interaction (eds Hutchison BA, Hicks BB), pp. 591(1996) A system to measure surface fluxes of momentum. 61. Reidel, Dordrecht.

1996 Blackwell Science Ltd., Global Change Biology, 2, 231-240

Das könnte Ihnen auch gefallen