Geophysical Research Letters

On the assignment of prior errors in Bayesian inversions of CO2 surface fluxes

Authors


Abstract

[1] For the estimation of surface CO2 fluxes from atmospheric concentration measurements, most often Bayesian approaches have been adopted. As with all Bayesian techniques the definition of prior probability distributions is a critical step in the analysis. However, practical considerations usually guide the definition of prior information rather than objective criterions. In this paper, in situ CO2 flux pointwise measurements made by the eddy-covariance technique are used to estimate the errors of prior fluxes provided by the prognostic carbon-water-energy model ORCHIDEE. The results contradict the usual convenient assumption of a multivariate Gaussian distribution. The errors of ORCHIDEE have a heavier-tail distribution with a linear temporal dependency after the second lag day and no particular spatial structure. Such error distribution significantly complicates the inversion of CO2 surface fluxes.

1. Introduction

[2] Quantifying the spatio-temporal variations of CO2 surface fluxes over continents has been a scientific target of primary importance since the end of the 1980s [Keeling et al., 1989; Tans et al., 1990]. Several bottom-up methods have been developed, based on local observations [e.g., Baldocchi et al., 2001], on ecosystem modelling [e.g., Krinner et al., 2005], or on a combination of both. Despite dramatic improvements, the uncertainty of each estimate is still too large for these estimates to be reliably used for detailed regional estimates of carbon fluxes. One alternative and complementary approach consists in inferring the fluxes from the atmospheric concentration measurements, knowing how the movement of air parcels link the former to the latter (the top-down approach implemented by, e.g., Enting et al. [1995], Gurney et al. [2002], and Rödenbeck et al. [2003]). The diffusive nature of atmospheric transport makes such an inversion problem mathematically ill-posed: the measurements have to be combined with some other information to regularize it, usually through the Bayes' formula. This essential extra information consists of what one knows about the CO2 surface fluxes prior to the examination of the concentration measurements. If one knew nothing, one should choose uniform prior distribution probabilities for the fluxes (i.e., any flux state is all equally likely). There actually exists some (limited) knowledge of the biogeochemical processes that govern the fluxes and which are gathered in numerical models of the terrestrial carbon cycle. In situ pointwise measurements of the ecosystem fluxes at flux towers, large scale inventories of fossil fuel emissions, of carbon stocks changes, and satellite-based observations of vegetation activity and disturbances also provide some prior information about the fluxes.

[3] Empiricism has dominated the assignment of prior flux errors in Bayesian inversions so far, which bears consequences on the reliability of the inferred fluxes. For convenience, errors are usually modelled by tuneable multivariate (space-time) normal (Gaussian) distributions. Two different strategies could improve on the current situation. The first one, called “marginalization”, consists in treating the unknown characteristics of the prior errors, like the standard deviations, as unknown variables in the Bayes' rule. Michalak et al. [2005] developed a simplified approach along this path. This method is still difficult to implement for large dimension problems. Another strategy, which is favoured here, consists in estimating the prior error characteristics based on actual flux observations.

[4] In this paper, surface flux measurements made by the eddy-covariance technique [Aubinet et al., 2000; Baldocchi et al., 2001] on a continuous basis are used to investigate the errors of prior fluxes of the terrestrial biosphere. Prior CO2 fluxes (i.e., fluxes prior to the analysis of any concentration observation) are provided here by a numerical carbon cycle model: the Organizing Carbon and Hydrology In Dynamic EcosystEms model (ORCHIDEE) described by Krinner et al. [2005]. The model and the observations are presented in the next section. Section 3 shows the results, which are discussed in the last section.

2. Model and Data

[5] Developed at Institut Pierre-Simon Laplace (IPSL), the carbon-water-energy model ORCHIDEE explicitly simulates the principal processes of the continental biosphere influencing the global carbon cycle and the fluxes of CO2 exchanged with the atmosphere, like photosynthesis and respiration of plants and soils. The model handles short-term (half-hourly) to long-term (yearly and beyond) flux and pool variations. It is fully described by Krinner et al. [2005]. We focus here on the configuration of the model which is being used as prior information for flux inversions at Laboratoire des Sciences du Climat et de l'Environnement (LSCE). It relies on prescribed atmospheric conditions, a static land-cover distribution, a prognostic observation-independent phenology and a simple two-layer hydrology module.

[6] The direct measurement of CO2 surface fluxes is provided by the eddy covariance method. This method deduces fluxes from the covariance between fluctuations in anomalous vertical wind velocity and CO2 mixing ratio [e.g., Aubinet et al., 2000]. Some limitations of the method in unsteady atmospheric conditions and over complex landscapes induce substantial uncertainty in the fluxes. Random errors are about 0.4 gC.m−2 for daily totals, based on Hollinger and Richardson [2005], i.e., of smaller amplitude than the departures between the ORCHIDEE simulations and the measurements presented below. Biases occur in some atmospheric conditions and are difficult to quantify. Despite their uncertainties, the flux towers are considered as the reference standard for CO2 flux measurement and a network of them has been developed across representative ecosystems [Baldocchi et al., 2001]. We use quality-controlled records obtained at 34 flux tower stations located in the Northern hemisphere, for which we assume negligible errors for daily (24-hour) averages in comparison to those of ORCHIDEE. These sites data were from the FluxNet archive at Oak Ridge [Baldocchi et al., 2001] and from a separate collection of European forest sites recently used by Ciais et al. [2005]. Each record spans several years between 1994 and 2004, and consists of observations of 2-meter temperature, 10-meter wind, precipitation and radiation fluxes in addition to the CO2 fluxes, with a time step of 30 minutes. Incomplete records of the CO2 fluxes have not been gap-filled whereas meteorological variables have been interpolated when needed. These meteorological variables have been used as a boundary condition for ORCHIDEE simulations at each site. Vegetation is distributed in the simulations according to the site characteristics. Note that most sites include more than one vegetation type. The initial plant and soil carbon reservoirs are not known and, following the common practice, have been set at the initial time step of each simulation so that the simulated ecosystems are carbon-neutral on a yearly basis.

[7] The eddy-covariance flux observations make it possible to investigate the errors of the simulated fluxes at various temporal resolutions. Most inversion studies up to now have inferred monthly fluxes [Gurney et al., 2002, and references therein]. However, the specification of fixed temporal flux patterns within a month necessarily induces spatial and temporal correlations of the model errors at the observation locations that are difficult to take into account, and are usually not. The technical limitations (i.e., computer memory and power) that prevented to infer fluxes at higher temporal resolutions are being circumvented thanks to the introduction of new formulations of the Bayesian inference problem [e.g., Chevallier et al., 2005; Peters et al., 2005]. Consequently, we focus here on 24-hour flux averages.

3. Results

[8] Altogether, the database of daily-mean eddy-covariance fluxes consists of 31,500 quality controlled samples. Figure 1 displays the correlation between the modelled and the observed daily fluxes as a function of the dominant plant functional type (PFT) on each site. The 12-PFT classification of ORCHIDEE is used. The scatter of the points illustrates the diversity of processes that are involved and that are reproduced in ORCHIDEE with various skills, with correlations ranging from about 90% for some forest sites to about zero for some crop sites.

Figure 1.

Correlations between the modelled and the observed daily fluxes as a function of the dominant plant functional type on each measurement site. The 12 plant functional types of ORCHIDEE are used. Note that a plant functional type may relatively dominate a site even though it only covers 40% of the site.

[9] Figure 2 shows the distribution of the model-minus-observation differences after combining the data from all sites. The negative bias of the distribution (0.6 gC.m−2 per day) was expected because vegetation at most eddy covariance sites is in growing phase, and therefore acts as sink for carbon whereas the model has been initialized to be carbon-neutral on the long-term mean. However, a mean 0.6 gC.m−2 per day sink over the whole vegetated land surface of the Earth, which is about 1014 m2, would translate into a global sink of 22 GtC per year. This excessive figure [Intergovernmental Panel on Climate Change (IPCC), 2001] indicates that the database of flux measurements over-represents productive vegetation stands and that different stages of ecosystem disturbance regimes are not covered by the FluxNet network well enough. This is consistent with the fact that most sites are affected by human activities, like cutting, planting and nitrogen fertilisation.

Figure 2.

Probability Density Function (PDF) of the ORCHIDEE-minus-observation departures for daily CO2 fluxes. The Gaussian distribution with the same mean and standard deviation, as well as the Cauchy distribution with location parameter 0.3 and scale factor 1.0, are also reported on the graph.

[10] The differences spread around the mean with a 2 gC.m−2 per day standard deviation. Note that such random errors translate to much less than 22 Gt C per year because uncorrelated random errors evolve as the square root of the time and space scales of the aggregation whereas the bias have a linear behaviour. The actual value of the impact of random errors depends on the time and space correlations. As a corollary, the much-looked-for land biospheric sink [IPCC, 2001] may be negligible compared to the daily 24-hours flux at the scale of a model grid point, even though it dominates the uncertainty of the global annual budget.

[11] In Figure 2, two theoretical distributions have been superimposed to the one of the model-data flux differences. The first one is the Gaussian distribution with the same mean and standard deviation. The second one is the Cauchy distribution (also called Lorentz distribution) with location parameter 0.3 and scale parameter 1. Obviously, approximating the ORCHIDEE errors by the Gaussian distribution is a poor approximation whereas the Cauchy distribution would be more appropriate. This is because the daily fluxes are well simulated at some sites during some periods but poorly in other cases (as seen in Figure 1). Errors at individual sites are more normally distributed (not shown). In other words, the simulation error is not purely random but depends on nuisance variables (in the statistical sense), like the start of the growing season or the plant hydric stress. Therefore a Cauchy distribution, with a relatively narrow peak compared to the tails, better fits the error distribution than a normal distribution. In order to keep Gaussian errors for individual fluxes, one would have to use different widths of the Gaussian distribution, depending on sites or periods.

[12] Correlations of the differences between the simulations and the measurements of the in situ daily fluxes are shown in Figure 3 as a function of time and space. Time correlations drop down to about 70% at lag-day 2 and behave rather linearly afterward, which is far from a Gaussian decay. The correlations at lag-day 30 are about 30%. Spatially, the correlation between the differences at distinct sites is below 50% in absolute value even when considering nearby sites. Given the large spread of the space correlations at any distance, no obvious spatial coherence can be identified.

Figure 3.

(top) Time and (bottom) space correlations of the differences between the ORCHIDEE simulations and the observations.

4. Discussion and Conclusions

[13] Assuming normally-distributed prior errors is common practice for Bayesian inversions for several reasons. First, this approximation makes the problem analytically solvable, either by matrix operations or by the minimization of a cost function [e.g., Lorenc, 1986]. Second, it is the least committal choice when one only knows the mean and the standard deviation of the actual distribution [Jaynes, 1957]. Third, under certain conditions, the distribution of the sum of a large number of independent variables is indeed approximately Gaussian, as stated by the central limit theorem. None of these reasons justifies a systematic use of Gaussian error distributions and there is a need to investigate the properties of the prior errors, at least to estimate the mean and the variance of their distribution. In this study, pointwise continuous measurements made by the eddy covariance technique have been used to estimate the characteristics of the prior errors for CO2 flux inversions. The prior is provided by the model of the terrestrial biosphere ORCHIDEE. This strategy is not fully exhaustive. First, the eddy covariance measurements are affected by some errors. Second, the observation towers consider areas of size typically about 1 km2, whereas ORCHIDEE is used at a lower spatial resolution to provide flux maps, typically about 100 × 100 km2 (see for instance http://www.lsce-orchidee.cea.fr/orchidee_gb.html). Furthermore, for large areas, the atmospheric forcing needed to run the model cannot be obtained from local observations but is provided by short-range weather forecasts of lower accuracy. Last, our flux database is biased toward temperate middle-aged forest ecosystems. However, in spite of these limitations, local observations by the eddy-covariance technique may be the only reliable benchmark and only they can currently provide some evidence about the structure of the simulation errors. The present study focuses on the daily CO2 fluxes simulated by ORCHIDEE and its results may not be valid for other time scales or other models, that require specific attention based on a similar methodology.

[14] This study indicates an error standard deviation of 2 gC.m−2 per day when considering all ecosystems together. Combined with the temporal correlations of Figure 3, this number corresponds to a monthly error budget (i.e., the square root of the sum of the covariances within a month) of about 60 gC.m−2 and a yearly one of about 200 gC.m−2 at one site (without any respect to the nature of its vegetation). As a comparison, Rödenbeck et al. [2003] guessed errors to be of similar or smaller amplitude (depending on latitude) for the prior information in their flux inversion (their Figures 9d and 9e) whereas Houweling et al. [2004] supposed twice as large values. These studies included biomass burning in addition to the biosphere photosynthesis and respiration CO2 fluxes, and hypothesised perfect knowledge of fossil fuel emissions. Both types of processes deserve a specific investigation. Our study also shows that the temporal correlations of the ORCHIDEE errors slowly decrease in a linear way after lag-day 2. No particular spatial structure could be identified. The absence of spatial correlation emphasizes the importance to perform flux inversions at horizontal resolutions as high as possible, since subgrid scale errors are correlated by construction. Obviously, the sparse eddy-covariance network may not reveal some correlation structure that may actually exist. In particular, the fact that each site contains a mixture of vegetation types prevented us to analyze the arguably-larger correlations within a given vegetation type. Nevertheless, there seems to be less justification in introducing spatial correlations than to ignore them. If the results hold for other models of the biosphere and other sites, one would wonder whether surface measurements of CO2 concentrations actually contain much information about the spatial distribution of biospheric fluxes. Indeed, in the absence of spatial correlations, inversion increments generated by the surface observations are mainly confined to the vicinity of the measurements [e.g., Bocquet, 2005]. Such issue highlights the importance of the forthcoming spatial instruments dedicated to the observation of atmospheric CO2, because they will provide measurements well above the surface and a much denser coverage of the globe.

[15] Finally, our results seem to indicate that there is no ground for choosing Gaussian prior error distributions in atmospheric inversion, at least when using daily fluxes from ORCHIDEE. Gaussian distributions can still be justified by their analytical properties from a pragmatic point of view. Indeed a distribution with heavier tails makes the Bayesian cost function non-quadratic and therefore increases the computational burden of the inversion. Further, there is no closed-form solution to the inversion problem any more. However, properly assigning the prior errors in flux inversions would make the flux inversions closer to what they are supposed to be: the best solutions given the evidence provided.

Acknowledgments

[16] Authors wish to thank F.-M. Bréon, P. Naveau, P. Peylin and P. Rayner (LSCE) for fruitful discussions about the topic. C. Rödenbeck (MPI-Jena) and two anonymous reviewers made useful comments on an earlier version of the paper. This study was made possible by the work of numerous scientists, students and technicians involved in data collection and analysis at the various Fluxnet sites. Most data have been downloaded from http://www.daac.ornl.gov/FLUXNET/fluxnet.html. The other ones have been kindly provided by M. Aubinet, J. Banza, C. Bernhofer, A. Carrara, A. Granier, W. Kutsch, D. Loustau, D. Papale, J. S. Pereira, K. Pilegaard, M. J. Sanz, G. Seufert, J.-F. Soussana and T. Vesala. This study was co-funded by the European Union under projects GEMS and GEOLAND.

Ancillary