Title: On the use of eddy-covariance and optical remote sensing data for biogeochemical modelling
1On the use of eddy-covariance and optical remote
sensing data for biogeochemical modelling
Carbon Fusion International Workshop Edinburgh,
May 2006
- Markus Reichstein, Dario Papale
- Biogeochemical Model-Data-Integration Group,
Max-Planck-Institute Jena - Laboratory of Forest Ecology, University of Tuscia
2BGC-Model-Data Integration Overview
Ecosystem models
provide system understanding promise
inter-/extrapolation capacity may include
historical effects are simplifications of the
world cant predict stochastic events
Ecosystem data
Remote sensing
Potentially high quality often high temporal
resolution data compatibility ? point
observations
objective/consistent observations spatially
and temporally dense data quality lower
processes not directly observable, no history,
no prediction
3Outline
- Introduction to eddy covariance data
- Bottom-up perspective of an ideal data
integration-validation process - Problems and obstacles in this process
4Overview
- General data assimilation example remote sensing
of a cut!! - Fluxnet as large data archive
- Particular problems with biospehere eddy error
and quality discussion, spatial scaling, model
structure, dynamic parameters - Proxel example of tracking parameter
- MODIS example of RUE model ? monthly RUE
structural update, problems with generalisation - Overview of inverse parameter estimation
approaches (multiconstraints) - Future better characterisation of errors,
spatial scaling, multiple constraints,
generalization from sites - Consider pools! Time scales, Error
5Observing ecosystem gas exchange eddy covariance
Flux speed x concentration
Photo Baldocchi
6Eddy covariance
Measures whole ecosystem exchange of CO2 and
H2O, Non-destructive continuous
time-scale hourly to interannual integrates
over large area - only on flat sites - relies on
turbulent conditions gt data gaps, stochastic
data - source area varying (flux footprint) -
only point measurements Does not deliver
compartment fluxes, but NEP GPP - Reco
7Half-hourly eddy covariance data
Evapotransp.
8Network of ecosystem-level observations
gt1000 site-years ? 1012 raw measurements (1013
bytes)
- Network and intercomparison studies
- Harmonised and documented data processing
- Aubinet et al. (2000), Falge et al. (2001), Foken
et al. (2002), Göckede/Rebmann/Foken (2004)
general set-up and methodology, quality
assurance, gap-filling - Reichstein et al. (2005), Glob. Ch. Biol.
u-correction, gap-filling, partitioning of NEE - Papale et al. (in prep), Biogeosciences Quality
control, eval. uncertainties - Moffat et al. (in prep) Gap-filling
inter-comparison - Online processing tool http//gaia.agraria.unitus
.it/lab/reichstein/
9Ideal model-data integration cycle (bottom-up)
10The bottom-up model PROXEL
11I. Model charaterization / forward model run
Drought stressed conditions
Reichstein, Tenhunen et al., Global Change
Biology, 2002
12II. Dual-constraint parameter estimation
Reichstein et al. 2003, JGR
13IIa. Inferred parameter timeseries
14III. Interpretation Generalization
Relative leaf activity
Relative soil water content
Reichstein et al. 2003, JGR
151.8
1.6
III. Interpretation and Generalization Keyp.
RUEmax
1.4
1.2
1
RUE gC / MJ APAR
0.8
0.6
- inter-PFT variability
- intra-PFT variability
- f(species, N, T???)
0.4
0.2
0
ENF
EBF
DBF
MF
Sav
Oshrub
Crop
16IV. Validation at larger scale
17GCB, in press
18The problems
19To consider with DA of eddy covariance data
- How is the error structure of the data itself?
- How to address mismatch of scales (point versus
pixel)? - Remote sensing
- Meteorological data
- How do perform up-scaling from tower sites?
- Representativity
- Generalization
20Errors in the data
21Error model influence on parameter estimates
Search strategy
I
II
Parameter estimate
Const. abs errors
Const. rel. errors
Simplified after Trudinger et al. (OPTIC)
22Errors in eddy covariance data
- Random errors
- 30 for the half-hourly flux, (turbulences !)
- Systematic errors
- can be largely controlled/avoided
- Selective systematic errors
- Conditions where the theory does not apply
- Low turbulent conditions (night-time)
- Advection
- good quality control necessary
- Better few unbiased data, than a lot of biased
data - Uncertainties mean NEE gt interannual variability
23Characterization of the random error
cf. Richardson et al. (2006)
24Quantifying uncertainties
NEE
NEE_sigma
Dec
Dec
Dec
Dec
Nov
Nov
Nov
Nov
Oct
Oct
Oct
Oct
Sep
Sep
Sep
Sep
Aug
Aug
Aug
Aug
Jul
Jul
Jul
Jul
NEE_sigma µmol m-2 s-1
Jun
Jun
Jun
Jun
May
May
May
Apr
Apr
Apr
Apr
Mar
Mar
Mar
Mar
Feb
Feb
Feb
Feb
Jan
Jan
Jan
Jan
0
6
12
18
24
0
6
12
18
24
0
6
12
18
24
0
6
12
18
24
25Error distribution of eddy covariance data
26Distribution of model error against eddy data
Chevalier et al. (in rev.)
27PDF only 10am-3pm and Jun-Sep
NEE error
28More complicated error structures
29Maximizing the likelihood?
Bayesian approach Cost function
Trust in data
Trust in apriori model parameters
30Spatial representation problem I
- Does the tower site represent the grid cell of
interest? - 0.25-2km km for MODIS/SEAWIFS remote sensing
- 30-100 km for meteorological fields
- 30-100 km for DGVMs, BGCs applied in global
context
31Spatial heterogeneity...
1 km
32Its not always so bad...
Dinh et al., subm.
33Spatial representation problem II
- Does the network of tower sites represent the
spatial domain of interest or are there chances
to generalize with scaling variables?
34fAPAR MODIS-RT)
? We have to have up-scaling strategies
Day of the year
35Conclusions
- Eddy covariance data contains a lot of
interpretable information on both carbon and
water cycle - Inclusion of pools and fluxes for system
understanding and for linking short and long
time-scales necessary - Major challenge within eddy data
- Characterization of the error (random, bias)
- Scale and representativeness problem
- Interpret. Generalization of site specific
parameters - Documentation of site dynamics, that may violate
model structure (e.g., soil water, management)
36Conclusions