Some challenges of modeldata integration a collection of issues and ideas based on model evaluation

1
Some challenges of model-data- integrationa
collection of issues and ideas based on model
evaluation excercises

• Martin Jung, Miguel Mahecha, Markus Reichstein,
• Model Simulations by Guerric Le Maire, Maarten
  Braakhekke, Sönke Zaehle, Mona Vetter

2
Models in steady-state (see contribution by Nuno
Carlvalhais)
Implications for data assimilation and model
evaluations!

• Carbon balance simulated by process models is
  most likely biased
• Models may be useful to study variability of the
  carbon balance (anomalies, processes, )
• Variability of the carbon balance results from
  variability of big constituent fluxes (GPP, TER,
  )
• Models need to be quite precise at the
  constituent fluxes to get variability of the
  carbon balance right

Model world
Real world
After Odum (1969), modified by from Alex Knohl
3
How to handle confounding effects?
Correlation of NEP residuals with GPP and TER
residuals (based on site-level runs, monthly data)
If NEP is wrong it can be because
of -GPP -TER If GPP is wrong it can be because
of -some parameter -LAI/fpar -soil water
dynamics -temperature sensitivity
function -sensitivity of gcan to VPD and soil
moisture -coupling of Gcan and photosynthesis ...
Isolating model components as much as possible
for evaluation/assimilation excercises?!
Sensitivity experiments?!
4
Agreement among models regarding inter-annual
variability of GPP
1-R2
Biome-BGC vs Orchidee LPJ
Models were run with the same input data!
Based on annual GPP from 1981-2000
5
Biophysical vs. ecophysiological control of GPP
interannual variability in the models
GPP APAR x RUE
APAR Absorbed Photosynthetic Active Radiation
MJ/m2/yr
RUE Radiation Use Efficiency gC/MJ
Interannual variations of radiation use
efficiency are the primary cause of GPP
interannual varibaility
Input Radiation
Fraction of absorbed radiation (FAPAR) 1 -
exp(-0.5 x LAI)
Simulated LAI
APAR
Jung et al., GBC, 2007
6
Correlation and sensitivity of summer (JJA)
meteorology with GPP
Reducing meteorological variable space
(radiation, temperature, vapour pressure deficit,
and precipitation) to principal components
PCA1 explains 84 of variance of the summer
meteorological data PCA1 weights RAD (-0.28),
TEMP (-0.28), VPD (-0.28), RAIN (0.24)
Does nitrogen dynamics influence interannual
variations of GPP?!
Effects of water stress on photosynthesis largely
control GPP interannual variability ? canopy
conductance and coupling with carbon
assimilation ? representation of soil, roots,
below ground processes
Jung et al., GBC, 2007
7
Do the models have a systematic bias during
drought?
Site-level runs
n.s.
n.s.
significant
Drought effect too weak
Drought effect too strong
(Model_site_month_DryYear Model_site_month_WetYe
ar) (Eddy_site_month_DryYear
Eddy_site_month_WetYear)
8
The models response to meteorology - How to
tackle equifinality?
Site-level runs

• 21 day sliding correlation window between
  C-fluxes and Temp, Rad, VPD, SWC

Response of simulated NEP to meteo is more
consistent with site data than the gross fluxes ?
equifinality or artifact of flux
separation? Largest differences with respect to
TER
Consistency
Confounding effects because meteo variabels are
co-linear
Consistency how often does the simulated flux
correlate with the same meteo driver as the
eddy-based flux sum(Var_maxR_site
Var_maxR_model)/sum(significant correlations)
9
Model RMSE as a function of time scale
RMSE (norm by data range)
High frequency components seasonal cycle work
better than inter- and intra-annual components
Inter-annual components of GPP vs Gcan
Significance of changing pools ecosystem
properties?
Mahecha et al. In prep.
10
What is an adequate model?

• scatter is ok, bias not (data are noisy,
  simulations not)
• RMSE, R2, ... are not really good measures of
  model performance
• Looking for robust patterns in the FLUXNET data!
• Can patterns be assimilated into models?

Jung et al 2007, Biogeosciences
11
What about using patterns from upscaled carbon
fluxes?

• Advantages noise goes away no issues of site
  specific pecularities no representation bias
  matches the scale of the models
• Disadvantages uncertainties from drivers (meteo
  data, remote sensing products) model specific
  sensitivity to meteo no effects from changing
  pools (? IAV)

12

• Comparison of European mean GPP pattern Process-
  vs. data-oriented models

R2
Mean annual GPP patterns from data-oriented
models are becoming sufficiently robust for
benchmarking process-oriented models
Process oriented models
Data driven models
13
2003 GPP anomaly from different data-oriented
models
Inter-annual variability from data-oriented
models is not sufficiently robust for
benchmarking process-oriented models
Jung et al., GCB, in press
14
Final Remarks/Questions

• How to deal with important input data that are
  usually not available (effective rooting depth,
  water holding capacity)?
• To what extent are parameters allowed to
  compensate for inadequate structure?
• What is an adequate model structure?
• How to identify not adequate structure
  components?
• Should we concentrate on patterns rather than
  on values?