steady state impacts in inverse model
parameter optimization
Carvalhais, N., Reichstein, M., Seixas, J., Collatz, G.J., Pereira, J.S., Berbigier, P.,
Carrara, A., Granier, A., Montagnani, L., Papale, D., Rambal, S., Sanz, M.J., and
Valentini, R.(2008), Implications of the carbon cycle steady state assumption for
biogeochemical modeling performance and inverse parameter retrieval, Global
Biogeochem. Cycles, 22, GB2007, doi:10.1029/2007GB003033.
motivation / goals
• CASA model parameter optimization
• spin-up routines force soil C pools
estimates
• impacts of the steady state in:
– model performance
– parameter estimates / constraints
• propagation of C fluxes estimates
uncertainties for the Iberian Peninsula
the CASA model
NEP  GPP  Ra   Rh
NPP  APAR  
APAR  fAPAR  PAR
   *  T  W
p
Rh   Ci  k i  Ws  Ts  (1  M  )
i
Aws Q10  * TOpt Bw
Potter et al., 1993
approach to relax
the steady state approach
• inclusion of a parameter that relaxed
the steady state approach: η
Cns = Css ∙ η
Fix Steady State
Aws Q10  * TOpt Bw

Relaxed Steady State
 1
 1
 1
experiment design
• significance of each parameter:
– removing one parameter at a time;
• alternatives to η:
– replacing by :
• soil C turnover rates;
• extra parameters on NPP and Rh temperature
sensitivity.
• Levenberg-Marquardt least squares
optimization
site selection and data
• CARBOEUROPE-IP:
– 10 Sites
• optimization constraints: NEP
• model drivers:
– site meteorological data;
– remotely sensed fAPAR and LAI;
– different temporal resolutions
IT-Non [sink: 542gC m-2 yr-1]
adding η
effect of η in optimization
determinants of parameter variability:
ANOVA
Bw
T
*
opt
2 4
2
17
site
39
33
40
parameter vector
6
17
9
39
2
temporal resolution
33
site x
parameter vector
5
16
Q
9
A
10
FST

ws
2 4
4
site x
temporal resolution 28
parameter vector x
temporal resolution
27
7 1
6 4
0 8
12
6
11
19
56
0
55
PRM
T MR
83
FST *PRM
FST *T MR
PRM*T MR
what drives η?
r2: 0.76; α < 0.001
model performance improvements
model performance in relaxed > fixed steady state assumptions.
differences in parameter
estimates and constraints
relaxed
relaxed
P/P
↑NPP
SE/SE
fixed
ε* Topt Bwε Q10 Aws
↓Rh
fixed
ε* Topt Bwε Q10 Aws
total soil C pools
measurements relaxed fixed
steady state approach impacts
• model performance
– relaxed > fixed
• parameter estimates
– biases
• parameter uncertainties
– relaxed < fixed
• soil C pools estimates
– relaxed closer to measurements
propagating parameters / uncertainties
spatial simulations
• Iberian Peninsula
• optimized parameters per site:
– optimization: naïve bootstrap approach
• no assumption on parameters distribution
– GIMMS NDVIg : 8km, biweekly;
• parameter propagation per PFT:
– estimating NEP / NPP / Rh
spatial impacts : NPP 1991
relaxed
fixed
relaxed - fixed
seasonality : NPP : IP
relaxed versus fixed
iav : NEP : IP
relaxed versus fixed
seasonality and iav : IP
inter annual
variability
var.
NPP
Min
max
seasonal
amplitude
min
max
uncertainties
min
max
-9%
62%
-11%
53%
-60%
-2%
Rh
-15%
74%
-39%
131%
-60%
-2%
NEP
-10%
63%
-10%
91%
-60%
6%
(relax – fix) / fix
remarks
• biases in optimized parameters lead
to significant differences in flux
estimates: seasonality and iav
• uncertainties propagation show
significant reductions under relaxed
steady state approaches
• impacts in data assimilation schemes
…