Model-Data Synthesis of CO2 Fluxes at Niwot Ridge

Model-Data Synthesis
of CO2 Fluxes at Niwot
Ridge, Colorado
Bill Sacks, Dave Schimel
NCAR Climate & Global Dynamics Division
Russ Monson
CU Boulder
Rob Braswell
University of New Hampshire
Motivation
Derive general process-level information
from eddy covariance data
• What processes do CO2 flux data contain information
about?
• Can we separate NEE into its component fluxes?
• Scale up CO2 fluxes in space and time
• Improve parameterization of regional & global models, like
CCSM
Outline
• Methods overview
• Which parameters/processes are constrained by NEE data?
• Exploration of optimized model-data fit: What do we get
right? What do we get wrong?
• Partitioning the net CO2 flux
• What do we gain by including an additional data type (H2O
fluxes) in the optimization?
• Using model selection to explore controls over NEE
• Scaling up (briefly)
SIPNET Model
Photosynthesis
Autotrophic
Respiration
PLANT WOOD
CARBON
PLANT LEAF
CARBON
Leaf
Creation
VEGETA TION
Leaf Litter
Wood Litter
Heterotrophic
Respiration
SOIL CAR BON
Precipitation
Interception &
Evaporation
No:
Snow
Yes:
Rain
Snow melt
Evaporation
Heterotrophic Respiration:
Surface Layer
Drainage
SOIL WATER:
ROOT ZONE
Root Zone
Drainage
Autotrophic Respiration:
f (Plant C, Tair)
Infiltration
SOIL WATER:
SURFACE LAYER
f (Leaf C, Tair, VPD, PAR, Soil Moisture)
SNOW PACK
Throughfall
Fast flow
(Drainage)
• Goal: keep model as
simple as possible
Photosynthesis:
Sublimation
Tai r > 0?
• Twice-daily time step
(day & night)
Transpiration
f (Soil C, Tsoil, Soil Moisture)
Data
• 5 years of half-hourly data from Niwot
Ridge, a 100 year-old subalpine forest
just below the continental divide
– Climate drivers (air & soil temp., precip.,
PAR, humidity, wind speed)
– Net CO2 flux (NEE) from eddy covariance
• Gaps in climate drivers and NEE filled
using a variety of methods
http://spot.colorado.edu/~monsonr/Ameriflux.html
• Half-hourly data aggregated up to day/night time step
– Optimization only uses time steps with at least 50%
measured data
Parameter Optimization
• 32 parameter values optimized to fit NEE data
–
–
–
–
Initial conditions (e.g. initial C pools)
Rate constants (e.g. max. photosynthetic rate, respiration rates)
Climate sensitivities (e.g. respiration Q10)
Climate thresholds (e.g. minimum temp. for photosynthesis)
• Optimization performed using variation of Metropolis
Algorithm: minimize sum of squares difference between
model predicted NEE and observations
• Each parameter has fixed allowable range (uniform dist’n)
• Ran 500,000 iterations to generate posterior distributions
Count
Count
Parameter Histograms
Initial guess
Initial guess
PAR attenuation coefficient
Optimum temp. for photosynthesis
Count
Count
Min. temp. for photosynthesis
Initial guess
Initial guess
Soil respiration Q10
Parameter Correlations
Base soil respiration rate
(g C g-1 C day-1)
C content of leaves per unit area
(g C m-2)
Some parameters can not be estimated well
because of correlations with other parameters:
Initial soil C content
(g C m-2)
PAR half-saturation point
(mol m-2 day-1)
Parameter Behavior
• 13 well-constrained parameters, 5 poorly-constrained
parameters, 14 edge-hitting parameters
• Initial conditions: mostly edge-hitting
• Parameters governing carbon dynamics: mostly wellconstrained. Exceptions:
–
–
–
–
PAR attenuation coefficient
Parameters governing C allocation/turnover rate
Base soil respiration rate
Soil respiration Q10
• Parameters governing soil moisture dynamics: mostly
poorly-constrained or edge-hitting
Nighttime NEE residual (g C m-2)
Daytime NEE residual (g C m-2)
Daytime NEE (g C m-2)
Observations
Model
Day of Year
Nighttime NEE (g C m-2)
Optimized Model: Range of Predictions
Day of Year
Observations
Model
Days after Nov. 1, 1998
Modeled nighttime NEE (g C m-2)
Cumulative NEE (g C m-2)
NEE (g C m-2)
Modeled daytime NEE (g C m-2)
Model vs. Data: Initial Guess
Observed daytime NEE (g C m-2)
Observed nighttime NEE (g C m-2)
Optimized nighttime NEE (g C m-2)
Optimized daytime NEE (g C m-2)
Unoptimized vs. Optimized Model
Unoptimized daytime NEE (g C m-2)
Unoptimized nighttime NEE (g C m-2)
Observations
Model
Days after Nov. 1, 1998
Modeled nighttime NEE (g C m-2)
Cumulative NEE (g C m-2)
NEE (g C m-2)
Modeled daytime NEE (g C m-2)
Model vs. Data: Optimized Parameters
Observed daytime NEE (g C m-2)
Observed nighttime NEE (g C m-2)
Model vs. Data: Optimized Parameters
5
Modeled Daytime
Observed Daytime
Modeled Nighttime
Observed Nighttime
4
NEE (g C m
-2
-1
day )
3
2
1
0
-1
-2
-3
-4
-5
J
F
M
A
M
J
J
A
S
O
N
D
NEE Interannual Variability (g C m
-2
-1
yr )
Model vs. Data: Optimized Parameters
60
40
Modeled Daytime
Observed Daytime
Modeled Nighttime
Observed Nighttime
20
0
-20
-40
-60
1999
2000
2001
2002
2003
Nighttime NEE (g C m-2 day-1)
Missing Variability in Nighttime Respiration
Observations
Model
Air temperature (°C)
Pool Dynamics
Optimized
Days after Nov. 1, 1998
Days after Nov. 1, 1998
Fractional soil wetness
Fraction of initial pool size
Initial Guess
Days after Nov. 1, 1998
Parameter Optimization
Incorporating knowledge of which
parameters/processes are not
well constrained by the data
• Used a single soil water pool
• Held about 1/2 of parameters fixed at best guess values; estimated 17
parameters
Fixed parameters for which:
– Value was relatively well known, and/or
– NEE data contained little information; and
– Fixing the parameter did NOT cause significantly worse model-data fit
This included:
–
–
–
–
Most initial conditions
Many soil moisture parameters
A few parameters that were highly correlated with another parameter
Turnover rate of wood
Almost all parameters
are now well-constrained
Fraction of initial pool size
New Parameter Optimization
Fractional soil wetness
Days after Nov. 1, 1998
Days after Nov. 1, 1998
600
500
GPP
Rtot
NEE (modeled)
400
300
200
100
0
-100
-200
Initial Opt. New Opt.
-1
700
600
-2
800
Mean Annual Flux (g C m yr )
Mean Annual Flux (g C m-2 yr-1)
Partitioning the Net Flux
500
400
RA
RH
NPP
300
200
100
0
Initial Opt.
New Opt.
Partitioning the Net Flux
Flux partitioning using the optimization with fewer free parameters
-1
3
Mean Monthly Flux (g C m
4
-2
day )
5
GPP
Rtot
NEE (modeled)
2
1
0
-1
-2
J
F
M
A
M
J
J
A
S
O
N
D
Partitioning the Net Flux
Flux partitioning using the optimization with fewer free parameters
GPP
Rtot
NEE (modeled)
800
Annual Flux (g C m
-2
-1
yr )
700
600
500
400
300
200
100
0
-100
-200
1999
2000
2001
2002
2003
Mean
Optimization on H2O Fluxes
Optimized simultaneously on H2O fluxes and CO2 fluxes
H2O fluxes also measured using eddy covariance
Hypotheses:
• Using H2O fluxes in the optimization would allow better
separation of NEE into GPP and R, since GPP is highly
correlated with transpiration fluxes
• Using multiple data types would allow better estimates of
previously highly-correlated parameters
Optimization on H2O Fluxes
Optimized CO2 fluxes: similar to optimization on CO2 only, although
slightly worse fit to observations when optimize on both fluxes
Observations
Model
H2O flux (cm precip. equiv.)
Opt. on CO2 only:
Days after Nov. 1, 1998
Fractional soil wetness
H2O flux (cm precip. equiv.)
Optimized H2O fluxes:
Opt. on CO2 & H2O:
Days after Nov. 1, 1998
Days after Nov. 1, 1998
Optimization on H2O Fluxes
Parameter correlations:
800
700
120
600
111
CO2 only
CO2 and H2O
100
500
400
GPP
Rtot
NEE (modeled)
300
200
100
Count
-2
-1
Mean Annual Flux (g C m yr )
Flux breakdown:
80
80
60
40
2525
20
0
9
15
4 8
4 1
0 2
3 1
1 3
.4 .5
.5 .6
.6 .7
.7 .8
.8 - .9 - 1
.9
0
-100
-200
CO2 only
CO2 &
H2O
.1 .2
.2 .3
.3 .4
|r|
1 2
Model Structural Changes
• Tested whether hypothesis-driven changes to model
structure improve model-data fit in the face of an optimized
parameter set
• Goal: learn more about controls over NEE
• Evaluated improvement using Bayesian Information
Criterion (BIC):
BIC = -2 * LL + K * ln (n)
(LL = Log Likelihood; K = # of free parameters; n = # of data points)
Model Structural Changes
Four changes:
• No longer shut down photosynthesis & foliar respiration
with frozen soils
• Separated summer and winter soil respiration parameters
• Split soil carbon pool into two pools
• Made soil respiration independent of soil moisture
Model Structural Changes: Results
• No shut down of photosynthesis & foliar respiration with
frozen soil: significantly worse fit
• Separate summer/winter soil respiration parameters:
slightly better fit
• Two soil carbon pools: slightly worse fit
• Soil respiration independent of soil moisture: little change
Scaling Up
Niwot Ridge
Flux Data
SIPNET
Flux Model
Satellite Data
SIPNET
Optimized for
Niwot Ridge
(e.g. MODIS LAI)
Spatially-explicit Estimate of GPP/NEE
Across Colorado Coniferous Forest Biome
Comparisons with Top-down Flux Estimates
(e.g. Flux Estimates from Airborne Carbon in the
Mountains Experiment (ACME), MODIS GPP)
Conclusions
• Eddy covariance CO2 flux data can be used to constrain most
model parameters that directly affect CO2 flux
Optimization yields better fit of CO2 flux data, but can force other
model behavior (e.g. pool dynamics) to become unrealistic
• Parameter optimization can be used to probe model structure
and learn about controls over NEE
In this ecosystem, it appears that photosynthesis, and possibly foliar
respiration, are down-regulated when the soil is frozen
• NEE partitioning:
GPP
Rtot
= 600 - 700 g C m-2 yr-1
= 550 - 600 g C m-2 yr-1
• Including H2O fluxes in optimization does NOT help us learn
more about controls over CO2 flux