Characterizing Uncertainty for Regional Carbon Cycle Modeling

1
Characterizing Uncertainty for Regional Carbon
Cycle Modeling
Lara Prihodko
2
Model-Data Fusion

• Predicted Flux Uncertainty
• Parameter Uncertainty
• Transport Uncertainty
• Observation Uncertainty

3
The Model SiB2

• Simple Biosphere Model, version 2 (SiB2)
• Sellers et al. 1996
• Biophysical land surface model
• Describes heat, water and carbon transfers
• in the soil, vegetation, atmosphere
  continuum
• Developed for general circulation models
• Useful at many scales, globe to point
• Single canopy layer scheme
• Highly non-linear
• Large number of input parameters

4
Part I. Parameter Sensitivity Analysis
Satellite Imagery
Biosphere Model
Predicted CO2 Flux
Model Parameters
Observed CO2 Flux
A model exhibits equifinality when compensation
among parameters can result in equally good
simulations across a wide range of parameter
values
Parameter Uncertainty Predicted Flux
Uncertainty Observation Uncertainty

• Sensitivity analyses help
• Determine the relative worth of parameters
• Analyze model structure and system behavior
• Reduce model uncertainty
• Guide research priorities

5
Parameter Sensitivity Analysis - Methods
46 input parameters 20,000 parameter sets
-Realistic ranges -Uniform distributions
20,000 sims spun up for 6 yrs (each time!)
(Franks, 1998)
6
Parameter Sensitivity Analysis - Methods
Good fit
Poor fit
7
Parameter Sensitivity Analysis - Methods

• Kolmogorov-Smirnov test statistic
• -Test whether the posterior distribution is
  significantly different than uniform
• Significant at p0.01
• Retained and tested top 10 of parameter sets
  for
• -each parameter
• -each month
• -all months combined

8
Parameter Sensitivity Analysis
Parameters can be ranked by the K-S test
statistic, indicating the relative influence
9
Parameter Sensitivity Analysis
10
Parameter Sensitivity Analysis

• Stable, normal distributions indicate well
  defined parameters and processes
• Subtle shifts in parameter ranges suggest
  mechanisms that might not be adequately
• captured by the model

11
Parameter Sensitivity Analysis - Results
Minimum error
12
Part II. Uncertainty Analysis
Satellite Imagery
Biosphere Model
Predicted CO2 Flux
Model Parameters
What choice would any of us make between an
honest but depressingly ambiguous prediction and
a faithfully promoted, complex and seemingly
accurate prediction?(Fedra et al, 1981)
Observed CO2 Flux
Parameter Uncertainty Predicted Flux
Uncertainty Observation Uncertainty
Uncertainty exists in Model structure Model
parameters Model driver data Observations that
we test against An uncertainty analysis provides
an estimate of this uncertainty
13
Uncertainty Analysis - Methods
Generalized Likelihood Uncertainty Estimation
(GLUE) (Beven and Freer, 2001)
Bayes equation
1) Take previously calculated likelihoods
(but only for the best sets!) 2) Scale them so
they sum to 1 3) Rank the hourly flux estimate
4) Sum the scaled likelihoods until you
reach the predictive uncertainty quantile.
Uncertainty quantiles
Result is hourly predictive uncertainty bounds as
conditioned on the input data, observations,
parameter sets and likelihood measure constrained
monthly and annually.
14
Uncertainty Analysis - Results
Diurnal Cycle of Predictive Uncertainty 2 days
July 29 and October 23, 1997 NEE, LE, H
15
Uncertainty Analysis
Hourly Predictive Uncertainty 2 months July and
October 1997 NEE
16
Uncertainty Analysis
Monthly Mean Diurnal Cycle of Predictive
Uncertainty All Months NEE
17
Part III. Spatial uncertainty
Satellite Imagery

• Biosphere models often rely on remotely sensed
  data to represent surface processes and/or to
  provide state variables
• Remote sensing provides a critical link between
  local, regional and global scales
• Restrictions on representation
• -Computational limitations
• -Spatial scale of available data
• Addressing mismatches in scale
• -Functional groupings of
  vegetation types
• -Averaging of surface
  properties and soil characteristics
• When we do this we lose heterogeneity

18
SiB2-RAMS

• Regional Atmospheric Modeling System (RAMS)
• Pielke et al., 1992
• Simulates weather
• Mesoscale atmospheric phenomena
• Useful at many scales, hemispheres to large
  eddies
• Nested grid capability
• RAMS to SiB2
• Boundary layer values of temperature, vapor
  pressure
• and wind speed
• Direct and diffuse shortwave and long-wave
  radiation
• CO2
• SiB2 to RAMS
• Fluxes of heat, moisture, momentum and CO2
• Upwelling radiation

19
Spatial Uncertainty - Methods
Coupled land surface-atmosphere model Simple
Biosphere Model version 2 (SiB2) Regional
Atmospheric Modeling System (RAMS) Two
representations of land surface heterogeneity
1 km 8 km Three time periods Early
spring (April) Leaf-on (Late May, early
June) Summer (July) 5 Clear days 4 year
off-line spin-up of the inner grid- for soil
temperature and moisture profiles NCEP gridded
weather reanalyses Fully-coupled prediction of
weather and surface fluxes and of water, energy,
and carbon
640x640 (16km) 152x152 (4km) 38x38 (1km)
20
Input data Soil texture and vegetation class

• Soil Texture
• Sand and clay calculated
• at 1km from the STATSGO
• soil database.
• Area averaged to 8km
• Soil thermal and hydraulic
• properties calculated

0 Sand
100 Sand

• Vegetation Class
• From Hansen et al. 2000
• remapped to SiB2 classes
• Assigned dominant class
• from 1km data to 8km area
• Time invariant model
• parameters assigned

0.0

• NDVI
• 1km NDVI (Teillet et al. 2000)
• Area averaged individual bands
• to 8km
• Calculate time varying parameters
• for SiB2 (Sellers et al, 1996b)

1.0
1 Kilometer
8 Kilometer
21
Spatial Uncertainty - Results
Net Ecosystem Exchange micromoles m-2 s-1
April 15, 1800 GMT Noon CT
June 3, 1800 GMT Noon CT
July 30, 1800 GMT Noon CT
5
0
0
0
-8
-14
-6
-18
-30
Y(m)
5
0
0
-8
-6
-18
X(m)
22
Model uncertainty April 2 mmols m-2 s-1 June
5 mmols m-2 s-1 July 10 mmols m-2 s-1
Spatial Uncertainty - Results
Net Ecosystem Exchange
April(0.95)
July(1.0)
May/June (1.05)
Top row 5th and 95th percentile deviations from
the mean difference Bottom row cumulative mean
deviations
23
Conclusions
Satellite Imagery
Compensation among parameters complicates
optimization Largest reductions in uncertainty in
shoulder seasons Parameterization uncertainty
greater than spatial uncertainty Model-data
mismatch used as a prior uncertainty
Biosphere Model
Predicted CO2 Flux
Model Parameters
Observed CO2 Flux
Predicted Flux Uncertainty Parameter
Uncertainty Observation Uncertainty
Recommendations Longer coupled model
runs Similar analysis with analyzed weather
fields Other geographic locations
24
Questions?
Funding
NASA Earth System Science Fellowship NIGEC NSF