1. Jean-François Exbrayat1 T.L. Smallman1, A.A. Bloom2, M. Williams1
Using a data-assimilation system to assess the influence of fire on simulated carbon fluxes and plant traits for the Australian continent
Jean-François Exbrayat1
T.L. Smallman1, A.A. Bloom2, M. Williams1
1School of GeoSciences and National Centre for Earth Observation, University of Edinburgh, UK
2Jet Propulsion Laboratory, California Institute of Technology, Pasadena, California, USA
EGU Assembly 2015 – 17 April 2015

2. Background
Fire emissions are not always considered in Global Vegetation Models
Vegetation-fire feedbacks regulate ecosystems structure and differentiate between savannas and forest
Figure 4 from Staver et al. (Science, 2011)

3. Background
Fire emissions are not always considered in Global Vegetation Models
Vegetation-fire feedbacks regulate ecosystems structure and differentiate between savanna and forest

4. Science questions
Can we identify large-scale fire-related differences in ecosystem functional properties using remote-sensing data?
How do these map onto existing land cover maps used in Global Vegetation Models?

5. DALEC: Data-Assimilation Linked Ecosystem Carbon model
GPP calculated with ACM model (Williams et al., Ecol Appl 1997)
20 process parameters to simulate pool allocation, phenology and turnover rates
Phenology based on Growing Season Index Jolly et al. (GCB 2005)
6 initials pool sizes
Williams et al. (GCB 2005)
Bloom and Williams (BG, 2015)

6. DALEC: Data-Assimilation Linked Ecosystem Carbon model
GFED Burned Area transformed to Burned Fraction
Fixed combustion rates and emission factors
Fire accelerates mortality and emits carbon

7. Markov-Chain Monte Carlo
CARDAMOM - CARbon DAta Model fraMework
DRIVERS
Weather data
GFED burned area
INITIAL CONDITIONS (TBC from Saatchi et al., PNAS 2011)
DA in each pixel
Ecological and Dynamic Constraints (Bloom and Williams, BG 2015)
No PFTs
Continuous maps of parameters
OBSERVATIONS
OUTPUT
PDFs of parameters, pools, fluxes, etc…
Update parameters
Markov-Chain Monte Carlo

8. Australian C balance Source Sink NEE = -GPP + Ra + Rh NBP = NEE + FIRE
1990 – 2009
Haverd et al. (2013, BG)
2000 – 2012
This study (mean +/- std)
GPP
-4110 Tg C yr-1
-4621
+/ Tg C yr-1
NPP
-2210 Tg C yr-1
-2258
+/- 777 Tg C yr-1
NEE = -GPP + Ra + Rh
NBP = NEE + FIRE

9. Regional fluxes

10. Phenology parameters
Northern regions tolerate a higher VPD before stopping leaf production
SW and SE crop regions have well-defined value of low VPD tolerance – reality or artefact due to lack of management in drivers?

11. Parameter maps vs land cover
Some land cover types can be identified from retrieved parameter maps although DA does not used any discrete land cover data
True for some parameters, but…

12. Carbon Use Efficiency
Fire-prone regions in NW have higher NPP:GPP ratio (lower uncertainty)
Ecosystems adapted to improve CUE to survive repeated fire impacts
Parameter maps does not correspond to land cover classification

13. Outlook & Conclusion
Australian fire-prone regions optimize CUE to cope with fire impacts.
Is it true globally?
PFTs may need to be refined as a function of fire-regimes
What aspect of fire dominates: intensity, frequency, both?
Australia is likely a sink of carbon but intense fire seasons may temporarily turn it into a source
Ways forward :
Integrate information about managed land area, e.g. soil moisture dynamics as a proxy for irrigation, harvesting dates, FORMA dataset in the wet tropics
Assimilate time series of biomass (Liu et al., NCC 2015) to constrain post-fire regrowth dynamics

14. Questions?
Another CARDAMOM application on the added value of repeated biomass measurements to reduce model uncertainty
Can we reliably estimate managed forest carbon dynamics using remotely sensed data?
T. Luke Smallman, J.-F. Exbrayat, A. A. Bloom, and M. Williams
EGU – 16:30 Room G5

15. Growing Season Index
Leaf production (Clabile  Cfoliar) and leaf fall (Clabile  Clitter) are scaled by a Growing Season Index.
Product of three linear functions of minimum air temperature, Vapour Pressure Deficit and Photoperiod
Red lines indicate optimal and limiting conditions and correspond to parameters that are optimized during the data-assimilation procedure
after Jolly et al. (GCB, 2005)

16. Model Data Fusion (MDF)
Random Sampling of DALEC parameters
Bayes’ Theorem
p(x|c) ∝ p(c|x) p(x)
A. p1, p2, …. , p23
B. p1, p2, …. , p23
C. p1, p2, …. , p23
Observation
likelihood, given parameters
Posterior parameter probability
Prior Parameter Probability
DALEC
Method = Metropolis Hastings MCMC
(1) Parameter value priors span across multiple orders of magnitude, BUT
(2) Only a subset of parameter space can be considered “ecologically consistent” A
We have a model, with a set of parameters. Data driven parameter estimation method - Determine model parameters, given data constraints. The challenge is can we determine the parameter values given the available data? B C
DATA (e.g MODIS LAI)

17. Data-independent Ecological and Dynamic Constraints (EDCs)
Turnover constraints
Dynamic constraints
CPOOL 1 2 3
years No
Yes
tSOM < tlitter & twood < tfoliar
Croot: Cfol ratio 1
0.2 5 p
CFOL : CROOT
Bloom and Williams, Biogeosciences 2015

18. Comparison to GFED

19. Correlation of median parameter values

20. Terrestrial Carbon Cycle
CARDAMOM - CARbon DAta Model fraMework
MODEL
DALEC: Data Assimilation Linked Ecosystem Carbon model
DATA
- Initial conditions (SOC, AGB),
- Biometric Satellite data, Eddy flux tower data, Plant trait data.
Terrestrial Carbon Cycle
Analysis
DRIVERS
Weather re-analyses
GFED Burned Area
LULCC
ASSIMILATION
- Metropolis-Hastings Markov Chain Monte Carlo
- Ecological and Dynamic Constraints (EDCs)
model and data knowledge across a range of spatiotemporal scales, by constructing a Carbon Model Data Framework.
components can be inter-changed to adapt for a wide range of carbon cycle problems.