Traditional Agroforestry System in India- Shifting Cultivation, Taungya, Home...
Constraining terrestrial carbon balance through assimilation of remotely sensed biomass data into CARDAMOM
1. Constraining terrestrial carbon balance
through assimilation of remotely sensed
biomass data into CARDAMOM
Mathew Williams
Luke Smallman and Jeff Exbrayat
University of Edinburgh and NCEO, Edinburgh, UK.
Anthony Bloom, JPL, Pasadena, USA
BIOMASS
2. Summary
• Challenge of understanding a complex system
• Opportunity provided by new observations
• Example of a model-data fusion system
• Observations reduce parameter uncertainty
• Tackling, fine scale dynamic
• Constraining forecasts using data
• The way ahead
3. C cycle projections differ significantly
Jones et al. 2013
Different
model
runs
Varying internal
dynamics
4. Williams et al. 2009
Improving process models with data
BIOMASS
5. Land surface models
Complex
Model
Climate
Land cover/
PFT
Process
parameters
Initial
conditions
C
stocks/fluxes
(mod)
C fluxes (obs)
[stocks]
compare
simulate
spin-up
forcing
Field
studies
Uncertainty?
8. Terrestrial ecosystem carbon cycle analysis
p(x|c) ∝ p(c|x) p(x)
Parameter probability p(x|c) at each pixel derived
using a Metropolis-Hastings MCMC algorithm
Biometric Data Constraints
DALEC
Drivers:
ERA-interim 1° x 1°
resolution 8-day
time-step 2001-2010
Ceff constraint:
100 Gt/yr < Global
GPP <150 Gt/yr
Posterior DALEC
Parameter
Probability
1° x 1° Pixel scale
parameter, flux &
carbon pool
estimates
MODIS LAI Pan-Tropical Biomass
HWSD Soil Organic C
Bloom et al., PNAS, 2016.
C state likelihood function
= observation likelihood &
parameter priors
No spin-up
No PFTs
No Steady state
9. Global retrievals of carbon residence times
Bloom et al., PNAS 2016
RT ~ Pool size / Input
10. Land cover classifications do not
match parameter mappings
AF – allocation fraction
RT - residence time
Retrieved C
state and
process
variables are
largely
explained by
eight modes of
spatial
variability
11. Reduction of uncertainty with biomass prior
Quegan et al. (in review.)
Single biomass prior (Avitabile et al. 2016) lead to
~50% reduction of uncertainty (CI95) in retrieved
turnover times
BIOMASS
12. • Aggrading forest, non steady-
state
• Duke Forest, NC (Loblolly Pine
planted in 1983)
• What information is required to
constrain model parameters?
Constraining local dynamics
• Coarse resolution quasi steady state analyses
are tractable
• At finer scale dynamical processes
(disturbance and aggradation) provide a
modelling challenge
13. Information content of observations
Data Reference Multi
Wood
One
Wood
Man
Only
No
Man
Foliar C/LAI In situ MODIS MODIS MODIS MODIS
Root C In situ - - - -
Wood C
Soil C
In situ
In situ
annual
HWSD
one
DWSD
age
HWSD
-
HWSD
Concept Field site BIOMASS Avitabile LC history Baseline
15. Challenges for high resolution DA
• Managing large data (capacity)
• Scale of assimilation (pixel, PDF…?)
• Dynamic systems (non steady state)
• Propagating and scaling error
Williams et al. 2013
Milodowski et al. 2017
RapidEye images, Brazil ALOS-PALSAR biomass estimates, Mozambique
16. Constraining forecasts through
reliability assessments
• Multi-model averages are typically used to
generate forecast best-estimates
• Model skill can be used as a constraint
x 5 climate
forcings
17. Reliability Ensemble Averaging
• Skill-based average of 30 ISIMIP projections of end of 21st century ΔNPP under
RCP8.5 conditional of models’ current performance (wrt CARDAMOM) and inter-
model convergence
• ΔNPPISIMIP = 24.2±36.3 Pg C y-1 ; ΔNPPREA = 24.6±8.5 Pg C y-1 (+2% mean ΔNPP, 68%
uncertainty reduction)
Exbrayat et al. (ESD 2018)
1 SD error
18. Differences between NPP in 2095–2099 compared to
2001–2005 from the REAC average and ISIMIP
ensemble mean (gCm−2 y−1).
Red indicates where the REA averages predict NPP greater than the
ISIMIP ensemble mean.
Blue indicates where the REA averages predict NPP less than the ISIMIP
ensemble mean.
Exbrayat et al. (2018)
19. Next steps
• Global C-water assimilation using a range of
model complexity, linking to new EO products
• REA analyses using broader range of C cycle
variables from CARDAMOM
• Links to global trait databases
• SIF, atmospheric CO2 data
• Global analysis of information content of
repeated biomass maps – mission preparation
20. Summary
• C cycle data is growing exponentially, higher
spatial/temporal resolution, more diverse
• Huge opportunity to calibrate and evaluate
process models
• Massive challenge to maximise data value and
develop knowledge
• Combined efforts of biogeochemical,
ecological, remote sensing, data science,
modelling communities required
21. Acknowledgements:
NASA MODIS team, Saatchi et al., Avitabile et al.,
Duke Forest team, Biomass MAG, HWSD team,
FLUXCOM partners
Funding: NERC, NCEO
22. Terrestrial
Carbon Cycle
Analysis
MODEL
DALEC: Data
Assimilation Linked
Ecosystem Carbon
model
DATA
- Pool sizes (SOC, AGB),
- Biometric Satellite data, Eddy flux
tower data, Plant trait data.
ASSIMILATION
- Metropolis-Hastings Markov
Chain Monte Carlo
- Ecological and Dynamic
Constraints (EDCs)
DRIVERS
Weather data
GFED Burned Area
(Deforestation)
CARDAMOM - CARbon DAta Model fraMework
23. C cycle models have major differences
Natural sink
Natural
sink and
LUC
Friedlingstein et al. 2014
Historical land
carbon fluxes
from 11 ESMS
24. Impact of fire on turnover time
after Exbrayat et al. (under review)
Climate, HWSD SOM, Avitabile biomass, (GFED burned area)
25. Ratio of the uncertainty in 21st century NPP from each REA
average to the uncertainty in the ISIMIP ensemble.
This reduction of uncertainty leads to a confidence on
the sign estimation of NPP in 86% of all the land pixels
(stipling) compared with 43% for the ISIMIP ensemble.
Exbrayat et al. (2018)