1. Ruth Doherty, Edinburgh University Adam Butler & Glenn Marion, BioSS
Impact of climate uncertainty upon trends in outputs generated by an ecosystem model
Ruth Doherty, Edinburgh University
Adam Butler & Glenn Marion, BioSS
ALARM meeting, Athens, January 2007

2. Acknowledgements LPJ code: Ben Smith, Stephen Sitch, Sybil Schapoff
CRU data: David Viner
GCM data: PCMDI
Statistical methods: Jonathan Rougier, Chris Glasbey
Uncertainty analysis: Bjoern Reineking, Stijn Bierman

3. Aim
Quantify uncertainties in projections of global & regional vegetation trends for the 21st century from the LPJ ecosystem model, based on future climate uncertainty within the SRES A2 scenario

4. The LPJ model http://www.pik-potsdam.de/lpj/
“…the Lund-Potsdam-Jena Dynamic Global Vegetation Model combines process-based, large- scale representations of terrestrial vegetation dynamics and land-atmosphere carbon and water exchanges in a modular framework”

5. Drivers
Vegetation
Dynamics
(annual)
Fluxes
(daily)

6. Sources of uncertainty in LPJ
Zaehle et al. (2005) analysed parameter uncertainty: used Latin hypercube sampling to sample uniformly over values of 14 functionally important parameters
Cramer et al. (2001) and Smith et al. (2001) analysed structural uncertainty, by looking at alternative parameterisations of processes within LPJ
Estimated uncertainty range (NPP, ):
43 –103 PgC/yr in Zaehle et al., 2005
44 – 66 PgC/yr in Cramer et al., 2001
Zaehle et al.: uncertainty range increases in future

7. Climate uncertainty LPJ is driven by climate, CO2 and soils data.
Future climate inputs are uncertain, due to:
Future emissions: choice of SRES scenario
Choice of GCM (climate model)
Intra-model uncertainty for each GCM

10. Climate model runs GCM simulations from the IPCC 4th Assessment
We consider only SRES scenario A2
We use 17 ensemble runs, from a total of 9 GCMs
GCMs with multiple ensembles
CCCMA-CGCM3, MPI-ECHAM5, NCAR-CCSM3
GCMs with a single run
CNRM-CM3, CSIRO-MK3, GFDL-MK2,
MRI-CGCM2-3, UKMO-HADCM3, UKMO-HADGEM

11. LPJ model runs We run LPJ 18 times at a global scale
Soil inputs: FAO global soils dataset, with 9 types
CO2 inputs
Climate inputs:
monthly temperature, precipitation, solar radiation
control run: gridded 0.5o x 0.5o CRU data for
other runs: GCM model runs for , with LPJ run at native spatial scale of the GCM
Spin-up period of 1000 years at start of each run
Run on average grid-cell basis with 1-year time-step

12. LPJ Outputs Daily: carbon and water fluxes
Annual: vegetation dynamics and competition amongst 10 Plant Functional Types
Spatial scale of outputs varies, depending on scale of the climate data / model used to provide the inputs
We analyse trends from 2002 to 2098 in global annual values of vegetation carbon, soil carbon & NPP
Adam to continue …

13. Systematic biases
LPJ runs using GCMs exhibit systematic biases – presumably related to coarse spatial scale
By calibrating against the LPJ control run we can use a statistical model to describe the statistical properties of these biases over the period
This model can then, along with the LPJ runs under scenario A2, be used to predict the response of the LPJ model to climate over the 21st century

14. Statistical methodology
Past
t = years 1900,…,2001
k = GCM run 1,…,17
We have data on:
xt = LPJ control run
ykt = LPJ run using GCM run k
bkt = xt - ykt (bias in run k)
Assume bkt = k + ekt, where:
ekt is AR(1): ekt ~ N(k ek,t-1 ,k2)
vague priors on k, k ,k,ek,1899
Future
t = years 2002,…,2098
k = GCM run 1,…,17
We have data on:
ykt = LPJ run with GCM k
Predict Xt = BKt + yKt
K is randomly chosen GCM run:
K = k with probability 1/17
BKt is predicted using the fitted
AR(1) model for {bkt}

15. Statistical assumptions
Historical biases between the control & GCM-forced runs can be described by a simple time series model
Future biases have the same distributional properties as historical biases
The future LPJ runs provide equal information about year-to-year variations in vegetation characteristics
The control run of LPJ rovides an error-free and unbiased representation of current vegetation

16. Results
Fit using LinBUGS ( free software for fitting a vast range of statistical models via Bayesian inference
Can obtain similar results using ARIMA() function in R: but this does not account for estimation uncertainty

17. Simulated data

21. Annual global NPP

24. Annual global soil carbon

27. Annual global vegetation carbon

30. Diagnostics Does the AR model describe historical biases well?
Model checking:
plot of residuals from model, sample autocorrelations, estimates for k
sensitivity of predictions to value of K
Possible extensions:
long-term linear or quadratic trends
higher-order terms in an ARIMA model
model responses y1t,…,y17,t as covariates

31. Annual global NPP

36. Future work Improve time series model for bias terms
Investigate possible reasons for systematic bias
Apply a similar analysis at the regional scale
Analyse outputs from the other SRES scenarios
Incorporate global satellite data on NPP…?

37. Open questions
How reasonable is the assumption that future biases are related to past biases?
Should we assign equal weights to model runs?
Should we run LPJ at the native spatial scale of the climate model that is being used to force it?
We use statistical post-processing – could we use statistical methods to generate climate inputs for LPJ?
LPJ can be run with stochastic modules – how could we incorporate uncertainty from these?

38. Contact us Adam Butler adam@bioss.ac.uk Ruth Doherty
Glenn Marion
File: created 11 December, last modified 13 December, author Adam Butler