Transcript Document

Statistical approach
• Statistical post-processing of LPJ output
• Analyse trends in global annual mean NPP based
on outputs from 19 runs of the LPJ model
• Runs forced using a total of 18 ensembles from 9
GCMs, and using gridded CRU data
• Analysis (partially) deals with climate uncertainty,
but does not deal with parameter or structural
uncertainties in the LPJ model
Motivating factors
• Statistical pre-processing of LPJ inputs is tough:
would need to describe month-to-month trends in
three climate variables for each location
• GCMs are each run at different spatial resolutions,
all of which differ from the resolution of the CRU data
• LPJ is computationally intensive to run
• No useful observational data to validate LPJ against
Time series model
Use a hierarchical time series model to draw
inferences about “true” response of LPJ model to
projected climate changes based on the 19 runs
Output from past year t using CRU data:
xt ~ N(t , vt )
Output for past or future year t using run i of GCM I:
yit ~ N(zIt , )
Assume conditional independence in both cases
Latent trends
Model trends in true signal t and GCM biases YIt - t
as independent random walks: e.g.
t ~ N(t 1 , s  t )
 allows process variability to change linearly over time
Can fit as a Dynamic Linear Model using the Kalman
filter – easy to implement in R (sspir package)
Parameter estimation by numerical max likelihood
Results - temperature
NPP
Assumptions
• Observational errors are IID and unbiased
• Inter-ensemble variabilities for a given GCM are IID
• Random walk model can provide a good description
of actual trends
• Levels of variability do not change over the course
of the runs (except for a jump at present day)
Inter-ensemble variability
Future work - methodology
Explore impacts of making different assumptions about
the biases in the GCM responses
Explore impacts of varying levels of inter-ensemble
variability and observation error
Explore links between this and a regression-based
(ASK-like) approach
Deal with uncertainty in estimation of parameters in
time series model – e.g. a fully Bayesian analysis
BUGS
BUGS: free software for
fitting a vast range of
statistical models via
Bayesian inference
Provides an environment for
exploring the impacts of
different assumptions
Allows for the use of
informative priors
[http://www-fis.iarc.fr/bugs/wine/winbugs.jpg]
http://mathstat.helsinki.fi/openbugs
http://www.mrc-bsu.cam.ac.uk/bugs
Bayesian analogue of the DLM
z It  t  bIt
t  2t 1  t 2 ~ N (0, )
Problems:
Lack of identifiability
Bias terms are not really AR(1)
bI ,t   I bI ,t 1 ~ N (0,  I )
A Bayesian ASK-like model
M
t   wI z It  bt
I 1
zIt  2zI ,t 1  zI ,t 2 ~ N (0,  I )
Problems:
Lack of fit
Unconstrained estimation leads
to weights outside range [0,1]
bt  bt 1 ~ N (0, )
Open questions
– statistical methodology
• What assumptions can we make about the biases
in GCM responses and in the observational data?
• How reasonable is the assumption that future
variability is related to past variability, and how
far can we weaken this assumption?
• How should we best deal with small numbers of
ensembles & unknown levels of “observational
error”? Can we ellicit more prior information?
Future work - application
Apply analysis to output from newer version of LPJ
Apply a similar analysis at the regional scale
Extend approach to other variables, especially PFT
Analyse outputs from multiple SRES scenarios
Open questions - application
Should LPJ be run at the native spatial scale of the
data/GCM that is being used to force it ?
LPJ includes stochastic modules – switched off here,
but how could we best deal with these…?
For a limited number of runs what experimental design
would enable us to best reflect the different elements
of climate and impact uncertainty?
Context: the ALARM project
Assessing impacts of environmental change upon
biodiversity at the European scale
Modules: climate change, environmental chemicals,
invasive species, pollination
Relies heavily upon climate and land use projections
Impacts assessed using either via mechanistic models
(e.g. LPJ) or through extrapolation from current data
Contact us
Adam Butler
[email protected]
Ruth Doherty
[email protected]
Glenn Marion
[email protected]