gfÖ 2013 talk: connecting dynamic vegetation models to data - an inverse perspective
DESCRIPTION
Talk at the GFÖ meeting 2013 in PotsdamTRANSCRIPT
Florian Hartig
Department of Biometry and Environmental System Analysis
Florian Hartig
Department of Biometry and Environmental System Analysis
Connecting dynamic vegetation models to data -
an inverse perspective
Florian Hartig, University of Freiburg
http://florianhartig.wordpress.com/ GFÖ, 2013, Potsdam
Figures by Ernst Haeckel, scans by Kurt Stüber, MPI Köln
Florian Hartig
Department of Biometry and Environmental System Analysis Purves, D. et al. (2013) Ecosystems: Time to model all life on Earth, Nature, 493, 295-297
Florian Hartig
Department of Biometry and Environmental System Analysis
For vegetation models, lots of data
available
► On plant traits
► On a large number of
vegetation distributions /
responses (Hartig et al., 2012)
► The real problem seems
to be to bring this data
together with models in
a meaningful way!
Page 3
Hartig et al. (2012) Connecting dynamic vegetation models to data -
an inverse perspective. Journal of Biogeography, 2012.
Florian Hartig
Department of Biometry and Environmental System Analysis
Statistical (correlative) approaches to
using vegetation data
► Response: distribution,
growth, …
► Relate response to other
factors (e.g. soil,
climate) with a simple
relationship
► Essentially inter /
extrapolation of
pattern; difficult to
translate between
different data types
Page 4
Thuiller et al. (2011) Consequences of climate change on the tree of
life in Europe Nature.
Florian Hartig
Department of Biometry and Environmental System Analysis
Dynamic (process-based)
vegetation models
Pioneer
Intermediate
Climax
FORMIND animation of a model parameterization for a forest in South Ecuador,
1900-2100 asl ; details see Dislich, C. et al., Simulating forest dynamics of a tropical
montane forest in South Ecuador, Erdkunde, 2009, 63, 347-364
Florian Hartig
Department of Biometry and Environmental System Analysis
Recent review
Bayes’ Formula
Direct information
on parameters
Inverse information
on parameters based on
data D on model outputs
Posterior
probability distribution
for parameters Q
Florian Hartig
Department of Biometry and Environmental System Analysis
Example: inverse calibration to stand data
► „Vague“ prior information
► Parameter estimation with stand
data across Europe
► Result: better parameters, model
comparison, averaged prediction!
Page 7
Florian Hartig
Department of Biometry and Environmental System Analysis
Example: inverse calibration to
distribution data
► Physiological model
fit to distribution of
22 European tree
species
► Predicts of course
distributions, but
also carbon / N
uptake …
Page 8
Florian Hartig
Department of Biometry and Environmental System Analysis
Example: inverse calibration to
distribution data
► Physiological model
fit to distribution of
22 European tree
species
► Predicts of course
distributions, but
also carbon / N
uptake …
Page 9
Florian Hartig
Department of Biometry and Environmental System Analysis
Interim summary
Page 10
► Bayes allows us to fit process-based
models with direct and inverse data
in a statistically meaningful way
► Because process-based models
couple to many outputs
► Data-translators!
► Data synthesizers – challenge:
meaningfull likelihood!
Florian Hartig
Department of Biometry and Environmental System Analysis
How to define the inverse term in Bayes
formula?
► As in statistical model, define the
probability of deviating from mean
model predictions by some probability
density function
► Fine for simple problems, problematic
for strongly stochastic models and
for fitting to heterogeneous data Hartig et al. (2012) Connecting dynamic vegetation models to data - an inverse
perspective. Journal of Biogeography, in press.
Florian Hartig
Department of Biometry and Environmental System Analysis
Generating complicated likelihood functions:
simulation-based likelihood approximation
Pioneer
Intermediate
Climax
FORMIND animation of a model parameterization for a forest in South Ecuador,
1900-2100 asl ; details see Dislich, C.; Günter, S.; Homeier, J.; Schröder, B. & Huth,
A., Simulating forest dynamics of a tropical montane forest in South Ecuador,
Erdkunde, 2009, 63, 347-364
Local biomass results of 1600
model runs
Field data D
Florian Hartig
Department of Biometry and Environmental System Analysis
Generating complicated likelihood functions:
simulation-based likelihood approximation
Pioneer
Intermediate
Climax
FORMIND animation of a model parameterization for a forest in South Ecuador,
1900-2100 asl ; details see Dislich, C.; Günter, S.; Homeier, J.; Schröder, B. & Huth,
A., Simulating forest dynamics of a tropical montane forest in South Ecuador,
Erdkunde, 2009, 63, 347-364
Local biomass results of 1600
model runs
Field data D
Florian Hartig
Department of Biometry and Environmental System Analysis
A practical example: fit to data from
Reserva Biológica San Francisco, Ecuador
Pioneer
Intermediate
Climax
FORMIND animation of a model parameterization for a forest in South Ecuador,
1900-2100 asl ; details see Dislich, C.; Günter, S.; Homeier, J.; Schröder, B. & Huth,
A., Simulating forest dynamics of a tropical montane forest in South Ecuador,
Erdkunde, 2009, 63, 347-364
Probability distr. for stem
size distributions
Probability distr. for growth
rates per size class
Florian Hartig
Department of Biometry and Environmental System Analysis
A practical example: fit to data from
Reserva Biológica San Francisco, Ecuador
Pioneer
Intermediate
Climax
FORMIND animation of a model parameterization for a forest in South Ecuador,
1900-2100 asl ; details see Dislich, C.; Günter, S.; Homeier, J.; Schröder, B. & Huth,
A., Simulating forest dynamics of a tropical montane forest in South Ecuador,
Erdkunde, 2009, 63, 347-364
Probability distr. for stem
size distributions
Probability distr. for growth
rates per size class
Hartig, F.; Dislich, C.; Wiegand, T. & Huth, A. (2013) Technical Note: Approximate
Bayesian parameterization of a complex tropical forest model Biogeosciences
Discuss., 10, 13097-13128
Florian Hartig
Department of Biometry and Environmental System Analysis
Conclusions
► Using Bayes allows coupling proces-
based vegetation models to a wide range
of data (on parameters and outputs)
► Option to use simulation-based
approximations; creates statistical model
based on the ecological processes ► Correlations between heterogeneous data
► Complicated error structures
► What this means for ecological research ► Process-based as data translators and data
synthesizers
► Test of our process-understanding with ALL
data instead of isolated hypothesis with
isolated data
Page 16
Florian Hartig
Department of Biometry and Environmental System Analysis
Thank you!