Huth, A.; Dislich, C.; Florian, H. & Thorsten, W. (2014): <b>Approximate Bayesian parameterization of a process-based tropical forest model</b>. <i>Biogeosciences</i> <b>11</b>, 1261-1272.
Resource Description
Title:
Approximate Bayesian parameterization of a process-based tropical forest model
FOR816dw ID:
1370
Publication Date:
2014-03-01
License and Usage Rights:
PAK 823-825 data user agreement. (www.tropicalmountainforest.org/dataagreementp3.do)
Resource Owner(s):
Individual:
Andreas Huth
Contact:
email:
Andreas.Huth <at> UFZ.de
P.O. Box 500 136
UFZ Centre for Environmental Research Leipzig-Halle
Dept. of Ecological Modelling
04301 Leipzig
Germany
Individual:
Claudia Dislich
Contact:
email:
claudia.dislich <at> ufz.de
UFZ Centre for Environmental Research Leipzig-Halle
Department of Ecological Modelling
Permoserstrasse 15
04318 Leipzig
Germany
Individual:
Hartig Florian
Contact:
email:
webmaster <at> tropicalmountainforest.org
Individual:
Wiegand Thorsten
Contact:
email:
webmaster <at> tropicalmountainforest.org
Abstract:
Inverse parameter estimation of process-based<br/>
models is a long-standing problem in many scientific disciplines.<br/>
A key question for inverse parameter estimation<br/>
is how to define the metric that quantifies how well model<br/>
predictions fit to the data. This metric can be expressed by<br/>
general cost or objective functions, but statistical inversion<br/>
methods require a particular metric, the probability of observing<br/>
the data given the model parameters, known as the<br/>
likelihood.<br/>
For technical and computational reasons, likelihoods for<br/>
process-based stochastic models are usually based on general<br/>
assumptions about variability in the observed data, and not<br/>
on the stochasticity generated by the model. Only in recent<br/>
years have new methods become available that allow the generation<br/>
of likelihoods directly from stochastic simulations.<br/>
Previous applications of these approximate Bayesian methods<br/>
have concentrated on relatively simple models. Here, we<br/>
report on the application of a simulation-based likelihood<br/>
approximation for FORMIND, a parameter-rich individualbased<br/>
model of tropical forest dynamics.<br/>
We show that approximate Bayesian inference, based on<br/>
a parametric likelihood approximation placed in a conventional<br/>
Markov chain Monte Carlo (MCMC) sampler, performs<br/>
well in retrieving known parameter values from virtual<br/>
inventory data generated by the forest model. We analyze<br/>
the results of the parameter estimation, examine its sensitivity<br/>
to the choice and aggregation of model outputs and<br/>
observed data (summary statistics), and demonstrate the application<br/>
of this method by fitting the FORMIND model to<br/>
field data from an Ecuadorian tropical forest. Finally, we discuss<br/>
how this approach differs from approximate Bayesian<br/>
computation (ABC), another method commonly used to generate<br/>
simulation-based likelihood approximations.<br/>
Our results demonstrate that simulation-based inference,<br/>
which offers considerable conceptual advantages over more<br/>
traditional methods for inverse parameter estimation, can be<br/>
successfully applied to process-based models of high complexity.<br/>
The methodology is particularly suitable for heterogeneous<br/>
and complex data structures and can easily be adjusted<br/>
to other model types, including most stochastic population<br/>
and individual-based models. Our study therefore provides<br/>
a blueprint for a fairly general approach to parameter<br/>
estimation of stochastic process-based models.
Keywords:
| FORMIND | aboveground biomass | forest model |
Literature type specific fields:
ARTICLE
Journal:
Biogeosciences
Volume:
11
Page Range:
1261-1272
Metadata Provider:
Individual:
Andreas Huth
Contact:
email:
Andreas.Huth <at> UFZ.de
P.O. Box 500 136
UFZ Centre for Environmental Research Leipzig-Halle
Dept. of Ecological Modelling
04301 Leipzig
Germany