Reduction of a complex biogeochemical model with data mining techniques
Tobias A. Sperr and Kai W. Wirtz
Ecological Modelling (to appear)
Abstract:
Recently build biogeochemical models of marine sediments are able to calculate the efflux of nitrous oxide, which is an important agent in the global climatic system.
Due to the complex nature of the model, parameterization uncertainty is high, in particular when being applied to various sites.
In order to reduce complexity without losing features relevant for nitrous oxide dynamics we first generated a large number of time series for important variables employing realistic Wadden Sea boundary conditions and random parameter sets.
Secondly, we used Kohonen-SOMs to map the data onto a set of prototype vectors which are further reduced by means of different clustering algorithms.
From the resulting set of basic states a Bayesian-type state-transition network was constructed to investigate the temporal behavior of the system. We found only a few constellations of the model with high gas efflux and also only a few number of antecedent states leading to them.
The results confirm the assumed large contribution of coastal waters to the global nitrous oxide budget.
In addition it is shown that there exists only a narrow domain in which those high emissions occur.
The dynamics of these emissions can be formalized in a reduced model frame without losing specific aspects of interest.
Finally we conclude that with our approach it is possible to visualize complex model data, take into account model uncertainty and to deduce new knowledge.
Keywords: Marine Sediments; Nitrous Oxide; Kohonen SOM; Model Reduction; Uncertainty
|
