For our lab meetings in the Williams Lab we’ve begun producing some Pecha Kucha presentations on recent papers in the literature. Every few weeks the grad students and post-docs will present a brief talk on a paper that’s caught their eye. This week I’m going to talk a bit about a new paper in the Journal of Biogeography (so new it’s still Online Only) by Florian Hartig and others about ways to connect dynamic vegetation models (DVMs) to raw data.
Hartig F, Dyke J, Hickler T, Higgins SI, O’Hara RB, Scheiter S, Huth A. Connecting dynamic vegetation models to data – an inverse perspective. Journal of Biogeography. Online Only
Hartig and co-authors summarize the utility of DVMs , to do so they set up a dichotomy:
- Species Distribution Models (SMDs), largely correlative, drawing on the realized niche of species under modern conditions to develop relatively easy to implement predictive models
- DVMs, bottom-up models based on model parameters for physiology, growth and other factors drawn from experimental or estimated data that are far more complex, ut potentially explore the potential niche more fully.
The SDMs can be considered inverse models, they predict vegetation based on a model that relates existing climate to vegetation. We ultimately know vegetation, but (in the simplest case) not how climate structures vegetation on the landscape. By predicting vegetation using a model calibrating current vegetation to known climate we can not only predict vegetation in other domains (or in the present domain), but, depending on the model, we begin to derive knowledge about the role of climate in structuring vegetation. Inverse models are the basis of pollen-based climate modeling, although this is changing somewhat (see Birks et al., 2010 for an excellent review).
The problem with DVMs is that they are both data hungry and that the types of data available for validation and constraint are not always directly useful in the model construction. For example, although models keep track of Net Primary Productivity (NPP), it is difficult to directly integrate NPP records for a region or through time, especially given the potential variability in NPP estimates at large temporal and spatial scales and its autocorrelation at small spatial scales. For this reason Bayesian methodologies provide an excellent solution. It is possible to link the data provided by (for example) paleorecords, given some basic understanding of the structure and nature of the parameters that would provide the links. Given what we know about how tree ring records represent NPP, we can link the data to the DVM through Bayesian models, providing a error model for the DVM that can then be used to assess the model output and to ultimately improve the model.
This is a similar approach to the modeling we are undertaking in PalEON. As mentioned elsewhere, I have been developing the pre-settlement vegetation data to be assimilated into the DVMs, but PalEON is also developing pollen-based estimates of past vegetation, dendrochronological records and flux tower measurements to further improve the DVMs. In these cases, the key is developing reliable estimates of error surrounding the estimates, in a Bayesian framework, so that they can be integrated into the modeling process. It’s all very exciting, and I’m looking forward to getting deeper into Bayesian models in the future.