Anthony's Science Blog

Combining process-based and spatially explicit statistical modeling

October 8, 2008 – 8:59 am

Yesterday, my fellow graduate students and I gave talks about some of our investigations of various ecosystem functions on Niwot Ridge. I talked about nitrogen cycling and others talked about microbes, plant physiology, and hydrology. Just as was planned by the conference organizers, I think our graduate work provided a nice contrast to the type of spatial modeling of plant distributions that the group in Lausanne has been using. Hopefully my nitrogen obsession didn’t bore the participants too horribly.

On Niwot Ridge, a lot of the research has focused on understanding ecological processes such as biogeochemical cycles and the biotic interactions of plants, which at this workshop we have been calling a “process-based” approach. In contrast, the group in Lausanne has been using a statistical modeling approach where they measure species distribution in the field and then attempt to back out the conditions under which they occur in order to make predictions about other places or times. This might be called a “statistical modeling” approach.

During discussions this week, we have talked about how these two approaches complement each other and how they might be combined in order to make better projections of species distributions and ecosystem processes. I might add predictions of ecosystem services, similar to what’s being done at the Natural Capital Project.

The advantage of using the process-based approach that has been the paradigm on Niwot Ridge is that mechanistic knowledge improves our power to make predictions. The disadvantage is that in order to focus on these mechanisms, we typically have to ignore spatial heterogeneity, which reduces our predictive power. In contrast, the spatial statistical modeling does a good job of incorporating differences over large areas, but does not have the mechanistic basis of the process-based models.

We have been discussing how to combine these approaches, and it’s a brain twister. But, basically, the idea put forth by Chris Randin, the organizer of the workshop, is that we can use process-based models (like CENTURY) to create spatially explicit layers that can be used in statistical modeling. This type of approach has not, to our knowledge, been used for ecosystem modeling before.

For example, we might use a soil biogeochemical model to predict nutrient availability across the landscape and then use the nutrient availability in a statistical model that predicts plant distribution. Nutrient availability would add a whole new dimension–and presumably more predictive power–that is not available in the usual statistical models that only use topoclimatic information. Basically, it would improve our estimate of the plant’s niche space by including an essential niche axis: nutrients. Conversely, once we have the model results, the areas of poor fit may then inform which ecosystem processes need further investigation. I’ve also been trying to implement this approach in my relatively small study area on Niwot Ridge to predict N cycling rates.

We’ve been chatting all day about how to rig all of this up in some kind of group-authored theory paper. There’s always the danger that too many cooks are spoiling the broth! But it’s definitely been an enlightening discussion. And once the writing actually begins, there will be just one lead author, Chris, who is a pretty smart and organized guy. I’m optimistic.

I’m psyched about tomorrow, when we will head to see some field sites in the Swiss Alps. It should be sweet and I’ll post some pictures.