Biodiversity measurements

Steinmann, Linder, and Zimmermann. 2009. Modelling plant species richness using functional groups. Ecological Modelling

In the ever-elusive goal of better plant richness modelling, these folk take on the challenge by using various higher order divisions in their models, to see what shakes out best for modelling plant species richness across Switzerland. They mainly attempted a middle-path sort of route- not a bottom-up, model the individual plant responses to environmental gradients and then overlay all those maps together (a common approach which completely disregards community interactions), and not a top-down, community model which doesn't model individual species responses (an obvious deficiency).

They found little benefit to their approach, which was sort of a let-down- but, in another way, refreshing. Non-results are rarely published, which is a shame. There was an interesting facet in that some functional groups were modeled far better than others, indicating some interesting differences (but, one of the best were trees, which usually model well because of their longevity and, therefore, sensitivity to environmental factors/relative immunity from stochastic events).

Rocchini, Chiarucci, and Loiselle. 2004. Testing the spectral variability hypothesis by using satellite multispectral imagery. Acta Oecologica

My interest in the spectral variability hypotheses (also described in an earlier post about Palmer, et al 2002) stems mainly from my dislike of classified imagery as such. It eliminates so much information from the image, and locks you into whatever errors are present at the time of the classification. The SVH is a way around it, but relatively unexplored. Rocchini and others seem to be the only group working on the idea, and they always use high spatial resolution stuff (they also have a kicking dataset, like Steinmann et al, which apparently they get mostly from other people. Lucky.). They've published a few more papers on the idea, but this 2004 is the first, so... start at the beginning.

They explain ~50% of the species richness at a 1 ha scale (over landscapes, of course), which is pretty impressive, since it's all done without looking at the individual spectral responses of the plants. The image is ground-truthed, of course, but only by count; i.e. anybody can do that, without any special tools. Plus, this is with Quickbird, so it's only four bands, and I'd bet good money you can improve on that 50% really easily with some higher spectral resolution stuff. At 30 m Landsat? That's a good question, and I'm sure it really depends on the scale of vegetiation present. People have used NDVI for this sort of thing via Landsat (Gould 2000), but I'm unaware of any attempts at full-on variability style methodologies.

VP free

Variance partitioning
In an attempt to re-understand VP, I’m going to attempt to explain it to myself in word-form…

Variance, in regressions, is always present. R2 values are given to illustrate the amount of variance the model (regression) explains. Multiple regressions use several variables (x1, x2, etc) in an attempt to explain the response variable (y); each x contributes some “explanatory power” to the model. That’s stuff we know already…

How much does each x contribute to the explanatory power of the model? Well, first note that the order of insertion of independent variables into the model doesn’t change the final result (x1b + x2b + x3b…). Intermediate results will differ, but the final result is the same. Secondly, the contribution (in terms of increasing r2, if that’s our scorecard) of the second variable entered will change depending on what was entered first. Not only that, some variables which don't add anything to r2 might still be important, as they could mediate the effect of another variable in some important way. Overlapping disturbances would be a candidate.

Basically, it appears that you look at the change in r2 when each independent variable is added last, and then call that difference the “unique variance.”
Nothing super new so far, it’s the commonality analysis which seems like it’ll be useful- is it possible to look at the influence of A on B even if A adds nothing to an explanatory model of B, but mediates the effect of variable C on B?

Perhaps it works this way: A -> B <- C


Or this: C -> A -> B <- C

In the second case, A would appear, in a regression, to add nothing to a predictive model of B, but perhaps it’s mediating the effect of C in some way. Since they are correlated, it would be difficult (impossible?) to separate the two influences specifically, but perhaps it would be possible to get a measure of the importance of C via subtraction…

VARIATION PARTITIONING OF SPECIES DATA MATRICES: ESTIMATION AND COMPARISON OF FRACTIONS
Pedro R. Peres-Neto1, Pierre Legendre, Stéphane Dray, and Daniel Borcard
Ecology 87(10)

Haven't been posting, but have been reading...

This morning-

Palmer, M., et al. 2002. Quantitative tools for perfecting species lists. Environmetrics. 13:121-137

Zimmermann, N., et al. 2007. Remote sensing based predictors improve distribution models of rare, early successional and broadleaf tree species in Utah. Journal of Applied Ecology. 44:1057-1067