OOS 81: "Hyperspectral Remote Sensing Data Supports 21st Century Ecological Research"
By the time this is posted I will have finished my ESA talk, and a link to the slides is posted here. I'm going to attempt to summarize what I talked about (technically what I'm planning on talking about) here.
One of the things I'm interested in is how imaging spectroscopy (== hyperspectral imagery) can inform fundamental questions in community ecology and biogeography. Since I now live in flat, agriculture-dominated Michigan, I'm curious if this particular landscape can tell us anything useful. Hence, this year at #ESA100 I talked about "Hyperspectral imagery for biodiversity mapping in a wildland-agriculture matrix."
Broadly, one of the questions many of us would like an answer to is "Why do plants grow where they grow?" and, critically, "Can we predict where plants will grow in the future?" One way I like to think about this is that when we have an observed landscape, we can divide its influences into a number of different categories, and we'd like to know which ones dominate a given landscape at a given scale. But five influences is too complicated, so I really like this diagram from Weiher & Keddy (1995, Oikos) putting environmental and competitive diversity on a single access. One way to think of this is then a random landscape would just look like noise, while a strongly environmentally filtered landscape would look extremely patchy (see: California. Also, caveat, things other than environmental filtering (like disturbance) can create a patchy landscape. This is not a perfect setup.) From the biogeography side we have the species-area relationship (SAR), which, while much beleaguered, can reveal interesting things about a landscape, especially when compared to some null expectation. Recently Smith et al (2013, Ecology) extended this idea to a functional-diversity-area relationship (FAR) and they suggest that when an actual FAR falls below their null expectation, that means it's a patchier landscape. And lots of other things (go read the paper!) Though subject to many of the same issues as the FAR, we can again use this null expectation or 'model landscape' approach to test hypotheses about the distribution of data in our landscape of interest.
In this talk used my own local landscape, a mix of agriculture and mesic northern forest, as two different but intermixed landscapes - one, the mesic forest, that we'd expect to be near random. It's real wet and real pleasant in Michigan, so these plants are likely more controlled by competition for light, priority effects, etc, than they are stratified by environmental gradients (though there are probably some of those, too), so we would expect to find high alpha diversity, low beta diversity. In contrast, green agricultural fields are an EXTREMELY patchy landscape in that each field is nearly homogeneous, while between field variability undoubtedly exists due to different crops being planted, different times of planting , and different management practices, so, low alpha diversity, high-ish beta diversity.
I located 'plots' (300 50x50 m plots) across this landscape, calculated principal components on the continuum removed spectral data, and looked at how the two landscapes compare. I'm essentially going to treat the first three PCs as traits. I'm fully aware that this negates a lot of the trait-based work that cares about the influence of traits on physiology, but for my purposes I'm just going to worry about spatial heterogeneity for the time being.
So, what to the PCs look like and is their variance different between forest and ag? Yes, indeed. As expected, there's a lot of variation within the forest plots, and very little within the agriculture. This is no surprise, but it should be reassuring to the community ecologists that remote sensing can say something they believe. The FAR curves also match what we would expect: given the size of my plots, the variance within a single plot is essentially equal to the variance across all plots when the data is randomized (gray line on Actual v. Predicted FARs slide looking at variance), so that line is totally flat, similar to the forest line (green) while the ag line (purple) rises steeply. In contrast, the CHV lines keep going up because the range of values for all the points is really big. Enjoy and feel free to ask questions!
ERSAM Lab Thoughts
This is where we occasionally post things that are not research, but more than 280 characters. If you want to know what we're working on now, this would be a good place to look. Code will be linked to here and posted on github.