“Ground Truthing” maxent model
Hi Keith, Todd, and fellow DLIA Science Advisors,
I did a GIS project using Maxent in the spring 2014 semester using the ten environmental layers and ATBI data I had access to.
I looked at hemlock forest decline as a function of community association, using Maxent models to delineate typical hemlock forest communities (e.g., hemlock-silverbell, hemlock-pine, etc.) and an ancillary dataset to quantify "decline" objectively.
The work is described here (you’ll notice I "borrowed" slides for class from my 2013 DLIA conference presentation):
My interest in this class project was partly motivated by a desire to "ground reference" hemlock distribution models.
For example, the maxent model can "ground truth" models based on remotely sensed photo interpretations, and vice versa.
I am unaware of how either UT, NPS or DLIA would go about "ground truthing" new predictive models produced with the new collection of 32 environmental layers.
However, the precipitous decline of Hemlock in the past decade in GRSMNP offers a macabre opportunity:
I propose that "discovering death" of hemlock is possible given the availability of high resolution satellite imagery captured by vendors supplying data to Google Earth. This imagery preserves the distinctive visual properties of hemlock skeletons that can be georeferenced.
The high mortality of hemlock in the Park may offer a unique opportunity reveal exact geographic coordinates for hemlock, revealed by their denuded skeletons. In contrast with the surrounding forest, I surmise hemlock are easily discernable – either manually or via automation – from high resolution satellite imagery available from Google Earth. I also am optimistic that the relatively small geographic area of the Park would pose a reasonable burden for manually georeferencing hemlock trees.
To collect this data, I would suggest we can use a free software package from the UN Food and Agriculture Organization called OpenForis Collect Earth <http://www.openforis.org/tools/collect-earth.html> to pinpoint locations of denuded hemlock from hi-res, 15m x 15m satellite imagery supplied by Google
This is the next best thing to "complete physical inventory," lying somewhere between exact locations provided by DLIA volunteers’ boots-on-the-ground collecting point coordinates for hemlock trees and the other side of the inventory spectrum: generalized locations extracted from 30 x 30 m remotely sensed imagery.
It is interesting to contrast point location data from the ATBI database against the Park’s official "Hemlock distribution" dataset – a point I made in my slide set for my class presentation on slide 24. This slide shows ATBI point locations for hemlock far afield from the "official" NPS hemlock polygons derived from remotely sensed photo interpretation.
Given hemlock’s unique situation in the Park, we might be able to construct a highly accurate, nearly comprehensive inventory of the locations of hemlock trees.
I believe we could then use this high-accuracy point inventory as a reference for comparison to gauge the strength of the maxent model’s ability to match the "truth on the ground" using its 32 environmental layers – with respect to hemlock as a model species, at least.
This approach would eliminate any sampling bias inherent in DLIA-supplied point locations used to inform the Maxent predictive model.
We could also compare and contrast predictive models using different datasets – a "comprehensive inventory" produced dataset, versus a "citizen science" produced dataset, and determine what, if any effect inherent "sampling bias" might have on the final product (again, with respect to hemlock trees).
Anyway, something to think about as we move forward thinking about ground-referencing the models produced from DLIA point locations and the new environmental layers.
Tanner M. Jessel
Information Technology Specialist
Center for Renewable Carbon
The University of Tennessee
Institute of Agriculture
Center for Renewable Carbon
Mail: 2506 Jacob Drive