The novelty of the techniques used here caused us to focus largely on data exploration
rather than specific conclusions on the study sites. The goal was to be as objective
as possible in determining vegetation groups. To accomplish this, we relied
largely on comparisons to a previous study using conventional community statistics
such as hierarchical divisive clustering such as TWINSPAN or hierarchical agglomerative
clustering. We also used these techniques to analyze cover data collected within
a week of the remote sensing flight. We decided to concentrate on the analysis
of the image of the Grant Creek Nature Preserve because it is considered the
least disturbed.
The image of the Grant Creek Prairie Nature Preserve was cropped and then georectified
using the polynomial procedure in ERDAS Imagine. The resulting image with plotted
ground reference points is shown in Figure
7. An unsupervised pixel clustering was done in MultiSpec. Thirteen clusters
were produced, each having spectral commonalities (Figure
8). The 13 clusters were initially grouped to represent the three TWINSPAN
communities in the original analysis of Sluis and Tandarich (2000): 4 –
mesic prairie, 5 – wet prairie, and 6 – sedge meadow (Figure
9). This appeared to result in a loss of data through the generalization
of 13 clusters to essentially three.
The 13 clusters were further examined by several visits in the field. Although
the vegetation was mostly dormant by that time, many of the features depicted
by MultiSpec were observable. Some clusters were eliminated for comparison to
vegetation because they consisted exclusively of unsampled areas such as roads,
mowed areas and patches of shrubs and trees. We also combined two clusters because
it appeared that they reflected the same community; it had been spectrally divided
due to the presence of standing water in part of the community. These adjustments
produced eight possible communities (Figure
10 and Table 4).
Ground reference photographs of three of the eight possible communities are
shown in Figure 11.
We examined the results of different unsupervised pixel clustering on a small
section of the Grant Creek image, about 100 square meters. Figure
12 shows a spreading shrub clone that resembles the number "4",
using different wavelength bands. Some wavelength band combinations bring out
the differences in plant communities in the image much better than others. Channels
90, 60, 40 are the default bands for displaying 120 channel images in MultiSpec.
We then chose the three bands 1, 71, and 120, as well as all color bands combined,
for two different analyses. The analyses involved the desired number of clusters,
a required a priori input parameter. In the top row of Figure
13 are shown the results of initially asking MultiSpec to produce 8 clusters.
MultiSpec combines clusters that were not distinct at a given threshold value,
resulting in 4, 5, or 6 clusters rather than 8. In the bottom row of Figure
13, we initially asked for 3 clusters, which were produced for each example.
Increasing the number of bands used greatly decreased the graininess of the
figures. This probably reduced the accuracy of results. However, we consider
the image produced using all bands to be much easier to examine and interpret,
especially at a large scale. It is apparent that results of clustering are dependent
on initial parameters. As a result, it is critical that ground reference data
validate results (see below) and caution be used in the interpretation of results.
As with other areas of ecology, subjectivity still plays a large role, but can
be reduced somewhat through statistical analyses.
Comparison of remote sensing clusters to vegetation was conducted in several
ways to determine if a particular technique was preferable to other techniques.
Agglomerative clustering was conducted on the complete data set, Set A, of 133
1m2 quadrat samples from Grant Creek originally collected in 1999 (Table
5). This included statistical estimates of the relative frequency of all
canopy and understory species. Sampling in early spring and late summer of 1999
were combined. These same quadrats were sampled again in 2000 within one week
of the acquisition of the remote sensing images. This data set, Set B, was limited
to species with canopy cover of greater than 24% to replicate the view of the
ITD Spectral Visions sensor. The visual cover estimates were then assigned to
one of five categories: 0-24%, 25-49%, 50-74%, 75-89%, and 90-100%. Another
data set, Set C, used only the latter three categories. A final data set, Set
D, used only species considered dominant in the cluster after further field
investigation. This is the most subjective and, therefore, least desirable data
set. Set D was used mainly to determine if subjective opinion was better than
numerical analysis at determining cluster composition. Agglomerative clustering
was done on Sets B (Table
6), C (Table 7) and
D (Table 8), and TWINSPAN
clustering was done on Sets B (Table
9) and D (Table 10).
Each analysis of the data was designed to produce 8 clusters to allow 1 to 1
comparison with the remote sensing clusters. Statistical analysis of agreement
among results was achieved using Cohen’s Kappa coefficient. Cohen’s
Kappa coefficient approaches 1 as agreement approaches 100%, but is also dependent
on the number of clusters used. The lower limit is unbounded. TWINSPAN communities
using Sets B and D showed the closest agreement to MultiSpec clusters (Tables
9 and 10).
There are several confounding factors when trying to interpret remote sensing
results. Among these are successional stage, hydrology, and rate of photosynthesis.
The rate of photosynthesis can be reflected in two ways. One is the difference
between C3 (cool season) and C4 (warm season) photosynthetic pathways. Our data
suggests the ability of this remote sensing to distinguish these groups by distinguishing
groups for C3 dominated species such as Carex, Festuca, and Poa spp. from warm
season prairie grasses. Rate of photosynthesis can also be affected by environmental
conditions. Excessive moisture or drought are among factors increasing or decreasing
photosynthetic rates. Any of these may affect the interpretation of results.The
remote sensing imagery also appears to distinguish predominantly wet areas from
predominantly dry areas, with several moisture levels between these two extremes.
This is indicated by the positioning of clusters adjacent to only certain other
clusters either slightly wetter or drier. The Interspersion and juxtaposition
index (IJI) in the Fragstats landscape metrics program (McGarigal & Marks
1994) measures such occurrences. The IJI approaches 0 when the corresponding
patch type is adjacent to only 1 other patch type and the number of patch types
increases. IJI = 100 when the corresponding patch type is equally adjacent to
all other patch types (i.e., maximally interspersed and juxtaposed to other
patch types). A few examples illustrate this.
The boundary that is actually off the image would be expected to be adjacent
of any cluster that is against the edge of the image. Such an artificial cluster
should have a high IJI value. Alternatively, a cluster such as standing water
within a community should be almost completely surrounded by the same community
type, except where the standing water and community edges coincide. Using the
Grant Creek clusters, the off-image boundary has the highest IJI and the standing
water area has the lowest (Table
11). The fractal dimension also reflects this relationship.
We did not have the opportunity to examine the other images in the detail that
we did for Grant Creek. After examining the other six images, we concluded that
it was a priority to investigate procedures for image analysis and apply them
to a selected image rather than doing a little analysis of each image. Because
the techniques are somewhat new, the capabilities and accuracies need to be
determined before conclusions can be made. During the training in image analysis
at LARS, some initial work was done on a second image that covered the Drummond
area. As the analysis unfolded it became clear that much remains to be done.
However, from the preceding analysis it is clear that this method has the potential
we expected it to have.