pilot_study by xiaoyounan

VIEWS: 2 PAGES: 31

									                                             BigFoot Pilot Study:
                              Test of the Vegetation Cover Component
                                   Characterization System (3CS)
                               Warren B. Cohen, Polly Thornton and Tom K. Maiersperger


Introduction

BigFoot objectives require that we accurately characterize fractions of vegetation cover components at
all four sites. Doing so facilitates several subsequent characterizations and analyses. Foremost among
these is the provision of cover maps based on ETM+ data that provide building blocks for a variety
of prescribed land cover classication systems. The BigFoot exible land cover classication system
enables our cover data layers to be compared with many alternate cover classications developed by
others for the same study sites. To accomplish the goal of providing fractional characterizations of
cover components at each site we developed the Vegetation Cover Component Characterization System
(3CS). In the summer and early autumn of 1999 we conducted a pilot study to test the system. Field
sampling is expensive, especially over four 25 km2 sites. Consequently, for BigFoot, we had to devise
a sampling strategy that would take full advantage of data from a limited number of plots. This meant
utilizing the spatial domain, in combination with modeling and more classical statistical analyses
involving the plot-level data. The sampling strategy we use (see Figure 1) involves measurements at
100 plots, 80 of which are arranged as a nested spatial series (Clinger and Van Ness 1976). Within
each plot, several types of measurements, described in the BigFoot Field Manual are made at several
points from which plot-level means are calculated. For most biophysical measurements (e.g., LAI,
biomass, NPP) protocols were developed in other studies (Gower et al. 1999), but for characterizing
cover fractions using 3CS there was no precedent. Critical to the accurate characterization of cover
components within a plot is the number of photo samples required.

                                                 An additional consideration for the sampling scheme was
                                                 the apparent cost of misregistration. To develop the BigFoot
                                                 ne-grained data surfaces of land cover, LAI, and NPP, it
                                                 is imperative that the plot measurements be accurately
                                                 aligned with the ETM+ imagery, and that any error associ-
                                                 ated with the misregistration of these measurement sets
                                                 be assessed. Over the relatively at terrain of the BigFoot
                                                 sites, misregistration of ETM+ imagery, on average, will
                                                 likely be on the order of one pixel or less. Each BigFoot plot
                                                 is 25 m on a side and we are resampling the ETM+ data to
                                                 that same resolution. To insure that the eld plots are geo-
                                                 referenced as accurately as possible, they were positioned
                                                 using a GPS with real-time correction and accurate to the
                                                 nearest 0.1 m. To assess the apparent cost of misregistra-
                                                 tion we focused on the vegetation cover component frac-
                                                 tions. As these represent the most fundamental properties
                                                 of the system, if misregistration is likely to be an issue, it
                                                 should be expressed most prominently in this dataset.
           Figure 1, BigFoot Sampling Strategy


Design

The rst two sites to be characterized were NOBS (7-12 July) and AGRO (10 September). For these two
sites, the design on the left side of Figure 2 below was used. For each cell in the 3 x 3 cell group (or
test plot) we sampled four points spaced evenly across each cell within the test plot, as shown. Twelve

                                                                                Cohen et al. - 15 June 2000 - Page 1
additional points were sampled in the center cell; these were evenly spaced within the bounds of four
corner points. At NOBS, one test plot was sampled in each of four cover types: 1) muskeg, 2) black
spruce, 3) mixed black spruce, jack pine, and aspen, and 4) wetland. These cover types are roughly
equivalent to the four cover types described in Appendix A (NOBS). At AGRO, one plot was sampled in
corn and one in soybeans (see Appendix B (AGRO)). The other two sites, HARV (27-28 September) and
KONZ (1-2 October), were sampled using the design on the right, in Figure 2 below. The only difference
in the design was a more intense sampling density in the 8 external cells of each test plot.




                                                              Figure 2, Pilot Study Sample Design


At each point in the NOBS pilot study plots, sampling involved acquisition of one “ground” and one
“canopy” photo. The tripod of the 3CS holds two cameras on a horizontal extension crossbar, with one
camera facing up and one facing down (see Appendix E for tripod placement photos). The focal plane
of the cameras was positioned roughly at 1.5 m above ground level. The ground camera was positioned
on the crossbar, 0.85 m from the tripod center and placed between two legs of the tripod to minimize
the occurrence of the legs in the pictures. The camera is a Canon PowerShot Pro70. The angle of view
was set to 53 degrees, which yielded a ground image area of approximately 1.5 m by 1 m. The images
were acquired using standard compression in high resolution mode (1.573 million pixels), and a remote
switch. The canopy camera was a Sony Mavica FD91, and had its angle of view set at 20 degrees.
Standard compression and high resolution mode (0.786 million pixels) were used. This camera was
situated on the opposite end of the crossbar from the ground camera, about 0.25 m from the tripod
center. Assuming an average tree canopy height of 10 m, the canopy projection area imaged was
approximately 3 m by 2.2 m. An accessory set of measurements were made at each of the four points
of the external cells of each plot and at each of the four corner points of the center cell. These
included: tree basal area, understory basal area, and ocular proportional estimates of major ground
cover components. Descriptions of these measurements are given in BigFoot Field Measurements. At
AGRO there was no overstory, thus only ground photos were acquired. One test plot was sampled in
each of the two major cover types, corn and soybeans. The corn was approximately 3 m tall, thus the
horizontal crossbar had to be placed on a vertical extension bar to acquire the images (3CS). At HARV
we sampled two test plots, one in a mixed-hardwood conifer site, and the other at a pure hardwood site
(see Appendix C (HARV)). The 3CS was set up the same as at NOBS. At the mixed site, the predominant
overstory species were deciduous hardwoods, whereas the understory was largely coniferous. At KONZ
we sampled a grassland cover type with no overstory and the gallery forest cover type (see Appendix
D (KONZ)). For all ground photos at these three sites, the camera was situated at a height above the
vegetation that would yield an approximate coverage of 1.5 m by 1 m at the surface height of the
vegetation (the same as at NOBS). At the gallery forest site the trees were approximately 30 m tall,
which yielded an imaged canopy surface area of roughly 10 m by 7.5 m. Accessory measurements were
acquired at AGRO, but not at KONZ or HARV.

                                                                          Cohen et al. - 15 June 2000 - Page 2
Analysis & Results

Each digital photograph was analyzed by laying a mylar sheet over a printed copy about 20 cm by 26
cm in size. The mylar had a grid with 99 line intersections spaced 2.5 cm apart. The vegetation over
component lying directly under each grid intersection was noted, and the proportion of total cover in
each component cover type was calculated. At NOBS (see Appendix A), the cover components that we
could characterize from the photos included conifer, hardwood, and dead tree canopy components, and
moss, lichen, herbaceous, shrub, ne litter, tree regeneration, coarse woody debris (CWD), water , and
unknown ground components. The unknown component is a result of either glare or deep shadow,
because of which, the true component could not be identied. At AGRO (see Appendix B) the
component set included corn, soybeans, and open ground. HARV plots (see Appendix C) included
canopy trees, and tree regeneration, shrub, fern, herbaceous, coarse woody debris, rock, moss, water,
and ne litter ground cover components. There were essentially no dead canopy trees in the plots
and we could not condently separate conifer from hardwood, due to a complex vertical mixing of
these components. In the grassland KONZ test plot (see Appendix D), the separable components were
grass and herbaceous plants; no bare ground was seen. In the gallery forest KONZ test plot, we could
characterize canopy tree cover, and the following ground cover components: grass, herbaceous plants,
ne litter, coarse woody debris, understory tree regeneration, shrub, and unknown.

A basic characterization of each test plot (n=48 for NOBS and AGRO; n=88 for KONZ and HARV)
reveals that at NOBS conifer cover was highly variable among test plots, with Plots 4 and 1 having the
lowest (18% and 23% respectively) and Plots 3 and 2 having the highest (35% and 52% respectively).
Hardwood canopy cover was low in all plots (<5%). Dead canopy cover (snags) was less than or equal
to 3% in Plots 1-3, but was 7% in Plot 4. Understory cover was also highly variable among NOBS test
plots: Plot 1 was dominated by moss, herbaceous, and shrub cover; Plot 2 was nearly 50% moss cover,
with tree regeneration and shrub cover being of secondary and tertiary importance; Plot 3 consisted
primarily of moss tree regeneration; and Plot 4 consisted mostly of herbaceous, shrub, and lichen
cover. In all plots there was a small proportion of the “unknown” class (7% in Plot 1, 2% in Plot 2,
12% in Plots 3 and 4). In contrast, the AGRO site was very uniform. Soybean cover was 100% in all
photos. Corn cover averaged 87%. At HARV, canopy cover was 95% in the mixed forest and 90% in the
hardwood forest. Ground cover in both test plots was predominately (61%-63%) litter, followed by tree
regeneration, ferns, and shrubs. The grassland at KONZ was 93% grass. Canopy closure in the gallery
forest was 90%. Ground cover was dominated by herbs (32%), litter (30%), and grass (28%).

To determine how cover type components were related within each test plot, correlation coefcients
were calculated for each plot (n=48 for NOBS; n=88 for KONZ and HARV). At NOBS, most cover
components were poorly correlated, with more than one-half of the correlation coefcients in each
test plot being below an absolute value of 0.25. The greatest correlation was between understory tree
regeneration and moss cover in Test Plot 2 (-0.64). All other coefcients within all test plots were less
than an absolute value of 0.50. At AGRO, correlation coefcients are zero, given that corn and soybean
did not coexist in the same test plot. At HARV, cover component correlations were likewise quite low,
with the only relationships above an absolute value of 0.5 being associated with tree regeneration and
litter cover in Test Plot 1 and tree regeneration and moss cover in Test Plot 2. At the KONZ gallery
forest plot, only grass and herbaceous cover exhibited a correlation coefcient above an absolute value
of 0.5. Correlations are meaningless in the grassland, because there was only grass and herbs.

Principal components analysis was performed to examine the multivariate structure of the data from
each test plot (n=48 for NOBS; n=88 for KONZ and HARV). At NOBS, the rst PC axis accounted
for 47%, 48%, 52%, and 30% of the variance in plots 1-4 respectively. In Test Plots 1-3, PC1 was
essentially related to the variance in conifer cover. In Test Plot 4, PC1 was a contrast of herbaceous
plants against shrubs, snags, and tree regeneration. PC2 of Test Plots 1-3, was a contrast of tree
regeneration with the ground components (herbaceous and moss in Plot 1, moss in Plot 2 and 3). In
Test Plot 4, PC2 as associated with the variation in conifer cover. The proportion of variance explained
by additional PC axes was <12% for all plots. No clear patterns among components emerged, with
the exception of Test Plot 4, where PC3 was a contrast of shrubs with several other ground cover

                                                                           Cohen et al. - 15 June 2000 - Page 3
components. The clearest result from this analysis was that, with the exception of Plot 4, at NOBS,
conifer canopy proportion is the greatest source of variation in the data set.

At AGRO and at the grassland KONZ plot, the PC analyses were not performed because of the within-
plot lack of cover component variation. For the gallery forest at KONZ, PC1 was essentially a function
of the variation in the amount of ground cover that was grass, whereas the second PC axis was a
function of the amount of tree canopy and litter. Together, PC1-PC3 accounted for 80% of the variation
in the data set, with PC3 being a function of contrast between the amount of herbaceous and litter
components. At HARV , the rst two PC axes from Test Plot 1 accounted for 86% of the variation
in the data set, with PC1 being a contrast of litter and tree regeneration and PC 2 being a contrast
of tree and litter with fern. In Test Plot 2, PC1 was a contrast between litter and shrub, PC2 was a
contrast between shrub and tree regeneration, and PC3 was a contrast between fern and regeneration
and shrub cover. The rst three PC axes accounted for 93% of the variation in Test Plot 2. The lack of
variation in tree canopy cover at this is evident in its lack of importance in the PC analyses.

For the next set of analyses, the data from all photographs within a cell of a given test plot (n=16 for
center cell; n=4 or n=9 for surrounding cells) were averaged to provide a set of mean cover component
proportions for that cell. From these data, Analysis of Variance (ANOVA) tests were run by cover
component to examine the variation among the 9 cells as compared to the variation within each
of the 9 cells of each test plot (see Figure 3 below). Greater variation within cells would suggest
the larger test plot area is homogenous and 1-2 pixel misregistration should not be an important
concern. Greater variation among cells would suggest the test plot neighborhood is not homogenous
and misregistration could be a problem. At the NOBS site, in Test Plot 1 (muskeg), only litter was
different among cells (p<0.05). In Test Plot 2 (black spruce), shrubs and coarse woody debris were




                                                                                      Figure 3, ANOVA Results


                                                                           Cohen et al. - 15 June 2000 - Page 4
different among cells. No cover components were different among cells in either Test Plot 3 or 4. It
is important to note that, for test plots 1 and 2, conifer canopy cover is the most important variable
in the PC analysis and it is not different among cells. Neither corn nor soybean cover was different
among cells at the AGRO site. In the grassland test plot at KONZ no cover components were different
among cells. In the gallery forest, all ground cover components were different among cells, but canopy
cover was not. Since this is a closed canopy environment, variation in ground cover will not be
detected by the ETM+ sensor and we need not be concerned with that variation for mapping purposes.
Similar results are seen at the HARV site; i.e., ground cover components were different among cells
and canopy cover was not. This is also a closed canopy site, and ground components will not be
detected. From these ANOVA results misregistration does not appear to be an important problem at
any of the four sites.

We experimented with two different sample sizes, since we had no previous experience using the 3CS
system and did not know what sample size was needed and logistics, particularly at NOBS, were a
major concern. It is not feasible to choose an adequate sample size for this project based only on
statistics, as cost and logistics are too constraining. However, we want to have an estimate of the error
as a function of sample size. Using the same data as for the ANOVA tests, i.e., cover proportions from
each photo, we calculated the standard error associated with varying sample sizes, from

Standard Error = population standard deviation / sqrt(n)

where n=sample size

This required an estimate of the true population standard deviation, for which we used the pooled
variance, or Mean Square Within, from the ANOVA results. At the NOBS site, conifer cover had the
largest standard error of all cover components, approximately 15% with a sample size of 4 at each
of the four test plots. At the AGRO site, soybean cover was 100% in all photos so the standard error
at any sample size is 0. For corn the standard error at a sample size of 4 is less than 1.5%. In the
grassland at KONZ standard error for both grass and herb is less than 3% with n=4. In the gallery
forest standard errors for all cover components are less than 10% with n=4. At HARV canopy standard
error is less than 15% at both test plots with n=4. Understory standard errors are higher in the mixed
forest than in the hardwood forest, but both are less than 5% with a sample size of 4 points. The
shapes of the curves indicate that an n of 4 is too small. Logistics and cost allow for a maximum n of 9.
Fortunately, an n of 9 is near the asymptote in the graphs of the standard error as a function of n and
provides a reasonably acceptable standard error for all components.

Conclusions

We wanted an in-depth characterization of the canopy and ground cover components in each cover
type at all sites, which we now have. In addition, we have examined the potential problem of satellite
image pixels not being accurately aligned with the ground. This was done by comparing the variance
within 25 m cells to the variance among 25 m cells for each cover component. At NOBS, some cover
components did vary more among cells than within cells. However, they were not the variables that
contributed the most to the overall variation in the dataset. Canopy conifer was the most important
variable, and it was not different among cells, suggesting that misregistration will not be a problem at
NOBS. The other three sites had less variability then NOBS, and again, the most important variables
did not vary among cells. Small amounts of misregistration should not be a problem at any of the
sites. Our primary goal was to determine if the Vegetation Cover Component Characterization System
would provide data we needed to develop continuous elds of cover components at each site, and if
so, what the appropriate sample size should be. AGRO is a simple system, consisting largely of corn,
soybeans, and fallow elds. Each eld is quite compositionally homogenous. As such, there is no need
to characterize cover components in any greater detail than to note the crop type present. At KONZ,
the situation is similar, in there is either grassland or gallery forest. There is spatial variability in
grass species present and in the mix of herbaceous species present (in relatively small quantities).
However, biomass measurements made at each plot will be sufcient to characterize spatial variations
in the proportions of these cover components. The gallery forest at KONZ is a relatively uniform, closed

                                                                           Cohen et al. - 15 June 2000 - Page 5
canopy of mixed hardwood species. Plot-level biomass measurements alone should be sufcient to
characterize relative proportions of canopy species present. Understory components will not be visible
to the ETM+ sensor, thus we need not concern ourselves with that level of detail for mapping with
ETM+. The forest at HARV is also closed canopy and there is little in the way of understory component
variability. The overstory canopy, however, contains considerable variability in hardwood and conifer
species that is important to characterize. Hardwood species appear always to be in the upper canopy
layer (dominant), but conifer species occupy the full range of relative canopy positions from dominant
and co-dominant to intermediate and suppressed. We found it difcult to characterize the relative
canopy positions of trees using 3CS, and this measure can be easily and accurately assessed for
all “in” trees when making the plot-level biomass measurements. Also the satellite will not detect
vegetation in the understory of these closed canopy forests. Using the 3CS in these environments will
not further our goal of developing land cover maps. Therefore, we will not use this system at HARV,
KONZ, or AGRO. NOBS is a very compositionally complex site, consisting of a relatively open, even-
aged forest overstory with a spectrally diverse and highly variable understory and set of ground cover
components. Without use of 3CS, or some comparable system, we could not adequately characterize
vegetation cover at this site. From the graphs of standard error versus sample size, and having a eld
season of experience working at the site, we have chosen to sample nine points in each 25 m cell.




                                                                         Cohen et al. - 15 June 2000 - Page 6
                              APPENDIX A - (NOBS)
                Major Cover Types at Northern Old Black Spruce

Major Cover Types Encountered                            Cover Type qualiers

 1.Muskeg (open-canopy black spruce)                       1.Burned
 2.Black spruce (closed-canopy black spruce)               2.Unburned
 3.Aspen
 4.Wetlands
 5.Jack pine

Cover Type Descriptions

Muskeg
Acronym: MSKG
Overstory: dominated by black spruce often mixed with eastern larch
Understory: sparse to heavy cover Labrador tea, Vaccinium spp., and willow spp.
Ground Cover: predominately sphagnum with feathermoss and reindeer lichen
Vegetation Structure: ground cover hummocky; canopy sparse; trees often stunted (1-6 m tall)
Land Form: at, low-lying occasionally ooded
Comments: this cover type is very abundant in NOBS; there exists a gradual transition between
muskeg and closed canopy black spruce- feathermoss forests; demarcation unavoidably arbitrary.

Black spruce
Acronym: BLSP
Overstory: dominated by black spruce occasionally mixed with eastern larch; low level occurrence of
balsam poplar, and jack pine
Understory: sparse coverage of Labrador tea, Vaccinium spp.
Ground Cover: predominately feathermoss
Vegetation Structure: ground cover at (not hummocky); canopy closed; trees not stunted (6-9 m
tall).
Land Form: at, low-lying, but never ooded
Comments: this cover type is very abundant in NOBS; transition between muskeg and closed canopy
black spruce-feathermoss forests is gradual; demarcation is unavoidably arbitrary

Aspen
Acronym: ASPN
Overstory: dominated by trembling aspen; low level occurrence of white spruce, balsam poplar, black
spruce, and jack pine
Understory: green alder and hazel spp.
Ground Cover: very little moss or forbs present
Vegetation Structure: canopy closed, trees often tall (12-15 m tall), hazel and alder often forming
second closed canopy at 1-2 m
Land Form: uplands
Comments: several patches occur at NOBS but they are small and infrequent

Wetland
Acronym: WTLD
Overstory: scattered bog birch and eastern larch
Understory: open water lined with willow, Labrador tea, and marsh grasses
Ground Cover: mosses
Land Form: ooded lowlands, creek margins, and beaver ponds
Comments: this is a difcult community to describe because it includes both ooded peatlands

                                                                        Cohen et al. - 15 June 2000 - Page 7
(oligotriphic fens dominated by aquatic sphagnum spp, Vaccinium, and Labrador tea) as well as the
marshy borders of creeks and beaver ponds (marshes containing willows and sedges); despite the range
of plant communities in this cover type they are grouped together because of their similar structure

Jack pine
Acronym: JKPN
Overstory: dominated by Jack pine; low level occurrence of white spruce, balsam poplar, black spruce,
and trembling aspen
Understory: sparse coverage of Labrador tea, vaccinium spp. and occasional patches of green alder
Ground Cover: sparse to complete coverage by reindeer lichen, sparse coverage by feathermoss
Vegetation Structure: canopy closed, trees often tall (10-12 m tall)
Land Form: uplands, sandy soils
Comments: very rare at NOBS except for regeneration stands in 1981 burn at southern edge of site




Cover Type Qualiers and Additional Comments

A large re burned a 150 km2 area on the southern boundary of the NOBS BigFoot study area in
1981. A few of the extensive plots on the south end of the 5x5 km grid occur in this burn. These plots
are classied according to their current plant community (i.e., MSKG, BLSP, WTLD, ASPN, or JKPN)
but their status as burned will also be recognized as a cover type qualier since the burn inuences
the species composition, LAI and NPP

A cover type map including the NOBS BigFoot study area was constructed from aerial photography by
the Manitoba Department of Natural Resources (MDNR) in 1988 and is available as raster map from
the BORIS data base. This is a high quality map that recognizes over 100 cover types. Based on our
own on-ground experience, the map is accurate.




                                                                         Cohen et al. - 15 June 2000 - Page 8
  NOBS Canopy Image Samples




NOBS Ground Cover Image Samples




                         Cohen et al. - 15 June 2000 - Page 9
       NOBS Test Plot 1: Muskeg




                                                           Figure 4




Correlation for NOBS Test Plot 1: Muskeg




                                                           Figure 5



                               Cohen et al. - 15 June 2000 - Page 10
       NOBS Test Plot 2: Black Spruce




                                                              Figure 6




Correlation for NOBS Test Plot 2: Black Spruce




                                                              Figure 7



                                  Cohen et al. - 15 June 2000 - Page 11
       NOBS Test Plot 3: Mixed Forest




                                                              Figure 8




Correlation for NOBS Test Plot 3: Mixed Forest




                                                              Figure 9



                                  Cohen et al. - 15 June 2000 - Page 12
       NOBS Test Plot 4: Wetland




                                                         Figure 10




Correlation for NOBS Test Plot 4: Wetland




                                                         Figure 11



                               Cohen et al. - 15 June 2000 - Page 13
Principal Component Analysis for NOBS

         Test Plot 1: Muskeg




                                                         Figure 12




      Test Plot 2: Black Spruce




                                                         Figure 13




                               Cohen et al. - 15 June 2000 - Page 14
Principal Component Analysis for NOBS

      Test Plot 3: Mixed Forest




                                                         Figure 14




        Test Plot 4: Wetland




                                                         Figure 15




                               Cohen et al. - 15 June 2000 - Page 15
Standard Error Estimates for NOBS


  Muskeg                   Black Spruce




               Figure 16                               Figure 17




Mixed Forest                 Wetland




               Figure 18                               Figure 19




                             Cohen et al. - 15 June 2000 - Page 16
                               APPENDIX B - (AGRO)
                     Major Cover Types at Midwestern Cropland

Major Cover Types Encountered

 1.Corn
 2.Soybean
 3.Fallow

Cover Type Descriptions

Corn
Acronym: CORN
Species: corn
Architecture: closed canopy row crop growing >2 m tall by late summer
Comments: roughly half of the row crops planted in the site will be corn

Soybean
Acronym: SOYB
Species: soybean
Architecture: closed canopy row crop growing 50-75 cm tall by late summer
Comments: roughly half of the row crops planted in the site will be soybean

Fallow
Acronym: FALO
Species: hay grasses
Architecture: grassland of variable height
Comments: only a small proportion of the site (<5%) is fallow




Cover Type Qualiers and Additional Comments

The BigFoot extensive research plots are stratied among many farms, each of which may have
its own planting and harvest date. The timing of planting and harvest for each study plot will be
recognized as a cover type qualier since crop phenology (especially early in the season) inuences
vegetation cover and LAI.




                                                                           Cohen et al. - 15 June 2000 - Page 17
AGRO Ground Cover Image Samples




                         Cohen et al. - 15 June 2000 - Page 18
AGRO Corn and Soybean Test Plots




                                                          Figure 20




 AGRO Standard Error Estimates

  Corn Test Plot                  Soybean Test Plot

                                There was no variance in
                                soybean cover, so the SE
                                is 0




                    Figure 21




                                Cohen et al. - 15 June 2000 - Page 19
                                APPENDIX C - (HARV)
                          Major Cover Types at Harvard Forest

Major Cover Encountered                             Cover type qualiers

  1.Eastern hardwoods                                 1.disturbed (clearcut)
  2.Eastern hemlock                                   2.undisturbed
  3.Red pine
  4.Oldeld meadow

Cover Type Descriptions

Eastern hardwood
Acronym: EHWD
Overstory: dominated by sugar maple mixed with red oak, ash, basswood, and beech
Understory: saplings of shade tolerant tree species, and Vaccinium
Ground Cover: grasses and forbs belonging to the “Canadian Carpet” community
Land Form: uplands
Comments: additional visits to HARV will allow us to better describe this community

Eastern hemlock
Acronym: HEML
Overstory: eastern hemlock with remnant red oak
Understory: hemlock saplings
Ground Cover: sparse cover of grasses and forbs belonging to the “Canadian Carpet” community
Land form: uplands to lowlands
Comments: additional visits to HARV will allow us to better describe this community

Red pine
Acronym: RDPN
Overstory: red pine
Understory: red pine saplings
Ground Cover: sparse cover of grasses and forbs and shrubs
Land Form: uplands
Comments: additional visits to HARV will allow us to better describe this community

Oldeld meadow
Acronym: OLDF
Overstory: none
Understory: grasses, shrubs
Comments: this cover type is largely the result of anthropogenic disturbance; additional visits to HARV
will allow us to better describe this community




Cover Type Qualiers and Additional Comments

A clearcut in 1999 removed the forest from a portion of the private land occurring within the Harvard
Forest study area, affecting one or more of the extensive plots. These plots will be classied according
to their current plant community but their status as clearcut will also be recognized as a cover type
qualier since cutting inuences the vegetation structure and function.



                                                                          Cohen et al. - 15 June 2000 - Page 20
  HARV Canopy Image Samples




HARV Ground Cover Image Samples




                         Cohen et al. - 15 June 2000 - Page 21
       HARV Test Plot 1: Mixed Forest




                                                            Figure 22




Correlation for HARV Test Plot 1: Mixed Forest




                                                            Figure 23




                                  Cohen et al. - 15 June 2000 - Page 22
       HARV Test Plot 2: Hardwood Forest




                                                             Figure 24




Correlation for HARV Test Plot 2: Hardwood Forest




                                                             Figure 25




                                   Cohen et al. - 15 June 2000 - Page 23
Principal Component Analysis for HARV

      Test Plot 1: Mixed Forest




                                                        Figure 26




    Test Plot 2: Hardwood Forest




                                                        Figure 27




                              Cohen et al. - 15 June 2000 - Page 24
Standard Error Estimates for HARV

       Test Plot 1: Mixed Forest




                                         Figure 28




     Test Plot 2: Hardwood Forest




                                              Figure 29




                                    Cohen et al. - 15 June 2000 - Page 25
                               APPENDIX D - (KONZ)
                          Major Cover Types at Konza Prairie

Major Cover Types Encountered                      Cover type qualiers

 1.Tallgrass prairie                                 1.Cattle grazed
 2.Shortgrass prairie                                2.Bison grazed
 3.Shrub community                                   3.Ungrazed
 4.Gallery forest                                    4.Burn frequency

Cover Type Descriptions

Tallgrass prairie
Acronym: TGPR
Species: Big bluestem, Indian grass, little bluestem, switchgrass, and other forbs
Architecture: 1-1.5 m tall at full ush
Land Form: bottomlands, deep soils, unexposed aspects
Comments: a wide, poorly dened, gradient exists between the tallgrass and shortgrass prairies

Shortgrass prairie
Acronym: SGPR
Species: blue-grama, hairy grama, xeric forbs
Architecture: 10-20 cm tall at full ush
Land Form: exposed ridge tops, shallow claypan soils
Comments: a wide, poorly dened, gradient exists between the tallgrass and shortgrass prairies

Shrub community
Acronym: SHRB
Species: smooth sumac and Cornus spp
Architecture: 1-2 m tall, very dense, thin stems, closed canopy
Land Form: exposed ridge tops, shallow claypan soils
Comments: forms patches in drainage gulches and seeps; shrub communities also occur adjacent to
creeks and as a transition between prairie and forest

Gallery forest
Acronym: GALF
Species: oaks, elm, hackberry, walnut, and hickory
Architecture: 15-20m tall closed canopy but lots of edge supports signicant understory with open
canopy at 3-5m
Land Form: lowlands, largely riparian
Comments: this is a diverse community that transitions into prairie either by way of open savanna or
by shrub communities; about 6% of Konza is gallery forest

Cover   Type Qualiers and Additional Comments

Konza is divided into over 60 managed experimental watersheds. The management practices vary with
respect to grazing regime and re frequency. Grazing treatments include cattle grazing, bison grazing,
and no grazing. Fire regimes vary by frequency (1, 2, 4, 10, or 20 year re cycles) and timing (winter,
summer, fall, and spring burning). While not all combinations of burning and grazing regimes are
practiced, many are making the Konza landscape very diverse. The BigFoot design can not sample
each of these management areas. The management history of each study plot will be recognized
as a cover type qualier since the management practice inuences species composition, vegetation
structure and function.

                                                                        Cohen et al. - 15 June 2000 - Page 26
  KONZA Canopy Image Samples




KONZA Ground Cover Image Samples




                         Cohen et al. - 15 June 2000 - Page 27
KONZA Test Plot 1: Grasslands




                                                   Figure 30




                         Cohen et al. - 15 June 2000 - Page 28
       KONZA Test Plot 2: Gallery Forest




                                                              Figure 31




Correlation for KONZA Test Plot 2: Gallery Forest




                                                              Figure 32




                                    Cohen et al. - 15 June 2000 - Page 29
Principal Component Analysis for KONZA

         Test Plot 2: Gallery Forest




                                                                 Figure 33




 Standard Error Estimates for KONZA
         Test Plot 2: Gallery Forest




                                               Figure 34




                                       Cohen et al. - 15 June 2000 - Page 30
APPENDIX E
3CS in Action




                Cohen et al. - 15 June 2000 - Page 31

								
To top