Digital Photograph Analysis for Measuring Percent Plant Cover in the Arctic

Abstract

Long-term satellite remote sensing data, when properly calibrated and validated against ground monitoring,
could provide valuable data sets for assessing climate change impacts on ecosystems, wildlife, and other important aspects of life in the Arctic. Percent plant cover is ideal for seasonal and long-term ground monitoring because it can be observed non-destructively and is closely related to other key ecosystem variables, such as biomass and leaf area index (LAI). Accurately measuring percent plant cover in the Arctic, however, has been a challenge. Advances in digital photography and image processing techniques have provided the potential to measure vegetation cover accurately. In this paper we report an adapted
method for quantifying percent plant cover based on plot digital photograph classification (PDPC). In this digital image analysis, the red, green, and blue image channels and the intensity, hue, and saturation image channels were used together to ensure more accurate cover measurement and labeling of plant species. We evaluated the accuracy of the PDPC method and two other techniques, visual estimate and digital grid overlay, by testing against artificial plots with known percent cover, by comparing with destructively measured LAI, and by comparing results of the three methods. Our evaluation indicates that the PDPC method is the most accurate. In addition, PDPC has the advantages of being objective, quick in the field, and suitable formeasuring percent plant cover in the Arctic at the level of functional types or species groups.

Full text

ARCTIC
VOL. 63, NO. 3 (SEP TEMBER 2010) P. 315– 326
Digital Photograph Analysis for Measuring Percent Plant Cover in the Arctic
ZHAOHUA CHEN,1,2 WENJUN CHEN,1 SYLVAIN G. LEBLANC1 and GREGORY H.R. HENRY3
(Received 11 August 2009; accepted in revised form 12 January 2010)
ABSTRACT. Long-term satellite remote sensing data, when properly calibrated and validated against ground monitoring,
could provide valuable data sets for assessing climate change impacts on ecosystems, wildlife, and other important aspects
of life in the Arctic. Percent plant cover is ideal for seasonal and long-term ground monitoring because it can be observed
non-destructively and is closely related to other key ecosystem variables, such as biomass and leaf area index (LAI). Accurately
measuring percent plant cover in the Arctic, however, has been a challenge. Advances in digital photography and image-
processing techniques have provided the potential to measure vegetation cover accurately. In this paper we report an adapted
method for quantifying percent plant cover based on plot digital photog raph classication (PDPC). In this digital image
analysis, the red, green, and blue image channels and the intensity, hue, and saturation image channels were used together to
ensure more accurate cover measurement and labeling of plant species. We evaluated the accuracy of the PDPC method and
two other techniques, visual estimate and digital grid overlay, by testing against articial plots with known percent cover, by
comparing with destructively measured LAI, and by comparing results of the three methods. Our evaluation indicates that the
PDPC method is the most accurate. In addition, PDPC has the advantages of being objective, quick in the eld, and suitable for
measuring percent plant cover in the Arctic at the level of functional types or species groups.
Key words: percent plant cover, Arctic, visual estimate, digital photograph, image classication, LAI
RÉSUMÉ. Lorsqu’elles sont bien calibrées et qu’elles sont validées contre le dépistage terrestre, les données résultant de
la télédétection satellitaire à long terme pourraient fournir d’importants ensembles de données en vue de l’évaluation des
incidences du changement climatique sur les écosystèmes, la faune et d’autres aspects-clés de la vie dans l’Arctique. Le
pourcentage de couverture végétale est idéal pour le dépistage saisonnier et le dépistage terrestre à long terme parce qu’il
peut être observé sans qu’il n’y ait de destruction et parce qu’il est étroitement lié à d’autres variables-clés se rapportant
aux écosystèmes, comme la biomasse et l’indice de surface foliaire (ISF). Toutefois, dans l’Arctique, la mesure exacte du
pourcentage de couverture végétale représente un dé. Les progrès réalisés dans les domaines de la photographie numérique
et des techniques de traitement d’images fournissent la possibilité de mesurer la couverture végétale avec précision. Dans
cette communication, nous faisons état d’une méthode adaptée permettant de quantier le pourcentage de couverture végétale
en fonction de la classication de photographies numériques de parcelles. Dans le cadre de l’analyse d’images numériques,
les canaux rouges, verts et bleus des images ainsi que les canaux d’intensité, de tonalité et de saturation des images ont été
utilisés pour donner lieu à la mesure plus exacte de la couverture végétale et à l’étiquetage des espèces végétales. Nous avons
évalué l’exactitude de la méthode de classication de photographies numériques de parcelles de même que celle de deux autres
techniques, soit l’estimation visuelle et la superposition de grilles numériques en faisant des essais à la lumière de parcelles
articielles dont le pourcentage de couverture végétale était connu et en les comparant avec des ISF mesurés de manière
destructive, puis en comparant les résultats des trois méthodes. Selon notre évaluation, la méthode consistant en la classi-
cation de photographies numériques de parcelles PDPC est la plus précise. La classication de photographies numériques
de parcelles a également l’avantage d’être objective, d’être rapide sur le terrain et de se prêter à la mesure du pourcentage de
couverture végétale dans l’Arctique en ce qui a trait aux types fonctionnels ou aux groupements d’espèces.
Mots clés : pourcentage de couverture végétale, Arctique, estimation visuelle, photographie numérique, classication d’images,
indice de surface foliaire
Traduit pour la revue Arctic par Nicole Giguère.
1
Canada Centre for Remote Sensing, Natural Resources Canada, 588 Booth Street, Ottawa, Ontario K1A 0Y7, Canada
2
Corresponding author: zhaohua.c@gmail.com
3
Department of Geography, University of British Columbia, Vancouver, British Columbia V6T 1Z2, Canada

316 • Z. CHEN et al.
INTRODUCTION
Climate change is occurring in the Arctic at double the
average global rate and is predicted to continue at an even
faster rate (Kattsov et al., 2005). A changing climate may
signicantly affect the quality of wildlife habitats in the
Arctic, the ecological integrity of northern national parks,
freshwater resources, and infrastructures of northern com-
munities, as well as global carbon and water cycles (PCA,
2000; McGuire et al., 2002; Klein et al., 2005; Anisimov
et al., 2007; Chen et al., 2009a). Satellite earth observation
data (such as Landsat and Advanced Very High Resolution
Radiometer, or AVHRR) could provide historical and on-
going data sets essential for assessing these climate-change
inuences if they are properly calibrated and validated
against ground observations, especially seasonal and long-
term ground monitoring data (Stow et al., 2004; Bunn and
Goetz, 2006; Olthof et al., 2008). Seasonal and long-term
monitoring also provides a more precise way to document
and assess climate-change effects. For example, Stow et
al. (2004) found through such monitoring that shrub dis-
tribution and density were responding to climate change at
Arctic sites.
One of the most commonly observed and potentially
most useful ecosystem variables for seasonal and long-
term ground monitoring is percent plant cover (Bliss et
al., 1984; Kennedy et al., 2001; Krebs et al., 2003; Bonham
and Clark, 2005; Chen et al., 2009b). As shown by Chen et
al. (2009b), other key ecosystem variables such as above-
ground biomass, foliage biomass, and leaf area index (LAI)
can be reliably measured in the Arctic using percent plant
cover and mean height, not only at the plot-total level, but
also at the levels of plant species or function groups. As a
result, seasonal and long-term monitoring of key ecosys-
tem variables can be achieved by repeated non-destructive
observations of percent cover and mean height at permanent
plots and thus provide essential data sets for calibrating and
validating remote sensing products. While measurement of
mean plant height is straightforward, accurate measurement
of percent plant cover has been a challenge.
Traditional techniques to measure the percent cover
are ground-based. Two methods are commonly used: vis-
ual estimate (VE) and point frame (PF) (Bliss et al., 1984;
Kennedy et al., 2001; Krebs et al., 2003; Bean and Henry
2003; Bonham and Clark, 2005; Chen et al., 2009b). The VE
method is quick and widely used. However, its reliability is
questionable because it is subjective in nature, the overval-
uation error increases with the view angle, and it is prone to
inter-observer bias (Meese and Tomich, 1992; Dethier et al.,
1993; Vanha-Majamaa et al., 2000). The point-frame method
is considered to be objective and is the recommended stand-
ard protocol for measuring the plant cover in the Interna-
tional Tundra Experiment (ITEX) (Walker, 1996; Bean and
Henry, 2003). However, it is time consuming, usually need-
ing 2 – 3 hours to complete measurements for one plot. This
can be a signicant constraint because doing eldwork in
the Arctic is very expensive (high helicopter transportation
cost) and the growing season is very short. More problem-
atically, this method is suspected to overvalue the percent
cover and miss rare species (Bråkenhielm and Liu, 1995;
Vanha-Majamaa et al., 2000). The recommended 10 × 10
grid over a 1 × 1 m plot for the PF method means that if a
point hits a plant, full plant cover over that 10 × 10 cm cell
of the grid is then assumed. In the eld, if a 1 × 1 m grid
divided into 10 × 10 cm cells is not appropriate for the scale,
then point density should be increased.
An increase in the grid resolution could retain the objec-
tivity of the point-frame method and signicantly increase
the accuracy of percent plant cover estimation. However,
the increase in time required to complete one measurement
for a plot would make this method impractical if it had to
be implemented in the eld. This impracticality can be cir-
cumvented by an image-based method, taking a nadir-view
photograph of the plot in the eld and analyzing the pho-
tograph later in the laboratory. Using digital grid overlay
(DGO), Booth et al. (2006b) measured percent cover over a
photograph. However, the contact area of each point (inter-
section) in the grid was larger than an image pixel in their
study. As a result, the large contact area was a factor inu-
encing measurement accuracy when using DGO.
Recent advances in digital photography and image analy-
sis software have provided the potential to measure percent
vegetation cover quickly and objectively and to repeat com-
parable measurements over time (Booth et al., 2006b; Lus-
cier et al., 2006; Laliberte et al., 2007). Several researchers
have investigated the feasibility of measuring percent plant
cover from a photograph taken by a digital camera over
a plot (Richardson et al., 2001; Booth et al., 2004, 2005,
2006b; Luscier et al., 2006; Laliberte et al., 2007). How-
ever, most such studies produced only general categories,
such as vegetation and non-vegetation (Laliberte et al.,
2007). Since wildlife species usually differentiate between
vegetation types in their diets (Krebs et al., 2003), catego-
rizing percent plant cover at a more detailed level (e.g., by
functional types or species groups, Chen et al., 2009b) is
required. With improved image analysis products and meth-
ods, such as VegMeasure, a software product developed by
Johnson et al. (2003); SamplePoint, developed by Booth et
al. (2006a); and eCognition (Deniens, 2003), several stud-
ies have measured percent cover for vegetation categories
on digital photographs, with promising results (Booth et
al., 2005, 2006a, b; Luscier et al., 2006; Laliberte et al.,
2007; Booth and Cox, 2008). However, classication error
resulting from spectral separation methods (VegMeasure,
ERDAS Imagine [ERDAS, 2010], eCognition) can cause
confusion among like-colored species.
Our goals were to develop a new, accurate method of
measuring cover of key species in the Arctic and to com-
pare its results over a small contact area to those of two
other methods: visual estimate and digital grid overlay.
Here we report on 1) how we modied standard image-
based methods to allow accurate measurement of cover by
species, and 2) how we tested the results obtained using
our plot digital photograph classication (PDPC) method,

MEASURING PERCENT PLANT COVER • 317
which incorporates the modication, and compared them to
the VE and DGO results. So far, no investigation of changes
in Arctic vegetation over time has been conducted using
plot photography.
METHODS AND MATERIALS
Study Area and Field Measurements
Our study on vegetation cover of Arctic tundra ecosys-
tems was conducted around Iqaluit (63˚46′ N, 68˚32′ W) and
Clyde River (70˚29′ N, 68˚30′ W) Nunavut, Canada. The
eld measurements were acquired during 16 – 27 July 2007.
To cover the whole spectrum of vegetation conditions for
the purpose of evaluating the three methods, we deliber-
ately selected three different types of sites (with relatively
low vegetation cover, medium growth conditions, and high
vegetation cover) around each of the two communities. A
site was at least 3 × 3 Landsat TM pixels (i.e., 90 × 90 m), to
meet the needs for calibrating and validating remote sens-
ing products. Five 0.5 × 0.5 m plots were used for each site,
with a central plot located randomly and the other four plots
located at 30 m east, south, north, and west of the central
plot. The plot was representative of local vegetation condi-
tions and relatively homogenous.
A nadir digital photograph (taken from directly above) at
the plot level was taken with a 10-megapixel Nikon D80 dig-
ital camera at a height of about 1.5 m above the frame, using
a tripod support. The resulting images had a size of 3872 ×
2592 pixels (the sensor array of the digital camera), cover-
ing the frame and four sides. We selected plots for analysis
on the basis of digital photograph quality. The plots were
not included in this study if their digital photographs were
poorly focused or severely skewed. The 26 plots selected for
percent plant cover analysis thus included 11 plots from the
three Iqaluit sites and 15 plots from the three Clyde River
sites.
In total, 13 vascular plant species were identied in these
plots: Salix arctica, Cassiope tetragona, Oxytropis arctica,
Vaccinium uliginosum, Poa sp., Empetrum nigrum, Dia-
pensia lapponica, Pedicularis hirsuta L., Dryas integrifo-
lia, Oxyria digyna, Salix reticulata, Saxifraga L., and Carex
membranacea. Not all the species were observed in each of
the 26 plots, and some species were more abundant than
others. For example, Salix arctica was found in 17 plots,
but Diapensia lapponica, Oxyria digyna, and Empetrum
nigrum each occurred in only one plot. The total leaf area
index (LAI) of a vascular plant species within a plot was
determined destructively (Chen et al., 2009b).
The Digital Grid Overlay Method
Since the point-frame method was not applied in the
eld, in this study we conducted measurements using an
improved DGO on the digital ground photos. Thus, we
could easily adjust the number of points used within a
digital frame with a GIS. The grid intersection was formed
over the original digital photograph (3872 × 2592 pixels
over 0.25 m2) on the computer screen, so the contact area of
the intersection was the area of a single pixel (0.025 mm2).
No matter how many intersections (10 × 10 grids or 1000
× 1000 grids) we used, the contact area did not change
because it was determined only by the resolution of the
original image, and the on-screen display allowed us to
zoom in to one pixel. By overlaying the frame with a plot
digital photograph, the analyst can identify the species of
vegetation at each intersection. From the counts of the veg-
etation intercepted, the percent cover of each species in the
plot can be calculated. For the present analysis, we counted
only leaf interception to make our results easy to compare
with those from other two methods.
The Visual Estimate Method
The VE method as used in this study was an estimate
of the percent plant cover of each vascular plant species by
nadir-view projection to the ground level within a xed plot
area of 0.5 × 0.5 m. The original estimates of percent plant
cover included all aboveground plant components (i.e.,
leaves, stems, and occasionally owers) and were made
layer by layer from the top to the bottom by averaging esti-
mates from the two to three observers on the eld team. To
make the results comparable with the LAI, as well as with
the measurements from the PDPC method and DGO method
that include only percent cover of leaves, we subtracted the
percent cover of stems to obtain the percent cover of leaves.
The percent cover of stems, visually estimated from the dig-
ital photos, was typically much smaller than that of leaves,
and occurred only for some woody shrub species (e.g., Salix
arctica, Vaccinium uliginosum). In the following analyses,
cover refers only to the percent cover of leaves.
The Plot Digital Photograph Classication Method
The PDPC method includes band transformation, image
segmentation, object classication, and species labeling.
1) Band Transformation: Most digital photos are taken
in the visible light wavelength region (red, green, and blue
or RGB). Different green vegetation may have similar spec-
tral reectance in the visible light region, thus species are
difcult to differentiate in an automatic image analysis
using imagery with only RGB bands. RGB images were
transformed to intensity-hue-saturation (IHS) images using
ERDAS software to reduce the high intercorrelation of RGB
bands and to improve the image analysis and vegetation rec-
ognition from photos in this study. Vegetation analysis using
digital images with IHS transformations increased accuracy
over the original RGB bands in several studies (Tang et al.,
2000; Hemming and Rath 2001; Karcher and Richardson,
2003; Laliberte et al., 2007). We used both RGB and IHS in
our image analyses. IHS images were used for image seg-
mentation and preliminary classication. RGB images were
used for visual interpretation in the labeling step.

318 • Z. CHEN et al.
2) Image Segmentation: Digital processing of a plot
photograph can be performed using pixel-based or object-
based image analysis. Pixel-based analysis is a traditional
method that assesses each pixel’s digital number individu-
ally. However, because of spectral similarity between spe-
cies and high color variation within pixels from the same
plant species, pixel-based classication results obtained
from spectral information are usually not satisfactory
(Vanha-Majamaa et al., 2000). Object-based image analy-
sis, a relatively new method that treats homogenous pixels
as objects with regard to spatial or spectral characteristics
(Ryherd and Woodcock, 1996), is an effective tool for clas-
sication of high-resolution satellite imagery (Herold et
al., 2003; Thomas et al., 2003; Laliberte et al., 2004). The
extraction of green-vegetation cover by species requires
clear boundaries between leaves of each species group.
To improve the capability of identifying various species
groups and reduce the spectral confusion between species
groups, both spectral information and geometric informa-
tion should be considered. Therefore, we applied a seg-
mentation technique—similar to that used by Luscier et al.
(2006) and Laliberte et al. (2007), with special emphasis on
the consideration of shape information in image analysis
with eCognition—to treat homogeneous areas as objects in
plot photographs of this study.
An image of IHS bands was segmented into polygons on
the basis of three parameters: scale, color (spectral informa-
tion), and shape. Considering most species in this study had
relatively small leaf sizes, we used the following parameter
values for all the images: 20 for the scale parameter, 0.6/0.4
for the color/shape parameter, and 0.9/0.1 for the smooth-
ness/compactness parameter. Setting a larger value for scale
parameter will result in larger image objects (Laliberte et
al., 2007). Using these parameters, an individual leaf can
mostly be treated as a polygon, although some small leaves
may be grouped into one polygon, and occasionally, a big
leaf may be split into several polygons.
3) Object Classication: In automated classication,
different green vegetation species with similar shape, color,
or both are very difcult to separate using the limited spec-
tral information contained in RGB bands. Frequently, the
result is classication confusion between vegetation species.
Therefore, automated classication for various species was
not adopted in this study. After segmentation, all objects in
the image can be classied into general categories by deter-
mining the threshold values in the IHS bands. In this study,
we rst separated two general categories, green vegeta-
tion and non-green vegetation. Although some non-green
vegetation areas may be included in the category of green
vegetation because of the spectral confusion or geometric
intermixing, these areas can be eliminated later on. From
this point forward, only green vegetation was of concern
and was further classied. Using a scale parameter of 20, a
large number of polygons were created. Since classication
and editing of such a large number of polygons can be time-
consuming even with only green vegetation species, further
grouping on the basis of unique vegetation characteristics
was necessary. Size and geometric characteristics (e.g.,
the length/width ratio) can be useful to distinguish various
green vegetation species. Ratio of length/width was calcu-
lated for each polygon. Since plant species with long leaves
had large length/width ratios and those with round leaves
had low values, each green vegetation object was then clas-
sied on the basis of a threshold value of length/width ratio.
4) Species Labeling: In addition to the spectral infor-
mation, RGB images may also contain other information
about the objects, such as shape, size, texture, pattern, asso-
ciation, and shadow. The human mind is good at recogniz-
ing and associating these complex elements in an image.
Information about color, texture, shape, and context and
expert knowledge of vegetation can be useful to identify
each polygon in an image. Overlaying the polygon layer
of green vegetation with the RGB composite images aided
with visual interpretation, we labeled each polygon by plant
species in a GIS environment. At this stage, we eliminated
the non-green vegetation polygons that had been misclas-
sied as green ones. Although the visual identication of
each polygon was time-consuming, the results were very
accurate, as the segmentation was conducted at a scale that
almost equaled individual leaf size. The labeling proce-
dures can be faster and easier in the plots that contain fewer
species or are more homogeneous, or after the analyst has
gained experience. Once all species were correctly labeled
and separated, layers containing different species were
combined to form a vegetation map of the plot. Finally, the
percent cover of each species and plot total percent cover
were calculated.
Estimation Error Analysis
Given that there is no proven method with which we can
obtain the true values of percent plant cover over a plot in
the eld with certainty, it is difcult to assess which one
of these three methods (VE, DGO, or PDPC) is the most
accurate. In this study, we evaluated the estimation errors
of these methods in the following ways.
Testing Against Articial Plots with Known Percent Plant
Cover
To mimic typical Arctic tundra ecosystems, which rarely
have more than 30% vegetation cover, we assumed eight
categories: 890, 1780, 2670, 5340, 7120, 8910, 10690, and
13360 oval shape leaves of 2 mm wide and 3 mm long, ran-
domly distributed over a 0.5 × 0.5 m monochromatic test
plot without overlap. These gures gave a percent plant
cover of about 1.68%, 3.36%, 5.03%, 10.06%, 13.43%,
16.78%, 20.14%, and 25.18%, respectively. Each given per-
cent plant cover was tested with ve random distributions.
For the DGO method, we investigated grids with different
resolutions: the ITEX-recommended 10 × 10 grids, as well
as 5 × 5 grids, 15 × 15 grids, 20 × 20 grids, 50 × 50 grids,
100 × 100 grids, 200 × 200 grids, 500 × 500 grids, and 1000
× 1000 grids. The PDPC method was also tested against the

MEASURING PERCENT PLANT COVER • 319
articial plots. The estimation error, E, was calculated as
follows:
(1)
Comparison with LAI Measured Destructively
According to its denition, the percent cover, C, is given
by
C = 1- P(θ) (2)
where P(θ) is the gap fraction at the view zenith angle θ =
0 and is related to the total projected LAI, Lt (Nilson, 1971;
Chen et al., 1997), as follows:
P(0) = exp (-G(0)ΩLt) (3)
where G(0) is the projection coefcient characterizing the
foliage angle distribution, and Ω is a parameter determined
by the spatial distribution pattern of leaves. When the foli-
age spatial distribution is random, Ω = 1. When leaves are
clumped, Ω is less than 1. Foliage in plant canopies is gener-
ally clumped, and hence Ω is often referred to as the clump-
ing index. Combining equation (2) and (3), we have
-ln (1 - C) = G(0)ΩLt (4)
Since the value of G(0)Ω is generally below 1, so the
value of –ln (1 - C) should be less than Lt. In other words,
Lt represents the maximum possible value of -ln (1 - C), in
which all leaves are positioned randomly, without clumping,
and parallel to the ground surface. In reality, most leaves
are clumped and are positioned at a certain angle relative to
the ground surface. Overlap of leaves exists in most plots.
Consequently, it can be concluded with certainty that if the
percent cover measured or estimated by a method is found
to be larger than the corresponding Lt, this method overval-
ues cover. Conversely, if an estimate of percent cover by a
method is found to equal zero, while the corresponding Lt is
larger than zero, it can be concluded with certainty that this
method makes a mistake of 100% undervaluation.
Besides making these magnitude comparisons, we can
also assess the LAI–percent cover relationship. For a given
species of vascular plant, the leaf distribution and overlap
should be similar between different plots, especially when
low, sparsely distributed tundra plots in the Arctic are con-
cerned. If the percent cover is measured or estimated accu-
rately, a strong linear relationship between Lt and -ln (1 - C)
is expected. As a result, if the relationship between Lt and
-ln (1 - C) estimated by a method has a low R2 value, the
method is likely to be less accurate.
Inter-comparison between Methods
Inter-comparison of different methods is a common prac-
tice, although it could be inconclusive if none of the methods
compared is accurate (Floyd and Anderson, 1987). Fortu-
nately, the two approaches noted above (articial plot evalu-
ation and comparison with total LAI) can determine which
one of the three methods is most accurate. If we consider the
estimates made using the most accurate method as the “true”
values, we can evaluate the estimates derived from digital
photographs of Arctic tundra plots using other methods.
RESULTS AND DISCUSSION
Errors Evaluated Using Articial Plots with Known
Percent Cover
We evaluated the errors in estimates of percent cover
against articial plots with known percent plant cover rang-
ing from 1.68% to 25.18%. Figure 1 shows that for the dig-
ital grid overlay method, the errors in percent cover range
from -100% to 257% using 5 × 5 grids (DGO5×5) and from
-100% to 78% using the ITEX-recommended 10 × 10 grids
(DGO10×10). The errors decrease to nearly zero for grids
reaching 1000 × 1000. The PDPC method, at the highest
resolution, with 3872 × 2592 grids, was found to be 100%
accurate in estimating percent cover for the articial plots
with known percent plant cover.
Comparison Results with LAI
For Salix arctica, 6 out of 17 samples investigated at the
species level for tundra plots around Iqaluit and Clyde River
by the VE method had an estimated –ln (1 - C) greater than
the total projected LAI (Fig. 2 and Table 1). Therefore, we
can conclude with certainty that in 6/17 or 35% of cases, the
VE method overvalues. The DGO method (DGO5×5 and
DGO10×10) also overvalued to a similar degree as the VE.
Using the PDPC method, we found no cases in which the
estimated –ln (1 - C) was greater than the total projected
LAI for Salix arctica.
Similar results were found for other vascular plant spe-
cies (Table 1), although the rates of overvaluation varied
somewhat from species to species. The PDPC method con-
tinued to be the most accurate method. However, in the
cases of Carex membranacea, Vaccinium uliginosum, and
Poa sp., we did nd 1/15, 1/7, and 2/9 cases, respectively, in
which the PDPC estimate for –ln (1 - C) was greater than
the total projected LAI. Several explanations are possible.
First, some leaves might be mistakenly merged and mis-
labeled in applying the PDPC method, so that leaves that
belonged to other species were misclassied as those of
Carex membranacea, Vaccinium uliginosum, and Poa sp.
In addition, during the destructive harvest of plants in the
eld, it was inevitable that some of the plant components
were omitted. As a result, the destructively measured LAI
was less than the actual LAI, which in turn could result in
the estimated percent cover of leaves surpassing the total
projected LAI measured destructively.
E=estimated percent cover - true percent cover
true percent cover
×100%

320 • Z. CHEN et al.
Nevertheless, when all 75 samples at the species level
were included, the PDPC method was found to have the
lowest overvaluation rate at 9%, in comparison to 38% for
the VE method, 33% for the DGO5×5 and DGO10×10, 31%
for the DGO15×15, and 27% for the DGO20×20 methods
(Fig. 3 and Table 1). When results were aggregated to the
plot total level, the errors caused by misclassication disap-
peared, resulting in 0% overvaluation for the PDPC method
(Fig. 4 and Table 1). However, the VE method and the DGO
using different grids still had overvaluation rates of 27%
to 58%. The overvaluation of vegetation cover by the VE
method and the DGO was also reported by Dethier et al.
(1993) and Vanha-Majamaa et al. (2000).
In contrast to the situation of overvaluation, the DGO
method could entirely miss a species. As shown in Table 2,
the DGO5×5 entirely missed a species in 43% cases when all
plant species were summarized. The rate of missing species
is only 15% for the DGO10×10 and 7% for the DGO15×15
and the DGO20×20. For the VE and PDPC methods, no
cases of missing a species occurred.
Signicant correlations at the 95% condence level
between LAI and –ln (1 - C) were found for the ve species
listed in Table 3, except for Vaccinium uliginosum using
the DGO10×10 and for Poa sp. using the DGO5×5 and
DGO10×10. Overall, values from the PDPC method show
the best correlation between LAI and –ln (1 - C), both at
the species level and at the plot total level (Figs. 3 and 4).
At the plot total level, the value of R2 was 0.88 for the PDPC
method, compared to 0.56 for the VE method and 0.58 to
0.79 for the DGO method using different grids.
On the basis of the testing results using articial plots
and comparison results with LAI, we concluded that the
PDPC method was the most accurate of the three methods
investigated here. Therefore, we used the PDPC values of
percent cover for tundra plots in Nunavut as the “true” val-
ues against which to evaluate the methods of VE and DGO
(using different grid resolutions).
Inter-comparison Results
Figure 5 shows the visual representation of plant cover
distribution for an example plot. It shows the original dig-
ital photograph and images of cover measured using the
PDPC method and the DGO method at different grid reso-
lutions. A close resemblance in plant cover distribution can
be found between the original digital photograph and the
measurement using the PDPC method. Much of the spatial
distribution detail is lost in the measurement using the DGO
method, especially with DGO5×5, resulting in overvalua-
tion of percent cover for some species and undervaluation
FIG. 1. Errors in percent cover estimates made by the digital grid overlay
(DGO) method using different grid resolutions and by the plot digital
photograph-based classication method (PDPC) for a rti cial plots with
known percent cover values.
FIG. 2. Comparison between leaf area index (LAI) and –ln (1 - C), where
C is percent cover of Salix arctica, measured or estimated using the visual
estimate method (VE), the DGO method with different grid resolutions, and
the PDPC method. Slopes are 1:1 lines.
TABLE 1. Fraction of samples for which –ln (1 - percent cover) was greater than the total LAI. The percent cover was estimated using the
visual estimate (VE) method, the digital grid overlay (DGO) using different grids, and the plot digital photograph classication method
(PDPC).
VE DGO5×5 DGO10×10 DGO15×15 DGO20×20 PDPC
Salix arctica 6/17 8/17 5/17 6/17 7/17 0/17
Dryas integrifolia 4/7 1/7 1/7 2/7 1/7 0/7
Vaccinium uliginosum 3/7 4/7 4/7 5/7 3/7 1/7
Carex membranacea 6/15 2/15 5/15 2/15 2/15 1/15
Poa sp. 4/9 3/9 3/9 2/9 2/9 2/9
All samples at the species level 29/75 25/75 25/75 23/75 20/75 7/75
At the plot total level 15/26 13/26 11/26 9/26 7/26 0/26

MEASURING PERCENT PLANT COVER • 321
for other species. For example, in relation to the “true”
value measured by the PDPC method, the DGO5×5 method
undervalued the percent cover of Carex membranacea in
the plot by 116% and that of Oxytropis arctica by 100%
(Table 4).
Even larger errors were found when all 75 samples were
analyzed at the species level (Fig. 6). The errors range from
-87% to 1941% for the VE method, -100% to 950% for
the DGO5×5, -100% to 2075% for the DGO10×10, -100%
to 770% for the DGO15×15, and -100% to 646% for the
DGO20×20. For percent cover measurement at plot total
level, the error range was smaller: -48% to 332% for the
VE method, -100% to 191% for the DGO5×5, -16% to 188%
for the DGO10×10, -22% to 155% for the DGO15×15, and
-11% to 93% for the DGO20×20. The main reason for the
decrease in error from the species level to the plot total level
is compensating errors in the aggregation process.
Such mutual cancellation of errors may also explain
why the error range decreases when percent cover value
increases for Salix arctica and all species in Arctic tundra
plots around Iqaluit and Clyde River. From Figure 7, we can
clearly see that the largest percentage errors occur when
the percent cover is less than 1%. This relationship also
partially explains why the error ranges for the actual tun-
dra plots in Figure 6 are much larger than that for the arti-
cial plots in Figure 1. For the articial plots in Figure 1, the
smallest percent cover used was 1.68%, while many plots
in Figure 6 had percent cover of less than 1%. When com-
parison was made at the plot level, we found similar magni-
tudes of error for both the actual plots in Nunavut and the
articial plots. In both cases, we found a decreasing trend
in measurement error as percent cover increased.
Technical Considerations for Using the Plot Digital
Photograph Classication Methods
We found that the combination of IHS and RGB images
was effective in identifying vegetation cover by species.
IHS transformation and object-based segmentation tech-
niques were effective for processing plot photographs and
segmenting the green leaves of vegetation species into rela-
tively homogeneous polygons. The adjustable IHS thresh-
old values performed well at mapping out green vegetation
polygons in this study. At a proper scale, a polygon may be
composed of an individual leaf (for big leaves) or a cluster of
leaves (for small leaves) from the same species. Along with
visual interpretation of RGB images, we used GIS editing
to correct boundary outlining and increase the accuracy of
species labeling for each polygon. This approach was use-
ful in calculating the percent cover from plot photographs
in this study. Since individual blades were identied for
most leaves of different species, we consider that the per-
cent cover measured using this method was close to the true
value.
FIG. 3. Comparison between LAI and –ln (1 - C) for all species, where C is
percent cover of a species measured or estimated using the VE method, the
DGO method with different grid resolutions, and the PDPC method. Slopes
are 1:1 lines.
FIG. 4. Relationships between plot-total LAI and –ln (1 - C), where C is
the percent cover at the plot total level measured or estimated using the VE
method, the DGO method with different grid resolutions, and the PDPC
method. Slopes are 1:1 lines.
TABLE 2. Fraction of samples for which the total LAI is greater than 0 while the estimate of percent cover = 0, for estimates of percent
cover made with the VE, DGO using different grids, and PDPC methods.
VE DGO5×5 DGO10×10 DGO15×15 DGO20×20 PDPC
Salix arctica 0/17 7/17 1/17 0/17 0/17 0/17
Dryas integrifolia 0/7 2/7 0/7 0/7 0/7 0/7
Vaccinium uliginosum 0/7 0/7 1/7 0/7 0/7 0/7
Carex membranacea 0/15 7/15 1/15 1/15 0/15 0/15
Poa sp. 0/9 5/9 2/9 1/9 0/9 0/9
All samples at the species level 0/75 32/75 11/75 5/75 5/75 0/75
At the plot total level 0/26 2/26 0/26 0/26 0/26 0/26

322 • Z. CHEN et al.
In plots containing Poa sp. and Carex membranacea, the
ratio of length to width was useful for separating these two
species from the rest of the green vegetation. Vegetation
with wider leaves, such as Salix arctica, Salix reticulata,
and Oxytropis arctica, can also be separated easily from
the surroundings. The boundaries of vegetation with small
leaves, such as the Dryas integrifolia and Vaccinium uligi-
nosum, can be hard to dene because the leaves are clus-
tered together.
The time for processing a plot photograph ranged from
30 to 90 minutes depending on the complexity of the veg-
etation. Compared to the quick VE method in the eld, the
PDPC method takes more time. However, analysis of plot
photographs can be faster once the method is developed,
and the results and photographs of different dates can be
used to assess vegetation dynamics for monitoring pur-
poses. In addition, photographs can be acquired quickly,
archived easily, and used for future change analysis. Indeed,
as several researchers have noted, one of the great advan-
tages of measuring cover from digital images is that analy-
sis can be done in the lab after the eld investigation. Field
research in the Arctic is time-consuming, labour-intensive,
and logistically expensive. In contrast to the results from the
VE method, which have high observer-to-observer variabil-
ity (Meese and Tomich, 1992), the results from the PDPC
method are relatively consistent between analysts because
the accuracy of the percent cover in plot photograph analy-
sis is largely dependent on polygons with clear boundaries
produced at the segmentation stage, which is a more stand-
ardized step.
However, application of the PDPC method in this paper
also raises a few issues for future study:
1) The results from the PDPC method may be more reliable
for vegetation of greenish appearance than for vegeta-
tion of less greenish appearance. For vegetation species
with grey or brown color, such as Cassiope tetragona
and Pedicularis hirsuta L., some polygons representing
leaves were masked out as non-green vegetation dur-
ing the preliminary digital classication using the IHS
threshold.
2) Shadow area varies in each image and affects the percent
cover measurement. Shadow is always a problem in very
high-resolution images because it occurs both on the
non-vegetation area and within the vegetation canopy,
making it hard to identify whether shadow area belongs
to vegetation or non-vegetation. Therefore, most shadow
areas were treated as non-vegetation areas using only
IHS threshold values during the segmentation. Shadow
affects the measurement of percent cover more for small-
leaf species than for big-leaf species. Some shadow may
be included in a polygon containing a few small leaves
during the segmentation stage and cause over-estimation
for species with small leaves. High dynamic range nadir
images may provide a solution for reducing the shadow
problem (Cox and Booth, 2008). In contrast, the green
vegetation in the shadow areas can be clearly identied
by the VE method or the point-frame method conducted
in the eld.
3) Limited error may be caused by mixed pixels. Color dif-
ferences in the boundaries caused mixed pixels to be cre-
ated in a captured digital image, and these pixels became
inherent error sources (Booth et al., 2006a). Fuzzy edges
of leaves may cause errors in image segmentation and
area calculation, but the size of such errors may be mini-
mal in an image of very high resolution.
4) Acquisition of plot photographs should be standardized
in order to achieve better results. At the beginning of this
study, some of the plot photographs were not well focused
and distorted and therefore were not usable. Also, during
the eld measurement, the analysts tied ribbons on the
TABLE 3. R2 values for the linear, intercept = 0 correlation between LAI and –ln (1 - C), estimated using the VE, DGO using different
grids, and PDPC methods. For other species that had a sample size less than 5, values of R2 were not calculated. Asterisks represent
statistical signicance at the 90% (*) and 95% (**) condence levels.
VE DGO5×5 DGO10×10 DGO15×15 DGO20×20 PDPC n
Salix arctica 0.67** 0.60** 0.67** 0.72** 0.81** 0.84** 17
Dryas integrifolia 0.67** 0.70** 0.79** 0.90** 0.87** 0.82** 7
Vaccinium uliginosum 0.70** 0.78** 0.50* 0.72** 0.66** 0.83** 7
Carex membranacea 0.69** 0.73** 0.75** 0.82** 0.78** 0.87** 15
Poa sp. 0.55** 0.01 0.24 0.93** 0.62** 0.80** 9
At the plot total level 0.56** 0.58** 0.66** 0.57** 0.79** 0.88** 26
FIG. 5. Visual illustrations of plant cover distribution in an example plot. (A)
shows the original digital photograph, and (B) shows distribution measured
by the PDPC method. The remaining blocks show distributions from the DGO
method with four different grids: (C) 20 × 20, (D) 15 × 15, (E) 10 × 10, and (F)
5 × 5. Percent cover values for each species are listed in Table 4.

MEASURING PERCENT PLANT COVER • 323
frames and put labels inside the plot for the convenience
of recording. Consequently, these ribbons and labels
were also included in the photos used for analysis. These
added objects obviously hindered the image interpreta-
tion of the objects underneath. Therefore, no information
was available from those areas using the DGO method
and the PDPC method.
5) Both the PDPC method and the DGO method used here
are two-dimensional measurements, while VE is a three-
dimensional estimation that takes multi-layer vegetation
into consideration. Because the photographs are taken
from above, only the surface part of the vegetation can
be viewed in the images, and the underlying vegetation
cannot be seen. As a result, image-based analysis cannot
identify multilayered species. Thus the PDPC method
may underestimate percent cover because leaves over-
lapped by other leaves cannot be treated individually,
while the VE method at the eld site can detect under-
lying vegetation and provide the better estimation. In
addition, the plot photograph may be more suitable for
identifying vegetation with spreading leaves, and less
suitable for species with tightly packed, pointy leaves,
such as Cassiope tetragona. However, a visual estimate
is one person’s opinion and cannot be validated, whereas
plot photographs are a permanent record that can be re-
analyzed if questions arise.
6) The articial plots with known quantities used in this
study are two-dimensional and therefore much simpler
than the real world. Such articial plots have many lim-
itations: constancy of shape and color, restriction of no
overlap, and non-random distribution at the edges, as sug-
gested by Schultz et al. (1961) and Booth et al. (2006b).
But even though articial plots do not accurately reect
the real world, they do serve the purpose of demonstrat-
ing the inuence of changing sampling density on accu-
racy when using the DGO and the PDPC methods.
7) The PDPC method could be useful in landscape monitor-
ing. However, landscape monitoring requires adequate
sample density and distribution. Many images may be
used to cover a study area. For this use, cover measure-
ments for an image would have to be completed in a very
short time (less than an hour processing time per image).
Note that in this study site selection was focused more on
the value of the three categories of plant coverage—low,
medium, and high plant cover—rather than the randomi-
zation. More randomly selected sites may be needed to
verify the relationship established here between plant
cover measured by PDPC and LAI. Inadequate sampling
at the landscape level affects accuracy as much or more
than limited sample points at the plot or sample level.
Although batch processing of hundreds of images is pos-
sible at the stage of image segmentation if large areas are
being monitored and hundreds of plots (samples) need
TABLE 4. Percent cover estimates (%) of ve species in an example plot, estimated by different methods: VE, DGO using different grids,
and PDPC.
VE DGO5×5 DGO10×10 DGO15×15 DGO20×20 PDPC
Salix arctica 12 8 7 7.6 10 6.3
Dryas integrifolia 20 12 3 6.2 4.5 11.4
Vaccinium uliginosum 15 16 3 5.8 4.7 4.7
Carex membranacea 5 8 4 4.9 7 3.7
Oxytropis arctica 0.1 0 0 0.8 1 0.8
FIG. 6. Errors in percent cover estimates for various species in tundra plots
around Iqaluit and Clyde River, Nunavut. Estimates from the VE method
and the DGO method using different grids are compared to the “true” values
obtained by the PDPC method.
FIG. 7. Measurement errors in relation to percent cover value for all species in
the Arctic tundra plots. Estimates from the VE method and the DGO method
are compared to the “true” values obtained by the PDPC method. All the
graphs show a clear trend: measurement error decreases as percent cover
increases.

324 • Z. CHEN et al.
to be analyzed, our suggestion in this study of species
labeling in a GIS, aided by visual interpretation of RGB
images, can be time-consuming.
SUMMARY
An object-oriented digital image analysis was adapted
for measuring plant cover by key species in Arctic ecosys-
tems. The innovation with respect to previous studies of
digital image analysis is that RGB and IHS images were
used together to ensure more accurate species labeling and
cover measurement.
We evaluated the accuracy of three methods (VE, DGO,
and PDPC) in three ways: by testing against articial plots
with known percent cover, comparison with destructively
measured LAI, and inter-comparison among methods.
The testing against articial plots with known percent
cover indicates, as others have reported, that the measure-
ment errors in percent cover decrease as the sample point
density (or DGO method grid resolution) increases. Using
the 3872 × 2592 grids, the sensor array of the digital cam-
era used for this study, the PDPC method was found to be
the most accurate of the three methods in percent cover
measurement.
Theoretically speaking, if percent cover of leaves C
is estimated correctly, the value of –ln (1 - C) should be
greater than the total projected LAI. For all 75 cases at the
species level, we found that the PDPC method has the low-
est overvaluation rate at 9%, compared to 39% for the VE
method and 27%–33% for the DGO method using various
grids. When estimates were aggregated to the plot total
level, compensating errors resulted in a 0% overvaluation
for the PDPC method (compared to 58% for the VE method
and 27%–50% for the DGO method using various grids).
On the other hand, the DGO method could entirely miss
a species, and the rate of missed species decreased as the
sampling density increased. Using the VE and PDPC meth-
ods, no species were missed. At the plot total level, the value
of R2 between LAI and –ln (1 - C) was 0.88 for the PDPC
method, compared to 0.56 for the VE method and 0.58–0.79
for the DGO method using various grids.
Inter-comparison of methods also revealed that errors
usually decreased as the percent cover value increased. All
these evaluations suggest that the PDPC method was the
most accurate of the three methods for vegetation cover
measurement in the Arctic ecosystem.
The PDPC method has clear advantages over previously
used methods for measuring vegetation cover in the Arc-
tic: it is accurate and objective; it reduces the time needed
in the eld; and it provides a permanent record that can be
re-analyzed. Image analysis can be performed in an ofce
or laboratory; however, it is more time-consuming than a
visual estimate in the eld. This method depends on the
quality of the initial image: correct and consistent methods
of taking digital photos must be followed to ensure images
will be usable. With further technical improvements that
will reduce the distorting effects of shadows and overlap,
the PDPC method has a promising future.
ACKNOWLEDGEMENTS
The work was nancially supported by the Canada Inter-
national Polar Year initiative, through the project entitled “Cli-
mate Change Impacts on Canadian Arctic Tundra Ecosystems:
Interdisciplinary and Multi-scale Assessments (CiCAT),” and by
Natural Resources Canada, through the Enhancing Resilience
to a Changing Climate program. The Nunavut Research Insti-
tute helped us in obtaining a research license. We are grateful to
Jackie Bourgeois (Climate Change Coordinator, Government of
Nunavut), Michelle Bertol (Director, Land and Planning, City of
Iqaluit), James Qillaq and Nick Illauq (Councilors, Clyde River,
Nunavut), Lee Ann Pugh and Jayko Ashvak (local residents), and
many members of the Clyde River Hunters and Trappers Associa-
tion for assistance with logistical arrangements and participation
in the eldwork in 2006. Drs. Yu Zhang and Junhua Li internally
reviewed the manuscript. We also thank the three anonymous
reviewers for their constructive comments and suggestions.
REFERENCES
Anisimov, O.A., Vaughan, D.G., Callaghan, T.V., Furgal, C.,
Marchant, H., Prowse, T.D., Vilhjálmsson, H., and Walsh,
J.E. 2007. Polar regions (Arctic and Antarctic). In: Parry,
M.L., Canziani, O.F., Palutikof, J.P., Van der Linden, P.J., and
Hanson, C.E., eds. Climate change 2007: Impacts, adaptation
and vulnerability. Contribution of Working Group II to the
Fourth Assessment Report of the Intergovernmental Panel on
Climate Change. Cambridge: Cambridge University Press.
653 – 685.
Bean, D., and Henry, G.H.R. 2003. CANTTEX Field Manual: Part
A – Setting up a basic monitoring site. Burlington, Ontario:
Ecological Monitoring and Assessment Network, Environment
Canada.
Bliss, L.C., Svoboda, J., and Bliss, D.I. 1984. Polar deserts, their
plant cover and plant production in the Canadian High Arctic.
Holarctic Ecology 7:305–324.
Bonham, C.D., and Clark, D.L. 2005. Quantication of plant cover
estimates. Grassland Science 51:129– 137.
Booth, D.T., and Cox, S.E. 2008. Image-based monitoring to
measure ecological change. Frontiers in Ecology and the
Environment 6:185– 190.
Booth, D.T., Cox, S.E., Louhaichi, M., and Johnson, D.E. 2004.
Lightweight camera stand for close-to-earth remote sensing.
Journal of Range Management 57:675 –678.
Booth, D.T., Cox, S.E., Field, C., Phillips, M., and Williamson,
N. 2005. Image analysis compared with other methods for
measuring ground cover. Arid Land Research and Management
19:91–100.
Booth, D.T., Cox, S.E., and Berryman, R.D. 2006a. Point sampling
digital imagery with ‘Samplepoint.’ Environmental Monitoring
and Assessment 123:97– 108.

MEASURING PERCENT PLANT COVER • 325
Booth, D.T., Cox, S.E., Meikle, T.W., and Fitzgerald, C. 2006b.
The accuracy of ground-cover measurements. Rangeland
Ecology & Management 59:179– 188.
Bråkenhielm, S., and Liu, Q. 1995. Comparison of eld methods
in vegetation monitoring. Water, Air and Soil Pollution 79:
75– 87, doi:10.1007/BF01100431.
Bunn, A., and Goetz, S. 2006. Trends in satellite-observed
circumpolar photosynthetic activity from 1982 to 2003: The
inuence of seasonality, cover type, and vegetation density.
Earth Interactions 10:1– 19.
Chen, J.M., Rich, P.M., Gower, S.T., Norman, J.M., and Plummer,
S. 1997. Leaf area index of boreal forests: Theory, techniques,
and measurements. Journal of Geophysical Research 102(D24):
29429 – 29443.
Chen, W., Blain, D., Li, J., Fraser, R., Zhang, Y., Leblanc, S.,
Koehler, K., Olthof, I., Wang, J., and McGovern, M. 2009a.
Estimating carbon release caused by land use changes over
Canada’s North during 1985–1990 and 1990–2000 using
satellite earth observation. Journal of Geophysical Research,
114, G01017, doi:10.1029/2007JG000631.
Chen, W., Li, J., Zhang, Y., Zhou, F., Koehler, K., Leblanc, S.,
Fraser, R., Olthof, I., Zhang, Y.S., and Wang J. 2009b. Relating
biomass and leaf area index to non-destructive measurements
in order to monitor changes in Arctic vegetation. Arctic
62(3):2 81 – 294.
Cox, S.E., and Booth, D.T. 2008. Shadow attenuation with high
dynamic range images. Environmental Monitoring and
Assessment 158:231 –241, doi:10.1007/s10661-008-0577-y.
Deniens Imaging. 2003. eCognition version 4.0.6 user guide.
Munich, Germany: Deniens Imaging GmbH.
Dethier, M.N., Graham, E.S., Cohen, S., and Tear, L.M. 1993.
Visual versus random-point percent cover estimations:
Objective is not always better. Marine Ecology Progress Series
96:93–100.
ERDAS. 2010. ERDAS Imagine. http://www.erdas.com/84/
currentid/1050/default.aspx.
Floyd, D.A., and Anderson, J.E. 1987. A comparison of three
methods for estimating plant cover. Journal of Ecology
75:221– 228.
Hemming, J., and Rath, T. 2001. Computer vision-based weed
identication under eld conditions using controlled lighting.
Journal of Agricultural Engineering Research 78(3):233–243.
Herold, M., Liu, X., and Clarke, K.C. 2003. Spatial metrics and
image texture for mapping urban land use. Photogrammetric
Engineering and Remote Sensing 69(9):991–1001.
Johnson, D.E., Vulfson, M., Louhaichi, M., and Harris, N.R. 2003.
VegMeasure Version 1.6 user’s manual. Corvallis: Department
of Rangeland Resources, Oregon State University.
Karcher, D.E., and Richardson, M.D. 2003. Quantifying turfgrass
color using digital image analysis. Crop Science 43:943–951.
Kattsov, V.M., Kallen, E., Cattle, H., Christensen, J., Drange, H.,
Hanssen-Bauer, I., Johannesen, T., et al. 2005. Future climate
change: Modeling and scenarios for the Arctic. In: Arctic
climate impact assessment. Cambridge: Cambridge University
Press. 100 – 150.
Kennedy, C.E., Smith, C.A., and Cooley, D.A. 2001. Observations
of change in the cover of polar grass, Arctagrostis latifolia, and
arctic lupin, Lupinus arcticus, in upland tundra on Herschel
Island, Yukon Territory. Canadian Field-Naturalist 115:
323–328.
Klein, D.R., Baskin, L.M., Bogoslovskaya, L.S., Danell, K.,
Gunn, A., Irons, D.B., Konas, G.P., et al. 2005. Management
and conservation of wildlife in a changing Arctic environment.
In: Symon, C., Arris, L., and Heal, B., eds. Arctic climate
impact assessment. Cambridge: Cambridge University Press.
597–648.
Krebs, C.J., Danell, K., Angerbjörn, A., Agrell, J., Berteaux,
D., Bråthen, K.A., Danell, Ö., et al. 2003. Terrestrial trophic
dynamics in the Canadian Arctic. Canadian Journal of Zoology
81:827 – 843.
Laliberte, A.S., Rango, A., Havstad, K.M., Paris, J.F., Beck, R.F.,
McNeely, R., and Gonzalez, A.L. 2004. Object-oriented image
analysis for mapping shrub encroachment from 1937 to 2003
in southern New Mexico. Remote Sensing of Environment
93:198 – 210, doi:10.1016/j.rse.2004.07.011.
Laliberte, A.S., Rango, A., Herrick, J.E., Fredrickson, E.L., and
Burkett, L. 2007. An object-based image analysis approach for
determining fractional cover of senescent and green vegetation
with digital plot photography. Journal of Arid Environments
69:1–14.
Luscier, J.D., Thompson, W.L., Wilson, J.M., Gorham, B.E., and
Dragut, L.D. 2006. Using digital photographs and object-based
image analysis to estimate percent ground cover in vegetation
plots. Frontiers in Ecology and the Environment 4(8):408–413,
doi:10.1890/1540-9295(2006)4[408:UDPAOI]2.0.CO;2.
McGuire, A.D., Wirth, C., Apps, M., Beringer, J., Clein, J.,
Epstein, H., Kicklighter, D.W., et al. 2002. Environmental
variation, vegetation distribution, carbon dynamics, and
water/energy exchange in high latitudes. Journal of Vegetation
Science 13:301–314.
Meese, R.J., and Tomich, P.A. 1992. Dots on the rocks: A
comparison of percent cover estimation methods. Journal of
Experimental Marine Biology and Ecology 165:59– 73.
Nilson, T. 1971. A theoretical analysis of the frequency of
gaps in plant stands. Agricultural Meteorology 8:25– 38,
doi:10.1016/0002-1571(71)90092-6 .
Olthof, I., Pouliot, D., Latifovic, R., and Chen, W. 2007. Recent
(1986–2006) vegetation-specic NDVI trends in northern
Canada from satellite data. Arctic 61(4):381– 394.
PCA (Parks Canada Agency). 2000. “Unimpaired for future
generations”? Protecting ecological integrity with Canada’s
national parks, Vol. II: Setting a new direction for Canada’s
national parks. Report of the Panel on the Ecological Integrity
of Canada’s National Parks. Ottawa: Minister of Public Works
and Government Services.
Richardson, M.D., Karcher, D.E., and Purcell, L.C. 2001.
Quantifying turfgrass cover using digital image analysis. Crop
Science 41:1884 –1888.
Ryherd, S., and Woodcock, C. 1996. Combining spectral and
texture data in the segmentation of remotely sensed images.

326 • Z. CHEN et al.
Photogrammetric Engineering and Remote Sensing 62(2):
181–194.
Schultz, A.M., Gibbens, R.P., and Debano, L. 1961. Articial
populations for teaching and testing range techniques. Journal
of Range Management 14:236 –242.
Stow, D.A., Hope, A., McGuire, D., Verbyla, D., Gamon, J.,
Huemmrice, F., Houston, S., et al. 2004. Remote sensing of
vegetation and land-cover change in Arctic tundra ecosystems.
Remote Sensing of Environment 89:281–308.
Tang, L., Tian, L., and Steward, B.L. 2000. Color image
segmentation with genetic algorithm for in-eld weed sensing.
Transactions of the American Society of Agricultural Engineers
43(4) :1019 – 102 7.
Thomas, N., Hendrix, C., and Congalton, R.G. 2003. A comparison
of urban mapping methods using high-resolution digital
imagery. Photogrammetric Engineering & Remote Sensing
69(9):963– 972.
Vanha-Majamaa, I., Salemaa, M., Tuominen, S., and Mikkola,
K. 2000. Digitized photographs in vegetation analysis: A
comparison of cover estimates. Applied Vegetation Science
3:89–94.
Walker, M. 1996. Community baseline measurements for ITEX
studies. In: Molau, U., and Mølgaard, P., eds. International
Tundra Experiment (ITEX) manual. Copenhagen: Danish
Polar Center. 5– 47.