TURFGRASS SCIENCE Quantifying Turfgrass Color Using Digital Image Analysis

advertisement
TURFGRASS SCIENCE
Quantifying Turfgrass Color Using Digital Image Analysis
Douglas E. Karcher* and Michael D. Richardson
ABSTRACT
on subjective data is debatable (Karcher, 2000) as the
data tend to be discrete and ordinal rather than continuous. Timely quantification of turfgrass color that uses
readily accessible equipment would strengthen the validity of study results without adding significant burden
to the evaluation process.
Several techniques have been used to objectively
measure turf color, including reflectance measurements
(Birth and McVey, 1968), chlorophyll and amino acid
analysis (Johnson, 1973; Nelson and Sosulski, 1984), and
comparison with standardized colors (Beard, 1973). All
of these methods have certain disadvantages compared
with subjective color ratings. Reflectance, chlorophyll,
and amino acid measurements all require relatively expensive equipment, and transport of samples to a laboratory for analysis. In addition, correlations between
color and chlorophyll or amino acid measurements are
either species or cultivar dependent. The use of standardized charts to measure turf color is effective, but
results in qualitative descriptions of color that are not
possible to statistically analyze with traditional ANOVA techniques.
Recently, Landschoot and Mancino (2000) demonstrated that the color of creeping bentgrass cultivars
could be successfully quantified with a colorimeter. Values from the colorimeter were significantly correlated
with visual color assessments averaged across five evaluators. Other researchers have successfully used colorimeters to evaluate varying turf color due to seasonal
changes (Kimura et al., 1989) or differences among cultivars and genetic lines (Thorogood et al., 1993). Although promising, a potential shortcoming of the colorimeter used in those studies is the relatively small
measurement area (⬍20 cm2). In the absence of extremely uniform surface conditions, numerous subsample measurements with the colorimeter would be necessary to accurately represent the color of typical turfgrass
field plots.
In recent years, digital photography has become a
common and affordable means for the scientific community to document and present images. Digital cameras,
in conjunction with image analysis software, are being
used to quantify wheat (Triticum aestivum L.) senescence (Adamsen et al., 1999) and canopy coverage in
wheat (Lukina et al., 1999) and soybeans [Glycine max
L. (Merr.)] (Purcell, 2000). Recently, digital image analysis was used to quantify turf coverage with increased
precision over more traditional evaluation methods
Color is a major component of the aesthetic quality of turf and
often evaluated in field studies. Digital image analysis may be an
improved, objective method to quantify turf color. Studies were conducted to determine if digital image analysis with SigmaScan software
(SPSS, Chicago, IL) was capable of: (i) accurately determining the
hue, saturation, and brightness (HSB) levels of Munsell Plant Tissue
color chips, (ii) quantifying visual color differences among zoysiagrass
(Zoysia japonica Steud.) and creeping bentgrass {Agrostis palustris
Huds. [⫽ A. stolonifera var. palustris (Huds.) Farw.]} plots receiving
various N treatments, and (iii) quantifying genetic color differences
among bermudagrass (Cynodon spp.) cultivars. Digital images of turf
plots were analyzed with SigmaScan software to determine average
HSB levels for each image. A dark green color index (DGCI) was
created from HSB values for direct comparison with visual ratings.
Digital image analysis accurately quantified the HSB levels (r2 ⫽ 0.99,
0.96, and 0.97, respectively) of Munsell color chips corresponding
to turf colors. Significant HSB differences were present among N
treatments in creeping bentgrass, while only significant hue differences
existed in zoysiagrass. Significant hue and saturation differences were
present among bermudagrass cultivars. There was strong agreement
between DGCI values and visual ratings. The relative variances of
the HSB and DGCI were significantly less than the variance associated
with multiple raters. This evaluation technique may facilitate objective
comparisons of turf color across researchers, locations, and years when
images are collected under equal lighting conditions (i.e., the use of
an artificial light source at night or in an enclosed system).
T
urf color is a key component of aesthetic quality
and a good indicator of water and nutrient status
(Beard, 1973). Therefore, color is often evaluated in
turfgrass experiments. Color is traditionally evaluated
by visually rating turf plots on a scale of 1 to 9, with 1
representing yellow or brown turf and 9 representing
optimal, dark green turf. Although color ratings provide
quick data acquisition without the need for specialized
equipment, they are a subjective measure from which
human bias is difficult to remove. As a result, inconsistencies often exist among raters when evaluating the
same turf plots. Relatively poor correlations existed
among experienced researchers (r ⬍ 0.68) when rating
the same turf plots for density, color, and leaf spot
(Skogley and Sawyer, 1992; Horst et al., 1984). Correlations this low would probably be considered unacceptable when using other evaluation tools (e.g., balances,
spectrometers, pH meters) to measure the same turf
sample. Furthermore, the applicability of standard ANOVA procedures and traditional means separation tests
Dep. of Horticulture, Univ. of Arkansas, 308 Plant Sci. Building,
Fayetteville, AR 72701. Received 4 April 2002. *Corresponding author (karcher@uark.edu).
Abbreviations: AS, ammonium sulfate; DAT, days after treatment;
DGCI, dark green color index; HSB, hue, saturation, and brightness;
NTEP, National Turfgrass Evaluation Program; PCU, polymer-coated
urea; RGB, red, green, and blue; SCU, sulfur-coated urea.
Published in Crop Sci. 43:943–951 (2003).
943
944
CROP SCIENCE, VOL. 43, MAY–JUNE 2003
(Richardson et al., 2001). Through digital photography,
researchers can instantaneously obtain millions of bits
of information on a relatively large turfgrass canopy.
For example, a digital image taken of a turf plot using
a 1280 ⫻ 960 pixel resolution contains 1 228 800 pixels,
with each pixel containing independent color information about the turf plot. Therefore, digital photography
and subsequent image analysis may be capable of quantifying turfgrass color in field experiments.
The information contained in a digital image includes
the amount of red, green, and blue (RGB) light emitted
for each pixel in the image. Although it may be intuitive
to use the green levels of the RGB information to quantify the green color of an image, the intensity of red and
blue will confound how green an image appears. To
ease the interpretation of digital color data, RGB values
can be converted directly to HSB values that are based
on human perception of color (Fig. 1). In HSB color
description, hue is defined as an angle on a continuous
circular scale from 0 to 360⬚ (0⬚ ⫽ red, 60⬚ ⫽ yellow,
120⬚ ⫽ green, 180⬚ ⫽ cyan, 240⬚ ⫽ blue, 300⬚ ⫽ magenta),
saturation is the purity of the color from 0% (gray)
to 100% (fully saturated color), and brightness is the
relative lightness or darkness of the color from 0%
(black) to 100% (white) (Adobe Systems, 2002). Among
HSB, hue has been found to be the best indicator of
the visual color of a turf (Landschoot and Mancino,
2000; Thorogood et al., 1993). However, preliminary
work at the University of Arkansas has demonstrated
that visual differences in turf color were sometimes the
result of color saturation differences between turf plots
rather than hue differences (Karcher, 2000, unpublished
data).
The objective of the following research was to determine if readily available equipment (a digital camera
and commercially available software) could accurately
quantify turfgrass color using an HSB color scale. Digital
images were taken of standard color objects (Munsell
Plant Tissue color chips) to determine the accuracy of
digital image analysis with regard to the quantification
of color parameters. Digital images were collected of
turfgrass field plots varying in visual color due to either
N fertility or genetically controlled differences to determine if digital image analysis was capable of quantifying
color differences.
MATERIALS AND METHODS
Color Quantification of Digital Images
The process used to determine the average color of a digital
image included: (i) acquiring an image with digital photography, (ii) obtaining the average RGB pixel levels for the image,
and (iii) converting the RGB levels to the more intuitive HSB
parameters. All digital images in these studies were taken with
an Olympus C-3030 camera (Olympus America Inc., Melville,
NY). The images were collected in the JPEG (joint photographic experts group, .jpg) format, with a color depth of 16.7
million colors, and an image size of 1280 ⫻ 960 pixels (≈260
kilobytes per image). Camera settings included a shutter speed
of 1/400 s, an aperture of f/4.0, and a focal length of 32 mm.
Images were downloaded to a personal computer for subsequent analysis.
The average RGB levels of the digital images were calculated using SigmaScan Pro version 5.0 software (SPSS, 1998).
The entire image was selected for analysis by including all
possible hue and saturation levels in the color threshold option
of the software. The average red, average green, and average
blue measurement settings were used to obtain the average
RGB levels for an image. The average RGB levels were then
pasted into an MS Excel spreadsheet (Microsoft Corporation,
1999) created by the authors to automate the conversion of
RGB to HSB values. The programmed formulas in the spreadsheet converted absolute RGB levels (measured on a scale of
0 to 255) to percentage RGB levels by dividing each level by
255. Percentage RGB levels were then converted to average
HSB levels by the following algorithms (Adobe Systems,
2002):
Hue
If max(R,G,B) ⫽ R, 60{(G ⫺ B)/[max(R,G,B) ⫺
min(R,G,B)]}
If max(R,G,B) ⫽ G, 60(2 ⫹ {(B ⫺ R)/[max(R,G,B) ⫺
min(R,G,B)]})
If max(R,G,B) ⫽ B, 60(4 ⫹ {(R ⫺ G)/[max(R,G,B) ⫺
min(R,G,B)]})
Saturation
[max(R,G,B) ⫺ min(R,G,B)]/max(R,G,B)
Brightness
max(R,G,B).
Camera Calibration
A series of digital images were taken of color chips from
Munsell Color Charts for Plant Tissues (GretagMacbeth LLC,
New Windsor, NY). Six images of varying hue were collected,
ranging from yellowish green to green (chip numbers 5Y 6/6,
2.5GY 6/6, 5GY 6/6, 7.5GY 6/6, 2.5G 6/6, 5G 6/6). Eight images
of varying saturation were collected, ranging from grayish
green to bright green (chip numbers 7.5GY 6/2, 7.5GY 5/2,
7.5GY 6/4, 7.5GY 5/4, 7.5GY 6/6, 7.5GY 5/6, 7.5GY 6/8, 7.5GY
5/8). Ten images of varying brightness were collected, ranging
from light green to dark green (chip numbers 7.5GY 8/4,
7.5GY 7/4, 7.5GY 6/4, 7.5GY 5/4, 7.5GY 4/4, 7.5GY 8/6, 7.5GY
7/6, 7.5GY 6/6, 7.5GY 5/6, 7.5GY 4/6). These Munsell color
chips were chosen because they covered a relatively broad
range of HSB levels and visually corresponded with plant
tissue HSB levels typical of turfgrass (Beard, 1973). Calibration images were taken under dark conditions using only the
camera flash as a light source. The images were analyzed for
HSB levels using the methods described above. To determine
the accuracy of HSB measurement with digital image analysis,
the actual HSB levels of the Munsell color chips were determined using Munsell Conversion software version 4.1 (Munsell Color, 2000).
Three separate linear regression analyses were performed
using PROC REG in SAS Statistical Software (SAS Institute.,
1996). The H, S, and B values from digital image analysis were
analyzed as the independent variables and the actual H, S,
and B values of the Munsell color chips were the dependent
variables. For each HSB parameter, digital image analysis
was considered to significantly detect color differences among
color chips when the slope of the regression line was signifi-
KARCHER & RICHARDSON: DIGITAL IMAGE ANALYSIS OF TURF COLOR
945
Fig. 1. Quantifying turfgrass color in the hue, saturation, and brightness (HSB) color space. (A) The hue is measured on a continuous scale
from 0 to 360ⴗ. Turfgrass hues are typically between 70ⴗ and 110ⴗ. (B) For a specific turfgrass hue, here 90ⴗ, the saturation and brightness
levels affect how dark green the color appears.
cantly different from zero (P ⬍ 0.05) (Freund and Wilson,
1993).
Nitrogen Fertility Color Differences
Two ongoing N fertility field studies were used to assess
the ability of digital image analysis to quantify visual color
differences among turf plots due to N treatments. The first
experimental area was established with ‘Meyer’ zoysiagrass
during the summer of 1996 on a silt loam (Typic Hapludult,
pH 6.2). Individual plots were 1.4 m2 and mowed at a height
of 1.9 cm. The second experimental area was a ‘Crenshaw’
creeping bentgrass putting green built in 1998 according to
USGA recommendations (United States Golf Association,
Fig. 5. Color analysis of various turfgrass plots. (A) The plot receiving the higher N rate has a darker green color as a result of an increased
hue angle and decreased brightness level. (B) ‘Shanghai’ bermudagrass has darker green color, the result of significantly decreased saturation
level when compared with ‘Mini-Verde’. H, hue; S, saturation; B, brightness.
946
CROP SCIENCE, VOL. 43, MAY–JUNE 2003
1993). Individual plots were 1.5 m2 and mowed at a height of
0.4 cm. Both experimental areas were located at the University of Arkansas Research and Extension Center in Fayetteville, AR.
The zoysiagrass study consisted of two treatment factors, N
source (7 levels) and N rate (3 levels). The N source treatment
levels included: (i) 100% ammonium sulfate (AS); (ii) 100%
polymer-coated urea (PCU); (iii) 100% sulfur-coated urea
(SCU); (iv) 33% AS, 67% PCU; (v) 33% AS, 67% SCU; (vi)
67% AS, 33% PCU; and (vii) 67% AS, 33% SCU. Each N
source was applied at three N rate levels: (i) 4.8, (ii) 7.2, and
(iii) 9.6 g m⫺2. Each of the resultant 21 fertility treatments
was replicated four times in a randomized complete block
design. Treatment applications were made in mid-May and
mid-August in 2000.
The creeping bentgrass study consisted of one treatment
factor, N rate (7 levels). The N rate treatment levels included
0, 1, 2, 3, 4, 5, and 6 g m⫺2. The N source for all treatments
was methylene urea. Each N rate was applied four times in a
completely randomized design. Treatment applications were
made monthly from June through September in 2000.
Digital images were collected from each plot on 28 Sept.
2000 on the zoysiagrass study [44 d after treatment (DAT)]
and on 16 Nov. 2001 on the creeping bentgrass study (55 DAT)
between 1300 and 1400 h during mostly sunny conditions (illuminance ≈ 50 000 lux). Images were collected by a researcher
standing immediately next to the plot while holding the camera
directly over the center of the plot ≈1.5 m above the turf
canopy. Care was taken to avoid casting shadows on the turf
inside plot. Concurrent to the collection of digital images, the
zoysiagrass and creeping bentgrass studies were visually rated
for color by five and three independent researchers (rater
experience ranged from a minimum of 2 yr to ⬎10 yr), respectively. Color ratings were based on a 1 to 9 scale where 1 ⫽
tan or brown turf, 6 ⫽ minimum acceptable color, and 9 ⫽
optimal dark green color. A DGCI was created from the HSB
values to obtain a single value from digital image color analysis
for comparison with values from subjective visual ratings. The
index was created to measure the relative dark green color of
an image using the following equation:
ance. Since the visual rating scale was unrelated to color values
obtained from digital image analysis, the relative variances
(coefficients of variation) were used for statistical comparison.
Sample variances were calculated as the within-plot mean
square for each color quantification method. Confidence bounds
(95%) were constructed for the sample means and the withinplot variances and were used to calculate confidence bounds
for the coefficients of variation. The relative variances of the
methods were determined to be significantly different if the
respective confidence bounds for the coefficients of variation
did not overlap.
DGCI value ⫽ [(H ⫺ 60)/60 ⫹ (1 ⫺ S) ⫹ (1 ⫺ B)]/3.
RESULTS
Camera Calibration
The color index was calculated from the average of transformed HSB parameters. Each transformed parameter measures dark green color on a scale of zero to one. Since the
hue of most turfgrass images ranges between 60⬚ (yellow) and
120⬚ (green), the maximum dark green hue was assigned as
120⬚. Therefore, the dark green hue transform was calculated
as (hue ⫺ 60)/60, so that hues of 60⬚ and 120⬚ would yield
dark green hue transforms of zero and one, respectively. Since
lower saturation and brightness values corresponded to darker
green colors, (1 ⫺ saturation) and (1 ⫺ brightness) were used
to calculate the dark green saturation and brightness transforms, respectively. The average of the transformed HSB values yielded a single measure of dark green color, the DGCI
value, which ranged from zero to one with higher values corresponding to darker green color.
Analyses of variance were performed using PROC GLM
in SAS Statistical Software (SAS Institute, 1996) on the visual
rating, HSB, and DGCI data sets. For a given color parameter,
treatment and/or interaction effects were determined significant when the corresponding ANOVA f test had a P value ⱕ
0.05. In such cases, a Fisher’s protected LSD test was performed to separate treatment means (Freund and Wilson,
1993).
Three digital images were taken on plots from the zoysiagrass and creeping bentgrass studies to compare the variance of digital image analysis with subjective visual rater vari-
Cultivar Color Differences
Plots from a bermudagrass cultivar trial were used to assess
the ability of digital image analysis to quantify visual color
differences among cultivars. The trial was established in the
summer of 1997 at the University of Arkansas Research and
Extension Center in Fayetteville, AR (silt loam, Typic Hapludults, pH 6.2), and was a test site for the 1997 National Turfgrass Evaluation Program (NTEP) bermudagrass trial (NTEP,
1999). Individual plots were 1.4 m2 and maintained at a 1.9-cm
mowing height. The study was replicated three times in a
completely randomized design.
Digital Images were taken as described previously on each
replication of four cultivars that varied in green color (NTEP,
1999): ‘Cardinal’ (strong yellow-green), ‘Shanghai’ (dark graygreen), ‘Mini-Verde’ (strong dark yellow-green), and ‘Tifway’
(typical bermudagrass green color). The plots were photographed on 21 Sept. 2000 between 1325 and 1335 h during
overcast conditions (illuminance ≈ 5000 lux).
One-way ANOVAs were performed using PROC GLM in
SAS Statistical Software (SAS Institute, 1996) on the HSB
and DGCI data sets, with cultivar as the treatment variable.
For a given color parameter, differences were determined
significant among cultivars when the ANOVA f test had a
corresponding P value ⱕ 0.05. In such cases, a Fisher’s protected LSD test was performed to separate cultivar differences
(Freund and Wilson, 1993).
Digital image analysis differentiated HSB levels of
the Munsell Plant Tissue color chips chosen for this
study (Fig. 2, 3, and 4). Hue and saturation measurements obtained through digital image analysis were statistically equal to the actual hue and saturation values
as the slopes and intercepts of the hue and saturation
regression lines were not significantly different (P ⬍
0.05) from 1 and 0, respectively. Brightness measurements were slightly less accurate, but could be effectively corrected (r2 ⫽ 0.96) by the following equation:
actual brightness ⫽ 0.60 (measured brightness) ⫹ 0.37.
Nitrogen Fertility Color Differences
Differences in turfgrass color resulting from various
N fertility treatments were quantified with digital image
analysis (Tables 1, 2). Although there were no differences among treatments with regard to saturation and
brightness levels in the zoysiagrass study, hue and DGCI
values were significantly affected by N source and N
rate treatments. In the creeping bentgrass study, HSB
and DGCI values were all significantly affected by N rate.
In both studies, similar treatment rankings were ob-
KARCHER & RICHARDSON: DIGITAL IMAGE ANALYSIS OF TURF COLOR
947
Fig. 2. Linear regression analysis between hue quantified by digital image analysis and the actual hue of Munsell plant tissue color chips.
tained by digital image analysis and subjective ratings
(Tables 1 and 2). The 100% PCU treatment had significantly lower DGCI and visual rating means than all
other treatments (with the exception the 67% PCU
mean for DGCI). In addition, there were significant
differences among all three N rate treatment means
(9.6 g m⫺2 ⬎ 7.2 g m⫺2 ⬎ 4.8 g m⫺2) with regard to
DGCI and visual ratings.
In both studies, the coefficients of variation for HSB
and DGCI ranged from 2 to 18 times less than that of
visual ratings (Tables 3, 4). All coefficients of variation
for the digital image analysis parameters were statistically smaller than the CV% for the visual ratings based
on the 95% confidence intervals.
Cultivar Color Differences
There were significant differences among bermudagrass cultivars with regard to hue, saturation, and DGCI
Fig. 3. Linear regression analysis between color saturation quantified by digital image analysis and the actual color saturation of Munsell plant
tissue color chips.
948
CROP SCIENCE, VOL. 43, MAY–JUNE 2003
Fig. 4. Linear regression analysis between color brightness quantified by digital image analysis and the actual color brightness of Munsell plant
tissue color chips.
(Table 5). Cultivar hue ranged from 71⬚ to 92⬚, while
saturation and DGCI levels ranged between 29 to 42%
and 0.39 to 0.55, respectively. ‘Cardinal’, with an average
hue of 76.2⬚, was ≈10⬚ (and significantly) lighter in hue
than the other three cultivars. This result was consistent
with ‘Cardinal’ appearing a lighter shade of green to
the eye than the other three cultivars. ‘Cardinal’ also
ranked lowest in genetic color among 28 cultivars in the
1997 NTEP trials when results were averaged across
18 trial locations (NTEP, 1999). ‘Shanghai’, which appeared darker to the eye than the other cultivars, had
a significantly lower saturation level than the other cultivars (Table 5). The dark color of this cultivar was apparently due to its grayish green color (less saturation),
rather than it being a darker shade of green (higher hue).
The ‘Cardinal’ DGCI mean ranked significantly (P ⬍
0.05) lower than the other three cultivars, which were
statistically equal. In addition, the increased DGCI for
Table 1. Color analyses by subjective visual ratings and digital image analysis of zoysiagrass turf fertilized with various N sources and
rates, 28 Sept. 2000 (44 d after treatment).
Visual rating†
Hue‡
Saturation§
Degrees
N source
Ammonium sulfate
Polymer-coated urea
Sulfur-coated urea
1/3 AS ⫺ 2/3 PCU
1/3 AS ⫺ 2/3 SCU
2/3 AS ⫺ 1/3 PCU
2/3 AS ⫺ 1/3 SCU
N rate, g m⫺2
4.8
7.2
9.6
ANOVA
Source (df)
N source (6)
N rate (2)
N source ⫻ N rate (14)
Error (60)
CV%
Brightness¶
DGCI#
%
6.5ab††
5.2c
6.7a
6.1b
6.7a
6.3ab
6.5ab
86.5ab
82.6c
86.3ab
82.7c
84.8b
85.5ab
86.7a
43.4a
43.6a
42.8a
43.5a
44.2a
44.5a
43.9a
58.3a
59.5a
58.6a
58.5a
58.6a
58.2a
58.7a
0.475a
0.449d
0.474a
0.453cd
0.462bc
0.466ab
0.473ab
5.3c
6.6b
6.9a
83.6c
84.9b
86.6a
44.0a
43.8a
43.2a
59.0a
58.5a
58.4a
0.454c
0.464b
0.476a
mean squares
0.04
0.05
0.06
0.06
5.6
0.02
0.03
0.03
0.03
7.4
15.89***
99.41***
1.12
1.92
22.1
36.03***
63.83***
1.59
5.10
2.7
0.811***
0.655***
0.119
0.226
4.2
*** Significant at the 0.001 level of probability.
† 1 ⫽ tan/brown turf, 6 ⫽ minimum acceptable color, 9 ⫽ optimal dark green color.
‡ 0ⴗ ⫽ red, 60ⴗ ⫽ yellow, 120ⴗ ⫽ green, 180ⴗ ⫽ cyan, 240ⴗ ⫽ blue, and 300ⴗ ⫽ magenta.
§ 0% ⫽ gray and 100% ⫽ fully saturated color.
¶ 0% ⫽ black and 100% ⫽ white.
# Dark green color index. A combination of HSB parameters for a single measurement of dark green color: DGCI ⫽ [(Hue ⫺ 60)/60 ⫹ (1 ⫺ Saturation) ⫹
(1 ⫺ Brightness)]/3.
†† Within each effect and column, means sharing a letter are not statistically different according to Fisher’s protected LSD test (␣ ⫽ 0.05).
949
KARCHER & RICHARDSON: DIGITAL IMAGE ANALYSIS OF TURF COLOR
Table 2. Color analyses by subjective visual ratings and digital image analysis of creeping bentgrass turf fertilized with various N rates,
16 Nov. 2001 (55 d after treatment).
N rate
g
Visual rating†
Hue‡
4.2d
5.2c
5.9c
6.8b
7.0ab
7.1ab
7.6a
Degrees
64.6e
70.3d
74.1c
77.5b
79.4ab
80.6a
80.9a
67.4a
67.2a
66.4a
65.9ab
64.7bc
64.5bc
63.3c
17.81***
0.90
15.3
447.7***
9.55
4.1
mean squares
0.27***
0.04
0.3
m⫺2
0.0
1.0
2.0
3.0
4.0
5.0
6.0
ANOVA
Source (df)
N rate (6)
Error (21)
CV%
Saturation§
Brightness¶
DGCI#
%
29.3a
28.3a
26.8b
24.8c
24.1cd
22.8de
22.3e
0.370e
0.405d
0.435c
0.462b
0.479ab
0.490a
0.497a
0.89***
0.03
0.6
0.027***
0.0005
5.1
*** Significant at the 0.001 level of probability.
† 1 ⫽ tan/brown turf, 6 ⫽ minimum acceptable color, 9 ⫽ optimal dark green color.
‡ 0ⴗ ⫽ red, 60ⴗ ⫽ yellow, 120ⴗ ⫽ green, 180ⴗ ⫽ cyan, 240ⴗ ⫽ blue, and 300ⴗ ⫽ magenta.
§ 0% ⫽ gray and 100% ⫽ fully saturated color.
¶ 0% ⫽ black and 100% ⫽ white.
# Dark green color index. A combination of HSB parameters for a single measurement of dark green color: DGCI ⫽ [(Hue ⫺ 60)/60 ⫹ (1 ⫺ Saturation)
⫹ (1 ⫺ Brightness)]/3.
†† Within each effect and column, means sharing a letter are not statistically different according to Fisher’s protected LSD test (␣ ⫽ 0.05).
‘Shanghai’ compared with ‘Tifway’ and ‘Mini-Verde’
was nearly significant (P ⫽ 0.07). These differences in
color are in strong agreement with results from the 1997
NTEP trials where all four cultivars were significantly
different: ‘Shanghai’ ⬎ ‘Tifway’ ⬎ ‘Mini-Verde’ ⬎
‘Cardinal’ (NTEP, 1999). Although ‘Tifway’ and ‘MiniVerde’ were not significantly different in DGCI using
digital image analysis, they only differed by 0.3 rating
units in the 1997 NTEP trial (LSD0.05 ⫽ 0.2).
DISCUSSION
Digital photography and image analysis were able
to quantify color differences among standard Munsell
Plant Tissue color chips, zoysiagrass and creeping bentgrass receiving various N fertility treatments, and bermudagrass cultivars of varying genetic color. When visual ratings and digital image analysis were both
performed, the statistical ranking of treatment means
were similar between the two methods. However, DGCI
variance was significantly lower than rater variance
when the same turf plots were evaluated multiple times,
probably the result of removing either rater bias or rater
error from the color evaluation process.
These results confirm that visual ratings can be used
to separate treatment effects on turf color. In most cases,
raters ranked the turf plots similarly although differences existed in their absolute rating values. Therefore,
color ratings remain a valid evaluation tool if data are
not compared across raters. However, the accuracy of
digital image analysis, demonstrated in the calibration
experiments, enables researchers to record reflected turfgrass color on a standardized scale rather than using
arbitrary rating values. Therefore, valid comparisons of
color data across researchers, locations, and years are
possible with digital image analysis.
Creeping bentgrass plots had significant differences in
HSB levels, whereas zoysiagrass plots were significantly
different only with regard to hue. This may be due to
a genetic difference in N uptake and utilization between
the two species. However, in both species, significant
DGCI differences existed due to N treatments. Therefore, the DGCI is a more consistent measure of dark
Table 3. Comparison of variance between subjective raters and digital image analysis for color evaluation of zoysiagrass turf, 28 Sept.
2000 (44 d after treatment).
Visual ratings†
Sampling information
Subsampling units
Experimental units
n
df
Statistics
x̄
95% confidence interval for ␮
s
95% confidence interval for ␴
CV%
CV% confidence bounds††
5
84
420
336
6.27
6.12–6.42
1.38
1.29–1.50
22.1
20.0–24.5
Hue‡
3
21
63
42
Degrees
83.76
83.31–84.20
1.52
1.25–1.93
1.8
1.4–2.3
Saturation§
Brightness¶
3
21
63
42
3
21
63
42
DGCI#
3
21
63
42
%
44.50
43.92–45.08
0.039
0.026–0.063
4.4
3.6–5.7
58.11
57.80–58.41
0.011
0.007–0.017
1.8
1.4–2.3
0.457
0.453–0.460
0.0001
0.0001–0.0002
2.6
2.1–3.4
† 1 ⫽ tan/brown turf, 6 ⫽ minimum acceptable color, 9 ⫽ optimal dark green color.
‡ 0ⴗ ⫽ red, 60ⴗ ⫽ yellow, 120ⴗ ⫽ green, 180ⴗ ⫽ cyan, 240ⴗ ⫽ blue, and 300ⴗ ⫽ magenta.
§ 0% ⫽ gray and 100% ⫽ fully saturated color.
¶ 0% ⫽ black and 100% ⫽ white.
# Dark green color index. A combination of HSB parameters for a single measurement of dark green color: DGCI ⫽ [(Hue ⫺ 60)/60 ⫹ (1 ⫺ Saturation) ⫹
(1 ⫺ Brightness)]/3.
†† CV% confidence bounds calculated as (lower ␴ bound/upper ␮ bound, upper ␴ bound/lower ␮ bound).
950
CROP SCIENCE, VOL. 43, MAY–JUNE 2003
Table 4. Comparison of variance between subjective raters and digital image analysis for color evaluation of creeping bentgrass turf,
16 Nov. 2001 (55 d after treatment).
Visual ratings†
Sampling information
Subsampling units
Experimental units
n
df
Statistics
x̄
95% confidence interval for ␮
s
95% confidence interval for ␴
CV%
CV% confidence bounds††
Hue‡
3
28
84
56
3
28
84
56
Degrees
75.34
75.17–75.52
0.70
0.59–0.86
0.9
0.7–1.1
6.23
6.05–6.43
0.75
0.63–0.92
12.0
9.8–15.2
Saturation§
Brightness¶
3
28
84
56
3
28
84
56
DGCI#
3
28
84
56
%
65.62
65.42–65.82
0.008
0.007–0.009
1.2
1.0–1.5
25.51
25.26–25.75
0.010
0.008–0.012
3.8
3.2–4.7
0.448
0.447–0.449
0.003
0.003–0.004
0.7
0.6–0.9
† 1 ⫽ tan/brown turf, 6 ⫽ minimum acceptable color, 9 ⫽ optimal dark green color.
‡ 0ⴗ ⫽ red, 60ⴗ ⫽ yellow, 120ⴗ ⫽ green, 180ⴗ ⫽ cyan, 240ⴗ ⫽ blue, and 300ⴗ ⫽ magenta.
§ 0% ⫽ gray and 100% ⫽ fully saturated color.
¶ 0% ⫽ black and 100% ⫽ white.
# Dark green color index. A combination of HSB parameters for a single measurement of dark green color: DGCI ⫽ [(Hue ⫺ 60)/60 ⫹ (1 ⫺ Saturation) ⫹
(1 ⫺ Brightness)]/3.
†† CV confidence bounds calculated as (lower ␴ bound/upper ␮ bound, upper ␴ bound/lower ␮ bound).
green color across species than the individual measurements of H, S, or B. Since N fertility significantly affected the HSB levels of creeping bentgrass and zoysiagrass (Fig. 5A), color measurement using digital
image analysis may be capable of assessing the N status
of plant tissues. For example, zoysiagrass plots exhibiting the darkest green N responses had hue angles near
90⬚ while the most chlorotic plots had hue angles near
70⬚. Other research has demonstrated that correlations
exist between the N content of creeping bentgrass tissue
and its color, measured by colorimeter (Landschoot and
Mancino, 2000).
The significantly larger CV% with visual ratings suggest that rating values are evaluator dependent and that
evaluators are likely to vary in how they rank different
shades of green (Skogley and Sawyer, 1992; Horst et
al., 1984). This may be a factor in multisite trials when
an individual cultivar is ranked inconsistently from location to location (NTEP, 1999). Color evaluation with
digital photography and image analysis may minimize
variations due to locations and years and would increase
the validity of comparing color data across both.
Table 5. Color evaluation among bermudagrass cultivars using
digital image analysis.
Cultivar
Hue†
‘Cardinal’
‘Mini-Verde’
‘Shanghai’
‘Tifway’
ANOVA
Source (df)
Cultivar (3)
Error (8)
CV%
76.2b#
88.1a
89.9a
86.6a
Saturation‡
Brightness§
DGCI¶
%
113.97**
12.01
4.07
40.3a
38.0a
30.0c
34.3b
61.0a
58.4a
58.4a
59.4a
mean squares
0.611***
0.045
0.016
0.025
3.62
2.68
0.419b
0.502a
0.538a
0.502a
0.0077***
0.00043
4.26
** Significant at the 0.01 level of probability.
*** Significant at the 0.001 level of probability.
† 0ⴗ ⫽ red, 60ⴗ ⫽ yellow, 120ⴗ ⫽ green, 180ⴗ ⫽ cyan, 240ⴗ ⫽ blue, and
300ⴗ ⫽ magenta.
‡ 0% ⫽ gray and 100% ⫽ fully saturated color.
§ 0% ⫽ black and 100% ⫽ white.
¶ Dark green color index. A combination of HSB parameters for a single
measurement of dark green color: DGCI ⫽ [(Hue ⫺ 60)/60 ⫹ (1 ⫺
Saturation) ⫹ (1 ⫺ Brightness)]/3.
# Within each column, means sharing a letter are not statistically different
according to Fisher’s protected LSD test (␣ ⫽ 0.05).
The ability to distinguish color differences among turf
plots as either H, S, or B differences is a significant
advantage of digital image analysis over subjective visual ratings. For example, a turf that has a darker color
because of grayish genetic color may not be as aesthetically desirable as a turf that is lighter in appearance but
is saturated with green color. Consequently, there exists
a potential for evaluator bias, which may have occurred
in the 1997 vegetative bermudagrass NTEP trails where
the dark grayish variety ‘Shanghai’ ranked among the
top three cultivars in genetic color in 13 of the 18 test
sites, while it ranked near the middle or bottom at the
other five sites (NTEP, 1999). Rather than ‘Shanghai’
exhibiting different genetic color at the various NTEP
locations, this discrepancy may have been due to varying
evaluator perceptions of optimal dark green color for
bermudagrass.
Digital image analysis was more time consuming than
visual color ratings, but far less labor intensive than
traditional laboratory methods that are used to quantify
turf color (amino acid and chlorophyll assays). Images
were collected in the field at a rate of ≈2 images per
minute and were analyzed with SigmaScan at a rate of
3 images per minute. Although subjective ratings require less time than digital image analysis, the color
data obtained from digital image analysis are free from
researcher bias and inaccuracies and include information on individual HSB parameters. Furthermore, SigmaScan macros have been developed for batch-analysis of
an unlimited number of images (Karcher, 2001, unpublished data).
Another advantage of digital image analysis over
other objective color evaluation methods is the ability
to measure large areas of turf in situ. The area of turf
that is possible to evaluate is limited only by the height
of the camera above the canopy and the subsequent
field of vision. An el-shaped monopod was designed at
the University of Arkansas that enables images to be
taken of turf areas in excess of 30 m2 (a remote control
releases the camera shutter). This is a significant improvement over standard colorimeters that typically
measure areas smaller than 20 cm2 (less area than a
KARCHER & RICHARDSON: DIGITAL IMAGE ANALYSIS OF TURF COLOR
35-mm slide). In addition, if a turf plot is not uniformly
green due to disease, injury, or dormancy, a color threshold technique can be used within SigmaScan to quantify
the color of only the green portions of an image which
correspond to healthy turf (Richardson et al., 2001).
Another advantage of digital image analysis is that once
images are obtained, they can be stored indefinitely
before analysis. For instance, images of field trials can
be collected regularly during the growing season, but
analyzed during the off-peak months. In contrast to
visual color ratings, trained, experienced researchers are
not needed to evaluate turf color using digital image
analysis.
Light conditions may affect the results from these
techniques, although successful results were obtained
under both sunny and overcast conditions in these studies. Digital color analysis may not be as effective during
dawn or dusk due to increased shadows within the turf
canopy. In addition, comparisons of color among turf
plots from different locations and times may only be
possible if the images are collected under equal light
conditions. This could be accomplished through the use
of standard artificial light sources while collecting images either at night or in an enclosed system.
A digital camera capable of acquiring high quality
images is becoming commonplace in turfgrass research
programs. The ability to capture extensive information
of turfgrass in situ makes it a viable tool to quantify
turfgrass parameters commonly of interest in field experiments. In addition to color quantification, digital
image analysis has been used successfully to quantify
percentage turfgrass cover (Richardson et al., 2001) and
may potentially be useful in quantifying turf parameters
such as weed infestation, disease incidence, herbicide
phytotoxicity, leaf area, and recovery from injury.
ACKNOWLEDGMENTS
The authors thank the O.J. Noer Foundation for the financial support of this research and SPSS, Inc., for the assistance
in the form of a copy of SigmaScan Pro software. Also, the
authors are grateful for the technical assistance of Yoshi Ikemura, John McCalla, Margaret Secks, and Chris Weight.
951
REFERENCES
Adamsen, F.J., P.J. Pinter, Jr., E.M. Barnes, R.L. LaMorte, G.W. Wall,
S.W. Leavitt, and B.A. Kimball. 1999. Measuring wheat senescence
with a digital camera. Crop Sci. 39:719–724.
Adobe Systems. 2002. Adobe Photoshop v. 7.0. Adobe Systems, San
Jose, CA.
Beard, J.B. 1973. Turfgrass: Science and culture. Prentice-Hall, Englewood Cliffs, NJ.
Birth, G.S., and G.R. McVey. 1968. Measuring the color of growing
turf with a reflectance spectrophotometer. Agron. J. 60:640–643.
Freund, R.J., and W.J. Wilson. 1993. Statistical methods. Academic
Press, San Diego, CA.
Horst, G.L., M.C. Engelke, and W. Meyers. 1984. Assessment of visual
evaluation techniques. Agron. J. 76:619–622.
Johnson, G.V. 1973. Simple procedure for quantitative analysis of
turfgrass color. Agron. J. 66:457–459.
Karcher, D.E. 2000. Investigations on statistical analysis of turfgrass
rating data, localized dry spots of greens, and nitrogen application
techniques for turf. Ph.D. diss. Michigan State Univ., East Lansing, MI.
Kimura, T., A. Misawa, and T. Ochiai. 1989. Measuring seasonal
changes in the leaf color of cool season turfgrass using a chroma
meter.
Landschoot, P.J., and C.F. Mancino. 2000. A comparison of visual vs.
instrumental measurement of color differences in bentgrass turf.
HortScience 35:914–916.
Lukina, E.V., M.L. Stone, and W.R. Raun. 1999. Estimating vegetation
coverage in wheat using digital images. J. Plant Nutr. 22:341–350.
Microsoft Corporation. 1999. MS Excel 2000. Microsoft Corp., Redmond, WA.
Munsell Color. 2000. Conversion program overview [Online]. [1 p.]
Available at http://www.munsell.com/Color%20Conversion.htm
[cited 26 March 2002; verified 1 Dec. 2002]. GretagMacbeth, New
Windsor, NY.
Nelson, S.H., and F.W. Sosulski. 1984. Amino acid and protein content
of Poa pratensis as related to nitrogen application and color. Can.
J. Plant Sci. 64:691–697.
National Turfgrass Evaluation Program. 1999. National Bermudagrass
Test—1997. NTEP Progress Rep. No. 00-4. USDA-ARS, Beltsville, MD.
Purcell, L.C. 2000. Soybean canopy coverage and light interception
measurements using digital imagery. Crop Sci. 40:834–837.
Richardson, M.D., D.E. Karcher, and L.C. Purcell. 2001. Quantifying
turfgrass cover using digital image analysis. Crop Sci. 41:1884–1888.
SAS Institute. 1996. The SAS system for Windows. Release 6.12. SAS
Inst., Cary, NC.
Skogley, C.R., and C.D. Sawyer. 1992. Field research. p. 589–614. In
D.V. Waddington et al. (ed.) Turfgrass. Agron. Monogr. 32. ASA,
CSSA, and SSSA, Madison, WI.
SPSS. 1998. Sigma Scan Pro 5.0. SPSS Science Marketing Dep.,
Chicago.
Thorogood, D., P.J. Bowling, and R.M. Jones. 1993. Assessment of
turf colour change in Lolium perenne L. cultivars and lines. Int.
Turfgrass Soc. Res. J. 7:729–735.
United States Golf Association. 1993. USGA recommendations for
putting green construction. USGA Green Section Record 31(2):1–3.
Download