Turf color is a key component of aesthetic quality

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 Color is a major component of the aesthetic quality of turf and data tend to be discrete and ordinal rather than continuoften evaluated in field studies. Digital image analysis may be an ous. Timely quantification of turfgrass color that uses improved, objective method to quantify turf color. Studies were conducted readily accessible equipment would strengthen the va- to determine if digital image analysis with SigmaScan software lidity of study results without adding significant burden (SPSS, Chicago, IL) was capable of: (i) accurately determining the to the evaluation process. hue, saturation, and brightness (HSB) levels of Munsell Plant Tissue Several techniques have been used to objectively color chips, (ii) quantifying visual color differences among zoysiagrass measure turf color, including reflectance measurements (Zoysia japonica Steud.) and creeping bentgrass {Agrostis palustris (Birth and McVey, 1968), chlorophyll and amino acid Huds. [ A. stolonifera var. palustris (Huds.) Farw.]} plots receiving analysis (Johnson, 1973; Nelson and Sosulski, 1984), and various N treatments, and (iii) quantifying genetic color differences among bermudagrass (Cynodon spp.) cultivars. Digital images of turf comparison with standardized colors (Beard, 1973). All plots were analyzed with SigmaScan software to determine average of these methods have certain disadvantages compared HSB levels for each image. A dark green color index (DGCI) was with subjective color ratings. Reflectance, chlorophyll, created from HSB values for direct comparison with visual ratings. and amino acid measurements all require relatively ex- Digital image analysis accurately quantified the HSB levels (r 2 0.99, pensive equipment, and transport of samples to a labo- 0.96, and 0.97, respectively) of Munsell color chips corresponding ratory for analysis. In addition, correlations between to turf colors. Significant HSB differences were present among N color and chlorophyll or amino acid measurements are treatments in creeping bentgrass, while only significant hue differences either species or cultivar dependent. The use of stanexisted in zoysiagrass. Significant hue and saturation differences were dardized charts to measure turf color is effective, but present among bermudagrass cultivars. There was strong agreement results in qualitative descriptions of color that are not between DGCI values and visual ratings. The relative variances of the HSB and DGCI were significantly less than the variance associated possible to statistically analyze with traditional AN- with multiple raters. This evaluation technique may facilitate objective OVA techniques. comparisons of turf color across researchers, locations, and years when Recently, Landschoot and Mancino (2000) demon- images are collected under equal lighting conditions (i.e., the use of strated that the color of creeping bentgrass cultivars an artificial light source at night or in an enclosed system). 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 colorim- Turf color is a key component of aesthetic quality and a good indicator of water and nutrient status eters to evaluate varying turf color due to seasonal (Beard, 1973). Therefore, color is often evaluated in changes (Kimura et al., 1989) or differences among cultiturfgrass experiments. Color is traditionally evaluated vars and genetic lines (Thorogood et al., 1993). Alby visually rating turf plots on a scale of 1 to 9, with 1 though promising, a potential shortcoming of the color- representing yellow or brown turf and 9 representing imeter used in those studies is the relatively small optimal, dark green turf. Although color ratings provide measurement area ( 20 cm 2 ). In the absence of exquick data acquisition without the need for specialized tremely uniform surface conditions, numerous subsamequipment, they are a subjective measure from which ple measurements with the colorimeter would be neces- human bias is difficult to remove. As a result, inconsisfield sary to accurately represent the color of typical turfgrass tencies often exist among raters when evaluating the plots. same turf plots. Relatively poor correlations existed In recent years, digital photography has become a among experienced researchers (r 0.68) when rating common and affordable means for the scientific commu- the same turf plots for density, color, and leaf spot nity to document and present images. Digital cameras, (Skogley and Sawyer, 1992; Horst et al., 1984). Correlaused in conjunction with image analysis software, are being tions this low would probably be considered unacceptcence to quantify wheat (Triticum aestivum L.) senes- able when using other evaluation tools (e.g., balances, (Adamsen et al., 1999) and canopy coverage in spectrometers, ph meters) to measure the same turf wheat (Lukina et al., 1999) and soybeans [Glycine max sample. Furthermore, the applicability of standard ANysis L. (Merr.)] (Purcell, 2000). Recently, digital image anal- OVA procedures and traditional means separation tests was used to quantify turf coverage with increased precision over more traditional evaluation methods Dep. of Horticulture, Univ. of Arkansas, 308 Plant Sci. Building, Fayetteville, AR 72701. Received 4 April 2002. *Corresponding author (karcher@uark.edu). Published in Crop Sci. 43:943 951 (2003). 943 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.

944 CROP SCIENCE, VOL. 43, MAY JUNE 2003 (Richardson et al., 2001). Through digital photography, The average RGB levels of the digital images were calcuresearchers can instantaneously obtain millions of bits lated using SigmaScan Pro version 5.0 software (SPSS, 1998). of information on a relatively large turfgrass canopy. The entire image was selected for analysis by including all For example, a digital image taken of a turf plot using possible hue and saturation levels in the color threshold option of the software. The average red, average green, and average a 1280 960 pixel resolution contains 1 228 800 pixels, blue measurement settings were used to obtain the average with each pixel containing independent color informa- RGB levels for an image. The average RGB levels were then tion about the turf plot. Therefore, digital photography pasted into an MS Excel spreadsheet (Microsoft Corporation, and subsequent image analysis may be capable of quan- 1999) created by the authors to automate the conversion of tifying turfgrass color in field experiments. RGB to HSB values. The programmed formulas in the spreadsheet The information contained in a digital image includes converted absolute RGB levels (measured on a scale of the amount of red, green, and blue (RGB) light emitted 0 to 255) to percentage RGB levels by dividing each level by for each pixel in the image. Although it may be intuitive 255. Percentage RGB levels were then converted to average to use the green levels of the RGB information to quan- HSB levels by the following algorithms (Adobe Systems, tify the green color of an image, the intensity of red and 2002): blue will confound how green an image appears. To Hue ease the interpretation of digital color data, RGB values can be converted directly to HSB values that are based If max(r,g,b) R, 60{(G B)/[max(R,G,B) on human perception of color (Fig. 1). In HSB color min(r,g,b)]} description, hue is defined as an angle on a continuous circular scale from 0 to 360 (0 red, 60 yellow, If max(r,g,b) G, 60(2 {(B R)/[max(R,G,B) 120 green, 180 cyan, 240 blue, 300 magenta), saturation is the purity of the color from 0% (gray) min(r,g,b)]}) to 100% (fully saturated color), and brightness is the If max(r,g,b) B, 60(4 {(R G)/[max(R,G,B) relative lightness or darkness of the color from 0% (black) to 100% (white) (Adobe Systems, 2002). Among min(r,g,b)]}) HSB, hue has been found to be the best indicator of Saturation the visual color of a turf (Landschoot and Mancino, 2000; Thorogood et al., 1993). However, preliminary [max(r,g,b) min(r,g,b)]/max(r,g,b) work at the University of Arkansas has demonstrated Brightness that visual differences in turf color were sometimes the result of color saturation differences between turf plots max(r,g,b). rather than hue differences (Karcher, 2000, unpublished data). Camera Calibration The objective of the following research was to determine if readily available equipment (a digital camera A series of digital images were taken of color chips from and commercially available software) could accurately Munsell Color Charts for Plant Tissues (GretagMacbeth LLC, New Windsor, NY). Six images of varying hue were collected, quantify turfgrass color using an HSB color scale. Digital ranging from yellowish green to green (chip numbers 5Y 6/6, images were taken of standard color objects (Munsell 2.5GY 6/6, 5GY 6/6, 7.5GY 6/6, 2.5G 6/6, 5G 6/6). Eight images Plant Tissue color chips) to determine the accuracy of of varying saturation were collected, ranging from grayish digital image analysis with regard to the quantification green to bright green (chip numbers 7.5GY 6/2, 7.5GY 5/2, of color parameters. Digital images were collected of 7.5GY 6/4, 7.5GY 5/4, 7.5GY 6/6, 7.5GY 5/6, 7.5GY 6/8, 7.5GY turfgrass field plots varying in visual color due to either 5/8). Ten images of varying brightness were collected, ranging N fertility or genetically controlled differences to deter- from light green to dark green (chip numbers 7.5GY 8/4, mine if digital image analysis was capable of quantifying 7.5GY 7/4, 7.5GY 6/4, 7.5GY 5/4, 7.5GY 4/4, 7.5GY 8/6, 7.5GY color differences. 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 MATERIALS AND METHODS tissue HSB levels typical of turfgrass (Beard, 1973). Calibration images were taken under dark conditions using only the Color Quantification of Digital Images camera flash as a light source. The images were analyzed for The process used to determine the average color of a digital HSB levels using the methods described above. To determine image included: (i) acquiring an image with digital photogra- the accuracy of HSB measurement with digital image analysis, phy, (ii) obtaining the average RGB pixel levels for the image, the actual HSB levels of the Munsell color chips were deter- and (iii) converting the RGB levels to the more intuitive HSB mined using Munsell Conversion software version 4.1 (Munsell parameters. All digital images in these studies were taken with Color, 2000). an Olympus C-3030 camera (Olympus America Inc., Melville, Three separate linear regression analyses were performed NY). The images were collected in the JPEG (joint photographic using PROC REG in SAS Statistical Software (SAS Institute., experts group,.jpg) format, with a color depth of 16.7 1996). The H, S, and B values from digital image analysis were million colors, and an image size of 1280 960 pixels ( 260 analyzed as the independent variables and the actual H, S, kilobytes per image). Camera settings included a shutter speed and B values of the Munsell color chips were the dependent of 1/400 s, an aperture of f/4.0, and a focal length of 32 mm. variables. For each HSB parameter, digital image analysis Images were downloaded to a personal computer for subsequent 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 360. Turfgrass hues are typically between 70 and 110. (B) For a specific turfgrass hue, here 90, 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 m 2 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 m 2 and mowed at a height of ance. Since the visual rating scale was unrelated to color values 0.4 cm. Both experimental areas were located at the University obtained from digital image analysis, the relative variances of Arkansas Research and Extension Center in Fayette- (coefficients of variation) were used for statistical comparison. ville, AR. Sample variances were calculated as the within-plot mean The zoysiagrass study consisted of two treatment factors, N square for each color quantification method. Confidence bounds source (7 levels) and N rate (3 levels). The N source treatment (95%) were constructed for the sample means and the within- levels included: (i) 100% ammonium sulfate (AS); (ii) 100% plot variances and were used to calculate confidence bounds polymer-coated urea (PCU); (iii) 100% sulfur-coated urea for the coefficients of variation. The relative variances of the (SCU); (iv) 33% AS, 67% PCU; (v) 33% AS, 67% SCU; (vi) methods were determined to be significantly different if the 67% AS, 33% PCU; and (vii) 67% AS, 33% SCU. Each N respective confidence bounds for the coefficients of variation source was applied at three N rate levels: (i) 4.8, (ii) 7.2, and did not overlap. (iii) 9.6 g m 2. Each of the resultant 21 fertility treatments was replicated four times in a randomized complete block Cultivar Color Differences design. Treatment applications were made in mid-may and mid-august in 2000. Plots from a bermudagrass cultivar trial were used to assess The creeping bentgrass study consisted of one treatment the ability of digital image analysis to quantify visual color factor, N rate (7 levels). The N rate treatment levels included differences among cultivars. The trial was established in the 0,1,2,3,4,5,and6gm 2. The N source for all treatments summer of 1997 at the University of Arkansas Research and was methylene urea. Each N rate was applied four times in a Extension Center in Fayetteville, AR (silt loam, Typic Haplucompletely randomized design. Treatment applications were dults, ph 6.2), and was a test site for the 1997 National Turf- made monthly from June through September in 2000. grass Evaluation Program (NTEP) bermudagrass trial (NTEP, Digital images were collected from each plot on 28 Sept. 1999). Individual plots were 1.4 m 2 and maintained at a 1.9-cm 2000 on the zoysiagrass study [44 d after treatment (DAT)] mowing height. The study was replicated three times in a and on 16 Nov. 2001 on the creeping bentgrass study (55 DAT) completely randomized design. between 1300 and 1400 h during mostly sunny conditions (illureplication Digital Images were taken as described previously on each minance 50 000 lux). Images were collected by a researcher of four cultivars that varied in green color (NTEP, standing immediately next to the plot while holding the camera 1999): Cardinal (strong yellow-green), Shanghai (dark gray- directly over the center of the plot 1.5 m above the turf green), Mini-Verde (strong dark yellow-green), and Tifway canopy. Care was taken to avoid casting shadows on the turf (typical bermudagrass green color). The plots were photo- inside plot. Concurrent to the collection of digital images, the graphed on 21 Sept. 2000 between 1325 and 1335 h during zoysiagrass and creeping bentgrass studies were visually rated overcast conditions (illuminance 5000 lux). for color by five and three independent researchers (rater One-way ANOVAs were performed using PROC GLM in experience ranged from a minimum of 2 yr to 10 yr), respecand SAS Statistical Software (SAS Institute, 1996) on the HSB tively. Color ratings were based on a 1 to 9 scale where 1 DGCI data sets, with cultivar as the treatment variable. tan or brown turf, 6 minimum acceptable color, and 9 For a given color parameter, differences were determined optimal dark green color. A DGCI was created from the HSB significant among cultivars when the ANOVA f test had a values to obtain a single value from digital image color analysis corresponding P value 0.05. In such cases, a Fisher s pro- for comparison with values from subjective visual ratings. The tected LSD test was performed to separate cultivar differences index was created to measure the relative dark green color of (Freund and Wilson, 1993). an image using the following equation: DGCI value [(H 60)/60 (1 S) (1 B)]/3. RESULTS The color index was calculated from the average of transformed Camera Calibration HSB parameters. Each transformed parameter mea- Digital image analysis differentiated HSB levels of sures dark green color on a scale of zero to one. Since the the Munsell Plant Tissue color chips chosen for this hue of most turfgrass images ranges between 60 (yellow) and study (Fig. 2, 3, and 4). Hue and saturation measure- 120 (green), the maximum dark green hue was assigned as 120. Therefore, the dark green hue transform was calculated ments obtained through digital image analysis were staas (hue 60)/60, so that hues of 60 and 120 would yield tistically equal to the actual hue and saturation values dark green hue transforms of zero and one, respectively. Since as the slopes and intercepts of the hue and saturation lower saturation and brightness values corresponded to darker regression lines were not significantly different (P green colors, (1 saturation) and (1 brightness) were used 0.05) from 1 and 0, respectively. Brightness measureto calculate the dark green saturation and brightness trans- ments were slightly less accurate, but could be effecforms, respectively. The average of the transformed HSB val- tively corrected (r 2 0.96) by the following equation: ues yielded a single measure of dark green color, the DGCI actual brightness 0.60 (measured brightness) 0.37. value, which ranged from zero to one with higher values corresponding to darker green color. Analyses of variance were performed using PROC GLM Nitrogen Fertility Color Differences in SAS Statistical Software (SAS Institute, 1996) on the visual Differences in turfgrass color resulting from various rating, HSB, and DGCI data sets. For a given color parameter, N fertility treatments were quantified with digital image treatment and/or interaction effects were determined signifi- analysis (Tables 1, 2). Although there were no differcant when the corresponding ANOVA f test had a P value ences among treatments with regard to saturation and 0.05. In such cases, a Fisher s protected LSD test was perbrightness levels in the zoysiagrass study, hue and DGCI formed to separate treatment means (Freund and Wilson, 1993). values were significantly affected by N source and N Three digital images were taken on plots from the zoyand DGCI values were all significantly affected by N rate. rate treatments. In the creeping bentgrass study, HSB siagrass and creeping bentgrass studies to compare the variance of digital image analysis with subjective visual rater vari- 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 statisti- cally 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 18 trial locations (NTEP, 1999). Shanghai, which appeared saturation and DGCI levels ranged between 29 to 42% darker to the eye than the other cultivars, had and 0.39 to 0.55, respectively. Cardinal, with an average a significantly lower saturation level than the other cultivars hue of 76.2, was 10 (and significantly) lighter in hue (Table 5). The dark color of this cultivar was apparhue than the other three cultivars. This result was consistent ently due to its grayish green color (less saturation), with Cardinal appearing a lighter shade of green to rather than it being a darker shade of green (higher hue). the eye than the other three cultivars. Cardinal also The Cardinal DGCI mean ranked significantly (P ranked lowest in genetic color among 28 cultivars in the 0.05) lower than the other three cultivars, which were 1997 NTEP trials when results were averaged across 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 Brightness DGCI# Degrees % N source Ammonium sulfate 6.5ab 86.5ab 43.4a 58.3a 0.475a Polymer-coated urea 5.2c 82.6c 43.6a 59.5a 0.449d Sulfur-coated urea 6.7a 86.3ab 42.8a 58.6a 0.474a 1/3 AS 2/3 PCU 6.1b 82.7c 43.5a 58.5a 0.453cd 1/3 AS 2/3 SCU 6.7a 84.8b 44.2a 58.6a 0.462bc 2/3 AS 1/3 PCU 6.3ab 85.5ab 44.5a 58.2a 0.466ab 2/3 AS 1/3 SCU 6.5ab 86.7a 43.9a 58.7a 0.473ab N rate, g m 2 4.8 5.3c 83.6c 44.0a 59.0a 0.454c 7.2 6.6b 84.9b 43.8a 58.5a 0.464b 9.6 6.9a 86.6a 43.2a 58.4a 0.476a ANOVA Source (df) mean squares N source (6) 15.89*** 36.03*** 0.04 0.02 0.811*** N rate (2) 99.41*** 63.83*** 0.05 0.03 0.655*** N source N rate (14) 1.12 1.59 0.06 0.03 0.119 Error (60) 1.92 5.10 0.06 0.03 0.226 CV% 22.1 2.7 5.6 7.4 4.2 *** Significant at the 0.001 level of probability. 1 tan/brown turf, 6 minimum acceptable color, 9 optimal dark green color. 0 red, 60 yellow, 120 green, 180 cyan, 240 blue, and 300 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).

KARCHER & RICHARDSON: DIGITAL IMAGE ANALYSIS OF TURF COLOR 949 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 Visual rating Hue Saturation Brightness DGCI# gm 2 Degrees % 0.0 4.2d 64.6e 67.4a 29.3a 0.370e 1.0 5.2c 70.3d 67.2a 28.3a 0.405d 2.0 5.9c 74.1c 66.4a 26.8b 0.435c 3.0 6.8b 77.5b 65.9ab 24.8c 0.462b 4.0 7.0ab 79.4ab 64.7bc 24.1cd 0.479ab 5.0 7.1ab 80.6a 64.5bc 22.8de 0.490a 6.0 7.6a 80.9a 63.3c 22.3e 0.497a ANOVA Source (df) mean squares N rate (6) 17.81*** 447.7*** 0.27*** 0.89*** 0.027*** Error (21) 0.90 9.55 0.04 0.03 0.0005 CV% 15.3 4.1 0.3 0.6 5.1 *** Significant at the 0.001 level of probability. 1 tan/brown turf, 6 minimum acceptable color, 9 optimal dark green color. 0 red, 60 yellow, 120 green, 180 cyan, 240 blue, and 300 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 Mini- Verde 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 vi- sual 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. There- fore, 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 Hue Saturation Brightness DGCI# Sampling information Subsampling units 5 3 3 3 3 Experimental units 84 21 21 21 21 n 420 63 63 63 63 df 336 42 42 42 42 Statistics Degrees % x 6.27 83.76 44.50 58.11 0.457 95% confidence interval for 6.12 6.42 83.31 84.20 43.92 45.08 57.80 58.41 0.453 0.460 s 1.38 1.52 0.039 0.011 0.0001 95% confidence interval for 1.29 1.50 1.25 1.93 0.026 0.063 0.007 0.017 0.0001 0.0002 CV% 22.1 1.8 4.4 1.8 2.6 CV% confidence bounds 20.0 24.5 1.4 2.3 3.6 5.7 1.4 2.3 2.1 3.4 1 tan/brown turf, 6 minimum acceptable color, 9 optimal dark green color. 0 red, 60 yellow, 120 green, 180 cyan, 240 blue, and 300 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 Hue Saturation Brightness DGCI# Sampling information Subsampling units 3 3 3 3 3 Experimental units 28 28 28 28 28 n 84 84 84 84 84 df 56 56 56 56 56 Statistics Degrees % x 6.23 75.34 65.62 25.51 0.448 95% confidence interval for 6.05 6.43 75.17 75.52 65.42 65.82 25.26 25.75 0.447 0.449 s 0.75 0.70 0.008 0.010 0.003 95% confidence interval for 0.63 0.92 0.59 0.86 0.007 0.009 0.008 0.012 0.003 0.004 CV% 12.0 0.9 1.2 3.8 0.7 CV% confidence bounds 9.8 15.2 0.7 1.1 1.0 1.5 3.2 4.7 0.6 0.9 1 tan/brown turf, 6 minimum acceptable color, 9 optimal dark green color. 0 red, 60 yellow, 120 green, 180 cyan, 240 blue, and 300 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 measure- The ability to distinguish color differences among turf ments of H, S, or B. Since N fertility significantly af- plots as either H, S, or B differences is a significant fected the HSB levels of creeping bentgrass and zoy- advantage of digital image analysis over subjective visiagrass (Fig. 5A), color measurement using digital sual ratings. For example, a turf that has a darker color image analysis may be capable of assessing the N status because of grayish genetic color may not be as aesthetiof plant tissues. For example, zoysiagrass plots exhib- cally desirable as a turf that is lighter in appearance but iting the darkest green N responses had hue angles near is saturated with green color. Consequently, there exists 90 while the most chlorotic plots had hue angles near a potential for evaluator bias, which may have occurred 70. Other research has demonstrated that correlations in the 1997 vegetative bermudagrass NTEP trails where exist between the N content of creeping bentgrass tissue the dark grayish variety Shanghai ranked among the and its color, measured by colorimeter (Landschoot and top three cultivars in genetic color in 13 of the 18 test Mancino, 2000). sites, while it ranked near the middle or bottom at the The significantly larger CV% with visual ratings sug- other five sites (NTEP, 1999). Rather than Shanghai gest that rating values are evaluator dependent and that exhibiting different genetic color at the various NTEP evaluators are likely to vary in how they rank different locations, this discrepancy may have been due to varying shades of green (Skogley and Sawyer, 1992; Horst et evaluator perceptions of optimal dark green color for al., 1984). This may be a factor in multisite trials when bermudagrass. an individual cultivar is ranked inconsistently from loca- Digital image analysis was more time consuming than tion to location (NTEP, 1999). Color evaluation with visual color ratings, but far less labor intensive than digital photography and image analysis may minimize traditional laboratory methods that are used to quantify variations due to locations and years and would increase turf color (amino acid and chlorophyll assays). Images the validity of comparing color data across both. were collected in the field at a rate of 2 images per minute and were analyzed with SigmaScan at a rate of Table 5. Color evaluation among bermudagrass cultivars using digital image analysis. 3 images per minute. Although subjective ratings re- quire less time than digital image analysis, the color Cultivar Hue Saturation Brightness DGCI data obtained from digital image analysis are free from % researcher bias and inaccuracies and include informa- Cardinal 76.2b# 40.3a 61.0a 0.419b tion on individual HSB parameters. Furthermore, Sigma- Mini-Verde 88.1a 38.0a 58.4a 0.502a Shanghai 89.9a 30.0c 58.4a 0.538a Scan macros have been developed for batch-analysis of Tifway 86.6a 34.3b 59.4a 0.502a an unlimited number of images (Karcher, 2001, unpub- ANOVA lished data). Source (df) mean squares Cultivar (3) 113.97** 0.611*** 0.045 0.0077*** Another advantage of digital image analysis over Error (8) 12.01 0.016 0.025 0.00043 other objective color evaluation methods is the ability CV% 4.07 3.62 2.68 4.26 to measure large areas of turf in situ. The area of turf ** Significant at the 0.01 level of probability. that is possible to evaluate is limited only by the height *** Significant at the 0.001 level of probability. 0 red, 60 yellow, 120 green, 180 cyan, 240 blue, and of the camera above the canopy and the subsequent 300 magenta. field of vision. An el-shaped monopod was designed at 0% gray and 100% fully saturated color. the University of Arkansas that enables images to be 0% black and 100% white. Dark green color index. A combination of HSB parameters for a single taken of turf areas in excess of 30 m 2 (a remote control measurement of dark green color: DGCI [(Hue 60)/60 (1 releases the camera shutter). This is a significant im- Saturation) (1 Brightness)]/3. # Within each column, means sharing a letter are not statistically different provement over standard colorimeters that typically according to Fisher s protected LSD test ( 0.05). measure areas smaller than 20 cm 2 (less area than a

KARCHER & RICHARDSON: DIGITAL IMAGE ANALYSIS OF TURF COLOR 951 35-mm slide). In addition, if a turf plot is not uniformly REFERENCES green due to disease, injury, or dormancy, a color thresh- Adamsen, F.J., P.J. Pinter, Jr., E.M. Barnes, R.L. LaMorte, G.W. Wall, old technique can be used within SigmaScan to quantify S.W. Leavitt, and B.A. Kimball. 1999. Measuring wheat senescence with a digital camera. Crop Sci. 39:719 724. the color of only the green portions of an image which Adobe Systems. 2002. Adobe Photoshop v. 7.0. Adobe Systems, San correspond to healthy turf (Richardson et al., 2001). Jose, CA. Another advantage of digital image analysis is that once Beard, J.B. 1973. Turfgrass: Science and culture. Prentice-Hall, Englewood Cliffs, NJ. images are obtained, they can be stored indefinitely Birth, G.S., and G.R. McVey. 1968. Measuring the color of growing before analysis. For instance, images of field trials can turf with a reflectance spectrophotometer. Agron. J. 60:640 643. be collected regularly during the growing season, but Freund, R.J., and W.J. Wilson. 1993. Statistical methods. Academic Press, San Diego, CA. analyzed during the off-peak months. In contrast to Horst, G.L., M.C. Engelke, and W. Meyers. 1984. Assessment of visual visual color ratings, trained, experienced researchers are evaluation techniques. Agron. J. 76:619 622. not needed to evaluate turf color using digital image Johnson, G.V. 1973. Simple procedure for quantitative analysis of turfgrass color. Agron. J. 66:457 459. analysis. Karcher, D.E. 2000. Investigations on statistical analysis of turfgrass Light conditions may affect the results from these rating data, localized dry spots of greens, and nitrogen application techniques, although successful results were obtained techniques for turf. Ph.D. diss. Michigan State Univ., East Lan- sing, MI. under both sunny and overcast conditions in these stud- Kimura, T., A. Misawa, and T. Ochiai. 1989. Measuring seasonal ies. Digital color analysis may not be as effective during changes in the leaf color of cool season turfgrass using a chroma dawn or dusk due to increased shadows within the turf meter. Landschoot, P.J., and C.F. Mancino. 2000. A comparison of visual vs. canopy. In addition, comparisons of color among turf instrumental measurement of color differences in bentgrass turf. plots from different locations and times may only be HortScience 35:914 916. possible if the images are collected under equal light 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. conditions. This could be accomplished through the use Microsoft Corporation. 1999. MS Excel 2000. Microsoft Corp., Redof standard artificial light sources while collecting im- mond, WA. ages either at night or in an enclosed system. Munsell Color. 2000. Conversion program overview [Online]. [1 p.] Available at http://www.munsell.com/color%20conversion.htm A digital camera capable of acquiring high quality [cited 26 March 2002; verified 1 Dec. 2002]. GretagMacbeth, New images is becoming commonplace in turfgrass research Windsor, NY. programs. The ability to capture extensive information Nelson, S.H., and F.W. Sosulski. 1984. Amino acid and protein content of Poa pratensis as related to nitrogen application and color. Can. of turfgrass in situ makes it a viable tool to quantify J. Plant Sci. 64:691 697. turfgrass parameters commonly of interest in field ex- National Turfgrass Evaluation Program. 1999. National Bermudagrass periments. In addition to color quantification, digital Test 1997. NTEP Progress Rep. No. 00-4. USDA-ARS, Belts- ville, MD. image analysis has been used successfully to quantify Purcell, L.C. 2000. Soybean canopy coverage and light interception percentage turfgrass cover (Richardson et al., 2001) and measurements using digital imagery. Crop Sci. 40:834 837. may potentially be useful in quantifying turf parameters Richardson, M.D., D.E. Karcher, and L.C. Purcell. 2001. Quantifying turfgrass cover using digital image analysis. Crop Sci. 41:1884 1888. such as weed infestation, disease incidence, herbicide SAS Institute. 1996. The SAS system for Windows. Release 6.12. SAS phytotoxicity, leaf area, and recovery from injury. 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, ACKNOWLEDGMENTS CSSA, and SSSA, Madison, WI. SPSS. 1998. Sigma Scan Pro 5.0. SPSS Science Marketing Dep., The authors thank the O.J. Noer Foundation for the finan- Chicago. cial support of this research and SPSS, Inc., for the assistance Thorogood, D., P.J. Bowling, and R.M. Jones. 1993. Assessment of turf colour change in Lolium perenne L. cultivars and lines. Int. in the form of a copy of SigmaScan Pro software. Also, the Turfgrass Soc. Res. J. 7:729 735. authors are grateful for the technical assistance of Yoshi Ikem- United States Golf Association. 1993. USGA recommendations for ura, John McCalla, Margaret Secks, and Chris Weight. putting green construction. USGA Green Section Record 31(2):1 3.