IMAGE-PROCESSING SOLUTION TO COTTON COLOR MEASUREMENT PROBLEMS: PART II. INSTRUMENT TEST AND EVALUATION

IMAGE-PROCESSING SOLUTION TO COTTON COLOR MEASUREMENT PROBLEMS: PART II. INSTRUMENT TEST AND EVALUATION J. A. Thomasson, S. A. Shearer, D. L. Boykin ABSTRACT. An experimental cotton color/trash meter was developed previously for the purpose of improving cotton color measurement by removing trash-particle effects on color measurement with image processing. In this work, the experimental meter was tested extensively on a large number of cotton samples varying widely in color and trash content. Testing involved: (1) comparing the measurement accuracy of the experimental meter to that of conventional meters that use diffuse reflectance for color measurement, and (2) comparing the experimental and conventional meters ability to predict clean lint color from that of uncleaned lint. Results indicated that basic cotton color measurement was as accurate with the experimental meter as it was with conventional meters. Additionally, the experimental meter s color measurements on uncleaned lint correlated better with cleaned lint color than did those of the conventional meters in every case. With the Z (blue-band reflectance) color measurement, the superiority of the experimental meter s correlation between uncleaned and cleaned lint color was statistically significant. The reduction in root-mean-square error with the experimental meter was about 14% for the Y (broad-band green reflectance) measurement and about 22% for Z. These reductions in prediction error had statistical as well as practical significance. Keywords. Color, Color/trash meter, Cotton, Fiber quality, Foreign matter, Image processing. Bulk cotton fiber properties like color, trash content, fiber length, fiber strength, etc., dictate to a large degree the price of cotton that a textile mill will purchase. Thus, cotton bales must be evaluated, or classed, with regard to these properties. Color and trash content have the largest effect on price, so it is especially important that measurements of them be accurate. Furthermore, these properties are important during the ginning process, if one is attempting to optimize the number of cleaning machines used. The initial cotton mass to be processed in ginning is called seed cotton, as it includes fiber, seed, and extraneous matter (trash). After the fiber-seed separation portion of the ginning process, the remaining cotton mass to be processed is called lint, and it includes only fiber and extraneous matter. To optimize the number of cleaning machines in the ginning process, measurements of color and trash content must be made early in the ginning process (before and/or after fiber-seed separation), and the number of cleaning machines must be adjusted in order to Article was submitted for review in June 2004; approved for publication by the Power & Machinery Division of ASAE in December 2004. Mention of tradenames does not indicate an endorsement by USDA. The authors are J. Alex Thomasson, ASAE Member Engineer, Professor, Department of Biological and Agricultural Engineering, Texas A&M University, College Station, Texas; Scott A. Shearer, ASAE Member Engineer, Professor, Department of Biological and Agricultural Engineering, University of Kentucky, Lexington, Kentucky; and Deborah L. Boykin, Area Statistician, USDA-ARS Midsouth Area, Stoneville, Mississippi. Corresponding author: J. Alex Thomasson, Department of Biological and Agricultural Engineering, Texas A&M University, 2117 TAMU, College Station, TX 77843-2117; phone: 979-458-3598; fax: 979-862-3442; e-mail: thomasson@tamu.edu. optimize color and trash content with respect to fiber loss and damage concomitant with additional processing. In order to be used effectively, the measurement of color made early in the ginning process must be capable of being used as a predictor of final color. Modern instrument-based cotton classing utilizes colorfiltered silicon sensors to measure color in terms of diffuse reflectance, and a black-and-white video camera to measure trash content. Since trash particles are typically much darker than cotton fibers, they are distinguishable from them with image analysis, so trash content can be measured with that method. However, trash particles impart a darker overall color to a fiber sample, and the higher the trash content, the more significant the effect. The cause of this darker color, which is not actually attributable to the fiber, cannot be discerned with diffuse-reflectance-based color measurement, so changes in color are evident when trash particles are removed (Nickerson, 1947, 1950; Phillips, 1980, 1982; Anthony, 1988, 1990, 1994; Thomasson, 1993). In the first part of this work, as an alternative to diffuse-reflectance-based cotton color measurement, Thomasson et al. (2005) designed, built, and operationally tested a wholly image-based color and trash measurement system. The system was intended to reduce the effects of trash on cotton color measurement, while maintaining the traditional practice of cotton color measurement in terms of units of measurement. The system was tested extensively and operated as designed and without any software or hardware failures. The second part of this work, described in detail herein, dealt with the evaluation of the instrument in terms of its effectiveness in reducing the effects of trash content on cotton color measurement. Transactions of the ASAE Vol. 48(2): 439 454 2005 American Society of Agricultural Engineers ISSN 0001 2351 439

OBJECTIVES The goal here was to compare the instrument developed by Thomasson et al. (2005) with instruments used in classing offices and gin process control systems. To accomplish this, the first objective was to compare them statistically in terms of basic measurement error. The second objective was to make the same type of comparison in relation to the predictive relationship between cleaned lint color measurements and those of uncleaned lint and seed cotton. In other words, if one were to measure the color of a number of high-trash-content cotton samples with each instrument, how well could those measurements be used to predict the color of a cleaned lint sample? MATERIALS AND METHODS INSTRUMENT DESIGN AND OPERATION As described in detail in Thomasson et al. (2005), the instrument s illumination system consisted of quartz-tungsten-halogen lamps in aluminum elliptical reflectors. Its sensor was a panchromatic video camera that acquired images through optical color filters on a rotating wheel. The sensitivities of the color and trash measurements were rigorously considered to maximize the dynamic range over which each measurement was made. The system s camera was connected to a computer through a frame grabber. Software was written to control the filter wheel, image acquisition, color/trash computations, and data recording. Image processing was employed to differentiate trash particles from cotton in the images. Color was calculated from only the image portion judged by image analysis to be cotton fiber. At the beginning of experiments, the color/trash meter was switched on and allowed to warm up for approximately 2 h. After that, it was not turned off during the period of data collection. A 2 h warm-up period has been typical for commercial color/trash meters to allow light intensity to stabilize. In addition, as with commercial meters, the experimental meter was calibrated every 2 h during periods of data collection. COMPARISONS WITH STANDARD INSTRUMENTS Measurement Error Cotton bales in the government loan program must be classed according to quality at classing offices of the USDA-AMS Cotton Division. Samples of roughly 85 g (3 oz) are cut from each side of a 220 kg (480 lb) bale (Verhalen and Banks, 1989). These samples are taken as representative of the bale and are measured on high-volume instrument (HVI) equipment for the various quality characteristics. A group of 242 actual bale samples, varying widely in color and trash content, was collected from the classing office at Dumas, Arkansas. Three ZU (for Zellweger Uster, Knoxville, Tenn.; now Uster Technologies AG) color/trash meters, referred to herein as A, B, and C, were used to measure the samples at the Dumas classing office. These instruments actually include two meter heads (each including light sources, sensors, and sample window); one is used to compress the cotton sample onto the other. Thus, two sets of readings are taken simultaneously, on opposite sides of the sample. Additionally, two sliding trays for cotton subsamples are used, and both subsamples are placed between the two meters consecutively. Therefore, the overall color/trash meter actually makes four measurements on a cotton sample and reports an average for the sample. Making these measurements with three instruments was done in an effort to differentiate among within-sample, within-instrument, and between-instrument variance in color measurement with the ZU meters. Whereas a sample face (the portion of a cotton sample being viewed) is not transferred directly from one instrument to the next, measurement of a sample on two instruments was considered to be measurements of two subsamples of the sample, one subsample for each instrument. The data collected were as follows: Meters A and B were used to measure the samples (subsample 1 for meter A, subsample 2 for meter B) two times without extra subsampling, i.e., once placed in the tray, the samples stayed in the same position for two full measurements. These data provided information regarding within-instrument variance. (Recall that these meters actually have two color/trash heads, and two subsamples in the sliding tray are presented to the meter. So each meter records a measurement that is the average of four sample faces.) Another reading on different subsamples was taken with meter A (subsample 3). These data added information regarding within-instrument variance plus within-sample variance. One reading was taken with meter C (subsample 4). These data, along with the measurements from meters A and B above, provided information regarding between-instrument variance plus within-sample variance. The experimental color/trash meter was used to measure two subsamples (subsamples 5 and 6) of each sample once. These data provided information regarding within-instrument variance plus within-sample variance on the experimental color/trash meter. Later, one subsample (subsample 7) of 20 of the samples was measured twice on the experimental color/ trash meter without moving the sample face between measurements. These data provided information regarding within-instrument variance on the experimental color/trash meter. The resulting data set included color measurements (Y and Z, see Thomasson et al., 2005) from two instrument types (ZU and experimental), four instruments (A, B, C, and experimental), 242 samples, seven subsamples, and two readings of subsamples 1 and 2 on meters A and B and subsample 7 (20 of 242 samples) on the experimental meter (collection methodology is summarized in table 1). The model statement for both color measurements (Y and Z) was as follows: y ijklm = µ i + a j( i) + sk( ij) + wl ( ijk) + em( ijkl) (1) where i = mean for the ith instrument type (ZU and experimental) y = measured value of Y or Z j = instrument (A, B, or C) k = sample l = subsample m = multiple readings on same subsample 440 TRANSACTIONS OF THE ASAE

Table 1. Methodology for collection of data for measurement error analysis. Subsample Meter [a] Measured No. of Samples No. of Readings per Sample 1 A 242 2 2 B 242 2 3 A 242 1 4 C 242 1 5 Exp. 242 1 6 Exp. 242 1 7 Exp. 20 2 [a] Meters A, B, and C are the three ZU meters that were used in the study; Exp. represents the experimental meter. Each of the 78 cottons used in the study had three replicates, thus 242 samples. Each sample was subsampled six times to provide subsamples for each measurement on an instrument, and a random selection of 20 of the samples were subsampled a seventh time to provide subsamples for an additional measurement with the experimental meter. and a j (i) s k (ij) w l (ijk) iid N(0, I2 ) 2 I = between instruments variance component iid N(0, S2 ) 2 S = between samples variance component iid N(0, ws2 ) 2 ws = within samples variance component e m (ijkl) iid N(0, e2 ) e 2 = within subsamples variance component where iid is used to mean is distributed as, and N indicates an assumed normal distribution. The SAS procedure PROC MIXED (SAS, 1996) was employed to estimate variance components (of Y and Z) according to the following equation: V 2 2 2 2 ( ) σ + σ + σ + σ y i = e ws S I (2) where V(y i ) is the variance of the sample mean for each instrument type. Prediction Error The most important aspect of this research was to determine if an effort to remove the trash effect on color measurements would improve the predictability of clean lint color from that of seed cotton or uncleaned lint. The plan was to collect a group of seed cotton lots (9.1 kg [20 lb] or more), varying widely in color and trash content, that could be sampled and ginned to produce corresponding samples of seed cotton, lint having passed through zero lint cleaners, lint having passed through one lint cleaner, and lint having passed through two lint cleaners (considered to be clean lint). A group of 78 seed cotton lots was amassed for the purpose of having a broad range of cotton colors and trash content levels. The primary factor in cotton color variation is weather exposure (Palmer, 1924; Nickerson, 1962; Kohel and Lewis, 1984; Aspland and Williams, 1991). Thus, in selecting seed cottons for this experiment, care was taken in finding cottons that had a variety of exposure levels, so that effects related to color variation could be taken into account. The first 21 lots were produced from the following seven cotton varieties, all grown under standard conditions in Stoneville, Mississippi, and harvested with a spindle picker: Deltapine 5409, Deltapine 5415, Delta Experiment Station 119, Deltapine 50, Suregrow 125, Stoneville 474, and Stoneville 453. Lots 1 through 7 included each of these varieties as harvested. Lots 8 through 14 included each of the varieties after the seed cotton had been placed in trays and left in the open to weather naturally for two weeks, during which several significant rains occurred. Lots 15 through 21 included each of the varieties after the seed cotton had been placed in trays, left in the open for the same two weeks, and also had been sprayed down with water almost every day during the two weeks. Lot 22 was mixed seed cotton, grown in Stoneville and picker-harvested, which had been in the end of a cotton storage trailer that was rained on numerous times over about a three-month period. Lots 23 through 28 were of the following six varieties, grown under standard conditions near Table 2. Cotton treatments in prediction error study. Seed cotton samples represented several varieties, production locations, harvest dates, and weathering durations. A broad variety of samples was used to indicate broad applicability with regard to color and trash levels. Treatment Variety Location Harvest Date Harvest Replication Weather or Harvest Regime 1 DP5409 Stoneville Normal 1 None 2 DP5415 Stoneville Normal 1 None 3 DES119 Stoneville Normal 1 None 4 DP50 Stoneville Normal 1 None 5 SG125 Stoneville Normal 1 None 6 ST474 Stoneville Normal 1 None 7 ST453 Stoneville Normal 1 None 8 DP5409 Stoneville Normal 1 Natural, 2 weeks 9 DP5415 Stoneville Normal 1 Natural, 2 weeks 10 DES119 Stoneville Normal 1 Natural, 2 weeks 11 DP50 Stoneville Normal 1 Natural, 2 weeks 12 SG125 Stoneville Normal 1 Natural, 2 weeks 13 ST474 Stoneville Normal 1 Natural, 2 weeks 14 ST453 Stoneville Normal 1 Natural, 2 weeks 15 DP5409 Stoneville Normal 1 Extra, 2 weeks 16 DP5415 Stoneville Normal 1 Extra, 2 weeks 17 DES119 Stoneville Normal 1 Extra, 2 weeks 18 DP50 Stoneville Normal 1 Extra, 2 weeks 19 SG125 Stoneville Normal 1 Extra, 2 weeks (continued) Vol. 48(2): 439 454 441

Table 2 (continued). Cotton treatments in prediction error study. Seed cotton samples represented several varieties, production locations, harvest dates, and weathering durations. A broad variety of samples was used to indicate broad applicability with regard to color and trash levels. Treatment Variety Location Harvest Date Harvest Replication Weather or Harvest Regime 20 ST474 Stoneville Normal 1 Extra, 2 weeks 21 ST453 Stoneville Normal 1 Extra, 2 weeks 22 Mixed Stoneville Normal 1 Wet in trailer 23 Cencot Lubbock Normal 1 Brush stripper 24 Acala 90 Lubbock Normal 1 Brush stripper 25 Pay145 Lubbock Normal 1 Brush stripper 26 ST474 Lubbock Normal 1 Brush stripper 27 Pay4526 Lubbock Normal 1 Brush stripper 28 ST132 Lubbock Normal 1 Brush stripper 29 Acala Arizona Normal 1 None 30 Deltapine Arizona Normal 1 None 31 HS200 Fayetteville Early 1 None 32 DP50 Fayetteville Early 1 None 33 H1244 Fayetteville Early 1 None 34 SG501 Fayetteville Early 1 None 35 H1330 Fayetteville Early 1 None 36 ST474 Fayetteville Early 1 None 37 HS200 Fayetteville Early 2 None 38 DP50 Fayetteville Early 2 None 39 H1244 Fayetteville Early 2 None 40 SG501 Fayetteville Early 2 None 41 H1330 Fayetteville Early 2 None 42 ST474 Fayetteville Early 2 None 43 HS200 Fayetteville Late 1 None 44 DP50 Fayetteville Late 1 None 45 H1244 Fayetteville Late 1 None 46 SG501 Fayetteville Late 1 None 47 H1330 Fayetteville Late 1 None 48 ST474 Fayetteville Late 1 None 49 HS200 Fayetteville Late 2 None 50 DP50 Fayetteville Late 2 None 51 H1244 Fayetteville Late 2 None 52 SG501 Fayetteville Late 2 None 53 H1330 Fayetteville Late 2 None 54 ST474 Fayetteville Late 2 None 55 HS200 Keiser Early 1 None 56 DP50 Keiser Early 1 None 57 H1244 Keiser Early 1 None 58 SG501 Keiser Early 1 None 59 H1330 Keiser Early 1 None 60 ST474 Keiser Early 1 None 61 HS200 Keiser Early 2 None 62 DP50 Keiser Early 2 None 63 H1244 Keiser Early 2 None 64 SG501 Keiser Early 2 None 65 H1330 Keiser Early 2 None 66 ST474 Keiser Early 2 None 67 HS200 Keiser Late 1 None 68 DP50 Keiser Late 1 None 69 H1244 Keiser Late 1 None 70 SG501 Keiser Late 1 None 71 H1330 Keiser Late 1 None 72 ST474 Keiser Late 1 None 73 HS200 Keiser Late 2 None 74 DP50 Keiser Late 2 None 75 H1244 Keiser Late 2 None 76 SG501 Keiser Late 2 None 77 H1330 Keiser Late 2 None 78 ST474 Keiser Late 2 None 442 TRANSACTIONS OF THE ASAE

Lubbock, Texas, and harvested with a brush stripper: Cencot, Acala 90, Paymaster 145, Stoneville 474, Paymaster 4526, and Stoneville 132. Lots 29 and 30 were an Acala and a Deltapine, grown under standard conditions in Arizona and harvested with a spindle picker. Lots 31 through 78 were of six varieties, grown under standard conditions in two Arkansas locations and pickerharvested on two dates, in two replications each: Hartz 200, Deltapine 50, Hartz 1244, Suregrow 501, Hartz 1330, and Stoneville 474. Lots 31 through 36 were grown near Fayetteville, Arkansas, harvested early, and first replication. Lots 37 through 42 were the second replication of the same treatment. Lots 43 through 48 were grown near Fayetteville, Arkansas, harvested late, and first replication. Lots 49 through 54 were the second replication of the same treatment. Lots 55 through 60 were grown near Keiser, Arkansas, harvested early, and first replication. Lots 61 through 66 were the second replication of the same treatment. Lots 67 through 72 were grown near Keiser, Arkansas, harvested late, and first replication. Lots 73 through 78 were the second replication of the same treatment. Table 2 lists all differences among the treatments. Each lot of cotton was ginned in the Microgin, U.S. Cotton Ginning Laboratory, USDA-ARS, Stoneville, Mississippi. The Microgin has the capacity to route cottons through various sequences of gin machinery. From each lot of seed cotton, three replicate samples were collected prior to ginning. The ginning machinery sequence included the following seed-cotton cleaning sequence: feeder, dryer, cylinder cleaner, stick machine, dryer, cylinder cleaner, extractor feeder, and gin stand. Lint-cleaning sequences of zero, one, or two lint cleaners were operated consecutively during the ginning of each lot, and three replicate samples were collected for each lint-cleaning sequence. So 12 samples, four types in three replications, were collected for each of the 78 lots of cotton, for a total of 936 samples. Each of the 936 samples was measured once for color and trash on the experimental color/trash meter and on commercial meters A and B by subsampling, as described above. Since subsamples needed to be measured only once, except for those few subsamples measured twice in place, maintaining constant trash content from one measurement to the next was not necessary. The model for statistical analysis was as follows: y i = β 0 + β 1 x + ε (3) where y i = measured color (Y or Z) value on uncleaned cotton with ith instrument type (ZU or experimental) 0 = intercept of the regression line 1 = slope of the regression line x = true color measurement after cleaning cotton with two lint cleaners, measured either with ZU or experimental meter depending on which is the current reference instrument = residual error. The data were analyzed with regression analysis by using the SAS procedure PROC REG (SAS, 1985). The data set was coded such that for each lot of cotton and replication, there were four measurements with the experimental color/ trash meter and four with each ZU color/trash meter. The initial regression analysis set cleaned lint color from each ZU instrument as the standard to determine how well it was predicted by the measurements of uncleaned lint and seed cotton. Later, each instrument s readings on uncleaned cotton were compared with clean-cotton readings from the same instrument. Since each regression concerned the same number of data points, it was reasonable to consider the instrument with higher R 2 values in these analyses to be the better predictor. Furthermore, an F-test based on a ratio of the mean square errors (MSEs) of the two meter types was conducted to determine the significance of the difference in prediction error (Steel and Torrie, 1980). There is a slight problem with using an F-test to compare the two MSEs from the two instrument types. The F-test based on the F-value of two MSEs assumes that the MSEs are independent. Since, in the initial analysis, these two MSEs use the same observations for the true color values, they are not independent. However, this means that the F-test would be conservative, because the MSEs would tend to be more different if they were based on different true color values. To further support this analysis, graphical representations of data from each instrument type were studied to determine where differences existed in prediction error. RESULTS INTRUMENT DESIGN AND OPERATION As described in detail in Thomasson et al. (2005), the experimental color/trash meter s design was adequate for using image analysis to measure cotton color in customary units, and in so doing to remove the effect of trash particles on the color measurement. The instrument could be operated in very much the same fashion as a commercial cotton color/trash meter. From the operator s perspective, the experimental instrument was slower because of its nature as a research prototype, but in most other aspects it was similar. The experimental color/trash meter worked as designed and without any software or hardware failures during about four weeks of intermittent operation. The calibration algorithm and measurement algorithm operated according to design, and no significant drift in the instrument s sensitivity was evident during periods of data collection. COMPARISONS WITH STANDARD INSTRUMENTS Measurement Error Results of the PROC MIXED analysis are shown in table 3. With both the Y and Z color measurements, differences among the ZU color/trash meters (between-instrument variance) accounted for a negligible amount of error. This is evident in that very little difference exists between cumulative sample error and cumulative instrument error; i.e., variations among the ZU instruments added very little error to the total. It is also notable that the measurement error (within-instrument variance) was negligible with the experimental color/trash meter. The associated values of S 2 are 0.16 for Y and 0.29 for Z. The measurement error with the ZU meters is higher (42 for Y and 61 for Z). However, it must be noted that ZU meters in the HVI system report color values only in terms of Rd rounded to the nearest whole number and +b rounded to the nearest tenth. This means that Y is reported in increments of 20, and the round-off with Z is of similar magnitude. Thus, it is reasonable to assume that much of the measurement error with the ZU meters is due to round-off. Vol. 48(2): 439 454 443

Table 3. Results of SAS PROC MIXED analysis for measurement error. [a] Y ZU Experimental ZU Experimental Mean Obs. Mean Obs. Mean Obs. Mean Obs. 1431 1452 1417 524 1413 1452 1389 524 Error source S 2 S S 2 S S 2 S S 2 S Instrument 9139 95.6 -- -- 14712 121.3 -- -- Sample 9121 95.5 9578 97.9 14614 120.9 13718 117.1 Subsample 219 14.8 272 16.5 290 17.0 391 19.8 Measurement 42 6.5 0.16 0.4 61 7.8 0.29 0.5 [a] Obs. denotes number of observations, S 2 represents mean square error (MSE), and S represents root-mean-square error (RMSE). MSE values are cumulative in the upward direction; i.e., subsample MSE includes measurement MSE, sample MSE includes subsample MSE, and instrument MSE includes sample MSE. Z And while measurement error appears to be much lower with the experimental meter, it cannot be readily concluded that the experimental meter is superior in this regard. It can, however, be inferred that the experimental meter is as good as the ZU meters in terms of measurement error. When one compares the error due to subsampling (within-sample variance), it appears that the ZU meters are less sensitive to the nonhomogeneity of cotton samples. For Y, the F value is 1.24 (experimental meter MSE over ZU meter MSE, degrees of freedom [df] = 241/241), which means that the subsampling error is lower for the ZU meter, with the difference being significant at the 10% level. For Z, the F value is 1.35 (df = 241/241), which means that subsampling error is again lower for the ZU meter, with the difference being significant at the 5% level. However, the reader should recall that the ZU meters actually report, as one subsample value, an average from four subsamples. Theoretically, if the experimental meter reported an average from four subsamples as one subsample value, its subsampling error would be reduced by a factor of four. Again, if the experimental meter were employed in a classing office in the same configuration as the ZU meters, four subsamples could readily be presented to it also. While these ideas do not prove that the experimental meter is superior in terms of insensitivity to subsampling, they do lend weight to the argument that the significance of these F tests is not to be given too much importance. In other words, there is nothing obvious to say that either instrument is better than the other in either measurement error or subsampling error. Comparing between-sample variance, which one would wish to maximize so that differences in color between samples could be measured, no statistically significant differences exist, even at the 10% level. The experimental meter appears to be slightly more sensitive to between-sample differences in Y (F = 1.05, df = 241/241), while the ZU meters appear to be slightly more sensitive to between-sample differences in Z (F = 1.07, df = 241/241). The differences are very small, however. On the whole, the question implied in the objectives, can a camera-based instrument measure the color of lint samples with as little error as commercial diffuse-reflectance-based instruments?, can be answered in the affirmative. Prediction Error Results of the PROC REG analyses are given in table 4. Upon measuring the color of the 936 cotton samples, the two ZU meters exhibited strong correlation, with R 2 values of 0.93 for Y and 0.92 for Z. The experimental color/trash meter was highly correlated with each of the ZU meters. With meter A, the R 2 values were 0.94 for Y and 0.93 for Z. With meter B, the R 2 values were 0.91 for Y and 0.93 for Z. Since the meters were not presented the same subsamples, most of the Table 4. Results of SAS PROC REG analysis for prediction error. [a] Standard Relative Cleaning Cleaning Y Z Meter Level Meter Level RMSE R 2 p>f RMSE R 2 p>f A 3 B 3 14.53 0.93 0.0001 19.31 0.92 0.0001 A 3 Exp. 3 13.82 0.94 0.0001 17.80 0.93 0.0001 A 3 A 0 56.62 0.00 0.6606 66.12 0.05 0.1009 A 3 A 1 28.28 0.75 0.0001 45.40 0.55 0.0001 A 3 A 2 18.56 0.89 0.0001 36.09 0.71 0.0001 A 3 Exp. 0 55.75 0.03 0.1118 65.04 0.07 0.0229 A 3 Exp. 1 26.22 0.79 0.0001 38.57 0.67 0.0001 A 3 Exp. 2 18.22 0.90 0.0001 28.92 0.82 0.0001 B 3 Exp. 3 17.53 0.91 0.0001 19.94 0.93 0.0001 B 3 B 0 59.15 0.03 0.1585 71.71 0.05 0.0449 B 3 B 1 31.84 0.72 0.0001 45.87 0.61 0.0001 B 3 B 2 25.37 0.82 0.0001 34.46 0.78 0.0001 B 3 Exp. 0 59.02 0.03 0.1289 71.35 0.06 0.0286 B 3 Exp. 1 29.17 0.76 0.0001 43.31 0.65 0.0001 B 3 Exp. 2 23.22 0.85 0.0001 33.86 0.79 0.0001 [a] Exp. refers to the experimental meter. The statistical values RMSE, R 2, and p > F (probability of non-significance) refer to the regression between the standard and relative measurements at the left; e.g., where meter A with cleaning level 3 (two lint cleaners) is the standard and the experimental meter with cleaning level 1 (zero lint cleaners) is the relative measurement, the RMSE is 26.22 for Y and 38.57 for Z. 444 TRANSACTIONS OF THE ASAE

Measured Y value of seed cotton 1300 1200 1100 1000 R 2 = 0.033 R 2 = 0.003 900 800 1250 1300 1500 1550 1600 Measured Y value of cleaned lint with ZU meter A ZU meter A Regression line, ZU meter A Figure 1. Y values of seed cotton, measured with ZU meter A and the experimental meter, vs. Y values of cleaned lint measured with ZU meter A. R 2 values are associated with the closest regression line in the graph. 1600 Measured Y value of lint, 0 lint cleaners 1550 1500 R 2 = 0.79 1300 R 2 = 0.75 1250 1250 1300 1500 1550 1600 Measured Y value of cleaned lint with ZU meter A ZU meter A Regression line, ZU meter A Figure 2. Y values of zero-lint-cleaner lint, measured with ZU meter A and the experimental meter, vs. Y values of cleaned lint measured with ZU meter A. R 2 values are associated with the closest regression line in the graph. Vol. 48(2): 439 454 445

1600 Measured Y value of lint, 1 lint cleaner 1550 1500 R 2 = 0.90 1300 R 2 = 0.89 1250 1250 1300 1500 1550 1600 Measured Y value of cleaned lint with ZU meter A ZU meter A Regression line, ZU meter A Figure 3. Y values of one-lint-cleaner lint, measured with ZU meter A and the experimental meter, vs. Y values of cleaned lint measured with ZU meter A. R 2 values are associated with the closest regression line in the graph. Measured Z value of seed cotton 1250 1150 1050 950 R 2 = 0.066 R 2 = 0.035 850 750 1300 1500 1550 1600 1650 Measured Z value of cleaned lint with ZU meter A ZU meter A Regression line, ZU meter A Figure 4. Z values of seed cotton, measured with ZU meter A and the experimental meter, vs. Z values of cleaned lint measured with ZU meter A. R 2 values are associated with the closest regression line in the graph. 446 TRANSACTIONS OF THE ASAE

1600 Measured Z value of lint, 0 lint cleaners 1550 1500 R 2 = 0.55 R 2 = 0.67 1300 1300 1500 1550 1600 1650 Measured Z value of cleaned lint with ZU meter A ZU meter A Regression line, ZU meter A Figure 5. Z values of zero-lint-cleaner lint, measured with ZU meter A and the experimental meter, vs. Z values of cleaned lint measured with ZU meter A. R2 values are associated with the closest regression line in the graph. 1650 Measured Z value of lint, 1 lint cleaner 1600 1550 1500 R 2 = 0.71 R 2 = 0.82 1300 1300 1500 1550 1600 1650 Measured Z value of cleaned lint with ZU meter A ZU meter A Regression line, ZU meter A Figure 6. Z values of one-lint-cleaner lint, measured with ZU meter A and the experimental meter, vs. Z values of cleaned lint measured with ZU meter A. R 2 values are associated with the closest regression line in the graph. Vol. 48(2): 439 454 447

Measured Y value of seed cotton 1300 1200 1100 1000 R 2 = 0.030 R 2 = 0.026 900 800 1250 1300 1500 1550 1600 Measured Y value of cleaned lint with ZU meter B ZU meter B Regression line, ZU meter B Figure 7. Y values of seed cotton, measured with ZU meter B and the experimental meter, vs. Y values of cleaned lint measured with ZU meter B. R 2 values are associated with the closest regression line in the graph. 1600 Measured Y value of lint, 0 lint cleaners 1550 1500 R 2 = 0.72 R 2 = 0.76 1300 1250 1250 1300 1500 1550 1600 Measured Y value of cleaned lint with ZU meter B ZU meter B Regression line, ZU meter B Figure 8. Y values of zero-lint-cleaner lint, measured with ZU meter B and the experimental meter, vs. Y values of cleaned lint measured with ZU meter B. R 2 values are associated with the closest regression line in the graph. 448 TRANSACTIONS OF THE ASAE

1600 Measured Y value of lint, 1 lint cleaner 1550 1500 R 2 = 0.82 R 2 = 0.85 1300 1250 1250 1300 1500 1550 1600 Measured Y value of cleaned lint with ZU meter B ZU meter B Regression line, ZU meter B Figure 9. Y values of one-lint-cleaner lint, measured with ZU meter B and the experimental meter, vs. Y values of cleaned lint measured with ZU meter B. R 2 values are associated with the closest regression line in the graph. Measured Z value of seed cotton 1250 1150 1050 950 R 2 = 0.061 R 2 = 0.052 850 750 1500 1550 1600 1650 1700 Measured Z value of cleaned lint with ZU meter B ZU meter B Regression line, ZU meter B Figure 10. Z values of seed cotton, measured with ZU meter B and the experimental meter, vs. Z values of cleaned lint measured with ZU meter B. R 2 values are associated with the closest regression line in the graph. Vol. 48(2): 439 454 449

1650 Measured Z value of lint, 0 lint cleaners 1600 1550 1500 R 2 = 0.61 R 2 = 0.65 1500 1550 1600 1650 1700 Measured Z value of cleaned lint with ZU meter B ZU meter B Regression line, ZU meter B Figure 11. Z values of zero-lint-cleaner lint, measured with ZU meter B and the experimental meter, vs. Z values of cleaned lint measured with ZU meter B. R2 values are associated with the closest regression line in the graph. 1650 Measured Z value of lint, 1 lint cleaner 1600 1550 1500 R 2 = 0.78 R 2 = 0.79 1300 1500 1550 1600 1650 1700 Measured Z value of cleaned lint with ZU meter A ZU meter B Regression line, ZU meter B Figure 12. Z values of one-lint-cleaner lint, measured with ZU meter B and the experimental meter, vs. Z values of cleaned lint measured with ZU meter B. R 2 values are associated with the closest regression line in the graph. 450 TRANSACTIONS OF THE ASAE

lack of fit can probably be attributed to sample nonhomogeneity. Based on a valid F-test with independent MSEs, no statistical differences at the 10% level were found in the strength of correlations. In other words, the ZU meters correlated no better with each other than they did with the experimental meter. The most important question was whether the experimental instrument improved prediction of cleaned lint color from that of uncleaned lint or seed cotton. Measurements are discussed as follows: Y and Z are subscripted with a number for level of ginning (0 for seed cotton, 1 for zero lint cleaners, 2 for one lint cleaner, and 3 for two lint cleaners) and a letter for the meter used (A, B, or N for experimental). For example, Y 0,A represents Y for seed cotton as measured by meter A, while Z 2,N represents Z of lint put through 1 lint cleaner as measured by the experimental meter. The first comparison was between the experimental meter and ZU meter A (fig. 1). No correlation existed between Y 0,A and Y 3,A. A marginal correlation (significant at 11%) existed between Y 0,N and Y 3,A. A relatively strong correlation (R 2 = 0.75) existed between Y 1,A and Y 3,A (fig. 2). A stronger correlation (R 2 = 0.79) existed between Y 1,N and Y 3,A. While this result appears to validate the idea that the experimental meter measures uncleaned cotton better, the improvement in correlation was not statistically significant. A stronger correlation (R 2 = 0.89) existed between Y 2,A and Y 3,A (fig. 3), but an even stronger correlation (R 2 = 0.90) existed between Y 2,N and Y 3,A. Again, the improvement in correlation was not statistically significant. Marginal correlation (significant at 10%) existed between Z 0,A and Z 3,A (fig. 4). The correlation (significant at 5%) was slightly higher between Z 0,N and Z 3,A. A stronger correlation (R 2 = 0.55) existed between Z 1,A and Z 3,A (fig. 5), but a much stronger correlation (R 2 = 0.67) existed between Z 1,N and Z 3,A. The F value of 1.39 (df = 76/76) concerning the improvement in correlation was significant at the 10% level. A stronger correlation (R 2 = 0.71) again existed between Z 2,A and Z 3,A (fig. 6), and again a much stronger correlation (R 2 = 0.82) existed between Z 2,N and Z 3,A. The F value of 1.56 (df = 76/76) concerning the improvement in correlation was significant at the 5% level. The next comparison was made between the experimental meter and ZU meter B (fig. 7). No correlation significant at the 10% level existed between Y 0,B and Y 3,B, or between Y 0,N and Y 3,B. A relatively strong correlation (R 2 = 0.72) existed between Y 1,B and Y 3,B (fig. 8). A stronger correlation (R 2 = 0.76) existed between Y 1,N and Y 3,B. However, the improvement in correlation was not statistically significant. A stronger correlation (R 2 = 0.82) existed between Y 2,B and Y 3,B (fig. 9), but an even stronger correlation (R 2 = 0.85) existed between Y 2,N and Y 3,B. Again, the improvement in correlation was not statistically significant. Correlation (significant at 5%) existed between Z 0,B and Z 3,B (fig. 10). The correlation (also significant at 5%) was slightly higher between Z 0,N and Z 3,B. The F value of 1.19 (df = 76/76) concerning the improvement in correlation was not significant. A stronger correlation (R 2 = 0.61) existed between Z 1,B and Z 3,B (fig. 11). A somewhat stronger correlation (R 2 = 0.65) existed between Z 1,N and Z 3,B. The F value of 1.12 (df = 76/76) concerning the improvement in correlation was not significant. Again, a stronger correlation (R 2 = 0.78) existed between Z 2,B and Z 3,B (fig. 12), but an even stronger correlation (R 2 = 0.79) existed between Z 2,N and Z 3,B. The improvement in correlation was again not significant. In all cases, the intercepts in the regression equations were statistically significant when seed cotton and zero-lint-cleaner lint were used to predict cleaned lint color. When one-lint-cleaner lint was used as the predictor, the intercepts were non-significant in every case. The question of whether the experimental instrument improved prediction of cleaned lint color from that of uncleaned lint or seed cotton can be answered generally in the affirmative. The correlations between uncleaned lint color and cleaned lint color were improved in every case with the experimental meter, with some of the improvements being statistically significant even with an overly conservative F-test. The correlations between seed cotton color and cleaned lint color improved in most cases, but the correlations were all low, and the improvements not of practical importance. Two points should be made here. First, all comparisons to this point between uncleaned lint or seed cotton color and cleaned lint color were made with the ZU instruments cleaned lint color measurements as the standard. When one considers that the experimental meter improved correlations across the board, that fact appears all the more significant, because the experimental meter was compared to the ZU meters while the ZU meters were compared to themselves. Since it was concluded previously that the experimental meter measured color with as little error as the ZU meters, it is reasonable also to consider prediction error wherein the experimental meter s cleaned lint color measurements are used as the standard for the experimental meter, and the ZU meters cleaned lint color measurements are used as the standard for only the ZU meters. Table 5 gives R 2 and RMSE values for comparisons between uncleaned lint color and cleaned lint color, with each meter s measurements compared to its own. In every case but one (comparing Y 2,A with Y 3,A ), predictions of cleaned lint color from uncleaned lint color were better with the experimental meter than with either ZU meter. The improvement in prediction was evaluated first by calculating the reduction in RMSE brought about by the experimental meter over the average of RMSE values of the ZU meters, then by a valid F-test with independent MSEs. The subscripts on the Y and Z color measurements in the dependent measurement column represent cleaning level: 2 means lint put through one lint cleaner, and 1 means lint put through zero lint cleaners. The R 2 and RMSE values describe regressions between the dependent measurement and a measurement on cleaned lint with the same meter. For example, the R 2 value for the regression between cleaned-lint Table 5. Comparison of cleaned lint color predictions between the experimental meter and the ZU meters. Exp. Meter ZU Meter A ZU Meter B Dependent R 2 RMSE R 2 RMSE R 2 RMSE Y 2 0.87 19.1 0.89 18.6 0.82 25.4 Y 1 0.77 25.8 0.75 28.3 0.76 31.8 Z 2 0.83 26.8 0.71 36.1 0.78 34.5 Z 1 0.69 36.1 0.55 45.4 0.65 45.9 Vol. 48(2): 439 454 451

Z and one-lint-cleaner lint Z both measured with the experimental meter is 0.83, while the same measurements made with ZU meter A had an R 2 value of 0.71. When predicting Y of cleaned lint from Y of lint put through one lint cleaner, the reduction in RMSE brought about by the experimental meter was 13.2%. This improvement was significant at the 10% level (F = 1.33, df = 76/76). When predicting Y of cleaned lint from Y of lint put through no lint cleaners, the reduction in RMSE brought about by the experimental meter was 14.0%. This improvement was again significant at the 10% level (F = 1.35, df = 76/76). When predicting Z of cleaned lint from Z of lint put through one lint cleaner, the reduction in RMSE brought about by the experimental meter was 23.9%. This improvement was significant at the 1% level (F = 1.73, df = 76/76). When predicting Z of cleaned lint from Z of lint put through no lint cleaners, the reduction in RMSE brought about by the experimental meter was 20.8%. This improvement also was significant at the 1% level (F = 1.60, df = 76/76). It was noted earlier that +b is more critical in price determinations than is Rd. Whereas +b depends strongly on Z measurements, the greater improvement in accuracy with Z emerges as an important development. The improvement in predicting cleaned lint color from uncleaned lint color appeared to be more pronounced when measurements from the different meters were compared against measurements from the same meter. The improvements in prediction were statistically significant in every Experimentalmeter ZU meter ZU meter Figure 13. Two-standard-deviation ellipses for ZU and experimental color/trash meters. It is demonstrated that the experimental meter is associated with higher prediction accuracy of cotton fiber s final color based on its uncleaned (zero lint cleaners) color. 452 TRANSACTIONS OF THE ASAE

case. Figure 13 shows the cotton color-grade chart with two ellipses representing 95% confidence regions about a true cotton color of Y = and Z = (Rd = 72.0 and +b = 8.92). The regions of error are simplified for visual presentation, but they are conservative in +b and are reasonably representative of cleaned lint color predictions from measurements on lint having been put through no lint cleaners. The figure makes it very clear that the experimental color/trash meter offers a significant reduction in error. The second point to be made here is that part of the difficulty in predicting cleaned lint color from seed cotton color was the wide range in trash contents of the seed cottons, and the fact that the seed cottons had not been cleaned after harvest. For instance, the six cottons from the Lubbock, Texas, area were stripper harvested, meaning that their trash contents were likely at least 5 times as great as those of the spindle-harvested cottons. The relatively high correlations (R 2 = 0.84 for Rd, R 2 = 0.24 for +b) between seed cotton color and cleaned lint color reported by Anthony (1988) involved only picker-harvested seed cotton that had been through a standard sequence of seed cotton cleaning equipment. It was thus much cleaner than the seed cotton used in this study. Whereas the ZU meters measured trash content to a maximum of only 5%, correlations between seed cotton color and cleaned lint color were evaluated on those samples with trash contents less than 5% as measured by the ZU meters. For those samples with trash content less than 5% as measured by meter A (25 of 76 sample averages), R 2 values for predicting cleaned lint color from seed cotton color were 0.57 for Y on the experimental meter, 0.56 for Y on meter B, 0.50 for Z on the experimental meter, and 0.39 for Z on meter B. For those samples with trash content less than 5% as measured by meter B (28 of 76 sample averages), R 2 values for predicting cleaned lint color from seed cotton color were 0.52 for Y on the experimental meter, 0.36 for Y on meter A, 0.44 for Z on the experimental meter, and 0.33 for Z on meter A. Once again, errors in predicting cleaned lint color were reduced with the experimental meter. Of course, the correlations between seed cotton color and cleaned lint color were fairly weak, even for this cleaner group of samples. SUMMARY AND CONCLUSIONS The experimental cotton color/trash meter reported by Thomasson et al. (2005) was developed for the purpose of improving cotton color measurement, particularly on trashy cotton, such as is found in the ginning process. The principle of improvement was the removal of trash-particle effects with image processing. In this work, the experimental meter was first shown to be as good as conventional color/trash meters in terms of measurement error. The measurements of the experimental meter also were found to be as well correlated with those of the conventional meters as were the conventional meters measurements with each other. It was thus concluded that basic cotton color measurement is as accurate with the experimental meter as it is with the conventional meters. Furthermore, correlations between the color values of trashy cottons and those of cleaned cottons were determined from measurements made with the experimental and conventional meters. Seed cotton color measurements were not well correlated with those of cleaned lint. It was concluded that, although the experimental meter improved the correlations, a very limited amount of information with which to predict cleaned cotton color is directly available from seed cotton color measurements with the current techniques. Since the primary function of gin process control with regard to color measurement is to determine the proper number of lint cleaners with which to process the cotton, predicting cleaned lint color from seed cotton color is not as important as predicting cleaned lint color from uncleaned lint color. Correlations between uncleaned lint color and cleaned lint color were fairly strong. In every case, the experimental meter s measurements correlated better with cleaned lint color than did those of the conventional meters. With Z (blue-band reflectance), the superiority of the experimental meter s correlation was statistically significant. It was thus concluded that prediction error was reduced with the experimental meter. When using the ZU meters as the standard for comparison, the reduction in RMSE with the experimental meter was about 10% for Y (broad band green reflectance), about 20% for Z (blue band reflectance) with meter A, and about 5% for Z with meter B. When using each meter as its own standard for comparison, and averaging RMSE values for the ZU meters, the reduction in RMSE with the experimental meter was about 14% for Y, and about 22% for Z. These reductions in prediction error had statistical as well as practical significance. PRACTICAL IMPLICATIONS In practical terms, the accuracy with which a gin process control system selects the proper number of lint cleaners depends on the accuracy with which it predicts cleaned lint color based on uncleaned lint color. If the process control system erroneously selects two lint cleaners rather than one, the second lint cleaner can cause a loss of about 1% of the mass of the bale, not to mention unnecessary damage to the fiber. Such a determination would be based primarily on the Y measurement. The experimental meter would offer a 14% improvement in the accuracy of such a determination, making the 1% loss that much less likely. Furthermore, if the process control system bypasses a lint cleaner based on an erroneous determination that the cotton is off-color (and thus cleaning will not improve its unit value significantly), the unit value of the cotton could be discounted as much as 10%. This determination would be based more on the Z measurement. The experimental meter would offer as much as a 22% improvement in the accuracy of such a determination, making the 10% discount that much less likely. Whereas about 18 million bales of cotton are produced annually nationwide, and the future of cotton ginning is moving toward automatic process control systems, a large amount of money could be lost to cotton producers through inaccurate sensing of cotton quality in the gin. It is thus very important to minimize the error with which cotton color is measured in gin process control systems. The experimental color/trash meter detailed in this article offers sizable reductions in color measurement error, and it, or another system that accounts for the effect of trash on color (such as that of Xu et al., 1997), should be strongly considered as a replacement for current color trash meters in process control systems. Vol. 48(2): 439 454 453