1300 individual grain samples (consisting of experimental breeding lines and ... analysis was carried out on 15 g whole grain samples in a ¼ cup sample cell.
Efficiency of Selection for Wheat Kernel Characteristics Using near-Infrared Reflectance Spectroscopy. Joshua D. Butler1, Scott D. Haley1 and Bradford W. Seaborn2,
(1)Colorado State University, Department of Soil and Crop Science, Fort Collins, CO 80523, (2)U.S. Dept. of Agriculture / ARS/ Grain Marketing & Production Research Center, Avenue, Manhattan, KS 66502-2736
Introduction
Model Validation
Results
Figure 1. Comparison of kernel weight determined by prediction model and by reference method.
Calibration Cross-validation Independent validation n Terms SEC Groups SECV R2 n SEP(C) R2 4 0.126 0.764 75 0.939 0.777 Kernel diameter (mm) 367 11 0.112
y = 0.7408x + 7.8696 R2 = 0.7412 n =75
45
Methods
Parameters
40
Single Kernel Characteristics 1300 individual grain samples (consisting of experimental breeding lines and released varieties) were harvested from five locations throughout eastern Colorado. Samples were collected from the 2004 and 2005 growing seasons.
25 25
30
35
40
45
50
55
A scanning monochromator NIRSystem 6500 (Foss NIRSystems, Inc., USA) was used to measure NIR diffuse reflectance spectra from 400 to 2500 nm at every 2 nm step. And the analysis was carried out on 15 g whole grain samples in a ¼ cup sample cell. The NIR spectra of whole grain were collected in 25 scans for each sample and were recorded as log(1/R).
Figure 2. Comparison of kernel diameter determined by prediction model and by reference method.
y = 0.8267x + 0.4895 2 R = 0.7766 n =75
3.3 Predicted Kernel Diameter (mm)
Test Weight 302 whole grain samples from variety trial plots throughout eastern Colorado were scanned with the NIRSystems 6500 as described above. Test weight of each sample was recorded using a ‘dry quart’ measuring cup and a Seedburo test weight device (Seedburo Equipment Co., USA).
11
2.53
4
2.82
0.79
75
2.79
0.741
103
3
9.03
5
11.37 0.94
22
10.25
0.921
Principal component analysis resulted in data sets of 449 samples for kernel characteristics and 184 samples for test weight with unique spectra that were used for further analysis. Approximately 20% of the samples were randomly selected from each of the data sets and were used to create a validation data set that was not included in calibration development. Extreme spectra and reference data outliers were eliminated as determined by WINISI software. Summary statistics for reference data for the calibration and validation data sets are reported in Table 1.
Results of independent validation of calibrations reveal only minor reductions in R2 when compared to the cross validation method. For example, results revealed reductions of 0.013 for kernel diameter, 0.056 for kernel diameter, and 0.020 for test weight. Results of regression of reference values on the NIR predicted values for each of the three traits are displayed in Figures 1-3. These were constructed using the independent validation data set.
3.5
Commercial spectral analysis software (WINISI III, Foss NIRSystems, Infrasoft International) was used to develop the calibration equations and evaluate the calibration performance. Various mathematical treatments were applied to the absorbance spectra to maximize the accuracy of the calibration model. Treatments included scatter corrections to minimize the nonlinear effect of light scatter due to particle size differences (none, standard normal variate + detrend, standard normal variate only, and detrend only) and data transformation via derivative mathematics that reduces the intercorrelation between the data points of a spectrum.
370
Test weight (kg/m3)
Results of calibration equations developed using various mathematical treatments were compared and the most effective equations were chosen. The results of prediction using the chosen equations are displayed in Table 2.
Reference Kernel Weight (mg)
Selected samples were analyzed for single kernel characteristics using a SKCS 4100 (Perten Instruments, USA) and average kernel weight (mg) and average kernel diameter (mm) were recorded.
Kernel weight (mg)
Calibration Development and Validation
35
30
Principle component analysis was carried out to eliminate outliers and spectrally redundant samples. Of the remaining sample spectra, 20% were randomly designated as an independent validation set and were not used in calibration development or cross-validation.
Table 2. Summary statistics for calibration development, cross-validation, and independent validation. SEC is standard error of calibration, SECV is standard error of cross-validation, and SEP(C) the standard error of prediction.
50
Predicted Kernel Weight (mg)
End-use quality improvement is an important objective in most wheat breeding programs. In the case of winter wheat, the short duration between harvest and planting represents a major challenge to efficiently and timely conduct quality evaluations to enable selection prior to planting. Furthermore, multiple test parameters, some being destructive of the grain sample, are commonly used as predictors of overall end-use quality which complicates the selection process. The large numbers of breeding samples typically handled even by programs of modest size represents another limitation in the process of quality evaluation and selection in a breeding program. Near infrared reflectance (NIR) spectroscopy is a rapid and non-destructive technique that could facilitate early generation screening in breeding programs. The precision and accuracy of an NIR equation for prediction purposes is dependent on the construction of a reliable calibration. Prediction models can be developed using spectral fingerprints and phenotypic reference data. In this way, a single NIR scan of a sample of wheat grain can provide non-destructive estimates for several different quality traits. The objective of this study was to develop and validate whole-grain calibrations using near-infrared reflectance (NIR) spectroscopy to estimate kernel weight, kernel diameter, and test weight.
Table 3. Realized heritability of NIR predicted values. Means are ± standard error. Parameter Overall mean Selected mean hR2 39.89 ±0.17 43.39 ±0.09 F4 Kernel weight (mg) 0.9350 37.15 ±0.39 40.43 ±0.13 F5 Kernel weight (mg)
3.1
2.9
2.7
2.5
F4 Kernel diameter (mm)
3.05 ±0.01
F5 Kernel diameter (mm)
2.91 ±0.02
3.22 ±0.01 3.05 ±0.01
F4 Test weight (kg/m3)
773.08 ±1.16
761.00 ±0.79
F5 Test weight (kg/m3)
742.34 ±2.45
762.42 ±0.77
0.8526
0.7205
Realized Heritability Realized heritability indicates the amount of genetic improvement that is realized by selection within a population.
Realized Heritability
2.3 2.3
2.5
2.7
Realized heritability estimates for single kernel characteristics were calculated from a subset of 200 early generation lines using grain from F4 and F5 lines.
2.9
3.1
3.3
3.5
Reference Kernel Diameter (mm)
Realized heritability was calculated as hr2 = R/S, where R is the response realized by selection, and S is the selection differential (difference between mean of selected individuals and mean of overall population).
Figure 3. Comparison of test weight determined by prediction model and by reference method.
Results
830
Table 1. Summary statistics for calibration and validation data sets.
Parameters
n
Mean
min.
max.
Calibration set Kernel diameter (mm) 374 Kernel weight (mg) 374
2.92
2.19
3.46
36.3
23.3
49.8
Test weight (kg/m3)
781.9
691.6
847.5
109
Validation set Kernel diameter (mm) 75 Kernel weight (mg) 75
2.97 37.7
Test weight (kg/m3)
769.5
22
y = 0.9082x + 5.5746 2 R = 0.9206 n =22
790 770 750 730
References
710
3.40
690
25.5
50.7
670 670
680.8
835.0
2.43
Conclusion The objective of our study was to develop and validate whole-grain calibrations using near-infrared reflectance (NIR) spectroscopy to estimate kernel weight, kernel diameter, and test weight. Use of whole-grain NIR predictions shows promise to provide rapid, non-destructive estimates of multiple end-use quality parameters, thus enabling expanded evaluation in earlier generations and evaluation from a broader collection of environments than may be feasible with conventional methods. Regression of NIR predicted values on reference values revealed R2 values of 0.92 for test weight, 0.74 for kernel weight, and 0.78 for kernel diameter. Realized heritability estimates for the traits were 0.72 for test weight, 0.94 for kernel weight, and 0.85 for kernel diameter. While the results of this study encompass data from only two years over several locations, results reveal the possible utility of whole-grain NIR as a tool in plant breeding programs.
3
Range
Predicted Test Weight (kg/m )
810
Results of the calculations for estimates of realized heritability are reported in Table 3. Heritability estimates appear high when compared to past studies on test weight and kernel characteristics. This may be due to the fact that the estimates are based on data from a single year and a single location. While the results are estimates of the heritability of the NIR predicted values, rather than actual reference values, these estimates would suggest that there is a heritable textural or structural property that is detected by the NIR.
Approved Methods of the Association, The American Association of Cereal Chemist, St. Paul, MN. 690
710
730
750
770
790 3
Reference Test Weight (kg/m )
810
830
Fehr, W.R. Principles of Cultivar Development. 1987. Macmillan Publishing, New York, NY. Williams, P., and Norris, K. 2001. Near-infrared Technology in the Agriculture and Food Industries. 2nd edition, The American Association of Cereal Chemists, St. Paul, MN.