| | Abstract | Introduction | Objectives | Materials and methods | Results and discussion | Conclusions
Abstract
The main objective of this study was to determine if Near Infrared Spectroscopy (NIRS) could be used effectively to screen whole-grain barley samples for feed quality characteristics. Three hundred samples of whole-grain barley were collected in different locations and three years (1993-95). These samples included nine hulless lines and one hulled barley control (Harrington). These samples were analyzed for: crude protein, protein digestibility, energy digestibility, digestible energy content and gross energy at the University of Alberta, in digestibility studies with pigs. The same samples were used to develop near infrared equations on a NIRSystems model 6500 (NIRSystems, Inc., Silver Springs, MD) monochromator using Modified Partial Least Squares Regression analysis. Various combinations of 1st and 2nd derivatives, gap, smooth and scatter corrections were used to maximize the mathematical treatment for each constituent. The Rē of the developed equations ranged from 0.98 for crude protein to 0.90 for gross energy content. All the equations developed had high Rē values and indicate that excellent quantitative information can be obtained from these. The equations will be invaluable as a tool for screening feed quality characteristics in a large breeding program, and will be able to separate samples into high, medium and low quality feed characteristics.
Introduction
The cost of feed is the most expensive input in any livestock operation. Deficiencies in feeding can affect animal performance by decreasing weight gain, lowering milk production or causing malnutrition and health problems, all of which affect the profitability of an operation. A new cereal grain cultivar with exceptional feed value would make a huge impact on the livestock industry, resulting in increased production and lower costs. However, the process to develop a new cereal cultivar is slow. It requires screening thousands of early generation lines to select only a few lines with the potential to become new cultivars. This process is further hindered by the lack of quick and inexpensive methods to determine feed quality characteristics. For this reason, producers, plant breeders and animal nutritionists spend a great deal of time trying to define 'feed quality' and how to accurately measure it.
Traditional methods for determining feed quality have many limitations. Many of the analysis require either chemicals in the methodologies, or involve digestibility trials with animals. In either case, the process is time consuming and expensive therefore limiting the number of samples that can be evaluated. In addition, neither of these methods are readily available to producers. Producers are left to guess at the quality of their feedstuffs. A new method is needed to speed up the process and make results of analysis easier to obtain.
Near Infrared Spectroscopy (NIRS) is a technology that has shown potential in the analysis of feed components. It has been in the process of development since the 1960's. Initially, it was used to predict moisture content in wheat, but today the method is used to predict hundreds of components, with more being added every day. Near Infrared Spectroscopy technology offers its users many advantages. It is a simple, accurate and fast procedure in which no chemicals are used and does not affect the quality of the sample. Although the initial cost of the equipment is quite high, these costs are quickly recovered in time. In addition there are labor-cost savings. Near Infrared Spectroscopy is being accepted by the industry as a method of measuring many different components; however, it still remains a secondary method of measurement, and requires accurate reference values for equation development.
The near infrared region of light spans 700nm-2500nm. Near Infrared Spectroscopy works by directing light that has been separated into specific wavelengths onto a sample and measuring the amount of light that is reflected back. Detectors within the machine are able to measure the quantity of light reflected, and which wavelengths of light that were reflected back. Using these measurements the sample can then be mathematically analyzed and the constituents determined based on the correlations with the actual measurements of the components using traditional methods in which animals are used.
Objectives
Although NIRS had been quite successful in determining some primary components of cereal grains, there was no information about its usefulness in predicting feed quality characteristics that up until now have required animal digestibility trials. The purpose of this study was to determine if NIRS could be used to accurately predict feed quality values for barley. The specific objectives of this study were to develop NIR equations for crude protein (CP), gross energy (GE), protein digestibility (PD), energy digestibility (ED) and digestible energy content (DEC).
Materials and Methods
Ten lines of barley, from the Hulless Barley Cooperative Test (Prairie Registration Recommending Committee for Grain, 1993, 1994, 1995), were grown at various locations in Western Canada (Table 1). The samples were collected over a period of three years (1993, 1994, and 1995). In total, 300 barley samples were collected. The three replications from each site were bulked and a 1 kg sample of cleaned seed was sent for determination of the quality traits, measured using pigs at the University of Alberta. The traits PD, ED and DEC were measured using digestibility trials with pigs. Crude Protein and GE were determined with the Kjeldahl procedure (N) and with an Adiabatic Bomb Calorimeter, respectively. Harrington was the only hulled barley cultivar (control) evaluated in the studies with pigs. All quality characteristics were determined on a dry matter basis.
Table 1. Samples and locations grown in each year.
NIRS scanning
A 100g sample of cleaned whole grain barley was used for NIRS scanning. The sample was kept as whole grain kernels, and scanned using a monochromator NIRSystems model 6500 (NIRSystems, Inc., Silver Springs, MD), equipped with a transport module. Samples were scanned in a large natural product cell with a removable back. The reflectance spectrum of 400 - 2500 nm was recorded at 2 nm intervals for each sample (Figure 1). We chose to include the visible region (400-700nm) in our scanning for future consideration. The machine was set up to take 32 scans of the sample and the values were averaged.
Figure 1. Variation in the reflectance spectrum of all 300 samples used in the study.
Equation development
Equations were developed using all 300 samples using WINISI software, version 1.02a (FOSS NIRSystems, Silver Springs, MD). The samples were sorted by reference data for each constituent, and every fifth sample was extracted for equation validation groups (60 samples total). Calibration equations for the remaining 240 samples were developed by using spectral data from 400-2500 nm and Modified Partial Least Squares regression (Shenk and Westerhaus, 1991a) . A repeatability file was used to minimize the effects of particle size and sample temperature. Combinations of 1st and 2nd derivative mathematical treatments, gap, smooth and scatter correction were used to maximize the equation results. Cross validation was used to prevent overfitting, and the software was set up for four outlier elimination passes. Equations were selected based on the optimum combination of a large Rē, a small standard error of calibration (SEC), and a low standard error of prediction (SEP) for the validation group. Unique validation samples, omitting any outliers, would be added to make the final equations.
Validation of equations
Equations were validated using the 60 sample external validation group developed earlier. The complete set of 300 samples (including any outliers) were then analyzed using the developed equations. The prediction results were analyzed with the original reference data on all 300 samples. Correlation analysis were carried out to determine correlation coefficients and to evaluate the significance of correlations between the NIR prediction values and the reference data (SAS, 1989).
Results and Discussion
When this project was undertaken, there was no information available on using digestibility studies with animals to develop near infrared calibrations. In total there were five equations developed in this study: CP, PD, GE, ED and DEC (Table 2). Using data from all three years, the Rē of the equations developed ranges from 0.98 to 0.90. Correlation coefficients and the significance of the correlation between the predicted values and the original reference data are summarized in Table 3.
Table 2. Near infrared reflectance spectroscopy equation statistics for feed quality characteristics in whole grain barley.
 | n | Mean | SD | SEC | Rē | Math Treatment
A, B, C1, C2 | Scatter
Correction |
| Protein | 227 | 13.05 | 2.12 | 0.297 | 0.980 | 2,4,4,1 | SNV&DET |
| Protein Digestibility | 148 | 78.72 | 5.50 | 1.622 | 0.913 | 1,4,4,1 | SNV&DET |
| Energy Digestibility | 142 | 84.95 | 5.00 | 1.436 | 0.918 | 2,4,4,1 | SNV&DET |
| Digestible Energy Content | 153 | 3407.93 | 202.78 | 56.390 | 0.923 | 1,4,4,1 | NONE |
| Energy | 156 | 4017.90 | 66.94 | 21.120 | 0.901 | 1,4,4,1 | NONE |
Table 3. Correlation analysis results of the NIR predicted values for the entire population of 300 samples (including outliers) and the original reference data (with and without Harrington).
| Constituent | Rē | Significance | Rē
(without Harrington) | Significance
(without Harrington) |
| Protein | 0.98188 | P</=0.001 | 0.98353 | P</=0.001 |
| Energy | 0.82408 | P</=0.001 | 0.83706 | P</=0.001 |
| Protein Digestibility | 0.65328 | P</=0.001 | 0.59920 | P</=0.001 |
| Energy Digestibility | 0.71767 | P</=0.001 | 0.40108 | P</=0.001 |
| Digestible Energy Content | 0.72186 | P</=0.001 | 0.50341 | P</=0.001 |
Crude Protein (%)
The best prediction equation was developed for CP content having an Rē=0.98, and a SEC of 0.30. Removal of Harrington from the calibration set, had no significant effect. The calibration population had an even distribution of reference data as shown in Figure 2.
It was expected that the calibration for CP content would be the most accurate prediction equation. Near Infrared Spectroscopy has been used extensively to determine protein in cereals since it's development as the N-H bonds in protein are high absorbers in the NIR region.
Digestible Energy Content (kcal/kg)
Digestible Energy Content gave the next best equation with an Rē= 0.92 and a SEC value of 56.39. The inclusion of the hulled control, Harrington, had a significant effect on the calibration for this constituent (Table 3). Harrington samples were lowest in DEC and are clearly visible as a distinct group at the low end of the graph (Figure 3).
Protein Digestibility (%)
The Protein Digestibility equation has an Rē value of 0.91 and an SEC value of 1.62. This equation was improved slightly with the inclusion of Harrington in the analysis (Table 3). Protein Digestibility also had a more even distribution of samples in the calibration population as compared with ED, GE and DEC (Figure 4).
Energy Digestibility (%)
The equation of ED was similar to PD with an Rē value of 0.92 and an SEC value of 1.44. The inclusion of the hulled control, Harrington, had a significant effect on the calibration for this constituent (Table 3). Harrington samples also had the lowest values for ED and are clearly visible as a distinct group at the low end of the graph (Figure 5).
Gross Energy (kcal/kg)
The calibration for GE was the least accurate of the five equations, although it still had an Rē value of 0.90 and an SEC value of 21.12. Harrington was eliminated from the calibration population before developing the equation for GE. Figure 6 shows that there are some samples from a specific location and year (1994, Olds) with a significantly higher GE content than most of the other samples.
General discussion
For PD, ED, GE, and DEC significantly less samples were used for the equation development process and overall the equations were less accurate than for CP content. It is logical to expect that using animals to determine digestibility values will produce a range of variability. Even though the samples were replicated in the digestibility studies, this will not eliminate all the variability. Near Infrared Spectroscopy is based on the chemical bonds present in the sample and any variation found in the digestibility studies used for the calibration caused samples to be thrown out of the calibration population as outliers.
Smaller calibration populations for PD, GE, ED and DEC were also due to the elimination of similar samples and the limited amount of genetic variability for some of the traits in this study. When developing a robust calibration for research applications, it is extremely important to start with the proper samples. The samples should have a range of composition and have a uniform distribution of reference data, without having unnecessary duplication of similar samples. All sources of variation should be identified and accommodated in the calibration population. Variables that should be considered include growing location, growing season, chemical composition, physical composition and spectral identity. These variables should be included at least three times in any calibration, and accumulation of sufficient samples for a stable calibration is often the most critical aspect (Williams, 1987). Without the right samples the true spectro-chemical relationship cannot be determined.
The GE components of this study showed the year was more important to increase the variability than location. The samples from Olds in 1994 demonstrate the importance of including seasonal variation in the calibration population in addition to genetic variation (Figure 6). Samples from 1993 and 1995 Olds fit within the average calibration population; however, with the inclusion of 1994 Olds samples, the range of reference values is increased and the equation is improved.
With the exception of CP content and PD content, the calibration populations contained large groups of samples with very similar values. Samples were eliminated from the calibration population that were repetitive in order to minimize any bias towards the mean of the population. There are also some obvious 'holes' in the populations that have no samples. Finding samples to fill these holes would expand these equations and would increase prediction accuracy.
As mentioned earlier, it is apparent in Figures 3 and 5 that Harrington (the only hulled variety in the study) accounts for a large amount of the variability in the parameters analyzed. By eliminating Harrington, the curves we obtained had high Rē and low SEP values, but the range of constituent data was limiting. Since it is our intention to use these equations for analyzing hulled in addition to hulless varieties, Harrington was left in the population set. When the GE equation was developed, we found that we could get a much better relationship by eliminating Harrington from the calibration population. This observation was validated with the SAS analysis that shows that the Rē increased slightly with the exclusion of Harrington (Table 3).
Conclusions
In conclusion, this study has shown that NIRS can be used successfully to predict feed quality characteristics such as CP, GE, PD, ED and DEC. All the equations developed had high Rē values and indicate that excellent quantitative information can be obtained from using these. The equations will be invaluable as a tool for screening feed quality characteristics in a large breeding program, and will be able to separate samples into high, medium and low feed quality groups. The protein calibration is accurate enough to be used for any application, not just for screening or quality assurance.
The future direction of this project should aim at expanding variation in these initial equations. By selecting additional unique samples and unique seasonal variance to the calibration population the equations will improve in accuracy and become more robust. The selected samples would need to be analyzed in digestibility studies for PD, ED and DEC data before they could be used to expand the equations. Only samples that are not yet represented by the existing population would have to be added. This process of expansion will result in robust, stable calibration equations that could easily be applied to commercial farm operations.
L.A. Oatway, J.H. Helm and P.E. Juskiw
Field Crop Development Centre
Helm, J.H., P.E. Juskiw, L. Oatway, and W.C. Sauer. 2000. Genetic and Environmental Effects on the Feed Quality of Hulless Barley. Part 2. Development of Near Infrared Spectroscopy to Screen for Feed Quality Characteristics in Whole Grain Barley. Final Report, Farming for the Future Project #94M616 (http://www.aari.ab.ca/index.cfm) and Alberta Barley Commission Project #60-057 (http://www.albertabarley.com/). |
|