| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
,1
* Department of Grassland Science, College of Animal Science and Technology, China Agricultural University, Beijing, P. R. China, 100094
Agriculture and Agri-Food Canada, Quebec, Quebec, Canada G1V 2J3
Agriculture and Agri-Food Canada, Sherbrooke, Quebec, Canada J1M 1Z3
Université Laval, Québec, Québec, Canada G1K 7P4
1 Corresponding author: gaetan.tremblay{at}agr.gc.ca
 |
ABSTRACT |
|---|
 TOP ABSTRACT INTRODUCTIONMATERIALS AND METHODSRESULTS AND DISCUSSIONCONCLUSIONSACKNOWLEDGMENTSREFERENCES |
|---|
of 0.92, 0.89, and 0.93, and a ratio of prediction to deviation (RPD) of 3.3, 3.1, and 3.6, respectively. The NDSF prediction was classified as moderately successful 
The NIRS prediction of OA was unsuccessful 
All related constituents were predicted successfully 
by NIRS except ethanol-insoluble residual OM, with 
Our results confirm the feasibility of using NIRS to predict NDSC, its fractions, and other related constituents, except for OA and ethanol-insoluble residual OM, in timothy and alfalfa forage samples.
Key Words: alfalfa • near-infrared reflectance spectroscopy • neutral detergent-soluble carbohydrate • timothy
|
INTRODUCTION |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTS AND DISCUSSIONCONCLUSIONSACKNOWLEDGMENTSREFERENCES |
|---|
-amylase. As a class, NDSC are considered to be approximately 98% digestible (Van Soest, 1967), but their diversity precludes their use as a uniform nutritional fraction. A better understanding of how dairy cattle digest carbohydrates could lead to improved animal performance and contribute to more efficient nitrogen use (Leiva et al., 2000).
|
Since the report by Norris et al. (1976) first appeared, near-infrared spectroscopy (NIRS) has been widely used to analyze the nutritive components (CP, NDF, etc.) of animal feed. This method combines the advantages of minimal sample preparation and rapid reporting times. Various carbohydrates have been estimated in feed by using NIRS (Roberts et al., 2004), including starch in corn silage (Welle et al., 2003) and grains (Kim and Williams, 1990), and pectin in legumes and grasses (Fairbrother and Brink, 1990). Fonseca et al. (1999) used NIRS to predict NDSF in alfalfa. Water-soluble carbohydrates and total NSC are frequently measured in alfalfa (Brink and Marten, 1986; Gossen, 1994) and in other grasses and cereal straws (Brown et al., 1987; Fairbrother and Brink, 1990). Batten et al. (1993) reported that NIRS can be used to predict NSC, which are usually determined with traditional chemical methods in shoot samples of rice and wheat crops. Mentink et al. (2006) studied the accuracy of NIRS prediction for the NFC concentration in TMR analyzed according to the method of NRC (2001). To date, there has been no report using NIRS to predict the different carbohydrate fractions determined by the method of Hall et al. (1999) in nonfermented dried forage samples. The objective of the current study was to evaluate the feasibility of using NIRS to predict concentrations of OA, starch, sugars, NDSF, and NDSC (and all related constituents used to calculate these fractions) in timothy and alfalfa forages.
|
MATERIALS AND METHODS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTS AND DISCUSSIONCONCLUSIONSACKNOWLEDGMENTSREFERENCES |
|---|
All dried forage samples (n = 1,008) were scanned over 400 to 2,498 nm at 2-nm intervals using an NIRSystems 6500 monochromator (Foss, Silver Spring, MD). For each spectrum, principal component analysis scores were calculated using WinISI III software (version 1.61, Infrasoft International LLC, Silver Spring, MD). Seventy-five samples were selected based on these scores, to form a calibration set (n = 60) and a validation set (n = 15) of samples for each forage species. The calibration set of timothy samples was subsequently combined with the corresponding set of alfalfa samples to form a new calibration set (n = 120). Validation sets from both these single species were also combined to form a new validation set (n = 30). The similarity among samples within the combined sample sets was analyzed with a 3-dimensional plot of the principal component scores and the Global H statistic (WinISI III software, version 1.61, Infrasoft International LLC). Although the 3-dimensional plot indicated a slight clustering of the 2 species, all but 5 samples had a Global H score of <3.0, indicating that alfalfa and timothy did not constitute 2 different clusters of data.
Chemical Analyses
Seventy-five samples of each species (timothy and alfalfa) selected for NIRS calibration and validation were analyzed for carbohydrate fractions by using wet chemistry procedures (Hall et al., 1999). Neutral detergent-soluble carbohydrates and all related constituents were analyzed in several steps: 1) whole forage samples were analyzed for DM, OM, CP, and crude fat (ether extract, EE); 2) whole forage samples were solubilized in an 80% (vol/vol) ethanol:water solution, and the ethanol-insoluble residues (EIR) were analyzed for OM (EIROM), CP (EIRCP), and starch, whereas the ethanol-soluble extracts (ESE) were analyzed for total 80% ethanol-soluble carbohydrates (TESC), which also represented sugars (monosaccharides and oligosaccharides); and 3) whole forage samples were solubilized in a neutral detergent solution, and the neutral detergent residues (NDR) were analyzed for OM (NDROM) and CP (NDRCP).
Dry matter was determined by drying forage samples at 105°C for 24 h. The OM was determined as the difference in sample weight before and after ashing at 200°C for 2 h and at 500°C for 4 h in a muffle furnace. Ether extract was determined using an Ankom XT Extractor (Ankom Technology Corp., Macedon, NY) in accordance with American Oil Chemists’ Society official procedure Am 5-04 (AOCS, 1998). The NDF concentration was determined according to the procedure of Goering and Van Soest (1970), using sodium sulfite, -amylase, and the Ankom Fiber Analyzer (Ankom Technology Corp., Fairport, NY). The NDR was kept and further analyzed for OM and total nitrogen concentration. The ESE and EIR were obtained according to Hall et al. (1999). The TESC concentration in the ESE was determined by the phenol-sulfuric acid method (Hall et al., 1999). The EIR remaining after extraction was washed twice with methanol and used for starch quantification as glucose equivalent by using the p-hydroxybenzoic acid hydrazide method of Blakeney and Mutton (1980) after gelatinization at 100°C and digestion for 90 min with amyloglucosidase (Sigma A7255, Sigma Chemical Co., St. Louis, MO). Subsamples of 0.1 g of whole forage samples, NDR, and EIR were mineralized by using a mixture of H2SO4 and H2SeO3 as described by Isaac and Johnson (1976), and then total nitrogen concentration of these extracts was measured (method 15-107-06-2-E, Lachat Instruments, 2008) on a Lachat QuikChem 8000 flow injection autoanalyzer (Zellweger Analytics, Lachat Instruments Division, Milwaukee, WI). Concentrations of CP were calculated as total nitrogen x 6.25. All constituents, except DM, were analyzed in duplicate for each sample.
The NDSF were calculated as follows: NDSF = (EIROM – EIRCP) – (NDROM – NDRCP) – starch. Organic acids are soluble in aqueous ethanol and they are found in the ESE. The majority of the fat and some CP were coextracted with OA; very little fat remained in the EIR after the ethanol solution extraction and acetone rinses. The OA was calculated as (OM – CP) – (EIROM – EIRCP) – EE – TESC. The NDSC is the sum of OA, TESC, starch, and NDSF (Hall et al., 1999), but by substituting OA and NDSF by the respective formulas noted above, we used the following calculation to estimate this fraction: NDSC = OM – CP – EE – (NDROM – NDRCP).
NIRS Calibration
A modified partial least squares regression method was used to develop calibration equations with the full spectrum (Marten et al., 1983) for NDSC, the 4 major carbohydrate fractions (OA, TESC, starch, and NDSF), and all other related constituents including DM, OM, CP, NDF, EE, NDROM, NDRCP, EIROM, and EIRCP. To account for possible operator errors, a repeatability file was created by collecting 20 spectra per sample, using independently filled cups for each of 3 randomly selected samples of each forage species. To improve the calibration models, 14 spectral pretreatments were tested (WinISI III software, version 1.61, Infrasoft International). Two criteria were used to select the best spectral pretreatment parameters: simultaneous low standard errors of cross-validation and high coefficients of determination in cross-validation (1-VR). Four cross-validation groups were selected when developing the NIRS equations so as to choose the optimal number of terms and avoid overfitting (Shenk and Westerhaus, 1991).
Data Analyses
The standard error of laboratory (SEL) was defined as the standard error of variance between duplicates analyzed by the reference method. The SEL was calculated by using the following equation:
|
|
|
|
Selection of the NIRS equations was based mainly on the coefficients of determination of prediction 
and standard errors of prediction (SEP) corrected for bias [SEP(C)]. The SEP(C) was calculated by the following formula (Naes et al., 2002):
|
|
|
|
and RPD were used to classify the performance of a given NIRS equation according to Sinnaeve et al. (1994) and Williams (2001): |
|
|
|
|
|
|
RESULTS AND DISCUSSION |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTS AND DISCUSSIONCONCLUSIONSACKNOWLEDGMENTSREFERENCES |
|---|
|
> 0.90 and RPD >3. The OA, starch, and NDSF concentrations in timothy were predicted unsuccessfully, with 
0.53 and RPD 1.4. In alfalfa, the only successful NIRS equation was obtained for starch, with an 
of 0.93 and an RPD of 3.8. The calibration equations in alfalfa were moderately successful for TESC 
and unsuccessful for OA, NDSF 
and NDSC 
The 3 successful NIRS equations based on individual forage species were strongly related to the relatively high SD values in the validation sets: 17.4 g/kg of DM for TESC and 44.2 g/kg of DM for NDSC in timothy, and 10.6 g/kg of DM for starch in alfalfa. Accurate NIRS predictions require 3 basic conditions: 1) intense absorbance or sensitive variation of characteristic chemical bonds, such as –NH, –CH, and –OH in the NIR region; 2) low laboratory errors, and; 3) a sufficient concentration (higher than 1 g/kg of DM) and range of constituent concentrations (Shenk and Westerhaus, 1994; Nie et al., 2008). In the current study, the first requirement was not the principal limiting factor because, characteristically, all the chemical bonds of most constituents were mainly composed of –OH, –NH, and –CH, which could be quantified exactly, as reported by Roberts et al. (2004). Furthermore, laboratory errors were controlled at an acceptable level, with a coefficient of variation between duplicate analyses of less than 5%. However, the NIRS predictions in the current study, based on separate equations for each species, were still unsuccessful for many of the fractions (Table 1).
NIRS Prediction with Combined Equations for Both Species
To improve the NIRS performance by increasing the composition range and the standard deviations, as suggested by Dunn et al. (2002) and Cozzolino and Morón (2006), the calibration sets for timothy and alfalfa were combined to form a new calibration set, and the validation sets of the 2 species were also combined to form a new validation set. Compared with the calibration set of each species, ranges of values for NDSF and NDSC were greatly increased in the combined calibration set (Table 2). For instance, the NDSC concentration ranged from 101.1 to 294.9 g/kg of DM in timothy and from 234.4 to 410.9 g/kg of DM in alfalfa (Table 1), and by combining these 2 calibration sets, the range of values varied from 101.1 to 410.9 g/kg of DM (Table 2). The ranges of values for 3 of the NDSC fractions (OA, TESC, and starch) in the combined species were close to those for each species. The standard deviation values of NDSF and NDSC in the combined calibration set were also greater than in the calibration set of each species. The standard deviation values of TESC and starch for the combined species were in between those for each species, whereas a lower standard deviation value was obtained for OA in the combined species (Tables 1 and 2).
|
|
Table 3). The NDSC prediction for the combined set was improved, compared with the unsuccessful NIRS equation for alfalfa (Table 1). Mentink et al. (2006) reported a moderately successful NIRS equation for predicting NFC calculated according to the method of NRC (2001), with a coefficient of determination of 0.83 and an SEP of 15.5 g/kg of DM in TMR. In the current study, the SEL associated with the analysis of NDSC in single calibration sets of timothy (14.9 g/kg of DM; Table 1), alfalfa (15.7 g/kg of DM; Table 1), and combined forage species (15.3 g/kg of DM; Table 2) were all calculated by using the formula described earlier (Harris, 1991) and were based on the SEL of OM, CP, EE, NDROM, and NDRCP. Therefore, the SEL of NDSC was not affected by the high SEL associated with the EIROM determination (24.1 g/kg of DM). This result could explain the more accurate prediction of NDSC than of OA and NDSF, which were both calculated from other fractions including EIROM. The NIRS prediction of TESC for the combined set of both species performed better 
Table 3) than the successful NIRS equation for timothy and the moderately successful NIRS equation for alfalfa (Table 1). Mentink et al. (2006) report an unsuccessful calibration for TESC concentration in TMR with 1-VR of 0.61 and SEP of 13.8 g/kg of DM; this result was attributed to the high level of laboratory error associated with TESC analysis. In the present study, however, although the 
value (0.89) of the NIRS prediction of starch concentration for the combined set was lower than 0.90, it was still classified as being successful, with an RPD of 3.1 and a high slope value of 0.96. The starch NIRS prediction based on the combined species was considerably better than the starch prediction in timothy but was worse than that of alfalfa (Table 1). Similar successful NIRS predictions of starch concentration have been reported for corn silage (Kim and Williams, 1990; Welle et al., 2003).
In the current study, the worst NIRS calibration equation with the combined set was obtained for OA, with an SEP(C) of 12.8 g/kg of DM, an 
of 0.38, and an RPD of 1.3 (Table 3); this result is similar to that obtained with the single species calibration sets. Because concentration of OA was calculated as (OM – CP) – (EIROM – EIRCP) – EE – TESC, its SEL (27.5 g/kg of DM) was calculated from the SEL of all these constituents by using the formula presented above. The high SEL value led to the low RER value (3.5) for OA in the combined calibration set, which explains the difficulty of obtaining an accurate NIRS equation for OA (Table 2).
With an 
of 0.88 and an RPD value of 2.8 in validation (Table 3), the NIRS prediction of NDSF concentration in both timothy and alfalfa samples was classified as being moderately successful. The NDSF prediction with the combined set was much better than with timothy 
and alfalfa samples 
Table 1). Fonseca et al. (1999) reported a successful NIRS prediction of NDSF (R2 = 0.97) based on their alfalfa calibration (n = 56) and validation (n = 18) sets from a given harvest, but the prediction was less accurate (R2 = 0.72 to 0.89) when NIRS equations were based on samples from 4 other individual harvests or from combined harvests. The possible reason for the less accurate prediction for the combined set was attributed to insufficient sample variation because the R2 of NIRS equations was highly associated with the standard deviation values of calibration samples (r = 0.99, P < 0.01, n = 216; Fonseca et al., 1999). In the current study, although the high SD (51.6 g/kg of DM) for NDSF was obtained by merging the 2 calibration sets of samples from 2 different forage species that were taken in 2 harvests at 2 sites, the 
of the NDSF prediction was still lower than 0.90. The SEP(C) of NDSF for the combined validation set (17.5 g/kg of DM; Table 3) was also higher compared with the SEP(C) value reported by Fonseca et al. (1999), whereas the mean value for the combined calibration set (137.5 g/kg of DM; Table 2) was lower than their mean NDSF concentration (180.1 g/kg of DM). The limiting factor for obtaining accurate NIRS predictions of OA and NDSF was the high SEL value of 27.5 g/kg of DM in both cases; this result was mainly due to the high SEL associated with the determination of EIROM (24.1 g/kg of DM; Table 2).
Compared with noncalculated constituents, the SEL of the carbohydrate fractions OA and NDSF were much higher (Tables 1 and 2), whereas that of NDSC was relatively low. Considering both the SEL values and the overall performance of NIRS equations for OA (unsuccessful), NDSF (moderately successful), and NDSC (successful) in a combined sample set, the SEL of NDSF might have been overestimated. As Harris (1991) explained, the aforementioned formula used to calculate this SEL provides a more probable estimate than the one achieved by simply adding all the individual errors, but it could be greater than the error calculated from duplicate analyses. The overestimated error of NDSF was mainly transferred from the SEL of EIROM in the calculations, similar to that for the SEL of OA.
NIRS Prediction of Related Constituents.
By using the combined calibration set that included both forage species, it was possible to generate robust NIRS equations, with an 
of 0.91 to 0.99 and an RPD of 3.3 to 8.4 (Table 3), to predict DM and NDF, and other constituents (OM, CP, EE, NDROM, NDRCP, and EIRCP) that are used to calculate 3 carbohydrate fractions (OA, NDSF, and NDSC). Compared with other results of NIRS applications in forages (Roberts et al., 2004), our current EE estimation was more accurate 
few studies have reported NIRS predictions of EE concentration with R2 > 0.90 (Murray and Hall, 1983; Roberts et al., 2004). The NDF NIRS model was also more accurate, as evidenced by an RPD of 8.4 and an 
of 0.99 in validation. The high RPD of NDF partly resulted from the highest standard deviation value (140.5 g/kg of DM) of the combined validation set.
The NDROM and NDRCP constituents of the combined set of alfalfa and timothy samples were predicted with a high degree of accuracy 
Table 3). There are few reports of NIRS prediction of NDROM, whereas studies on the prediction of NDRCP have reported differing degrees of accuracy. Hoffman et al. (1999) and Mentink et al. (2006) observed low-quality calibration (R2 = 0.71 and 0.52, respectively) of NDRCP for grass silage and TMR samples. However, Valdés et al. (2006) reported that neutral detergent-insoluble nitrogen in heterogeneous permanent meadows was successfully predicted by NIRS, with 1-VR of 0.91 and RPD of 3.33. Nie et al. (2008) obtained a poor NDRCP prediction 
in alfalfa based on partial least squares regression, but greater accuracy 
was achieved with the support vector regression method. The high accuracy of the NDROM prediction observed in the current experiment may be due to its strong correlation (r = 0.97, P < 0.001, and n = 150) with NDF. Accurate NIRS equations for NDRCP possibly resulted from the wide range of 6.1 to 59.7 g/kg of DM and relatively low SEL of 1.2 g/kg of DM for the calibration set (Table 2).
The EIROM was less accurately predicted, with an 
of 0.75 and an RPD of 1.9 (Table 3). This was probably due to the higher laboratory error (SEL = 24.1 g/kg of DM; Table 2) of EIROM compared with all other noncalculated carbohydrate fractions. However, EIRCP, which is another EIR-based constituent, was accurately predicted, with an 
of 0.93 and an RPD of 3.2; this successful prediction could be partly explained by the strong correlation between EIRCP and CP concentration (r = 0.91, P < 0.001, and n = 150).
RER for Calibration Versus NIRS Performance
Williams (2001) calculated the RER and RPD for his validation set of samples and classified his NIRS predictions based on these statistics. He reported that a validation RER value greater than 13 indicated good NIRS performance. In the present study, we suggested calculating the RER statistic, which is equal to the range divided by the SEL, for the calibration set. Our results for combined species showed that all constituents with calibration RER values greater than 13 were successfully predicted by NIRS (Tables 2 and 3). The predictions for OA and EIROM were unsuccessful, with low calibration RER values of 3.5 and 9.0, respectively (Table 2). Although the calibration RER value was only 7.7 (Table 2) for NDSF, the NIRS prediction of this carbohydrate fraction was moderately successful 
Table 3). Successful NIRS prediction of carbohydrate fractions can therefore be obtained when the calibration RER values are greater than 13; when RER values are less than 13, NIRS prediction could be either moderately successful or unsuccessful. The relationships between the RER of the calibration set and RPD and 
of the validation set of combined alfalfa and timothy samples are presented in Figure 2. Two exceptions appear in the RER-RPD graph (Figure 2a): NDF and starch. Because of the much lower SEL and the wide range of values for starch compared with other constituents, the RER value for starch was inflated, whereas the NIRS performance values maintained a normal degree of accuracy. The exception of NDF in the RER-RPD graph (Figure 2a) was caused by its high RPD value (8.4), which resulted from the greatest standard deviation value (140.5 g/kg of DM) in the validation set (Table 3). When these exceptions were excluded from the RER-RPD graph (Figure 2a), both coefficients of determination of the 2 regressions in Figure 2 equaled 0.89 (P < 0.001). Based on separate equations for each species, the statistics are lower; the coefficient of determination of the relationship between the calibration RER and the RPD of the validation set was 0.75 for timothy and 0.67 for alfalfa, whereas the coefficient of determination of the relationship between calibration RER and the 
of the NIRS validation was 0.85 for timothy and 0.87 for alfalfa (data not shown). The calibration RER value is a quantified expression of the influence on NIRS performance of laboratory error and the range of concentrations of a given constituent in the calibration set of samples. It is therefore a potentially useful indicator of subsequent NIRS performance.
|
|
CONCLUSIONS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTS AND DISCUSSIONCONCLUSIONSACKNOWLEDGMENTSREFERENCES |
|---|
and RPD and could therefore be used to evaluate the possibility of obtaining successful equations before any NIRS work. |
ACKNOWLEDGMENTS |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTS AND DISCUSSIONCONCLUSIONSACKNOWLEDGMENTSREFERENCES |
|---|
Received for publication July 31, 2008. Accepted for publication December 12, 2008.
|
REFERENCES |
|---|
|
TOPABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTS AND DISCUSSIONCONCLUSIONSACKNOWLEDGMENTSREFERENCES |
|---|
This article has been cited by other articles:
|
|
 |
|
  G. F. Tremblay, Z. Nie, G. Belanger, S. Pelletier, and G. Allard Predicting timothy mineral concentrations, dietary cation-anion difference, and grass tetany index by near-infrared reflectance spectroscopy J Dairy Sci, September 1, 2009; 92(9): 4499 - 4506. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
for the validation set (b) for all constituents (solid circle) predicted by NIRS by using the combined alfalfa and timothy sample set. Solid lines represent regressions, excluding triangle symbols.