Effect of Sample Preparation on Prediction of Fermentation Quality of Maize Silages by Near Infrared Reflectance Spectroscopy*

,


INTRODUCTION
Corn silage is an important forage in dairy feeding, not only as a winter rations but also as a supplemental feed during the grazing season.Silage is the material produced by the controlled microbial fermentation of crops with high moisture content in a process known as ensilage.Quality of corn silages is commonly evaluated on the basis of pH, NH3-N, and concentrations of certain short-chain organic acids.However, traditional analytical methods for determining nutritive value of corn silage can be costly and time consuming.
Near-infrared reflectance spectroscopy (NIRS) offers advantages over wet chemistry in terms of simplicity, speed, reduced chemical waste, and a more cost-effective prediction of product functionality.Analysis of feeds and silages by wet chemistry require drying and milling of the sample.Although prediction of composition of dried samples has been carried out for a number of years, the direct analysis of undried silages by NIRS would provide additional gains in reduced time and labor for sample preparation and faster reporting of the results (Shenk and Westerhaus, 1995; De la Roza et al., 1996;Kennedy et al., 1996).
Relatively few studies have been published on the use of NIRS to determine fermentation quality parameters of undried samples, with most results reported for grasses and silages (Daniel et al., 1999;Park et al., 2002).In the case of wet silage samples, heat drying could result in losses of volatile substance, such as short-chain organic acids and alcohols (McDonald et al., 1991).In order to determine the amount of short chain organic acids in wet silage using NIRS, freezing with liquid nitrogen has been adopted in some laboratories to prevent loss of these volatiles from fresh forage samples.
Another problem with using NIRS directly on fresh silages may be increased errors due to differences in sample particle size, temperature and water content.These problems can be overcome by grinding silages in a frozen state with dry ice or liquid nitrogen.However, such procedures are time-consuming and inconvenient due to the cleanup required between samples and the need to thaw the sample for subsequent use.
These issues may be resolved in part by improvements in chemometrics software, spectral data transformation for scatter correction and partial least squares regression.Such advances have minimized some of the interference of particle size variation and water absorption presented by wet silage samples (Baker et al., 1994;Shenk and PARK ET AL. Westerhaus, 1995;Gordon et al., 1998).Thus, this experiment was conducted to assess the effect of sample preparation (drying or liquid nitrogen treatment vs. fresh) methods on prediction of fermentation quality of corn silage, and to select an acceptable sample preparation method for wet silage.

Silage preparation for NIRS scanning
Corn silage samples (n = 112) were collected from dairy farms in Kyunggi-do, Korea.The samples were frozen as soon as they arrived at the laboratory, and stored frozen (-20°C) until analyzed.Prior to NIRS scanning, the silage samples were thawed overnight at 4°C in a refrigerator.
Each sample was subdivided into three sub-samples of ca.200 g that were either oven-dried (OD), ground under liquid nitrogen (LN), or held intact fresh (IF).Oven-dried sub-samples were dried at 65°C for 24 h and then ground in Wiley mill with a 1 mm screen.Intact fresh sub-samples were measured immediately upon opening the silo with no sample preparation.Liquid nitrogen sub-samples were immersed in liquid nitrogen (-196°C) for 30 min, broken up with a wooden mallet and sub-sequently finely chopped in a kitchen food chopper.The milled samples were thawed at ambient room temperature before being presented for NIRS scanning on a NIRSystem Model 6500 spectrophotometer (Perstorp Analytical Silverstone, USA).

Chemical analysis
All chemical analyses were conducted on the undried samples.Dry matter (DM) was determined by drying for 24 h at 105°C in a forced-draft oven.To determine pH, 10 g of plant tissue were macerated in a blender with 100 ml of distilled water.Silage pH was measured with an electrometric pH meter (HI 9024; HANNA Instrument Inc., UK).
For analysis of short-chain organic acids, 20 g of fresh silage were transferred to 250 ml wide-necked bottle and 100 ml of distilled water were added to each.The bottles were capped and shaken mechanically for 1 h.Resulting solutions were then filtered through Whatman No. 1 filter paper.Five ml of filtrate were combined with 1 ml of a 2.5 g/L solution of pivalic acid (as internal standard) and 2.5 ml of 0.12 M oxalic acid in a 10 ml calibrated flask and diluted to volume.Flasks were centrifuged at 2,600 g for 5 min and the supernatants were collected for injection into a GC.
Short-chain organic acids were determined by gas chromatography (6890N, Agilent Co., USA) on an 80/120mesh Carbopack B-DA/4% Carbowax (Supelco Inc., Bellefonte, PA, Catalog No. 1-1889) with 20 M column treated with trimesic acid in methanol.Oven conditions included a gas flow-rate of 24 ml/min and oven temperature of 200°C with 1 卩l injections of 0.03 M oxalic acid made prior to use.

Analysis of spectral data
Spectra of silage samples were collected in NIRSystems 6500 scanning monochromotor.The OD and LN sub samples were placed in Quarter cups, and IF sub-samples were placed in Natural Product sample cups which were inserted in a transporting sample device.Absorbance data were collected as log 1/R, from 400 to 2,500 nm.Data analyses were performed using WinISI II-version 1.02a software (Foss NIRSystems/Tecator, Infrasoft International, LLC).In order to determine the best sample preparation of data, the 112 spectra for each of the 3 sample preparation methods were divided into a calibration set (n = 86) and a validation set (n = 36) using the SELECT facility as developed by Shenk and Westerhaus (1991) within WinISI II-version software.This selected a specified number of samples ( 86), on the basis of their spectral H-distance (equivalent to Mahalanobis distance), which represented the spectral characteristics of the whole population.
Calibrations were developed using modified partial least squares (MPLS) regression with internal cross-validation after scatter correction using standard normal variate (SNV) and detrending.Cross-validation was used to avoid over fitting of the equations.A 1,4,4,1 curve-smoothing mathematical treatment was applied to the NIRS output.The first number indicates the order of derivative, the second number is the gap in nm over which the derivatives were calculated, the third number is the number of data points used in the first smoothing, and the fourth number refers to the number of nm over which the second smoothing was applied.Calibration statistics included the standard error of calibration (SEC), the coefficient of multi determination in calibration (R2), and the standard error of cross-validation (SECV).Optimal calibrations were selected on the basis of minimizing the SECV.An independent validation set (OD = 48, LN = 29 and IF = 27) was used with samples not included in the calibration set.Standard error of prediction (SEP), squared simple correlation coefficient (RSQ), and slope were calculated using WinISI software, version 1.02a.

RESULTS AND DISCUSSION
In developing optimum systems for sample preparation it is important that the spectral data produced, within each method, are fully explored to produce the best prediction relationships.The silage samples selected for this study varied widely in their DM, pH and most short-chain organic acid parameters (Table 1).A number of the corn silage samples contained little or no detectable concentration of butyric acid, which should be undetectable in good silage.A  total of 7 parameters were examined in the calibrations.
Using MPLS regressions, with the firs (1,4,4,1) derivatization in conjunction with SNV-D, produced a calibrations for each of the parameters are given in Table 1.
Average spectral data from 400 to 2,500 nm for each treatment are shown in Figure 1(a).Lines of OD and LN treatment are obviously not superimposed, which indicates some differences between treatments, although they show similar absorption bands, rendering the curves almost parallel.Baseline shifts may occur due to factors such as differences in sample holder glass, or differences in sample compression, temperature, and particle size (Shenk and Westerhaus, 1995).In order to compensate for baseline shift and to enhance the relevant spectral information, spectra were transformed to their first derivative.The average differentiated spectra for each treatment are presented in Figure 1(b).Although differentiation removed most differences between processing methods, at least two absorption bands could be detected.These differences between treatments suggest that some chemical bonds are present in different concentrations as a result of processing method.
In order to acquire optimum calibration equations, Adesogan et al. (1998) suggested that equations with the largest R2, smallest standard error of calibration (SEC), and lowest number of spectral terms (to avoid overfitting) should be selected.When developing MPLS equations, cross validation was used to select the optimum number of factors and thus avoid overfitting.For purposes of comparison, the best calibrations were selected on the basis of minimizing the SECV.The standard error of cross validation (SECV) and 1-VR (equivalent to R2 of cross validation) statistics represent a truer prediction of how the calibration will perform when unknown samples are predict.
For estimates of DM, the NIRS calibration statistics (Table 2) of the LN treatment were slightly better (lower SEC, higher R2) than those obtained by the OD and IF treatments.The values obtained here provided good calibration statistics, in contrast with the findings of Cozzolino et al. (2000).Accurate determination of moisture (or DM) in forages has been hampered both by the lack of a primary reference method that is specific for water and by uncertainties in the change in moisture content during sample preparation.The inability of this calibration to accurately analyze the DM content in our sample set lies in the lack of agreement between the reference method (oven drying) and the NIRS measurement.Many authors (Windham, 1987;Windham et al., 1987) have concluded that NIRS could be used to accurately analyze forages for concentration of DM or moisture when calibrated with laboratory methods that truly define their water content.The coefficients of determination for estimates of silage pH were slightly greater for the LN and OD treatments than for the IF treatment (0.81 vs. 0.74; Table 2).Lower R2 for the IF treatment is likely due to differences in particle size and packing characteristics because these samples were not ground.Results of the LN treatment were consistent with reports of Gordon et al. (1998) The effects of sample preparation method on the accuracy of short-chain organic acid prediction were similar to those for pH and DM.Performance statistics for the prediction of acetic and lactic acids in the present study were excellent, with R2 for the LN treatment being slightly greater than for OD and IF treatments These performance statistics are also better than those obtained by Park et al. (1998) and Reeves and Blosser (1989) from NIRS when using liquid nitrogen treatment and dry ice sample preparation methods.Across treatments, NIRS calibration statistics for butyric acid were the worst in all parameters.This is likely due to the low concentrations of butyric acid in the silages used for this study.
The SD/SECV ratio ranged from 0.07 for butyric acid (OD treatment) to 3.13 for DM (OD treatment).It has been suggested that a ratio >2.5 indicates that the calibration is adequate for quality screening purposes and >3.0 indicates that the calibration should perform well for quantitative analyses (Sinnaeve et al., 1994).
In the present study, the LN treatment produced a relatively robust calibration (R2 = 0.81-0.94,1-VR = 0.63-0.87)for all parameters except butyric acid.Short-chain organic acids for dried samples cannot be detected directly with NIRS.In the case of wet silage samples, heat drying could result in losses of volatile substances, such as short chain organic acids and alcohols (McDonald et al., 1991).However, fermentation end-products in the oven-dried Short chain organic acids (%, DM) samples may be detected indirectly because their presence in organic complexes affects H bonds (Shenk, 1992;Givens et al., 1997) or the concentrations of organic constituents (Watson et al., 1976).
The results obtained on the independent validation set for the multiple correlation coefficient of validation (R2v), standard error of prediction (SEP) and bias are shown in Table 3.The samples in the validation set are normally different to those that are used to develop the prediction equation, and are usually a smaller set than the calibration set.Williams (1987) has provided rules for interpreting values for bias, SEP and correlation between predicted and reference values.He recommended that the SEP should not be more than 3% of the mean reference value.Predictions for all parameters (DM, pH and short-chain organic acids excluding butyric acid) had good validation statistics in present studies.The LN treatment gave the best results among the three sample-preparation methods.Validation of these equations with 26 samples produced SEP of 1.05 and 0.06 and validation coefficients (R2v) of 0.81 and 0.85 for DM and pH, respectively.For acetic, propionic, lactic and total acids, the best analytical results were obtained with IF (SEP = 0.13, R2v = 0.89), LN (SEP = 0.08, R2V = 0.65), OD (SEP = 0.58, R2v = 0.80) and IF (SEP = 0.69, R2v = 0.73), respectively.Sinnaeve et al. (1994), using fresh silages, have shown that it is possible to derive successful calibrations for pH, acetic acid, and lactic acid (R2v = 0.90, 0.85 and 0.86 respectively).Considering the diversity of this population of corn silages the NIRS estimated fermentation end-products such as pH, acetic, lactic and total acids to an acceptable degree.
The results of this study have shown that NIRS analysis of undried silages can provide accurate prediction of a wide range of fermentation products.Sample preparation treatment clearly influenced NIR spectra and accuracy of fermentation products analysis.The time required to carry out the sample preparation procedures is important not only in a commercial laboratory system but also in relation to changes in sample conditions.Although the accuracy of OD and IF treatments were lower than for the LN treatment, NIRS could be used as a screening tool to predict fermentation quality in wet corn silages.However, because little preparation or handling of samples is required, analysis of IF samples is worth consideration given the speed and convenience.Development of NIRS calibrations for silage samples prepared with OD methods will require more research on volatile substances such as short-chain organic acids in dried samples.NIRS offers considerable potential for analysis of wet silage in routine advisory systems using OD and IF preparation methods.

Figure 1 .
Figure 1.Average NIR spectra of silage either as log1/R (a) or as a first derivative (b) for oven-drying (OD), Liquid nitrogen (LN) and Fresh treatments.

Table 1 .
Average±SD and ranges for DM, pH and short-chain organic acids of corn silages according to drying method on DM basis Sample preparation1 1 OD = Oven-dried ground, LN = Frozen in liquid nitrogen and ground, IF = Intact fresh.

Table 2 .
NIRS calibration statistics for DM, pH and short-chain organic acids of corn silage Oven-dried ground, LN = Frozen in liquid nitrogen and ground, IF = Intact fresh. 2 R2 = Multiple correlation coefficient of determination.SEC = Standard error of calibration.SECV = Standard error of cross validation.
1-VR = Multiple correlation coefficient of cross validation.

Table 3 .
Validation statistics1 for NIRS of DM, pH and short-chain organic acids of corn silage samples Oven-dried ground, LN = Frozen in liquid nitrogen and ground, IF = Intact fresh.SEP = Standard error of prediction.