Prediction of gas production kinetics of maize stover and ear by near infrared re fl ectance spectroscopy

This study was conducted to evaluate the potential of NIRS to predict gas production (GP) kinetics for maize ear and stover. Two approaches were applied to characterize GP kinetics: 1. describing GP profi les by fi tting an exponential model and 2. by calculating GP for various incubation intervals. In the fi rst approach, NIRS pre dic tion equations could explain a high proportion of variability related to the fermen ta tion rate (c), the asymptotic GP (A), and the lag phase (L) of ear samples in the calibration sub set (R2: 0.81-0.87), while for stover predictive ability was inferior. Validation showed satisfactory results only for the fermentation rate (R2: 0.81-0.88). In the se cond approach NIRS calibration was acceptable for GP in several incubation inter vals, but validation showed satisfactory results for intervals 3-7, 7-12, and 7-16 h of sto ver samples only (R2>0.74). Further research is required to elucidate potential sources of error before obtaining robust NIRS equations.


INTRODUCTION
The possibility to accurately predict the nutritive value of forage crops is a prere quisite for designing rations based on the animal's requirements, but also for di rec ting forage crop breeding.The most precise determination of a crop´s nutritive value de rives from in vivo feeding studies.Routine analysis, however, is hampered by the need of laboratory facilities and cannulated animals, large quantities of feed and high costs.To this account, in vitro techniques were developed.The in vitro gas produc tion (GP) technique (Menke et al., 1979) is a simple and ro bust method, based on the close relation ship between gas produced from anaerobic fer men tation with rumen liquor and feed degradation, which allows to estimate various fo rage quality character istics.It has some advantages over in vitro methods that are based on meas uring fer men tation residues, since for instance, not only information on the absolute extent of feed digestion is provided, but also on the degra dation ki netics (Getachew et al., 2004).The effi ciency of sample evaluation brought about by au tomation (Theo do rou et al., 1994) makes the GP technique appli cable to plant bree ding programs (Getachew et al., 2005).Nevertheless, the time need for in cuba tion and the necessity of rumen fl uid donor animals is a dis ad vantage.This obstacle might be overcome by the near infrared refl ectance spectroscopy (NIRS) technology, which has shown great potential to predict for age quality para meters in a rapid and non-destructive way (Stuth et al., 2003).For fo rage grasses, NIRS gave pro mising results in predicting static gas volumes, while for ki netic para meters re sults were less accurate (Herrero et al., 1997;An drés et al., 2005).With respect to maize silage, NIRS prediction of fermentation ki netics has shown only moderate success so far (Lovett et al., 2004).This inability has primarily been attributed to the inhomogene ity of maize, which represents a mixture of con cen trate and roughage, i.e. a high starch content in grain and mostly fi brous components in the stover.
The objective of this study therefore was to explore the potential of NIRS for pre dic ting GP kinetics separately for ear and stover.This is not only interesting from a scientifi c point of view, but also has practical relevance, because maize stover is an important ruminant feed in many small-holder crop-livestock production systems of developing countries and maize-cob-mix is a common con cen trate in in ten sive cattle pro duction.

Feed samples
The study is based on data collected in a 3-year (2001-2003) fi eld experiment con duc ted at an experimental farm of the University of Kiel, northern Germany.Eight si lage maize varieties covering a wide range of matu ri ty groups (early to mid-late) and different maturation types were sown in early May.For further details con cer ning the crop management see Kruse et al. (2008).To obtain a broad sample popu la tion, crop samples were taken on six dates, which were chosen to be in line with developmental stages of a mid-early reference hybrid, scheduled to phenological stage of BBCH 32 (Meier, 2001), and ear dry matter contents of 20, 30, 40, 50 and 55%.On each sampling date, ten plants per plot were har ves ted by hand clipping.The plants were weighed, separated into ear and sto ver (in cluding husks), and chopped.Re pre sen tative sub-samples were oven-dried at 105°C to de ter mine dry matter con tent.Another sub-sample of both ear and stover was stored at -18°C for forage quality deter mi na tion.After freeze-drying, the samples were ground in a Cyc lo tec mill (Foss Tecator AB, Sweden) to pass a 1 mm sieve.In total, 480 samples of ears and 640 samples of stover were available for NIRS subset selection and subsequent determination of gas production and chemical composition.

Chemical analysis
Neutral detergent fi bre (NDF), acid detergent fi bre (ADF) and acid detergent lignin (ADL) were determined according to Goering and van Soest (1970;cited in Naumann and Basler, 1976), where ear samples were pre-treated with heat-stable amylase to ensure starch degradation.Nitrogen was determined in a CN-analyser (elementar, vario Max CN; Hanau, Germany) and multiplied by 6.25 to obtain crude protein (CP).Starch was analysed on the basis of an enzymatic method by Brandt et al. (1987), and water soluble carbohydrates (WSC) according to a modifi ed anthrone method as described by van Handel (1967) and McAllen (1985).

Gas production
The in vitro gas production was analysed according to Menke and Steingass (1988).Approxi mately 200 mg of each sample was placed in a glass syringe.Buffer mineral so lution was prepared along with modifi cations proposed by Liu et al. (2002).The so lution was placed in a water bath at 39°C under continuous CO 2 fl ushing.Rumen fl uid was obtained from two ruminally fi stulated German Red Pieds steers before mor ning feeding.The steers received a mixed diet of perennial ryegrass hay and con cen trates (2:1, wt/wt).The rumen fl uid taken from both animals was mixed, fi ltered through cheesecloth and fl ushed with CO 2 , as all laboratory handlings of rumen fl uid were carried out.Subsequently, the rumen fl uid was added to the buffered mineral so lu tion (1:2, v/v) and mixed.Thirty ml of the incubation medium were pi pet ted into each syringe, which was placed in a rotor (1 rpm) within an incubator at 39°C.Three blanks containing only 30 ml of medium were included in each assay, as were triplicates of standard hay and standard concentrate obtained from the Institute of Animal Nutrition, Hohenheim University (Germany).Gas pro duction was recorded after 1, 3,5,7,12,16,24,48 and 72 h of incubation.The runs were replicated three times.Fermentation kinetics were cha racterized by fi tting the data to the model of France et al. (2000) to derive curve para meters, which then were predicted by NIRS.This common approach involves two errors, one related to curve fi tting and one resulting from NIRS cali bra tion.We there fore applied an additional approach to circumvent the curve fi tting error by cha rac terizing the fermentation dynamics by the gas volume produced during specifi c in ter vals (3-5, 3-7, 5-12, 7-12, 5-16, 7-16, 12-16, 12-24, 24-72 and 16-72 h), which may be related to the de gra dation of specifi c chemical components (Cone et al., 1994;Chai et al., 2004).
Near infrared refl ectance spectroscopy of gas production kinetics NIRS prediction equations were developed separately for ear and stover for (i) the curve parameters of the model proposed by France et al. (2000), i.e. parameter A (asymptotic gas production, ml), c (fractional rate of fermentation, h -1 ), and L (lag time, h), and (ii) the gas produced in 10 different incubation intervals (3-5, 3-7, 5-12, 7-12, 5-16, 7-16, 12-16, 12-24, 24-72 and 16-72 h).To this end, all crop samples available were scanned with two replicates using a NIRSystems 5000 scanning mo no chro mator (FOSS GmbH, Germany).Mathematical treatment of the spectra was performed using the ISI-NIRS2 Version 3.1 software (Infrasoft International ® , Port Matilda, PA, USA).Absorbance was recorded as log (1/refl ectance) = log (1/R) at 2 nm in ter vals throughout the near-infrared region (1100-2500 nm) to give a total of 700 da ta points.Subsequently, samples were checked for erroneous measurements and out liers, using the option 'centre samples', which provides a ranking of the spec tral da ta on the basis of the standardized Mahalanobis distance (H) from the average spec trum.The sample population boundary was defi ned H = 3 and outliers were removed.
Samples were chosen to represent the whole spectral and chemical variability of the sample population in the calibration and validation subsets, based on the pooled 2001 and 2002 data, and extended by the 2003 samples.The H-value was used as a criterion for selecting those samples in the population as being more variable on the basis of spectra features.The option 'select samples' on the basis of H-value 0.6 was used to select calibration subsets representing the whole sample population, while the validation subsets were ran domly selected after ranking the spectral data according to their H distance.A total of 88 representative samples of ear and 210 samples of stover were selected for measuring and calibrating gas production kinetics.For validation, 53 samples of each ear and stover were selected.The fi nal number of samples included in NIRS analysis was variable due to missing values and outliers eliminated during the mathematical calibration process (Tables 2, 3 and 5).
Parameters in the following mathematical processing creating the pre dictive equations were sought to mi ni mize the standard errors of calibration (SEC) and to maximize the coeffi cients of de termination (R²) using the Modifi ed Par tial Least-Squares (MPLS) method (Shenk and Westerhaus, 1991).The minimum F statistics for terms included in the equation was 8.0.Spectral data was analysed using different mathematical treatments as provided in Table 1.Pearson's correlation coef fi cient was com puted to quantify the relation between chemical composition and gas volume pro duced during various incubation intervals.
Table 1.Mathematical treatments for calibrating gas production kinetics with approach I (parameters A, c, L according to model by France et al., 2000), and approach II (gas produc tion in different incubation intervals) Item Mathematical treatments 1,2,3,4   Scatter correction for particle size ear stover ear stover

RESULTS
Gas production kinetics from curve fi tting.Due to sequential harvesting, samples were available over a wide range of maturity sta ges, with DM content varying between 156 to 571 g kg -1 DM for ear and between 96 and 303 g kg -1 DM for stover.This was refl ected in the large varia tion of the laboratory determined kinetic parameters, which for instance ranged between 0.03 and 0.13 h -1 for the fractional fermentation rate (c) of the ear calibration subset (Table 2).Differences in kinetic Table 2. Calibration and validation statistics for the prediction of gas production kinetics by near infrared refl ectance spectroscopy (NIRS), where A is the asymptotic gas production (ml), c the fractional rate of fermentation (h -1 ) and L is the lag time (h).Range, mean and standard deviation (SD) refer to the laboratory determined values of the calibration and validation subsets DM; 4 standard deviation; 5 standard error of calibration; 6 coeffi cient of determination; relationship between NIRS-and laboratory-determined values; 7 variation coeffi cient referred to the mean (CV parameters between ear and stover were most apparent for the lag phase (L).NIRS calibration demonstrated satisfactory ability (R²>0.81) to predict gas pro duc tion parameters for ear samples, but was inferior for the stover as indicated by lower R² for asymptotic GP (A) and lag phase (L), and higher standard errors of calibration (SEC) for the fermentation rate (c) and lag phase (Table 2).However, the SEC as a absolute, difference based statistic has on ly limited suitability for comparing the accuracy of NIRS prediction across po pu la tions, if those are characterized by large differences in GP cha rac te ris tics, as for instance ear and stover, which differ substantially in the overall level of gas production and fermentation rates.We therefore additionally provided the varia tion co effi cient referring to the mean (CV M =SEC•100/mean), which should not ex ceed 10% of the mean de termined by the reference method (Hruschka, 1987).CV M -values basically confi rm the lower predictive ability for stover.The accuracy of prediction achieved in validation pointed out diffi culties with respect to the asymptotic GP and lag phase for both ear and stover, while fermentation rate could be satisfactorily predicted.The mean values, ranges and standard deviations were in a similar range for the calibration and validation set, indicating that the sample sets were comparable.
Gas production in incubation intervals.In addition to the curve fi tting approach, which showed only moderate success so far, we in vesti gated the potential to predict the GP in dif ferent incubation intervals.This is based on the assumption that the GP in specifi c intervals is closely related to the degradation of specifi c chemical con sti tuents (Cone et al., 1994;Chai et al., 2004).Incubation of ear samples re sul ted in laboratory determined gas volumes ranging between 2.5 ml 200 mg -1 DM (3-5 h) and 66.2 ml 200 mg -1 DM (5-16 h), (Table 3).Stover was characterized by a sub stan tial ly lower gas pro duc tion.The relationship between GP recorded in differ ent intervals and chemical com position was described by correlation coeffi cients provided in Table 4. Gas pro duced in intervals 3-5, 3-7 and 12-16 h of ear samples was strongly related to WSC, CP, starch, cellulose and hemicellulose, where as in stover relevant relation ships were detected for in ter vals 3-5, 3-7, 16-72 and 24-72 h.ADL content seems to have no impact on GP in ear, while high sto ver ADL reduced GP in early incubation and increased GP later on.

DISCUSSION
A comparison of the derived kinetic parameters with published values is diffi cult due to the limi ted number of studies available, which furthermore differ in crop maturity sta ge, sample pro cessing, incubation length, and fi tted models.Fer men tation of ear resulted in higher total GP (asymptote) compared to values re por ted by DePeters et al. (2007) for grain.Ear fermentation rate was consistent with the study by Carro and Ranilla (2003) on grains.For stover, we de tec ted a higher asymp totic value than Tolera and Sundstøl (1999), probab ly due to an earlier harvest in our trials.Fermen ta tion rates, however, were in agree ment.Compared to the work by Tang et al. (2006) on grain maize substantial differences be came evi dent for all para me ters.With respect to NIRS prediction of GP dynamics, the results of the present study do cu ment an insuffi cient performance for both approaches in vesti gated.Consistent to the study of Lovett et al. (2004) on maize silage, prediction of the asymptotic GP was inferior compared to the fermentation rate and lag phase.
Several factors may in fl u ence the success of NIRS equation development, such as the accuracy of laboratory de ter mined data and the errors related to NIRS calibration.Crucial to improving NIRS pre dictions is the degree of accuracy asso ciated with the reference method, since the va lidity of NIRS-predicted data will never be better than the reference me thods used for establishing the NIRS equations.Para me ters measured by biological me thods such as the gas production technique are sub ject to a multitude of error sources, which may increase variability.
With respect to the accuracy of gas production measurements, the repeatability within a run and the reproducibility between runs on different days have to be consi dered.While Getachew et al. (2002) reported a high degree of reproducibility, Van Laar et al. (2006) pointed out considerable differences among incubation runs.Pell and Schofi eld (1993), evaluating the variation of measurements over incubation time, de tected a slight decline in repeatability.In the pre sent study, however, no clear trend in prediction accuracy over incubation time became evident.
Another potential reason for poor NIRS statistics refers to unexplained or systematic errors (bias), which may be introduced when NIRS equations are used that are not pro perly represented in the calibration data set.This does not apply to the present study, where calibration and validation subsets originated from the same fi eld experiment.It may there fore be assumed that all sources of variation encountered in data samp ling and rou tine analysis were covered in the calibration as well as in the validation sub sets.With respect to the curve fi tting approach, however, a 72-h incubation period possibly was not long enough to ensure an accurate estimation of the asymptotic GP.
It is usually assumed that the calibration and validation subsets should represent the range of expected variation in the desired trait, i.e. a high degree of inho-mogeneity with respect to chemical composition of the subsets is advantageous (Stuth et al., 2003).Lovett et al. (2004), however, studying the ability to predict GP kinetics of maize silage, iden tifi ed the inhomogeneity of maize samples, which contain both fi b rous and starchy fractions, as a main error source in prediction perfor mance of NIRS.It is well known that feeds rich in non-structural car bo hy drates produce a lower acetate/propionate ra tio compared to feeds rich in fi bre, resulting in lower yields of gas volumes (Beuvink and Spoelstra, 1992).Lovett et al. (2004) regard this in direct effect of che mi cal com po sition on the short chain fatty acids (SCFA) profi le as the main cause for the in abi li ty of NIRS to predict GP kine tics.Ob viously, this effect also accounts for the results of our study.Although the crop was separated into ear and sto ver, the wide range of ma turity stages has caused a substantial inhomogeneity in che mical com position within each group (Table 4).In the case of the ear, the ra tio of vitreous to fl oury en do sperm, which increases with advancing ma turity and reduces ru mi nal starch availa bility (DePeters et al., 2007), may have affected the SCFA pro fi le and GP.Moreover, most plant car bo hydrate types ab sorb in similar spec tral regions (Deaville and Givens, 1998), which also may li mit NIRS prediction of GP for different in cubation inter vals.
These assumptions are supported by the results of the GP dynamics detected in dif fe rent incubation intervals (Tables 3 and 5).Prediction of GP was successful (R²≥0.74)only for inter vals 3-7, 7-12 and 7-16 h of stover samples, for which it may be assumed that large part of the GP resulted from degradation of one chemical constituent.With respect to inter val 3-7 h, the coeffi cient of correlation indicates that mainly WSC were fer men ted, since stover starch content is low.Easily fermentable cell wall is known to be degraded within 5-15 h incubation (Cone et al., 1994).Afterwards, less fermen table cell wall and microbial turnover are the main drivers of GP (Cone et al., 1994), as indicated by rising R-values for cellulose and hemi cel lu lose with incubation time.In ear samples, overlapping fermen tation of various com pounds, as for instance CP and WSC in early incubation, may have ham pered NIRS prediction.Coeffi cients of correlation, however, have to be interpreted with care as spurious correlations between fi bre fractions and GP may occur, caused by maturity acting as lurking variable.

CONCLUSIONS
The use of near-infrared refl ectance spectroscopy (NIRS) for predicting gas produc tion kinetics of maize silage has thus far been of limited validity, which was pri ma rily ascribed to the in homogeneity in substrate composition.Our attempts to increase homogeneity by de ve loping calibration equations separately for ear and stover showed promising results for NIRS calibration, while validation statistics were not fully convincing.For stover, predictive ability by the curve fi tting approach might be enhanced with longer incubation time.Altogether, NIRS prediction of GP kinetics remains a considerable challenge.On the one hand, a large variation in sample population is required in order to obtain representative calibrations.On the other hand, a wide variation in substrate composition and degradability, as obtained in our study by differences in crop developmental stage, seems to be a chief cause for the limited potential of NIRS to predict GP kinetics.
validation; 7 coeffi cient of determination; relationship between NIRS-and laboratory-determined values; 8 variation coeffi cient referred to the mean (CV M =SEC•100/mean or CV M =SEV•100/mean)Table 3. Calibration statistics of the gas production in different incubation intervals.Range, mean and standard deviation (SD) refer to the laboratory determined values of the calibration subset Interval, included in the calibration; 2 minimum-and maximum-volume of gas produced, ml 200 mg -1 DM; 3 mean volume of gas produced, ml 200 mg -1 Validation statistics of the gas production in different incubation intervals.Range, mean and standard deviation (SD) refer to the laboratory determined values of the validation subset Incubation interval

Table 4 .
Ranges of chemical compounds and coeffi cients of Pearson´s correlation between chemical composition and the gas volume produced in different incubation intervals, with CP: crude protein, WSC: water soluble carbohydrates, HZ: hemicellulose, Cell: cellulose, ADL: acid detergent lignin Item