Hay samples with 227 alfalfa and 360 timothy were randomly collected from imported hay bales in different batches. Approximately 2 kg of each hay sample was divided into two portions to simulate sample preparations, such as ground (conventional forage sample preparation) or unground (minimum sample destruction) hay samples. Both portions of hay samples were dried at 55°C in a forced-air drying oven for more than 72 hours. The ground sample preparation was to mill hay samples into 2-mm particle size using a Wiley mill (Thomas Scientific, Swedesboro, NJ, USA) and then milled into 1-mm particle size using a cyclone mill. The unground hay was prepared by cutting hay samples roughly less than 10 cm long.

Hay samples milled to 1-mm particle size were used for wet chemistry analysis. Approximately 0.25 g of samples were used to determine total nitrogen concentration by Bremner [9] using an automated nitrogen analyzer (Euro Vector EA3000; EVISA Co., Ltd, Milan, Italy). The total N concentration was multiplied by 6.25 to estimate crude protein (CP) in samples. Neutral detergent fiber (NDF) and acid detergent fiber (ADF) were determined using an Ankom²⁰⁰⁰ fiber analyzer (Ankom Technology, Fairport, NY, USA) following the method of Van Soest et al [10]. The in vitro dry matter digestibility (IVDMD) was determined after the 48-h incubation of samples with a buffered-rumen fluid in an Ankom Daisy^II incubator (Ankom Technology, USA), according to Goering and Van Soest [11]. Rumen fluid was obtained from Holstein steers (Bos taurus).

Samples prepared by the conventional method (milling to 1-mm particle size) were packed evenly in a ring cup (diameter 85 mm) provided by the NIRS manufacturer (SpectraStar XT; Unity Scientific; KPM Analytics, Westborough, MA, USA) for NIRS scan. The unground hay samples were scanned using a large sample holder (diameter = 115 mm) provided by the same manufacturer. The spectrum was collected between 680 and 2,500 nm wavelength range with 2 nm interval. The absorbance was measured in reflectance (R) mode and then transformed into reciprocal natural logarithm reflectance (log 1/R), which is the first spectral pre-processing phase that results in better calibrations relative to reflectance [12]. Plots were created of log 1/R versus wavelength.

The local calibration model of NIRS for the nutrient value parameters was developed using UCal software (version 4; Unity Scientific of KPM Analytics, USA). The sample population was structured based on Mahalanobis distance at 3.0 to detect outliers in spectral data [12]. For outlier detection in the laboratory dataset, the t-statistic value was set at 2.5. This statistic indicates that a value is an outlier if the difference between predicted and observed values exceeds 2.5 standard errors. The standard normal variate and detrend were used to correct the spectral scattering effect. Detrend correction has been reported to minimize the variations in the physical properties of NIRS spectra, including particle size and environment noise such as temperature [13]. Therefore, various mathematical treatments were also applied to correct the spectral imperfections. The mathematical treatment sets such as 1, 4, 4; 1, 8, 8; 1, 16, 16; 2, 4, 4; 2, 8, 8; 2, 16, 16; 3, 4, 4; 3, 8, 8; 3, 16, 16; 4, 4, 4; 4, 8, 8; and 4, 16, 16 were employed to the spectral data in the order of derivative order, derivative gap, and smoothing. The best spectral pretreatment was chosen for the lowest standard error and the highest coefficient of determination (R²) of cross-validation. The calibration models were generated using partial least squares regression, which has been identified as a regression method that retains most of the variation in the original values and prevents the overfitting of a calibration model [6]. The maximum number of factors was set at 16, and the cross-validation statistics included R², standard error, and the ratio of prediction to deviation (RPD, calculated as the standard deviation of laboratory (observed) data divided by the standard error of prediction). The acceptable performance of the developed NIRS calibration models was determined based on R² of cross-validation (>0.80) and SECrV close to the standard error of calibration (SEC). The RPD greater than 2 was considered acceptable as calibration models [14].

Table 1 summarizes nutrient value parameters in alfalfa and timothy hays. The maximum moisture concentrations were less than 200 g/kg in hays, meaning all imported hay fell within the safe moisture range for storage without possible spoilage. The moisture ranges of alfalfa were slightly greater than timothy hay but not to a substantial degree. The alfalfa hay presented greater feed values of high CP concentration and IVDMD than timothy hay but low fiber concentrations, which agrees with Yu et al [15]. The standard deviations of the nutrient parameters were greater in alfalfa hay samples than in timothy, indicating the alfalfa quality had greater variation. Also, some hay's extremely low CP values indicate insufficient quality of the corresponding hay to support CP requirements for heifers or milking cows. The current CP, NDF, and ADF value ranges are similar to those reported for alfalfa samples collected in Spain [16]. However, the nutrient value in hay samples reflected significant differences from alfalfa samples collected from central, western, and northern New York (USA), as reported Berzaghi et al [17]. These differences could be reasoned from specific agronomic conditions such as growth stage, cutting height, genetics, and growing conditions [15]. For example, Rego et al [16] suggested differences in harvest number and growth stage as factors affecting leaf-to-stem ratio, resulting in significant variations in alfalfa CP content. Alfalfa grown under high temperatures produced lower digestibility biomass than grown under low temperatures due to greater accumulation potential of ADF and acid detergent lignin, while increasing temperatures resulted in a decreased N concentration in timothy but had an opposite effect in alfalfa [18]. Except for dry matter (DM), standard deviations were greater with the other nutrient parameters in alfalfa hay than timothy hay, indicating a possibility of more robust calibration model development from the alfalfa hay dataset.

The NIR spectra of alfalfa and timothy hays in whole or ground samples are shown in Figures 1 and 2, respectively. The hay spectra in unground hay exhibited more scattered spectrum patterns than those in ground samples. This pattern is consistent with the findings of Prananto et al [19], identifying more consistent NIRS spectral pattern of dried, ground leaf versus fresh, intact leaf. Peaks were detected at wavelengths around 1,200 nm, 1,400 to 1,500 nm, and 1,900 to 2,000 nm.

The calibration and cross-validation statistics of the nutritive value parameters are presented by unground and ground alfalfa (Table 2) and timothy hay (Table 3). The partial least squares regression of spectrum against wet chemistry data indicated higher parameter prediction accuracies of alfalfa and timothy hays when spectral data were obtained from ground sample preparation. This suggests that the NIRS detects the variations in nutrient parameters related to molecular bondings of alfalfa and timothy hays with less interruption than unground hay samples due to dense and uniform sample packing in the scanning device when samples were prepared with conventional 1 mm particle size. The more significant R² drops between the comparison of calibration and cross-validation of unground hay samples indicated insufficient calibration with the reduced sample preparation and the necessity for including more samples and some standardized packing density at NIRS scanning [20].

Except for the moisture, the R² of the nutrient parameters was equal to or greater than 0.9 in alfalfa hay. Timothy hay was above 0.9, except for IVDMD when the calibration models were developed with the spectral data acquired with ground samples. Among the nutrient parameters, moisture concentration was least predictable in alfalfa hay, while IVDMD was least in timothy. In all aspects, the nutrient parameter prediction models developed with unground hay indicated unreliable performance in prediction accuracy [6,14]. According to Stuth et al [21], the difference in forage particle size accounts for most spectral noise affecting calibration statistics. Particle size influences diffuse reflection, which is radiation reflection from a sample, and has been identified as a false principal factor in light scattering [20]. Barnes et al [22] also reported that differences in sample particle size altered surface scattering and radiation path length. Sample grinding resulted in more uniform distributions of plant materials, which probably exposed more consistent representations of chemistry in samples [12].

The current study demonstrated superior CP estimation performance in both hay species. Andueza et al [23] explained the high accuracy of CP by NIRS analysis with the increased diversity of the reference values. Another research also reported high accuracy of CP estimation (R²>0.95) due to the strong absorptions in N–H band in the NIR region [16]. As reported in [21,24], the calibration model development could be advantageous for a fitting when a wide range of CP concentrations is applied. There was also a report taking residual moisture in forages as another interfering factor with the quantification CP using NIRS. Moisture overshadowing may be prevalent at wavelengths where –N–H band vibration, a major component of proteins, may occur and potentially lower the precision of CP prediction in forages using NIRS [6]. In this experiment, the calibration statistics for CP were more accurate with ground timothy than ground alfalfa hay (R² of cross-validations, 0.77 vs 0.95). Perhaps the R² of moisture in the alfalfa calibration model may be related to the relatively lower accuracy of alfalfa CP calibration. This difference in CP calibration model performance between alfalfa and timothy hays agrees with earlier reports that confirmed the greater accuracy of the calibration statistics for grasses than legumes, primarily because of the variability in the leaf-to-stem ratio, which is also associated with variations in CP concentrations [25].

The non-uniform nature of fiber resulted from different chemical entities making up ADF and NDF in forage. Therefore, the absorption may occur across a wide range of spectral regions, potentially complicating the development of a partial least squares regression between laboratory and spectral data to define fiber fractions in forages [26]. A review of previous studies reporting forage quality prediction using NIRS indicated that fiber is one of the most estimated components after nitrogen [21,24]. Although NDF is not a simple chemical entity like nitrogen, variations in –O–H and –C–H bonds in NDF take DM of forage from 30% to 80%, which has been suggested as the main reason for the high performance of fiber prediction using NIRS even with its complicated entity [21]. In the current study, the R² of the NDF was slightly lower than ADF due to the contribution of the hemicellulose component in NDF. However, the prediction models of the parameters were above 0.9, similar to a previous study with Italian ryegrass [27].

Digestibility is a more combined parameter affected mainly through cell wall components such as NDF, ADF, and lignin concentrations in samples. Therefore, even the prediction of digestibility in IVDMD is more complex than other nutrient parameters defined by chemical constituents [21]. However, like fiber and nitrogen, IVDMD prediction using NIRS has been achieved at a reasonable level in NIRS calibration modeling based on absorptions at 2,270 nm infrared region relevant to cellulose and lignin [24]. The prediction performance of IVDMD is usually more accurate in samples with a moisture level below 8% because of potential interference from moisture in forage samples with NIRS spectra [24]. Because organic compounds in the samples may contain O–H bond, which is also related to moisture, this bond potentially affects the spectra of hydrogen bonding in samples changing band position and width [28]. Therefore, the alterations of hydrogen bonding may result in substantial changes in the related NIR spectrum.

Both characteristics of nutrient value parameters and wet chemistry analysis potentially cause errors in IVDMD prediction of forage and feeds using NIRS [20]. Since NIRS calibration is based on accurate wet chemistry analysis results, consistency of the lab analysis procedure would be critical. For example, rumen inoculum can be a source of analysis errors in IVDMD caused by donor animals (species or genotype), daily feed, and rumen fluid collection time [29]. The complication of biochemical factors such as fiber and nitrogen digestibility and anti-quality factors may also contribute to variations in IVDMD analysis. The developed IVDMD prediction models exhibited low accuracy when compared to the accuracy (R² cross-validation = 0.82) of Lundberg et al [30]. Since well-developed NIRS calibration models are possible only when accurate wet chemistry analysis results and properly collected spectral data are available, appropriate sample preparation and consistent sample scanning technique must be the critical components in forage sample analysis using NIRS [12].

As anticipated, the NIRS calibration model developed with ground alfalfa or timothy hay samples demonstrated greater R² values when compared with unground hay samples (Tables 2 and 3). Since the study did not standardize the packing density of the unground hay samples in the sample scanning container of NIRS, there must be substantial variations in the spectral data of unground hay samples. The variations in the moisture concentration in samples can be another possible error source. Although the current study’s calibration models were developed with the spectral data of hay samples collected from multi-country origins at the calibration development, the NIRS calibration for imported hay may need to include broader quality hay samples. Therefore, consideration will be necessary to include more diverse hay samples produced across different origins, years, and production batches to upgrade the calibration models in the future continuously.