Development and validation of near-infrared spectroscopy for the prediction of forage quality parameters in Lolium multiflorum

Materials

A total of 403 Italian ryegrass samples were collected from 34 accessions (15 cultivars and 19 breeding lines, Table 1) at different forage development stage and different locations (Ya’an and Chengdu) from 2014 to 2016. For each sample, a bulked strategy was applied. The fresh samples were inactivated at 105 °C for 30 min, and then oven dried at 65 °C to a constant weight. Finally, the dried samples were ground into powders through a 2 mm sieve and stored in a dry container until use to analysis CP, fiber fractions (NDF and ADF) and WSC.

Near infrared spectra (NIRS) collection

Near infrared reflectance spectroscopy analysis was performed using a Bruker MPA Fourier Transform near infrared (FT-NIR) spectrophotometer (Bruker, Bremen, Germany), equipped with a quartz beamsplitter and a PbS detector. It was also equipped with an integrating macrosample sphere and a rotating sample cup, allowing the scanning of large areas of the samples. NIR spectra of ground samples were obtained with the following procedure: aliquots of around 25 g of dried samples were placed in rotating sample cup, and scanned on reflectance mode in the spectral range from 4,000 to 12,500 cm⁻¹ at room temperature (∼20 °C). In each of the reflectance measurements, 64 scans were run and the resolution used for spectral analysis was 8 cm⁻¹. Spectrum were produced with 2,203 date points per sample (Table S1). Background corrections were made before each sample was scanned. Samples were measured in triplicate, which increased the scanned surface of samples for reducing errors. The spectral absorbance values were recorded as log1/R, where R is the sample reflectance.

Biochemical analysis

The Kennard-Stone algorithm was used to select a subset of 123 samples (out of the 403 samples) for biochemical analysis. This algorithm selects a defined number of representative samples that systematically cover the spectral variation of all samples. Samples from the selected subset were analyzed for CP, NDF, ADF, and WSC by standard wet chemical analytical techniques (Table S1). The data generated from biochemical methods will act as the reference data for future analysis. All experiments were performed with three biological replicates.

Determination of Crude Protein content

The CP content were determined by wet chemistry analysis according to the Kjeldahl method (AOAC Official Method 984.13.15) (AOAC, 1990) with the Kjeltec™ 8400 analyzer unit (FOSS, Hoganas, Sweden). The ground samples (0.5 g) were added into a 250 ml TKN digestion tube with 10 ml concentrated sulfuric acid and two digestive tablets (Beijing Jinyuanxingke Technology, Beijing, China). Blanks containing all these reagents were simultaneously processed. All tubes were digested in the preheated digestion block (Tecator™ digestor auto; FOSS, Hoganas, Sweden) at 420 °C for 90 min or until the samples were green and clear. The Kjeldahl digests procedure was carried out using a Kjeltec™ 8400 analyzer unit (FOSS, Hoganas, Sweden). The CP content was calculated using the following equation: ${\displaystyle \text{CP}\left(\text{\%}\text{DM}\right)=\frac{\left(V1-V2\right)\times C\times 1.4007\times 6.25}{M}\times 100}$ Where: V1 = volume (ml) of standard HCl required for sample; V2 = volume (ml) of standard HCl required for blank; C = molarity of standard HCl; 1.4007 = milliequivalent weight of N × 100; 6.25 = average coefficient of nitrogen conversion into proteins; M = sample weight in grams.

Determination of neutral detergent fiber and acid detergent fiber contents

The NDF and ADF contents were measured by using the methods described by Goering & Van Soest (1970) and Van Soest, Robertson & Lewis (1991) with 0.5 g pulverized samples in Automatic Fiber Analyzer (ANKOM 2000 Fiber Analyzer; ANKOM Technology, NY, USA). The contents of NDF and ADF were calculated using the following equation: ${\displaystyle \text{NDF}\left(\text{\%}\text{DM}\right)=\frac{\left(M2-\left(M1\times C1\right)\right)}{M}\times 100}$ ${\displaystyle \text{ADF}\left(\text{\%}\text{DM}\right)=\frac{\left(M3-\left(M1\times C1\right)\right)}{M}\times 100}$ Where: M = Sample weight; M1 = Bag tare weight; M2 = Weight of organic matter after extraction by neutral detergent; M3 = Weight of organic matter after extraction by acid detergent; C1 = Ash-corrected blank bag factor (a running average of the loss of weight after extraction of the blank bag/original blank bag); C2 = Ash-corrected blank bag factor (a running average of the loss of weight after extraction of the blank bag/original blank bag).

Determination of water soluble carbohydrates content

The WSC content were determined as follows: 0.1 g dried samples were ground with 1 ml distilled water, and the homogenate were transferred to a 2 ml tubes and then water bath for 10 min at 95 °C. After centrifugation at 8,000× g for 10 min at 25 °C, all supernatants were collected and the final volume was adjusted to 10 ml with distilled water. The content of WSC was detected by using assay kits (Suzhou Comin Biotechnology Co., Ltd, Suzhou, China). Briefly, 40 µl supernatants were put in a 1.5 ml eppendorf tube with 40 µl distilled water, 20 µl mix solution and 200 µl concentrated sulfuric acid. The reaction mixture was shaken and incubated in a water bath for 10 min at 95 °C and then cooled to room temperature. A total of 200 µL reaction mixture was transfered to a 96 well EIA/RIA plate and read the absorbance at 620 nm using a Thermo Scientific Multiskan™ Go (Thermo Scientific, Waltham, MA, USA).

The contents of WSC were calculated as follows: ${\displaystyle \text{WSC}\left(\text{\%}\text{DM}\right)=\frac{2.34\times \left(\Delta A+0.07\right)}{W\times 10}}$ Where: ΔA = The absorbance of test tube; W = Sample weight.

Calibration and validation of near infrared spectra models

All 123 selected samples were used to construct the near infrared spectra models in the OPUS software (Bruker, version 5.5) by using partial least square (PLS) regression throughout the process to calculate the correlation between spectral data and laboratory data. The raw spectral data were transformed by several pretreatments before the calibration process, including standard normal variate (SNV), standard multiple scatter correction (MSC), minimum-maximum normalization (MMN), first derivative (FD), second derivative (SED), straight line subtraction (SLS), constant offset elimination (COE), and a combination of FD with MSC, SLS, SNV to remove artifacts and imperfections from the date. For the validation of the models, all 123 samples were split in 93 and 30 for cross validation and external validation sets, respectively, according to the range of the chemical values. Cross-validation is conducted when developing NIRS models by using PLS regression, which attempted to screen the optimal ranks and avoid over-fitting of the data (Shenk & Westerhaus, 1991). Furthermore, external validation subsets were applied to evaluate and validate the potential accuracy of the models.

The statistical methods applied in this study included the coefficient of determination calculated in cross-validation (R²_CV) and external validation (R²_V), the root mean square error of calibration (RMSEC), the root mean square error of cross-validation (RMSECV), and the root mean square error of prediction (RMSEP). Moreover, the ratio of prediction to deviation (RPD), which indicated the correlations between the SD of the standard wet chemical analyzed data and prediction data by NIRS model (RMSECV or RMSEP) (Williams & Sobering, 1996), was applied to estimate the prediction ability of the model. Besides, the prediction error relative to laboratory (PRL) and the range error ratio (RER, calculated as the ratio between the range of standard wet chemistry values and the RMSECV or the RMSEP of the parameters) were calculated in this study, and were considered as an additional criteria for determined the prediction ability of each of the models.

Chemical characteristics of forage quality attributes

The content of CP, NDF, ADF and WSC were detected by standard wet chemical analytical techniques in laboratory and the CP expressed in %DM ranged from 4.45 to 30.6, NDF ranged from 21.29 to 60.47, ADF ranged from 11.66 to 36.17, and WSC ranged from 3.95 to 51.52% in %DM, respectively (Fig. 1). The variability of NDF and WSC contents were observed highest (±9.35 and ±9.40 %DM), while CP content was lowest (±5.71 %DM). Besides, the calibration (93 samples) and the validation (30 samples) data sets were also analyzed separately, including the values between maximum and minimum values, mean, and SD (Table 2). The variation of the CP, NDF, ADF and WSC content could be considered acceptable and broad enough for development of the aimed NIRS calibration models.

Features of NIRS spectra

The raw spectral data of 123 selected samples exhibited the general spectral features that one expects from dried plant samples (Fig. 2). It is obviously shown that peaks and valleys were presented in the spectra, which indicated the different chemical component characteristics of Italian ryegrass samples. In the wavenumber region 4,000–12,500 cm⁻¹, there were five main absorption peaks located at wavelength approximately 4,240, 4,740, 5,170, 5,800, and 6,800 cm⁻¹, respectively. After comparison to the origin of near-infrared absorption bands, we found that these wavelength related to critical functional groups as carbon atoms and hydrogen (C–H), hydrogen atoms and oxygen (O–H) and ammonia (N–H) in CP, NDF, ADF and WSC.

NIRS model calibrations and cross validation

The PLS regression of NIRS spectra and laboratory values constructed good calibration models for CP, NDF, ADF and WSC content of Italian ryegrass with the cross-validation processing simultaneously. When a wide range of possible combinations of different spectral pretreatments were tested, the optimal combinations with lowest RMSECV values were applied to construct the models. The results showed that FD+MSC combinations could well improve the linearity relation between reference and spectral values of CP, the SNV pretreatment improved the linearity relation of NDF, FD improved the linearity relation of ADF, and MMN improved the linearity relation of WSC (Table 3). Besides, we also got the optimal wavenumber based on these spectral pretreatments, including the calibration model for CP was developed at 4,247–6,102 cm⁻¹, NDF at 4,247–5,450 cm⁻¹, ADF at both 5,446–6,102 cm⁻¹ and 4,247–4,602 cm⁻¹ wavenumber, and the WSC was same as ADF (Table 3). The regression coefficients contributing for the CP, NDF, ADF and WSC models are shown in Fig. 3. In these graphs, wavelengths within the horizontal line have zero contribution to the models. The coefficients for CP model presented abundant spectral variable within the optimal wavenumber regions, which provide efficient contribution in the calibration process. The coefficients for NDF, ADF, and WSC models showed an obvious positive contribution at the peak of 4,405 cm⁻¹, which indicated that the distinct spectral region is possibly related to a C–H+O–H combination band attributed to cellulose and sugar.

Based on the constructed optimal model for each trait, a comparison of four models was conducted. The results showed that the best model with lowest RMSECV and the highest R²_CV values (RMSECV = 0.68, R²_CV = 0.99) was the CP calibration model, while the highest RMSECV and the lowest R²_CV value was obtained from the WSC calibration model, indicating that the NIRS model for predicting CP content was well suitable in Italian ryegrass (Table 3). Correspondingly, the PRL_C (RMSECV/SEL_C) ratio scales the adjusted prediction error (RMSECV) relative to the precision of the standard wet chemistry method (SEL_C). For calibrations approaching laboratory precision in this study, the value of PRL_C ranged from 0.43 (in CP model) to 1.88 (in WSC model), which the prediction errors all located whin 2 times of the standard wet chemistry precision, indicating that four models were sufficient for application. The RPD_C (SD/RMSECV) ratios ranged from 3.10 for WSC to 8.58 for CP, which represented the relationship between the natural variation of the calibration population and the prediction errors of the calibration model (Table 3). Furthermore, the value of RER were calculated in the calibration models and all the values were higher than the recommended (RER > 10) for screening purposes.

External validation of NIRS calibrations models

External validation of the NIRS calibrations models was carried out using the validation data subset (including 30 samples), which had similarly broad distribution of CP, NDF, ADF and WSC content with the calibration data set (Table 4 and Fig. 4). For all individual parameters, robust and parsimonious calibration models were identified and their predictive ability in the external validation was retained. Corresponding to the results of calibration model, the value of the determination coefficient for CP content in external validation model still kept the highest (R²_V = 0.99) compared to other parameters, which revealed the powerful prediction ability of the NIRS model for CP content in Italian ryegrass. All R²_V values of ADF, NDF and WSC model reached the acceptable thresholds of 0.90, which efficiently validated the reliability of the models for these parameters. On the other hand, the calculated RPD_P values ranged from 9.37 for CP to 3.39 for WSC, and the value of PRL ranged from 1.56 for WSC to 0.41 for CP, which could well reflect the accuracy of the models. Additionally, all the RER values (ranged from 12.03 to 35.72) in this study exceeded the recommended threshold values for screening purposes. In total, all results characterized before in external validation proved that the NIRS models constructed had an excellent quantitative ability and a powerful prediction for further Italian ryegrass forage quality evaluation in the field, especially for the CP content.