A single calibration of near-infrared spectroscopy to determine the quality of forage for multiple species

Near-infrared spectroscopy (NIRS) is an efficient and chemical-free technique for quickly assessing forage quality. However, calibration curves are usually validated for the forage of a single species, while few studies have reported on the forage of multiple species. Therefore, this work aimed to develop a broad system of calibrating curves by NIRS to predict neutral detergent fiber (NDF), acid detergent fiber (ADF) and crude protein (CP) values from single and mixed forage. To accomplish this, single and mixed forage (32 forage species) were sampled over six years (2013 to 2019) from different regions of Santa Catarina state in southern Brazil. Forage samples were chemically analyzed for NDF, ADF and CP levels, followed by performing spectroscopy. Next, calibration curves were calculated as Second Derivative for NDF, First Derivative + Multiplicative Scattering Correction for ADF, and, Multiplicative Scattering Correction for CP. Approximately 200 sample forage, resulted in determination coefficient (R) values of 0.94, 0.95, and 0.98 and validation values of 0.94, 0.95, and 0.97 for NDF, ADF, and CP, respectively. Thus, calibration curves were properly developed for quality assessment of single or mixed forage for multiple species, resulting in a chemical-free and time-saving tool for routine laboratory use.


Introduction
Correct management of pastures, including forage planning, will ensure the self-sufficiency of feed supplies for livestock (Monrroy, Gutiérrez, Miranda, Hernández & García, 2017). Feed-related costs are always a significant part of variable costs for all types of livestock production. Accordingly, recent research focuses on feeding and nutritional requirements as part of pasture management (Modroño, Soldado, Martínez-Fernández & Roza-Delgado, 2017).
Nutritional values of pastures have a direct effect on livestock productivity. Accordingly, legumes mixed with grass forage may improve dry matter content and nutrient levels during the year (Barcellos, Ramos, Vilela & Junior, 2008) and may ensure sustainability for milk production systems in southern Brazil (Fernandes, 2012). Consequently, using a wide variety of forage species, including grass and legumes, single or mixed, to develop calibration curves is a good strategy to reach the required variability among pasture management, seasonality and soil fertility to ensure accuracy in predicting the quality of samples. Some precedent-setting studies have already reported a single quality calibration curve for multiple species. For example, Lobos et al. (2013) reported eight forage species from pastures over two years in Chile. Molano, Cortés, Ávila, Martens & Muñoz (2016) reported about 60 forage species from four different farms with different pasture management systems in Colombia. Parrini et al. (2018) reported 13 forage species over two years in Italy, and Norman, Hulm, Humphries, Hughes & Vercoe (2020) reported more than 100 forage species cultivated in plots from different places in Australia.
The present study aimed to develop broad calibration curves by near-infrared spectroscopy (NIRS) to predict values of neutral detergent fiber (NDF), acid detergent fiber (ADF) and crude protein (CP) from single and mixed forage.

Sampling
To conduced this research, characterized as experimental (Pereira, Shitsuka, Parreira & Shitsuka, 2018) Forages consisted of randomized samples collected from five points at each paddock using a 0.25 m2 square area.
Samples were cut close to the ground, and green mass forages were stored in a paper bag. A composite sample of approximately 500 g per paddock was obtained, weighted and dried in a forced-air oven at 55 o C for 72 h, and then ground (1 mm) in a Wiley mill.

Reference analyzes
Forage samples were also chemically analyzed, and these values were used as reference data for NIRS. Crude protein (CP) determination was analyzed by the Kjeldahl method (Association of Official Analytical Chemists [AOAC], 1995).
Neutral detergent fiber (NDF) and acid detergent fiber (ADF) determinations were analyzed according to Van Soest, Robertson & Lewis (1991) based on the pre-dried values of samples.

Calibration curves and validation
Approximately 15 g of ground forage sample were transferred to a quartz background sample holder attached to a MPA FT-NIR device (BRUKER® OPTIK GmbH, Rudolf Plank Str. 27, D-76275 Ettlingen), and spectra were performed three times with 64 different scanned points with resolution of 16 cm -1 from 4.000 to 12.500 cm -1 of wavelength. Following the manufacturer's instructions, a background was done every 20 samples.
Reference values of CP, NDF and ADF were added to the spectra of forage samples. Construction of data pretreatment and chemometric models, i.e., development of calibration curves, was performed by Opus 7.5 software using the partial least squares (PLS) model (Bjorsvik & Martens, 2001). The calibration model was adopted based on the smallest standard error of validation (root mean square error of cross validation (RMSECV) and the best value of determination coefficient (R2). Samples classified as outliers in graphs were detected and excluded from models. A sample set not included in the calibration step was used for external validation of curves.

Results
The original spectral set (Figure 1) used to develop NDF, ADF and CP calibration curves presented similarity, even considering that these spectra referred to a diversity of forage species (n = 32) among the grass and legumes collected. By applying mathematics and statistical modeling, spectral regions providing data about the analyzed nutrients were identified and selected for multivariate calibration to build regression models using PLS.   After optimization, the calibration set (Table 2) presented a determination coefficient (R 2 ) of 0.95% and a RMSECV of 2.06 (mg.g -1 ) for NDF. ADF calibration presented a R 2 of 0.96% and a RMSECV of 1.08 (mg.g -1 ), and CP calibration presented a R 2 of 0.98% and RMSECV of 0.855 (mg.g -1 ). Outliers numbered fewer than 10%, and RPD varied from 4.34 to 7.23 %. The external validation step (Table 3) was also performed with a sample set independent from the calibration set (n=20).
For this sample set, the calibration step presented a coefficient of correlation (r) of 0.94, 0.96 and 0.98 and RMSEP of 2.94, 2.07 and 1.78 mg g -1 for NDF, ADF and CP, respectively. RPD varied from 2.83 to 4.12%.

Discussion
Instrument interference and baseline deviations were found to cause errors in the spectral set used for calibration ( Figure 1). Therefore, it was necessary to apply pre-processing and spectral mitigation methods to rectify and standardize the impact of undesirable factors, but without altering spectroscopic information (Siesler, Ozaki, Kawata & Heise, 2008;Azzouz et al., 2003). In this case, mathematics can be used to process data ( Figure 2) to assist in the development of calibration curves and, moreover, improve their reliability (Neves, Soares, Morais, Costa, Porto & Lima, 2012). Also, the difference between regions of selected spectra (Figure 1) for each component is possible because absorbance is measured by different molecular links at specific wavelengths, mainly C-H, O-H and N-H, which are basic components of organic compounds from vegetable tissue (Bokobza, 2002).
A large sample range for each analyzed component is very desirable for developing a forage calibration curve (Table   1). Different species, such as grass and legumes, as well as other factors for the same species, such as geographic location, soil structure, environmental conditions, and grazing management, present variations in nutrient concentration (Kirkpinar & Açikgöz, 2018). In addition, since external validation is a recommended tool for performance evaluation of prediction models (Lobos et al., 2013), the validation set should present an amplitude similar to that of the calibration set.
To adjust parameters (Table 2), during optimization, outliers from spectral limits of NIRS, or extreme values of calibration equations, were removed from the dataset, which did not surpass 10% of total evaluated spectra. This is a relevant step since outlier detection is an important task in multivariate calibration because quality of the calibration model is determined by the quality of calibration data (Li, Xu, Wang, Du, Cai & Shao, 2016 As mentioned before, according to Kirkpinar and Açikgöz (2018), pastures present highly variable concentrations of nutrients, depending on many factors. Therefore, obtaining information about the composition of forage nutrients is essential to decision-making in pasture management and, consequently, higher animal performance. In the laboratory, NIRS can be a chemical-free and time-saving tool for routine use, and in the field, it can be a tool to quickly determine the values similar to those reported in this work in order to improve feed resource management. In this sense, NIRS has been adopted as the official method by the Association of Official Analytical Chemists (AOAC) (1984) to predict CP concentration of forages. However, the use of this technique depends on an authenticated reference method, such as the Kjeldahl method (AOAC, 1995) because NIRS must correlate spectral characteristics with values obtained by an authenticated reference method (Pasquini, 2003).
Using the external validation step (Table 3) with a sample set independent from the calibration set is recommended to predict accurate values (Manley & Baeten, 2018). Performing this step, we reached a coefficient of correlation (r) above 0.94, and RPD values were completely satisfactory because, according to the guidelines of Williams (2014), an RPD of 2.83% for NDF offers a fair screening potential, 3.52% for ADF is very good and 4.12% for CP is deemed excellent.
A study by Yang et al. (2017)  Our results demonstrated that a unique calibration curve for multiple forage species is possible. To achieve this, according to Norman et al. (2020), it is necessary to include spatial, temporal and management diversity within the dataset, especially for feed testing laboratories where the diversity of growing sites and seasonality for forages, as well as forage management regimes, would be very high. Therefore, our dataset included samples from 2013 to 2019 from farms with different pasture management at different seasons and regions from Santa Catarina State in Brazil.
Finally, it is important that NIRS calibration curves be routinely modified over time to update results (Andueza, Picard, Martin-Rosset & Aufrère, 2016).

Conclusion
Near-infrared spectroscopy can be used to develop models of calibration and validation for neutral detergent fiber, acid detergent fiber and crude protein of single or mixture forage, including grasses and legumes analyzed simultaneously. It allows the prediction of values with high correlation, and it is a chemical-free, nondestructive tool that is easy to use. It is suggested that new calibration curves be developed with a larger number of forage species, more than the 32 used in this experiment, to include as many species as possible.