Near-Infrared Spectroscopy Integration in the Regular Monitorization of Pasture Nutritional Properties and Gas Production

Abstract

Nutrition has a very significant impact on animal performance. Given the limited agricultural area of the Azores, the optimization of forage quality, quantity, and availability is key for the local livestock industry’s ability to respond to the challenges of an increasingly globalized market. This work’s goal was to evaluate the use of near-infrared spectroscopy to determine several chemical and biological parameters of pastures under the agroclimatic conditions of the Azores, and to compare its predicative ability when applied to dry homogeneous samples and to fresh inhomogeneous samples, so that we can assess the feasibility of using it to predict new samples on-site in the future. Infrared spectra of 400 fresh and dried grass samples were collected and associated with the corresponding reference values, determined through conventional methods. Mathematical models were created that established relationships between these readings and the values of the properties of interest. Predictive capacity proved especially good for crude protein, neutral detergent fiber, acid detergent fiber, ash, and dry matter, but insufficient for the biological parameters included in the study related to gas production. Near-infrared spectroscopy proved to be useable on-site as a quick, non-destructive, and cost-effective technique to monitor forage quality on a regular basis, enabling forage management and diet design optimizations.

1. Introduction

Around the globe, dairy cow diets heavily rely on pastures as the primary source of feed [1]. Also in the Azores, semi-natural pastures, resulting from the conversion of natural forests, have formed the basis of livestock farming for many centuries [2]. These pastures are typically located at medium (200–400 m) and high (>400 m) altitude, which leaves them subject to the effect of winds, precipitation, and low temperatures. This situation favors the existence of a period of higher grass production in spring and two periods of shortage in summer (especially in August–September) and winter (between November and February) [3]. More recently, in order to meet high stocking demands, intensively managed pastures located at low altitude have become dominant.

According to [2,4], intensively managed pastures located at low altitude (<200 m) are typically sown with perennial ryegrass (Lolium perenne) and white clover (Trifolium repens), although annual ryegrass (Lolium multiflorum), red clover (Trifolium pratense), and oat ryegrass (Bromus catharticus) can also be found. In the case of semi-natural pastures located at medium and high altitudes, it is frequent to find tansy grass (Holcus lanatus), as well as Poa trivialis L., Agrostis castellana Boiss. et Reut., Anthoxanthum odoratum L., Poa annua L., Trifolium repens L., Lotus pedunculatus Cav., Plantago lanceolata L., Juncus effusus L. and Rumex spp. [4].

Nutrition is among the aspects that most influence animal production [5]. Each plant presents specific characteristics in terms of nutrients, digestibility, and palatability, with these aspects impacting the consumption rate. Consequently, the choice and handling of the forage to be used influences the animal’s performance [6,7]. Close monitoring of forage quality can support the development of optimized diets that satisfy nutritional requirements [8] and pasture management improvements [7], while minimizing the costs and environmental impact [5].

Yet, these developments depend on typically time-consuming and expensive laboratory analysis, not to mention the negative effects on the environment that stem from the chemicals on which the analysis techniques depend. All this complicates the process of setting-up quality monitoring systems capable of providing timely and affordable results on a regular basis [9].

It was in this context that the application of NIR spectroscopy emerged, which combined with chemometric methods makes it possible to obtain very accurate quick estimates of several food parameters [10].

The purpose of this work was to evaluate the potential of NIR spectroscopy to predict several chemical and biological parameters of pastures of the Azores, and the effect of dry homogenous and fresh inhomogeneous samples on its predictive ability, so as to determine the feasibility of using it on-site in the scope of a close, timely, and cost-effective monitoring process.

2. Materials and Methods

2.1. Experiment Design

This study was conducted on Terceira Island, Azores. The Autonomous Region of the Azores is the outermost region of the European Union and has unique features such as the territorial discontinuity resulting from being an archipelago made up of 9 islands. The Azores are located in the northern hemisphere, between latitudes 37° and 40° N and longitudes 25° and 31° W, 1500 kilometers from Lisbon, and 3900 kilometers from North America, and form part of the bioregion of Macaronesia.

Being of volcanic origin, the soils of the Azorean islands can be characterized as andosols [11], with Terceira Island standing out for its affinity to agriculture (Figure 1), with large portions of land used for this purpose, as a result of the higher quality of its soils for this activity.

The climate of the Azores is temperate and characterized by high relative atmospheric humidity, which can reach up to 95% in high-altitude native forests. The peak rainfall happens during the months of January and February, while the lowest is usually registered in July [12]. The minimum temperature is typically reached in February and the maximum in August (Figure 2).

Pasture samples were collected over a period of 3 years to ensure the variability and heterogeneity of the samples used in the calibration procedure, a requirement for a good predictive capacity. Four hundred grassland samples were taken between the years 2019 and 2021 at different points of the island and during different seasons. Each pasture sample consisted of collecting 1kg of sample at various points on the grassland in order to be representative of the entire pasture. The pasture samples were manually collected at roughly 15 cm above the soil.

The samples consisted mostly of perennial ryegrass (Lolium perenne) and white clover (Trifolium repens), although annual ryegrass (Lolium multiflorum), red clover (Trifolium pratense), and oat ryegrass (Bromus catharticus) were also found.

The procedure for calibration always followed the same outline: after determining the parameters of interest via conventional methods, the chemometric model was created and checked with external validation. After evaluating the quality of calibration, the instrument was ready for routine analysis with unknown samples (Figure 3).

2.2. Conventional Methods

2.2.1. Determination of the Chemical Parameters

To determine the chemical parameters, a part of each collected sample was taken to be dried at 65 °C in an oven with forced air circulation until constant weight and then ground in a Retsch mill with a 1mm sieve. Analytical characterization was conducted resorting to the scheme of Weende [13] for the determination of the DM (dry matter, method 930.15), ash (crude ash method 942.05), EE (ethereal extract method 920.39), and crude protein (method 954.01), whereas the methods proposed by [14] were used to measure the NDF (neutral detergent fiber), ADF (acid detergent fiber), and ADL (acid detergent lignin).

2.2.2. Determination of Biological Parameters

Both in vitro dry matter digestibility and organic matter digestibility were determined according to the method described by [15], modified by [16].

Gas production was calculated by incubating 200 mg of each sample’s dry matter in a 100 mL calibrated glass syringe and adding 30 mL of a rumen juice and inoculant medium mixture (Menke medium mixture), mixed in a ratio of 1:2 v/v, and held in CO₂, as described by [17]. Buffer solutions were used to prepare the inoculant medium (reduced and mineral solutions). The glass syringe was then incubated at 39 ± 0.5 °C in an electrically heated isothermal oven equipped with a rotor, which rolled continuously at 1–2 rpm. The gas production measurements were taken via multiple manual observations of the syringe at 4, 8, 12, 24, 48, 72, and 96 h after the onset of incubation. Three parallel measurements were made in order to account for any differences in the composition and activity of the rumen liquor, consisting of a blank test and the incubation of a roughage and a concentrate standard.

The gas production constants used were based on the model of McDonald (1981) and fitted to the gas production kinetics curve of [18].

y = a + b(1 – e^(−ct))

(1)

where y represents the gas production at time t (in hours); a and b correspond to the gas production of the immediately soluble and insoluble fractions, respectively (mL 200 mg⁻¹ DM); and c is the insoluble fraction’s gas production rate (mL h⁻¹).

The rumen fluid used in the digestibility and gas production determinations was collected from the local slaughterhouse as described by [19] and according to the conditions specified by [20]. The rumen for each experiment was collected from 5 healthy dairy cows. Each cow was only fed ryegrass (Lolium multiflorum) and silage-based corn in the preceding days. Furthermore, the rumen fluid was collected in the first 10 min after slaughter. A cheesecloth was used to filter it and it was then preserved at 39 °C under anaerobic conditions until being delivered to the animal nutrition laboratory [21].

2.3. Statistical Analysis

Statistical analyses were performed using SPSS Statistics Software v.27 (IBM SPSS, Inc., Chicago, IL, USA). The statistical significance for the difference between the distributions was evaluated as follows: verification or not of the assumptions of normality and homogeneity of variance, using the Shapiro–Wilk and Levene’s tests, respectively. For the comparison of multiple independent groups with normally distributed data, we used one-way ANOVA, followed by an appropriate post hoc test, in this case using Duncan’s Multiple Range Test to determine which particular groups differed. Values were expressed as mean ± standard mean error (SEM), with comparisons being considered statistically significant when the p-value was lower than 0.05.

2.4. Non-Destructive Method—NIR Spectroscopy

In order to assess the predictive capacity of NIR spectroscopy applied to fresh inhomogeneous samples, each collected sample was separated in two parts: one was preserved without any processing at −20 °C (preserved part) and the other was dried at 65 °C, homogenized, and used for the determination of the reference values using conventional methods (dry part). Two independent calibrations were performed for each parameter for each preservation state so the predictive performance could be compared (Figure 4).

2.4.1. Spectra Collection

A LabSpec Pro Portable spectrophotometer from ASDinc. (Boulder, CO, USA), with an InGaAs detector, available at the Animal Nutrition and Feeding Laboratory of the Azores University, was used to collect the spectra. The collection was carried out in reflectance mode and in the 450–2500 nm range, with 50 scans per spectrum. However, the calibration was limited to the near infrared zone (780–2500 nm).

In order to collect the spectra, each sample was placed in a Petri box and processed by the spectrometer. Before and after processing each sample, a white ceramic plate was used to establish a reference. A conversion was made between reflectance (R) and absorbance (A) values. Finally, the values of the parameters being studied that were determined for each sample using conventional methods were associated with the corresponding spectrum.

2.4.2. Calibration and Validation

With a total of 400 pasture samples available, a subset of 300 samples was selected for calibration purposes, while the remaining 100 samples were designated for validation. The calibration process itself followed the same contours for all parameters under study, for both fresh and dry samples, with spectral data being collected and independent calibration models being developed for each parameter.

The calibrations were performed over the untreated spectra (log (1/R)) where several mathematical treatments and combinations of mathematical treatments were applied in order to understand which was the most advantageous (first derivative, SNV, SNV + Det, MSC).

The chosen regression method was PLS (Partial Least Squares Regression). One of its key parameters is the choice of the number of factors used. A range of factors were tested until the optimal number of factors to minimize the error was found, taking care to avoid overfitting or underfitting the model.

So as to ensure a more robust model, cross-validation was used on the calibration set. Mahalanobis distance was also applied on the calibration set to identify the outliers. Less than 10% of outliers were identified and removed.

A first evaluation of the predictive ability of the model was performed using the R² (coefficient of multiple determination of calibration), SEC (standard error of calibration) and RPD (residual prediction value, calculated on the calibration set).

After the development of the PLS model and with the choice of the best calibration equations, the external validation was performed, based on the values of R², SEC and RPD calculated over the validation set.

3. Results

3.1. Conventional Methods

3.1.1. Chemical Parameters

The calibration process begins with the identification of reference values obtained by conventional methods.

Table 1. Descriptive statistics of the different parameters studied for the pastures from the total set of samples.

	Pasture (n = 400)
	Min	Max	Mean	Sd	CV (%)
DM (%)	6.73	36.69	14.50	5.95	41.06
CP (%DM)	6.38	33.25	20.29	5.72	28.19
NDF (%DM)	33.22	86.79	63.48	7.80	12.29
ADF (%DM)	13.80	44.74	30.44	4.81	15.81
ADL (%DM)	1.10	10.94	3.56	1.80	50.52
EE (%DM)	1.00	4.17	2.55	0.64	25.11
Ash (%DM)	5.07	20.67	11.56	2.71	23.42
DMD (%)	40.27	86.04	60.78	9.48	15.59
OMO (%)	32.84	84.12	56.05	10.22	18.23

DM = dry matter, CP = crude protein, NDF = neutral detergent insoluble fiber, ADF = acid detergent insoluble fiber, ADL = acid detergent lignin, EE = ether extract, DMD = dry matter digestibility, OMD = organic matter digestibility, Min = minimum value found, Max = maximum value found, Sd: standard deviation, CV: coefficient of variation given by the formula CV = (Sd/Med) × 100.

presents the statistical characterization of the chemical composition and digestibility determination of the 400 samples. In general, the differences were found to be between the minimum and maximum expected values for virtually every parameter under study, which confirms that the chemical composition database is representative of a wide range of conditions. The variety range for each component was also found to be similar for both the calibration and validation sets.

The maximum values of NDF (86.79%DM) and ADF (20.67%DM) are especially high, while the minimum of PB (6.38%DM) is low. The mean values are close to the reference ranges. The standard deviation and coefficient of variation suggest that the pasture samples are heterogeneous, showing great variability (according to the coefficient of variation).

As the different pasture samples were collected at different times of the year (spring, summer, autumn, and winter), it was possible to analyze the effect of the season on their properties (Figure 5). The trend observed is a decrease in the values of the different parameters during the winter and an increase during the summer.

3.1.2. Gas Production

The gas production values obtained for the sample set can be observed in

Table 2. Descriptive statistics for the different parameters of in vitro gas production for the pastures.

	Pasture (n = 400)
	Min	Max	Med	Sd
Kinetics of reaction
a (mL/0.2 g DM)	−9.75	0.62	−4.18	2.10
b (mL/0.2 g DM)	19.99	94.91	46.43	7.82
c (mL/h)	0.0035	0.0538	0.0271	0.0091
Lag t (h)	0	7.2	3.52	1.26
RSD	0	46.99	1.90	2.39
Gas production (mL/0.2 g DM)
4 h	−2.44	5.82	0.48	1.26
8 h	0.59	10.25	4.57	1.85
12 h	2.94	17.32	8.21	2.74
24 h	6.64	27.54	16.88	4.63
48 h	12.07	42.53	27.59	5.90
72 h	14.77	45.61	33.30	5.85
96 h	16.23	54.05	36.54	5.61

Min = minimum value, Max = maximum value, Sd = standard deviation; a = gas production of the immediately soluble fraction (mL/0.2 g DM), b = gas production of the insoluble fraction (mL/0.2 g DM), c = gas production rate constant for the insoluble fraction (mL/h); Lag t = time it takes to produce gas (h); RSD = residual standard deviation.

. In general, varied results were obtained. The values obtained for 𝑎 lie in the range between −9.75 mL/0.2 g DM and 0.62 mL/0.2 g DM, while for 𝑏, which also shows high standard deviation, they ranged between 19.99 mL/0.2 g DM and 94.9a mL/0.2 g DM.

Globally, the pastures were, on average, responsible for the cumulative production of 36.54 mL/0.2 g DM after 96 h, also showing a variability of no more than 5.61.

The distribution of the different mean values determined for the fermentation kinetics curves are shown in the following graph (Figure 6). It can be observed that the gas production was slightly higher during the spring.

3.2. Non-Destructive Method—NIR Spectroscopy

3.2.1. Spectra Collection

After analyzing the samples using conventional methods, the spectra of every sample were taken for both the fresh and dried parts of each sample (Figure 7).

It can be observed that the fresh sample readings have a higher amplitude than those observable in the dry samples. This can be explained by the higher noise induced by light scattering as a result of the inhomogeneity of the samples (dry samples are milled and therefore homogenized and fresh samples are not). This characteristic will tend to be minimized by the mathematical treatments applied at the time of calibration. Overall, good homogeneity can be observed between the calibration and validation sets.

3.2.2. Calibration

The statistical values that were used to measure the calibration quality are summarized in

Table 3. Mathematical treatments and statistical performance of calibration and validation for the pastures (dry and fresh) for the different chemical parameters and digestibility.

		Calibration (n = 300)				Validation (n = 100)
Parameter	State	Mathematical Treatment	F	R2c	SEC	R2v	SEP	RPD	SEP/SEC
DM (%)	Dry	SNV	12	0.85	0.75	0.79	1.72	3.46	2.29
	Fresh	SNV + Det	8	0.9	1.12	0.87	1.75	3.4	1.56
CP (%DM)	Dry	SNV	6	0.98	0.3	0.9	0.42	13.62	1.4
	Fresh	SNV + Det	10	0.87	0.23	0.85	0.98	5.84	4.26
NDF (%DM)	Dry	MSC	8	0.95	1.12	0.93	1.46	5.34	1.3
	Fresh	MSC	8	0.88	1.72	0.87	1.65	4.73	0.96
ADF (%DM)	Dry	SNV	8	0.93	0.75	0.91	1.35	3.56	1.8
	Fresh	SNV + Det	7	0.77	1.25	0.75	1.4	3.44	1.12
ADL (%DM)	Dry	SNV	8	0.8	0.3	0.79	0.46	3.91	1.53
	Fresh	SNV	7	0.79	0.63	0.78	0.51	3.53	0.81
EE (%DM)	Dry	SNV	6	0.55	0.12	0.55	0.25	2.56	2.08
	Fresh	SNV	10	0.63	0.18	0.62	0.23	2.78	1.28
Ash (%DM)	Dry	SNV	8	0.94	0.44	0.92	0.54	5.02	1.23
	Fresh	SNV + Det	8	0.87	0.71	0.84	0.6	4.52	0.85
DMD (%)	Dry	SNV	7	0.85	2.12	0.82	2.51	3.78	1.18
	Fresh	SNV	8	0.76	2.8	0.75	3.2	2.96	1.14
OMD (%)	Dry	SNV	11	0.78	3.02	0.74	3.87	2.64	1.28
	Fresh	SNV	6	0.65	2.27	0.62	3.1	3.3	1.37

DM = dry matter, CP = crude protein, NDF = neutral detergent insoluble fiber, ADF = acid detergent insoluble fiber, ADL = acid detergent lignin, EE = ether extract, DMD = dry matter digestibility, OMD = organic matter digestibility, SNV = normal standard deviation, SNV + Det = SNV + “detrending”, MSC = multiplicative sign correction, F = number of factors of the PLS model, R²c = coefficient of multiple determination of calibration, SEC = standard error of calibration, R²v = coefficient of multiple determination of validation, SEP = standard error of validation, RPD = residual prediction value (SD/SEP).

and

Table 4. Mathematical treatments and statistical performance of calibration for the different pastures (dry and fresh) for the different parameters of in vitro gas production.

		Calibration (n = 300)				Validation (n = 100)
Parameter	State	Mathematical Treatment	F	R2c	SEC	R2v	SEP	RPD	SEP/SEC
a (mL/0.2 g DM)	Dry	SNV	15	0.68	2.25	0.65	0.98	2.14	0.44
	Fresh	SNV	10	0.65	2.52	0.58	0.87	2.41	0.35
b (mL/0.2 g DM)	Dry	SNV	8	0.43	1.5	0.36	5.1	1.53	3.4
	Fresh	SNV + Det	9	0.5	1.23	0.45	4.48	1.75	3.64
c (mL/h)	Dry	SNV + Det	8	0.53	1.89	0.46	0.005	2	0.0026
	Fresh	SNV	7	0.48	1.56	0.41	0.007	1.43	0.0045
Lag t (h)	Dry	SNV	6	0.78	1.87	0.72	0.52	2.42	0.28
	Fresh	SNV	4	0.73	1.69	0.66	0.61	2.07	0.36
4 h	Dry	SNV + Det	3	0.67	1.21	0.61	0.98	1.29	0.81
	Fresh	SNV + Det	4	0.62	1.12	0.55	1.51	0.84	1.34
8 h	Dry	SNV + Det	5	0.63	1.14	0.56	0.98	1.89	0.86
	Fresh	SNV	4	0.61	1.76	0.57	1.7	1.09	0.96
12 h	Dry	SNV + Det	7	0.77	1.75	0.71	1.94	1.42	1.11
	Fresh	SNV + Det	7	0.65	1.78	0.58	2.01	1.37	1.13
24 h	Dry	SNV	5	0.8	1.48	0.73	1.6	2.89	1.08
	Fresh	SNV	5	0.74	1.63	0.67	1.72	2.69	1.06
48 h	Dry	SNV + Det	6	0.67	3.21	0.6	3.84	1.54	1.2
	Fresh	SNV + Det	6	0.63	3.98	0.58	4.34	1.36	1.09
72 h	Dry	SNV + Det	4	0.74	1.73	0.67	2.15	2.72	1.24
	Fresh	SNV + Det	3	0.62	2.42	0.55	2.63	2.22	1.09
96 h	Dry	SNV + Det	4	0.87	1.76	0.79	1.87	3	1.06
	Fresh	SNV	6	0.72	1.83	0.65	1.98	2.8	1.08

a = gas production of the immediately soluble fraction (mL/0.2 g DM), b = gas production of the insoluble fraction (mL/0.2 g DM), c = gas production rate constant for the insoluble fraction (mL/h); Lag t = time it takes to produce gas (h); SNV = normal standard deviation, SNV + Det = SNV + “detrending”, F = number of factors of the PLS model, R²c = coefficient of multiple determination of calibration, SEC = standard error of calibration, R²v = coefficient of multiple determination of validation, SEP = standard error of validation, RPD = residual prediction value (SD/SEP).

, alongside the mathematical treatment that was found to provide the best results. The corresponding external validation performance is also presented.

The results obtained for most of the parameters in this study can be considered good, with the calibration equations showing greater capacity to accurately predict CP and NDF values (Table 3) for both fresh and dry samples. For these two parameters, R² values were always equal to or greater than 0.87 (reaching 0.98 in the case of CP) and the RPD varied between 3.08 and 13.62. ADF predictions also proved to be particularly good according to these statistics.

The calibration models (Table 4) did not prove capable of predicting gas production with great accuracy, with the best result being obtained for fresh pastures, with values of 0.87 and 3.00 for R² and RPD, respectively, although, overall, and considering the average values of R² and RPD, it can be observed that the best predictive ability was achieved for the dry samples.

4. Discussion

Whether used directly or in the form of silage, the predominant source of dairy cow feed worldwide is pastures [1,22,23].

Pastures have low production costs, are easy to use, and are the most efficient sources of nutrients for ruminants. This leads to ruminant production systems in the world being forage-based, with feed obtained predominantly from pastures [24].

However, the composition and nutritive value of pastures are extremely variable [25]. Given how feed directly influences the productive and reproductive performance of animals, optimizing the design of diets and pasture management can lead to relevant commercial gains. A timely understanding and monitoring of pastures can help with this effort [23].

Looking at the geographical distribution of the useful agricultural area of Terceira Island, also represented in Figure 1, it is possible to observe how it is mostly distributed in low- (intensively managed) and medium-altitude areas. Production in the medium- and high-altitude ranges tends to be carried out in semi-natural pastures, characterized by the presence of endemic plant species, which are maintained because grazing is less intensive in these areas and the introduction of fertilizers is limited [2,4].

4.1. Chemical and Biological Parameters

Close monitoring of pasture quality can support the design of a diet that meets animal nutrition needs [8], improves pasture management [7], and simultaneously minimizes its cost and environmental impact [5].

Until the 1860s, the way to assess the nutritive value of a feed was based on the concept of a hay equivalent—a subjective method that used standard hay as a reference [26]. In 1865, Wilhelm Henneberg and Friedrich Stohmann defined a chemical system of analysis to quantify the nutritional value of foods, which became known as the Weende System [26] and is still used today. Over the years, researchers have sought to establish more coherent systems for fractioning the organic matter of plant foods, particularly the carbohydrate fraction. However, a great qualitative leap forward would only come in the 1960s, when Peter van Soest proposed a revolutionary approach based on a system of detergent solutions for food analysis, and Tilley and Terry suggested in vitro methods for the determination of feed digestibility. At the end of the second half of the 20th century, NIR spectroscopy came to offer a complement to chemical analysis in the laboratory, allowing us to obtain several potentially complex indicators in a fast and non-destructive way [23].

The variations observed in the chemical parameters of pastures can be the direct result of variability in their components (botanical composition), growth stage, climatic and management factors, and soil type, among others [27].

The mean and coefficients of variation (Table 1) of DM, CP, NDF ADF, ADL, EE, ash, DMD, and OMD were 14.50% (41.06%), 20.29%DM (28.19%), 63.48%DM (12.29), 30.44%DM (15.81%), 3.56%DM (50.52%), 2.55%DM (25.11%), 11.56%DM (23.42%), 60.70% (15.59%) and 56.05% (18.23). Moreover, [28] reports similar mean values to these, in a study of the nutritional quality of permanent pastures over 11 years, as well as variation between for DM, CP, and NDF. In another study, when evaluating different pasture renewal strategies, [29] found concentrations between 58.3% and 78.3% for OMD, similar to those observed in this work. In addition, [30], in a study to build NIRS calibration models to evaluate the nutritional quality of pastures, found DM (15.1%), CP (23.4%), NDF (44.1%) DMD (83.5%) and OMD (73.2%).

Since the different pasture samples were collected at different times of the year (spring, summer, fall, and winter), it was possible to analyze the effect of the season on their properties (Figure 5). The observed trend is a decrease in the values of the different parameters during the winter and an increase in the summer. This is explained by the existence of edaphoclimatic factors that cause pasture growth to be faster in the summer, which, in turn, leads to a higher energy consumption and, consequently, to the existence of a higher amount of fiber in relation to the cell content. For DM, a significant difference (p < 0.05) can be observed across the seasons, whereas, for CP no significant difference (p < 0.05) was found between spring and summer, although there was an increase during autumn and winter. NDF showed lower values for samples collected during the winter as well as ADF and ADL. EE values peaked in the winter with significant differences (p < 0.05) between this and the other seasons. The ash shows significant differences (p < 0.05) between summer/spring and fall/winter, probably due to contamination from adverse weather conditions. The trend in DMD and OMD was an increase during the fall and winter (peaking in the fall), with significant differences (p < 0.05) between the spring/summer sets compared to the fall/winter sets.

The amount of gas produced in in vitro fermentation reflects the extent of fermentation and digestibility of forage [31] and is directly proportional to the rate at which the substrate is degraded [32]. According to [33], during the initial incubation period, fermentation of the fast fermentable soluble fraction of the substrate (i.e., soluble carbohydrates) and microbial protein synthesis take place. Once this phase is over, fermentation of the insoluble but potentially degradable components, such as the NDF fraction, begins.

The fermentation kinetics were described based on the model of [34]. Table 2 presents the in vitro gas production kinetics. As can be observed, with the exception of 72 h, the cumulative gas production values were significantly different (p < 0.05) at all incubation times.

The 𝑎 constant ranged between −9.75 mL/0.2 g DM and 0.62 mL/0.2 g DM. A positive 𝑎 is an indicator that the component has begun to degrade rapidly, while a negative value indicates that there was an initial phase in which no cell-wall degradation occurred, called the lag phase [35], and this was the behavior observed for the overwhelming majority of samples.

The constant 𝑏 of the reaction kinetics, in turn, varied between 19.99 mL/0.2 g DM and 94.91 mL/0.2 g DM for the pastures.

In the case of the constant 𝑐, it was found to vary between 0.0035 mL/h and 0.05 mL/h.

Lag t, on the other hand, which indicates the time that elapsed before gas production began, ranged between 0 h and 7 h 20, values in line with those expected given the corresponding 𝑎 value.

Figure 6 presents the cumulative gas production between seasons for the pastures, with the highest cumulative gas production in spring being evident.

4.2. Spectra Collection

Agriculture has benefited from NIR spectroscopy as an efficient tool for determining a large number of parameters and criteria, with unquestionable advantages as a universal technique that is fast, non-destructive, uses no chemical reagents, is easy to use, given the flexibility in the way the sample is placed in the apparatus [36], and that allows both qualitative (for identification) and quantitative (for dosage) assessments [37]. The technique is currently considered essential in several seemingly disparate fields, ranging from medical applications to food analysis, among many others [23]. Furthermore, the available instruments even enable the application of this technique outside the constraints of the laboratory, making it possible to obtain immediate information on-site, which enables more timely and informed decision making [38].

Defining a model to apply NIRS analysis to samples consists of the following steps: 1. Collection of spectra and reference data; 2. pre-processing of spectral data; 3. definition and validation of calibration equations; 4. use in routine analysis (Figure 3).

The calibration process started by collecting the spectra of all samples, both for the calibration set and the validation set (Figure 7) and found that they present, on average, a contour consistent with that presented by [39].

The average spectral profiles of the different forages are shown in Figure 7. These profiles exhibit a similar pattern across the tested range, which suggests the various samples have similar components.

Additionally, one can note how the dried samples are much more uniform than the fresh samples. This happens because the NIR spectra, in addition to containing information about the chemical composition of each sample, also relate some physical properties with the surface morphology and refractive index affecting the scattering properties of solid materials. Consequently, inhomogeneous particles with, e.g., significant variations in their size, compaction, or surface finish can result in a misalignment of the baseline [40], which can induce significant differences between the spectra. Naturally, the dried and ground samples are much more homogeneous than the corresponding fresh samples, which leads to the baseline being closer. This highlights the importance that using the appropriate mathematical pre-treatments has in obtaining consistent results, especially when dealing with untreated samples.

For the collected spectra (Figure 7), it can be observed that the most intense bands are located at around 1450 nm, due to the first O-H overtone, and 1950 nm, due to combinations of O-H vibrations, which reflect the sample’s moisture content [41]. These readings are consistent with the ones reported by [42,43], who report peaks at 1450 nm and 1970 nm. It should be noted that these bands have higher intensity in the fresh samples than in the dried samples. The fat content is reflected in the peaks around 1198 nm, associated with the second C-H overtone, and the peaks at 2266 nm and 2430 nm, which emerge from the combination of vibrations. This is also consistent with [42], who observed these peaks at 2312 nm and 2352 nm. The less intense peak, situated at around 1685 nm, is assigned to the first C-H overtone and directly correlated to protein content. An absorption band can also be observed at around 1500 nm (N-H first overtone), which is related to protein, and is more pronounced in the dry samples.

4.3. Calibration Models

The calibration models were built using the Partial Least Squares Regression (PLS) algorithm [44]. To achieve the best possible calibration, several alternative models were tested using different combinations of pre-treatments. It was found that, in general, the SNV pre-treatments gave the best results with fresh samples, which is consistent with the fact that these treatments are especially good at removing noise related to diffuse reflectance, which the inhomogeneity of fresh samples provides. The assessment of the predictive ability of the model was carried out using the R² (calculated over the calibration set), the SEC, and the RPD. In this study, a reliable prediction for a given parameter was only considered possible when the corresponding coefficient of determination R² was equal to or greater than 0.9 and the RPD values were simultaneously equal to or greater than 3 [44]. Note that the combination of a high R² and a high error can be understood as an indicator of weakness in the predictive capacity of the calibration, which is usually reflected in the corresponding RPD. In this work, the prediction accuracy (SEP) was found to be very similar to the calibration error (SEC). This suggests heterogeneous calibration and validation sets were used and a robust calibration model was achieved.

The obtained results indicate that the calibration equations can predict with great accuracy the values of PB, NDF, ADF and CB in the pastures studied, both with fresh and dry samples. Of the main indicators of forage quality, the only one for which the obtained equations do not show optimal results was DM. The results for the biological parameters were noticeably worse than those obtained for the chemical parameters, and no calibration equations with good predictive ability were obtained, which is consistent with what is reported in the literature for pastures.

Since the end goal of this calibration is its use in the field, so as to make the most of NIRS as a timely monitoring tool, the fresh samples were purposely not subjected to any previous homogenization step, making the calibration more robust and allowing for greater flexibility in sample preparation, as mentioned by [45].

To facilitate the reading of the above, the SEC value will be presented in parentheses next to the corresponding R² value, when applicable.

In the case of DM (Table 3), the R²c (cross-validation) obtained via the best calibration method for dry samples was 0.85 (0.75%), with the SNV treatment, which corresponded to an RPD of 3.46, while for fresh samples, the SNV + Det pre-treatments resulted in an R²c of 0.9 (1.12%), with an RPD of 3.40. By comparison, [46] found an R²c of 0.98 (7.50 g/Kg), with RPD of 7.15, for fresh samples, values much higher than those achieved in this work, but with a cross-validation error of 7.5, also much higher than that found in this study. Ref. [30], on the other hand, reported an R²c of 0.94 (1.13%) and an RPD of 3.7 for fresh pastures, values very similar to those found in this work.

For CP (the set where we obtained the best results) (Table 3), the R²c for the dried samples reached 0.98 (0.3%DM), with an RPD of 13.62 (best RPD value), while with the fresh samples, an R²c of 0.87 (0.23%DM) and RPD of 5.84 were obtained. For both the dried and fresh samples, the results found for CP were robust. These results are in line with those obtained for dry pastures by [47], who obtained an R²c of 0.90 (0.57%), and by [39], who achieved an R²c of 0.99 (0.95%DM). In the case of fresh pastures, [30] achieved an equal R²c of 0.87 with a higher SEC (2.04%) and a lower RPD (2.0). [46]; on the other hand, they reported an R²c of 0.93 (16.76 g/kgMS) for fresh pastures with an RPD of 3.69.

For the prediction of NDF (Table 3), an R²c of 0.95 (1.12%DM) and RPD 5.34 were obtained in dry samples, while for the fresh samples, the R²c was 0.88 (1.72%DM) and RPD of 4.73, these being the best results obtained for this parameter. For ADF, the R²c was 0.93 (0.75%DM) for an RPD of 3.56 in the dried samples, while in the fresh samples, it was 0.77 (1.25%DM) for an RPD of 3.44. ADL, on the other hand, shows an R²c of 0.8 (0.3%DM) and RPD of 3.91 for the dried samples, while for the fresh samples, it was 0.79 (0.63%DM) and 3.53 RPD. These results are better than those obtained by [39] for the prediction of NDF in dry samples (R² of 0.95 (2.79%DM)) and [46] for fresh samples, where they achieved an R² of 0.80 (33.58 g/KgMS) and RPD of 2.22 for NDF and 0.90 (13.96 g/KgMS) with an RPD of 3.20 for ADF. The values achieved by [48] with fresh natural pasture samples are lower, reporting R² values of 0.78 (1.63) and RPD 1.81 for ADL, 0.84 (2.60 g/KgMS) and RPD 2.76 for ADF, and 0.84 (3.49 g/KgMS) and RPD 1.76 for NDF, still maintaining the trend of better results for the latter two than for ADL.

In EE (Table 3), the R²c for dry samples was 0.55 (0.12%DM) and for an RPD of 2.56, while for fresh samples, it was 0.63 (0.18%DM), for an RPD of 2.78. These values are in line with those obtained by [48], who report an R² of 0.54 (0.26 g/gMS) and RPD of 1.27 for fresh natural pasture samples.

In the case of ash (Table 3), an R²c of 0.94 (0.44%DM) and RPD 5.02 were achieved for dry samples, while the fresh samples show an R²c of 0.87 (0.71%DM) and RPD 4.52. These indicators are substantially better than those presented by [48], who obtained an R² of 0.73 (0.91 g/KgMS) and RPD 2.12 for fresh pastures.

As shown in Table 3, the coefficient of determination of the calibration equations for predicting DMD in dry samples was 0.85 (2.12%), with an RPD of 3.78, while for fresh samples, it was 0.76 (2.8%), for an RPD of 2.96. The OMD in the dried samples shows an R²c of 0.78 (3.02%) and RPD of 2.64, while for fresh samples, the R²c was 0.65 (2.27%) and RPD 3.30. By comparison, [39] achieved better results for OMD prediction in dry samples, with an R² of 0.90 (3.37%).

There are not many papers that explore the application of NIR spectroscopy to predict parameters related to gas production in grasslands, and those that exist usually present poor results, such as those obtained in this work. We can note, however, that reasonably robust results were found for predicting gas production in the set of dry samples (Table 4) at 24 h (R²c of 0.80 (1.48 mL/0.2 g DM) and RPD of 2.89) and 96 h (R²c of 0.87 (1.76 mL/0.2 g DM) and RPD of 3.00).

Regarding the gas production kinetics, the calibration quality indicators for the parameters under consideration were below the target threshold, suggesting the calibration was generally not good enough for practical use. The best results were achieved using SNV + Det pre-treatment, which is in line with the studies in rumen degradation carried out by [49,50]. This work’s results compare with those achieved by [51], who found low cross-validation coefficients of determination (R² between 0.60 and 0.78) for pastures, [52] and who reported an R² between 0.60 and 0.80 in predicting the parameters of gas production kinetics in a study on corn silages. In contrast, [53] found that NIR spectroscopy could be used to accurately predict gas production for some legumes incubated in vitro, reporting an R² of 0.95 (7.2 mL/g).

5. Conclusions

In addition to being a method that can simultaneously determine several qualitative and quantitative parameters of pasture samples, in a fast, inexpensive, non-destructive, and consistent way, NIR spectroscopy can be used with portable devices that allow for these analyses to be conducted directly in the field, enabling the retrieval of immediate on-site information, which facilitates more timely and informed decision making.

From this study, it was established that it is possible to apply this technique in situ to inhomogeneous fresh pasture samples to accurately predict parameters such as CP, NDF, ADF and crude ash, without any significant loss in accuracy compared to using homogenized dry samples. It was also possible to obtain good approximations for other parameters, such as DM.

On the other hand, although the results were generally better than those found in the literature, sufficiently good calibrations were not obtained for the biological parameters, namely those related to gas production, with either fresh or dried samples.