1 Introduction

The quality of fresh forage, together with biomass availability, is a critical factor in pasture-based production systems, significantly influencing animal performance, stocking rates and methane emissions, as well as the health, well-being, and safety and quality of the products they provide (Bewley 2010; FAO 2010; Shorten et al. 2019). Due to its essential role in sustaining productivity and profitability in meat and milk production, understanding the variability of nutritional quality is crucial for effective animal feeding management (Rivero and Lee 2022).

Globally, livestock production systems heavily depend on cultivated or naturally occurring pastures, as well as temporary and permanent meadows, which serve as the primary feed source for animals (Bernués et al. 2011). In 2022, the global area of permanent meadows and pastures was estimated at 3208 million hectares (approximately one-quarter of the Earth's land surface) supplemented by an additional 300 million hectares used as temporary meadows and pastures within crop rotation systems (FAO 2024). Despite the widespread use of pasture, ensuring consistent forage quantity and quality remains a complex challenge. Variability in forage quality arises from factors such as plant species composition, phenological stage, weather conditions, grazing intensity, and plant health, all of which affect the nutritional profile of the pasture (Wróbel et al. 2023). In fact, key nutritional components (e.g., crude protein, fibre, minerals, ash, digestibility of organic matter, sugars and metabolisable energy) are shaped by a range of environmental and management variables, including species diversity, plant maturity, soil characteristics, topography, and local climate conditions, among others (Holmes 2007; Pullanagari et al. 2012). Additionally, grazing management practices are a key issue for preserving pasture quality over time and maintaining its ecological functions (Teague and Kreuter 2020). Overgrazing can degrade plant diversity and soil health, reducing forage yield and nutritional value, whereas underutilisation may lead to biomass accumulation and reduced palatability (Dunn et al. 2010). Climate change further compounds these challenges, with shifting precipitation patterns, rising temperatures, and extreme weather events altering pasture productivity and composition. This variability highlights the need for meticulous and targeted pasture management to optimise animal nutrition. In this context, feedstuff analysis emerges as a crucial tool for assessing the quality of grazed forages. Therefore, it is fundamental to determine, ideally with a monthly sampling frequency, the nutritional value of pasture to implement alternative or complementary grazing/feeding practices (e.g., supplementation) to meet animal requirements (Martin et al. 2020), leading to both cost savings and a reduced environmental impact (Shalloo et al. 2018).

While conventional methods (i.e., wet chemistry) have been employed for determining the chemical composition of feedstuff, they pose significant limitations due to their cost, time-consuming nature and unsuitability for real-time analysis (González et al. 2018; Parrini et al. 2018). These factors not only increase production costs but also hinder the development of efficient quality monitoring systems. The slow turnaround time of traditional methods makes it difficult to obtain timely results that are necessary for dynamic adjustments to feeding practices in response to pasture variations in both space and time (Molle et al. 2008). Moreover, their reliance on expensive equipment, skilled personnel and chemical reagents contributes to environmental concerns (Samadi et al. 2018; Evangelista et al. 2021). Consequently, there is an increasing interest in developing alternative analytical methods that provide faster, more cost-effective, and environmentally sustainable solutions, in line with the broader objective of improving agricultural efficiency and sustainability. In this context, it is important to consider additional aspects. Specifically, the analytical technique must operate as a multi-analytical tool, allowing the simultaneous acquisition of multiple results in a single step while minimising the use of wet-chemistry-based methods, except when their use is indispensable. In this context, while near-infrared (NIR) spectroscopy may not achieve exceptional accuracy across all measured parameters, it serves as an effective screening tool to identify samples that require further laboratory analysis. By distinguishing between samples that necessitate additional testing and those that do not, NIR can significantly reduce the overall laboratory workload.

From this perspective, NIR spectroscopy enables the rapid estimation of several feed quality parameters (Danieli et al. 2004; Cheli et al. 2012). The integration of Fourier transform near-infrared (FT-NIR) spectroscopy with chemometric models may offer high accuracy and precision in predicting the nutritional parameters of pastures (Iqbal et al. 2023). This approach allows for real-time feed analysis, providing data that are critical for adaptive pasture management. The implications of such advancements are substantial for improving animal nutrition, welfare, and overall production efficiency. By leveraging these tools, pasture-based systems can achieve greater sustainability and resilience, addressing the dual challenges of feeding a growing global population and mitigating the environmental impacts of livestock production (Dwivedi et al. 2023; Draghi et al. 2024).

Given the need for rapid and reliable fresh forage quality assessment tools, this study aims to evaluate the performance of FT-NIR spectroscopy in predicting key nutritional parameters of pasture samples. Specifically, we compare different preprocessing techniques and chemometric approaches to optimise model accuracy and robustness. Additionally, the present study aims to identify the most relevant spectral bands to simplify the model complexity as a crucial step for reducing the cost associated with the development of a field prototype. In summary, the ultimate goal is to facilitate the adoption of FT-NIR as a practical tool for routine forage analysis, enabling real-time pasture management and improved decision-making in livestock production systems.

2 Materials and Methods

2.1 Green Forage Sampling and Sample Preparation

The study was conducted on three beef cattle farms located in central Italy, Latium region (Figure 1), where pastures were composed of alfalfa meadows, permanent meadows, and mixed grasslands. The grass/legume ratio varied from 10:90 to 80:20, reflecting the diversity of forage composition across the sampled areas.

To ensure representative sampling across different biomass conditions, satellite-derived data were employed. Normalised difference vegetation index (NDVI) values were obtained from the most recent available Sentinel 2 A and 2B satellite images (10 × 10 m resolution), free of cloud cover before each sampling event, and processed using ArcGIS Pro 3.2.0. These NDVI values, ranging from 0.25 to 0.75, were used to stratify the study area and guide the selection of sampling locations, ensuring variability in biomass availability, phenological stages, and, consequently, nutritional quality characteristics. This approach improves the accuracy of the pasture quality assessment (Magney et al. 2016). The lower end of this range corresponds to sparse vegetation or early growth stages, while the upper end reflects denser and more vigorous plant cover. Following the selection of the NDVI range, spatial variability was analysed to identify distinct zones representing differences in phenological phases and vegetation conditions. To ensure an even distribution of potential sampling locations across the NDVI gradient, the range (0.25–0.75) was divided into nine equal-width intervals, each spanning 0.055 units. Within each interval, sampling locations were randomly selected to represent the corresponding vegetation conditions, ensuring balanced coverage of biomass variability. This stratified approach reduced the risk of under- or over-representation of specific NDVI values, thereby maximising the representativeness of the dataset. Sampling points for each NDVI interval were selected within the Sentinel grid, and the geographic coordinates of the centroid of each grid cell were recorded for accurate field navigation. To account for seasonal changes, sampling was conducted at different phenological stages, capturing the full range of biomass conditions observed throughout the growing season.

Field sampling followed a geo-referenced, plot-based approach (Primi et al. 2016), using 1 × 5 m plots where herbaceous biomass was harvested. Sampling was carried out under dry conditions from May 2023 to May 2024, with 9 samples collected from 2 farms on each sampling day between May 2023 and February 2024, while a third farm was sampled from March 2024 to May 2024. In total, 324 samples were collected throughout the study period (18 sampling days). Fresh biomass weights were recorded in situ using a Kern HDB dynamometer (accuracy: ± 5 g; Kern & Sohn GmbH, Germany), and cuts were made 2 cm above the ground using a double-bladed shear.

From each plot, approximately 500 g of fresh forage was placed in a sealed plastic bag, with air removed, and stored in a cool box with ice packs for transport to the laboratory. Upon arrival, samples were stored at 4°C in a cold room (Dale et al. 2017) and processed within 24 h. Finally, the material was chopped into approximately 1-cm pieces using sharpened knives, brought to room temperature, and subsequently analysed using NIR spectroscopy.

3 Results and Discussion

Data from wet chemistry analysis of forage quality attributes and spectra from FT-NIR spectroscopy constitute the two datasets for the development of PLS prediction models. Descriptive statistics of data from wet chemistry analysis are reported in Table 2 and are in line with pasture composition. The extended DM range from 7.57 to 90.56 g 100 g⁻¹ FW could be an effect of fluctuations in forage composition in terms of species, phenological stage, phytosanitary status and environmental condition. The high mean value of DM, close to 30% on FW (29.53 g 100 g⁻¹ FW), coupled with the high variability (SD = 15.68 g 100 g⁻¹ FW), underscores the effect of seasonality, particularly during summer droughts. Furthermore, DM has the highest kurtosis value (2.74), indicating heavier tails compared to a normal distribution and suggesting a higher probability of extreme values. The positive skewness (1.69), which results in a right-skewed distribution, further indicates the tendency for a higher DM content in some samples.

The high mean CP value of 16.06 g 100 g⁻¹ DW is indicative of a relatively high protein content in the forage, which is essential for animal growth and milk production. Furthermore, the first quartile (Q1) is 9.64 g 100 g⁻¹ DW, while the third quartile (Q3) is 21.23 g 100 g⁻¹ DW. These statistics indicate that half of the samples fall within a protein range that is suitable for cattle feeding (Pereira-Crespo et al. 2022). Differing from DM, a skewness of 0.44 suggests a more symmetrical distribution, indicating that the protein content is relatively consistent across samples. Prolonged drought conditions have been shown to accelerate forage maturation, leading to increased DM content and reduced CP levels, which in turn negatively affect digestibility and nutrient availability for livestock (Muzzo et al. 2025). These shifts in forage quality can impact livestock feeding strategies, requiring adjustments in dietary supplementation to maintain optimal animal performance.

aNDF, ADF and ADL generally follow normal distributions, with some asymmetry observed in ADF and ADL (0.47 and 0.92, respectively), suggesting greater variability. ADF is about half the concentration of aNDF, implying relevant digestible content. The mean ADF/ADL ratio (4.2) suggests relatively high levels of lignin, which could suggest potential digestibility issues. The data show a balance between mineral and fibre content, expressed as CA and CF, respectively. The high ranges and skewness in CA and CF suggest that different forage types could affect the overall quality, leading to differences in nutritional content and digestibility. As expected, EE (i.e., crude fat) (1.91 ± 0.68 g 100 g⁻¹ DW) was low, indicating it does not contribute substantially to the overall energy intake.

The second set of data acquired for the development of the PLS prediction models is composed of the spectra from FT-NIR spectroscopy. In Figure 2, the mean FT-NIR spectra for fresh and dried forage are reported, organised by sampling day. The spectra of both kinds of samples show a baseline shift, which is typical in NIR spectroscopy. In addition, a multiplicative effect is evident in the spectra of the dried samples in the 10,000–7000 cm⁻¹ region. This effect could arise from physical phenomena like light scattering caused by the particle size distributions of the powdery sample, highlighting the importance of sample preparation. Furthermore, the spectra are significantly influenced by the presence of water. In particular, the absorbance varies in intensity and is generally higher in the spectra of the fresh forage samples. Additionally, some peaks appear only in the spectra of the dried samples.

To enhance the visualisation of the differences between the overall mean spectra of fresh and dried samples, the absorbance units in Figure 3 are normalised using a min–max scaling approach (corresponding to the actual value minus the minimum value, all divided by the difference between the maximum and minimum values). This normalisation procedure emphasises the impact of moisture content on the bands observed at approximately 6900 cm⁻¹ (1st overtone of O─H stretching) and 5180 cm⁻¹ (combination band). Characteristic lipid-related peaks are also visible in the mean spectrum of dried forage at ~4320–4250 cm⁻¹ (2nd overtone of C─H bending; C─H stretching and C─C stretching combination bands, respectively) and ~5785–5680 cm⁻¹ (1st overtone of C─H stretching), even though lipids are present in modest amounts (Kays 2004; Workman and Weyer 2007; Barbin et al. 2014).

The best results of PLS models for predicting the quality attributes of green forage are reported in Table 3. To compare model performances, the NRMSECV was used to normalise RMSECV by considering the range of the Y variable (i.e., the quality attribute) and expressing the error as a percentage of its variability. Therefore, the two best models for fresh forage were DM and aNDF, with NRMSECV of 3.8% and 9.3%, respectively. For dried samples, the two best models were CP and aNDF, with NRMSECV of 5.6% and 7.9%, respectively (Table 3).

The Savitzky-Golay second derivative (2ndD) was the preprocessing step most frequently used in the best models. The 2ndD method employs different window sizes (ranging from 31 to 71, with increments of 10), followed by either mean centring or autoscaling. This method effectively removes irrelevant baseline signals from the spectra. Additionally, it suggests that the variables may be strongly correlated with each other and that adjacent variables might contain similar correlated signals. The EMSC was adopted for the best DM models of fresh samples and CP models of dried samples. EMSC corrected multiplicative effects in the spectra, which were primarily caused by light dispersion resulting from the physical characteristics of the samples. When considering fresh samples, the irregular arrangement of the approx. 1-cm fragments may have led to misalignment or overlapping during FT-NIR analysis, potentially creating empty space at the bottom of the sample cup, which could have influenced the measurements. In contrast, for dried and ground samples, the scattering effect can probably be attributed to their coarse and dusty texture, which results in a more substantial multiplicative effect observed in the spectra.

PLS models based on selected variables obtained through the two adopted spectral variable selection methods, iPLS and GA, generally outperformed models that included the full set of variables, as evidenced by the superior metrics in both CV and external set prediction (Table 3). Noteworthy, the iPLS method was particularly effective for fresh forage models, achieving the best results in 6 out of 8 cases. In contrast, the GA method was more effective for dried forage models, achieving the best outcomes in 5 out of 6 cases. Figure 4 compares the results of all variable selection methods applied to obtain the four best PLS models in terms of NRMSECV, highlighting the wavenumber bands selected. The GA_30 and GA_40 methods were set to provide a variable number of wavenumber bands, which may highlight the presence of synergies between bands but may also lead to more complex models. In contrast, the iPLS methods were set to provide a fixed number of bands: nine bands for iPLS_120, six for iPLS_60, and three for iPLS_30. The comparison of the results of the five variable selection methods revealed a trend in the selection of wavenumber bands in some spectral regions. Except for the aNDF model for fresh samples (Figure 4b), GA yielded the best results, represented by blue bars intersecting the red spectrum. Regarding DM (Figure 4a), two regions selected by at least four methods are approximately at 5500–5440 cm⁻¹ and 4230–4150 cm⁻¹, which may be associated with lignin, typically found at ~6250–5550 cm⁻¹ and ~4550–4150 cm⁻¹ (Shenk and Westerhaus 1994). Similar observations can be made for the aNDF model of fresh samples (Figure 4b), with the two regions ~4400–4120 cm⁻¹ and ~5850–5620 cm⁻¹ that could be associated with lignin and cellulose. Specifically, cellulose is one of the constituents of fibre, which absorbs in the ~4350–4270 cm⁻¹ region. CP for dried samples (Figure 4c) presents a common region at ~4770–4460 cm⁻¹, which could be associated with proteins, which absorb at ~4610–4590 cm⁻¹ (combination bands of N─H bend and C─H/C═O stretch) (Shenk and Westerhaus 1994). Finally, aNDF for dried samples (Figure 4d) presents interesting absorption bands in the ~5000–4000 cm⁻¹ region, specifically at ~4150–4000, ~4460–4290 and ~4920–4690 cm⁻¹, which can be associated with lignin and cellulose. Finally, to identify a set of “core” spectral regions that can be used for the development of portable instruments, the results of the best models obtained by the iPLS_30 variable selection method are reported in Table 3. Regarding CP in dried samples, the iPLS_30 model presented prediction performance perfectly in line with the best model with variable selection GA_40, with RMSEP of 1.865 and 1.841 g 100 g⁻¹ DW, and R²_P of 0.937 and 0.939 for the iPLS_30 and GA_40 models, respectively. In contrast, the predictive error performance of the other three iPLS_30 models (DM and aNDF for fresh samples, and aNDF for dried samples) were slightly lower. These results are a good compromise between accuracy and a large reduction of variables (from 3112 to 3 bands of 30 variables). Furthermore, this large selection of variables significantly reduces the instrumental and computational complexity required, making the models ideally implementable on compact sensors and devices, with benefits in terms of cost, power consumption and acquisition speed. The wavenumber bands of the respective iPLS_30 predictive models are reported in Figure 4.

The measured versus predicted values for the top four PLS models, previously discussed, are presented in Figure 5. Among these, DM for fresh samples and CP for dried samples (Figure 5a,c, respectively) stand out in terms of metric performances. The DM model achieves a coefficient of determination of prediction (R²_P) of 0.96 and a root-mean-square error of prediction (RMSEP) of 2.95 g 100 g⁻¹ FW, while the CP model yields an R²_P of 0.94 and an RMSEP of 1.84 g 100 g⁻¹ DW. Moreover, the CP model, which is characterised by only two latent variables, is simpler and more robust. Regarding aNDF, for fresh and dried samples (Figure 5b,d, respectively), the performance of PLS models was worse compared with DM and CP models. The dried samples model yielded better results, with fewer latent variables (five vs. seven) and superior metrics, achieving an R²_P of 0.87 versus 0.76 for fresh samples and an RMSEP of 4.20 g 100 g⁻¹ DW versus 6.08 g 100 g⁻¹ DW. The relatively lower prediction performance for aNDF may be caused by the intrinsic structural variability of the forage samples and its complex composition, which consists of a heterogeneous fraction of cellulose, hemicellulose and lignin. This can result in overlapping spectral absorption characteristics, affecting model performance. Furthermore, a high water content in fresh samples may cover or dilute the naturally variable spectral characteristics of the aNDF components. Finally, sample preparation, for example, drying and grinding, can influence the physical properties of the heterogeneous fibrous material, introducing spectral differences (e.g., shifts in peak positions, changes in peak intensity or spectral offsets).

In a recent study, Murphy et al. (Murphy et al. 2022). developed calibration models for DM and CP from FT-NIR spectra of fresh perennial ryegrass from Irish pastures. The homogeneity of the sample, composed of a single grass species, may have positively influenced the results. Prediction errors were expressed as standard error of prediction (SEP), which is not directly comparable to the RMSE that includes systematic error (BIAS). The SEP resulted in 2.038 g 100 g⁻¹ DW for CP and 0.946 g 100 g⁻¹ for DM. When normalised across the prediction measurement range, the errors corresponded to 9.6% for CP and 7.9% for DM. These findings are comparable to the NRMSECV values obtained in the current study, 12.3% for CP of fresh green forage and 3.8% for DM, indicating a slightly worse result for CP and a better result for DM (Table 3). In a previous study, Danieli et al. (Danieli et al. 2004). analysed dried forage from dairy sheep farms in central Italy over a period of 3 years. The results from their cross-validation models, expressed as the standard error of cross-validation normalised within the observed range of the quality attribute (NSECV), are comparable to those of the present study. However, the number of outliers removed was not reported. The error for CP was equivalent to that observed in the present study (NSECV and NRMSECV are equal to 5.6%). Regarding aNDF, the error was slightly better (NSECV of 6.2%, vs. an NRMSECV of 7.9%) (Table 3).

Overall, PLS models provided good predictive performance for some nutritional quality attributes of green forage. This result could lead to the development of real-time monitoring systems, which could improve livestock management. However, it is important to acknowledge some limitations of the study: (i) the relatively small dataset and (ii) the limited geographical sampling area only partially represent the diversity of forages, which is also determined by environmental conditions and species variability. These limitations reduce the generalisability of the predictive models. Furthermore, also considering the numerosity of the wavenumber bands of some prediction models, these limitations could contribute to overfitting in practical applications. Future research should focus on long-term studies supported by deep learning techniques that could improve the performance of predictive models. In addition, this approach could be combined with complementary spectroscopic techniques to provide better insights into forage nutritional quality dynamics.

4 Conclusions

The present study provided predictive models for the evaluation of the nutritional quality of green forage by NIR spectroscopy as a rapid and cost-effective alternative to traditional wet chemistry. The best PLS models resulted for DM in fresh forage (R²_P = 0.96, RMSEP = 2.95 g 100 g⁻¹ FW) and CP in dried samples (R²_P = 0.94, RMSEP = 1.84 g 100 g⁻¹ DW). Good results were also obtained for aNDF. Variable selection methods (interval PLS and genetic algorithms) significantly improved model accuracy; interval PLS performed better with fresh forage, while genetic algorithm performed better with dried samples. Furthermore, interval PLS demonstrated its potential in selecting a minimal set of “core” spectral bands with acceptable accuracy, suitable for the development of compact and cost-effective portable NIR sensors for use in the field. Sample characteristics and preparation, particularly the heterogeneity of the forage species and their growth stage, and sample preparation, as drying and milling processes, influenced the NIR spectra and consequently the performance of the predictive models. Future research should focus on expanding the dataset to include larger geographical areas with specific climatic conditions and related forage species at different phenological stages to improve the robustness and generalisability of the predictive models for wider adoption in forage quality management.

The study illustrates the potential of NIR spectroscopy in improving forage and pasture quality management, which impacts livestock sustainability. The use of FT-NIR spectroscopy provided the basis for low-cost, real-time forage and pasture monitoring, facilitating adaptive grazing practices.

Author Contributions

Alessandro Benelli: conceptualisation, formal analysis, investigation, methodology, software, validation, visualisation, writing – original draft, writing – review and editing. Riccardo Primi: conceptualisation, funding acquisition, methodology, software, writing – original draft, writing – review and editing. Chiara Evangelista: investigation, methodology, writing – original draft. Raffaello Spina: investigation, writing – original draft, writing – review and editing. Marco Milanesi: resources. Daniele Pietrucci: writing – review and editing. Bruno Ronchi: supervision. Umberto Bernabucci: conceptualisation, funding acquisition, project administration, supervision, writing – review and editing. Roberto Moscetti: conceptualisation, funding acquisition, methodology, software, supervision, writing – review and editing.

Conflicts of Interest

The authors declare no conflicts of interest.