Skip to main content
NCBI home page
Journal of Animal Science logoLink to Journal of Animal Science
. 2025 May 4;103:skaf151. doi: 10.1093/jas/skaf151

Spectral sensing for forage nutritive value determination of cool season, grass pastures during the grazing season

Ryan K Wright 1, Riley K Thompson 2, Chun-Peng James Chen 3, Robin R White 4,✉
PMCID: PMC12202004  PMID: 40326371

Abstract

Management surveys suggest that few cow-calf producers in the Southeastern United States submit forage samples for laboratory analysis due to time and labor constraints. Although tools like near infrared reflectance spectroscopy have helped reduce costs associated with nutritive value determination in stored feeds, their performance for pasture analysis has been limited. Our objective was to explore the efficacy of spectral sensing in predicting the dry matter (DM), acid detergent fiber (ADF), neutral detergent fiber (NDF), and crude protein (CP) of fresh forages during the growing season. Weekly from May through October, two random samples were collected from each of 12 fields. Spectral readings were taken above canopy level in-field and again in-lab, followed by bench chemistry analyses of DM, ADF, NDF, and CP. Chemistry results and spectral readings were aligned by field, sample, and date. The 18 individual light spectra and lidar-measured distance were used as features in a random forest regression fit to predict each nutrient and separate models were developed for in-field and in-lab spectral readings. Data were randomly split for hyperparameter tuning (15%), model training (55%), and independent evaluation (30%). The root mean squared prediction error (RMSPE), calculated on the independent evaluation data, was used to explore the viability of this system to predict forage nutritive value. The in-field and in-lab models performed similarly for each forage nutritive value. To evaluate the prediction capability of the system under various atmospheric conditions, cloud cover was added as a feature in each in-field regression. The RMSPE of DM, ADF, NDF, and CP with cloud cover were 21.8%, 9.88%, 10.1%, and 21.9%, respectively. These models were also evaluated on new, unseen data from nine subplots and used to explore the implications of the prediction errors. The NASEM (2018) Beef Cattle Nutrient Requirements model was used to simulate diet nutritional adequacy using forage nutritive value estimated from the spectral sensor compared with forage nutritive value measured by bench chemistry. These forage nutritive value estimation methods resulted in a 4.48% and 3.03% difference in metabolizable energy and metabolizable protein allowable gain, respectively. Considerable future data collection and model refinement efforts are necessary to determine the value of the spectral sensing system in supporting low-cost, in-field nutritive value monitoring.

Keywords: forage nutritive value, regression, sensing, spectroscopy

This study evaluated the viability of using a spectral sensing system to predict the nutritive value of fresh, cool season, grass pastures during the grazing season. Results indicate that spectral sensing is a promising management tool for monitoring the nutritive value of grazed forages despite various forage species, forage maturity, and atmospheric conditions.

Introduction

Traditionally, the nutritive value of feedstuffs is determined by lab-based bench chemistry analyses that are constrained by both time and cost as well as labor availability (Berzaghi et al., 2021). These challenges are amplified for extensive beef producers where feedstuff nutritive value determination in fresh forages grazed by livestock is considerably more time and labor intensive. According to National Animal Health Monitoring System (NAHMS) surveys, only 15.8% of United States cow-calf producers submit forage samples for laboratory analyses (NAHMS, 2017). More specifically, the estimate for smaller scale operations commonly observed in the Southeastern United States suggests that only 9.2% of producers are obtaining laboratory forage analyses (NAHMS, 2017). Producers often cite time and labor challenges as barriers to adopting forage sampling and analysis practices. As a result, there is a need for low-cost and easy-to-use (e.g., in-field) technologies which can monitor forage nutritive value. Such technologies could also support key advancements toward more precise supplementation strategies for grazing animals, and for improved pasture management.

Although near infrared reflectance spectroscopy (NIRS) has been advanced as a low-cost, nondestructive, rapid method of nutrient composition monitoring in stored feeds and fresh forages, its performance in estimating nutritive value of fresh forage has been limited (Ikoyi and Younge, 2020). The NIRS technology estimates forage nutritive value in a matter of seconds and has been developed to operate not only from a laboratory setting, but also as a handheld device (Acosta et al., 2020; Cherney et al., 2021). Although laboratory NIRS systems still require samples to be dried prior to scanning, handheld systems for use on-farm require less sample preparation (Fregulia et al., 2021). The lack of sample processing is advantageous for reducing the time needed to obtain forage nutritive value results; however, it can also present challenges because feed moisture levels and particle sizes impact prediction capabilities (Ikoyi and Younge, 2020). Similarly, calibration libraries have historically focused on specific feeds or species and studies on forages have mixed recommendations about the impacts of calibration method and sample preparation on performance (Andueza et al., 2011; Norman et al., 2020; Parrini et al., 2021; Catunda et al., 2022). Because NIRS and its calibration libraries were initially developed for in-laboratory use, these challenges limit the in-field applicability. Furthermore, although handheld NIRS systems offer a real-time method of analysis, they require high upfront costs that are impractical for many small-scale producers. Consequently, there is a need for technologies that share the rapid read time of NIRS, but also focus on in-field use and low-cost.

More recently, spectral sensing systems have been proposed as a more flexible strategy for determining forage nutritive value (Beeri et al., 2007; Smith et al., 2020; Geipel et al., 2021). There is a direct relationship between forage nutritive value indicators such as cell wall components or chlorophyl and plant color (Hu et al., 2010). Because of these relationships, it is possible to use spectral systems, which sense the reflection of various light frequencies, to estimate chemical composition of forages. Similar to NIRS, spectral sensing uses a range of wavelengths of light to estimate forage nutritive value. In addition to the near infrared spectrum, spectral systems may also include wavelengths in the ultraviolet (UV) or visible light spectra (Askari et al., 2019; López-Calderón et al., 2020). Spectral sensing systems typically scan fewer wavelengths than NIRS, and although this refinement is likely to compromise accuracy in forage nutritive value estimation, it does support reduced technology cost. Recent studies have shown that spectral data are effective in predicting forage nutritive value on a range of vegetation types and under varying weather and biomass conditions (Durante et al., 2014; Ferner et al., 2015), making them a promising solution for in-field monitoring across a range of forage species and environmental conditions. The objective of this study was therefore to assess the viability of a small-form, inexpensive spectral sensing system to determine forage nutritive value of mixed-grass pastures throughout the growing season in southwestern Virginia. Although the system is immediately proposed as a handheld tool assessing forage nutritive value, its small form and lower cost make it a viable candidate for a wearable sensor with opportunities to passively monitor forage quality. We hypothesized that the spectral sensing system would be capable of predicting key chemical composition indicators such as crude protein (CP), neutral and acid detergent fibers (NDF and ADF), and dry matter (DM) concentrations of forages.

Materials and Methods

The work conducted in this project included implementation of a spectral sensing system and testing of that system on 2 farms with 6 fields each, over a 27-wk forage growing season. In each week forage samples were collected and analyzed using traditional bench chemistry methods. The spectral readings and bench chemistry analyses were aligned in a database and used to derive random forest regressions which served to utilize spectral readings for prediction of forage chemical composition. These regressions were then evaluated on 30% held-out data, and on new samples obtained from in-field spatial scanning of forages. Details of each step in the process are enumerated in the below sections.

Sensor design

All sensors used in this experiment were open source. Two sensor suites were implemented utilizing identical sensor components with different methods for mounting the sensors, one for in-laboratory analysis and one for in-field analysis. These two systems were constructed to explore how the spectral sensor would work in a highly controlled (best-case) scenario (i.e., in-laboratory) compared with the more variable, in-field conditions. Both the in-field sensor suite and the laboratory sensor suite were comprised of a SparkFun ESP32 Thing Plus microprocessor, a SparkFun Triad Spectroscopy Sensor, and a SparkFun Garmin LIDAR-Lite v4 LED Distance Measurement Sensor (SparkFun Electronics, Niwot, CO). The spectral triad is capable of detecting 18 individual light frequencies over the range of 410 nanometers (nm) (visible light) to 940 nm (infrared) with a resolution of 28.6 nW/cm2. The LIDAR was included to account for any variation in distance between the sensor and the forage.

The sensor suites were programmed using the open-source Arduino IDE v 2.0.4 (Arduino Core Team, 2023) to activate the spectral triad, blink an LED to illuminate the sample and report the resulting readings via universal serial bus C to a logging computer. Sensor data were logged via PuTTY v 0.78 serial console (Simon Tatham, 2022). This scanning process was conducted at a rate of 1 hertz, or once per second. The in-field sensor suite was mounted to a solid plastic plate to ensure that the sensors maintained the same position for each reading (Figure 1). The laboratory sensor suite was mounted on the inside of a box to maintain sensor position and provide a uniform environment for readings to occur.

Figure 1.

Figure 1.

Open in a new tab

Image of the in-field spectral sensing system mounted on a plastic plate for structure and consistency.

Experimental design

Two farms were used for this analysis, with six fields selected at each farm. Thus, a total of twelve fields were used for sample collection. These fields were selected to represent a range of grazing species and qualitatively evaluated forage qualities (Table 1). At each location, two fields represented poor quality, two represented moderate quality, and two represented good quality. The quality of each field was determined by the presence of bare ground and apparent forage density as determined by canopy cover. Poor-quality fields had an abundance of bare ground and a forage density less than 60% while moderate-quality fields had relatively little bare ground and 60% to 85% forage density. Good quality fields were defined as having little to no bare ground and a forage density of 85% or greater. Fields were grazed by a variety of livestock including horses, cattle, sheep, or cattle and sheep. Fields were sampled weekly from May 1, 2023, to October 30, 2023, totaling 27 wk. Two random samples were collected per field, resulting in 24 samples each week and 648 samples over the course of the 27 wk. Forage species varied within and across fields but included common species prevalent in southwest Virginia (Kentucky 31 (tall fescue), orchardgrass, and ryegrass). Mixed-species pastures were selected and scanning/sampling irrespective of species was conducted to explore calibration of models and resulting accuracy that could be expected with common grazing/pasture management within the local geographic area.

Table 1.

Field classifications by species, quality, and location

Field ID Grazing species Pasture quality Latitude Longitude
P1 Horses Poor −80.4 37.2
P2 Cattle Poor −80.4 37.2
P3 Horses Moderate −80.4 37.2
P4 Cattle/Sheep Moderate −80.4 37.2
P5 No Animals Good −80.4 37.2
P6 Horses Good −80.4 37.2
K1 Cattle Moderate −80.6 37.2
K2 Cattle Moderate −80.6 37.2
K3 Cattle Good −80.6 37.2
K4 Cattle Good −80.6 37.2
K5 Cattle Poor −80.6 37.2
K6 Cattle Poor −80.6 37.2
Open in a new tab

Samples were collected at random using the “hula hoop method” (Mesa et al., 2023). In this sampling method, two random samples were selected each week by tossing an object of consistent circumference and sampling the area within said circumference wherever the object landed. For each sample, spectral data were recorded, a forage sample was clipped, bagged, and returned to the laboratory for further scanning and bench chemistry analysis. In the field, spectral data were recorded on the randomly chosen sample by holding the sensor-mounted plate roughly 15 cm above the tallest point of the forage. It is important to highlight that this may be a limitation of our study, as the sensor readings are collected above canopy height and are only informed by one dimension of the forage. Spectral data were recorded for 10 s in each location. The sample was then collected by clipping to a height of 5 cm with a 30 cm circumference circle. The time of sample collection and the clipped height was recorded for each sample. Each sample was then bagged, labeled, and returned to the lab for further analysis.

Samples arrived for laboratory analysis between twenty and ninety minutes after cutting. In the lab, spectral data were collected again using the laboratory sensor suite. This design included the box that the sensor suite was mounted to and was placed on a white plate containing the forage sample being analyzed. The purpose of this design was to create a controlled environment within the lab and eliminate outside influence by factors in-field such as sun intensity, shadows, rain, dew, etc. Much like the field strategy, these data were collected for a total of 10 s on each sample. The samples with the greatest time since collection often began wilting as in-lab processing began while the samples with the least amount of time since collection did not usually reach wilting during in-lab scanning and before being placed in the oven.

Ground truth determination

After the in-lab sensing, forage nutritive values were determined by traditional bench chemistry analyses. These bench chemistry analyses served as the ground truth measurements for comparison with spectral data. Each sample was analyzed for DM, ADF, NDF, and CP. Samples were dried for 72 h in an oven at 55 °C to determine DM according to AOAC method 930.15 (Horwitz and Latimer, 2000). Samples were then ground in a Wiley mill and passed through a 1-mm screen before being used to analyze NDF and ADF sequentially using the ANKOM 200 fiber analyzer system (ANKOM Technology, Macedon, NY) (Van Soest et al., 1991). Analysis of CP was conducted in an Elementar vario EL cube CHNS elemental analyzer (Elementar Americas Inc., Ronkonkoma, NY) using AOAC method 990.03 (AOAC, 2000).

Data preparation

All data preparation and analyses were conducted in R v 4.3.1 (R Core Team, 2022). Data from each sensor reading were logged and stored as individual text files labeled with the sample identifying number. These text files were iteratively read into R and the 10 readings were averaged to yield one spectral signature for each sample. Data from the bench chemistry analyses were also codified by sample identifying number and stored in a spreadsheet. These spreadsheet data were ingested into R using the readxl package (Wickham and Bryan, 2023) and merged into the averaged spectral sensor data such that each row within the data frame contained details on a single sample, including columns for the spectral-read wavelengths plus the chemical composition information obtained from bench chemistry analysis. Thus, each row contained the chemical composition and the spectral readings from either the in-field scan or the in-lab scan. We expected a total of 1,296 observations representing both spectral scans for all 648 samples. Data were cleaned by removing missing values and visual screening for outliers. Our final data frame had a total of 1,229 observations, resulting in a loss of 5.17% of data. Of the final observations, 603 were from in-field spectral readings, and 626 were from in-lab spectral readings.

Data analysis

A series of linear models was conducted to evaluate the effect of sensing conditions (in-field versus in-lab) on spectral readings. An analysis of variance (ANOVA) was performed for each model with significance set at P < 0.05 and a tendency at 0.05 < P > 0.10. Estimated marginal means (EMM) were also calculated using the emmeans package (Lenth, 2023).

The bench chemistry analyses were used to train and test random forest regressions to estimate DM, NDF, ADF, and CP from the spectral sensed data. Separate models were performed within each forage nutritive value to evaluate the in-field prediction capability versus the in-lab prediction capability. As a result, two random forest regressions were performed for each forage nutritive value. Data were randomly split into three parts with 55% for model training, 15% for hyperparameter tuning, and 30% for independent evaluation. The tuneRF function from the randomforest package (Liaw and Wiener, 2002) was used to identify the best hyperparameters for each random forest regression performed based on a 3-fold cross-validation. Hyperparameters for this regression included the number of decision trees and the number of variables considered at each split (Probst et al., 2019). Those hyperparameters were then used to fit a random forest regression with the randomForest package (Liaw and Wiener, 2002) utilizing the 18 wavelengths of spectral data and LIDAR measurement as features to predict each of the forage nutritive value parameters (DM, ADF, NDF, and CP). This parameter estimation used the 55% of data reserved for model training. The resultant model was then evaluated against the 30% independent data reserved for evaluation. In each evaluation, the residuals for each regression were analyzed to explore the presence of mean or slope bias related to different sampling dates or fields. The root mean squared prediction error (RMSPE; Hodson, 2022) of each model, indicating a percent error relative to ground truth, represented the ability of the sensor system to predict the individual forage nutritive value parameter. Because environmental influence can cause variation in spectral data, this data analysis approach was repeated for the in-field spectral data with cloud cover included as a feature in addition to the original features. Cloud cover was noted on each sampling day using data from the Virginia Tech Montgomery Executive Airport Weather Station. Cloud cover was defined by this weather station as one of the following: fair, partly cloudy, mostly cloudy, cloudy, light rain, or fog. The fit of the two approaches was compared to explore the extent to which cloud cover may influence the capacity of the spectral system to estimate forage composition.

A statistical analysis was also performed to evaluate the residuals of each in-field cloud cover model by date, field, and species using linear models. These residual analyses were conducted to explore whether the residuals of each model could be explained by differences in sampling date, field sampled, and grazing species. Ideally, the residuals from each model are the result of the different levels within the explanatory variables rather than mean or slope biases. A separate linear model was derived for the residuals of each forage nutritive value. Date, field, and species were the possible explanatory variables, and the residuals were the response variable. An analysis of variance (ANOVA) was performed for each linear model with significance set at P < 0.05 and a tendency set at 0.05 < P < 0.10. The P-values of each individual interaction were also noted to identify particular fields, dates, and species that resulted in residuals that differed significantly from other date and location combinations. These significant differences suggested an inhibited ability of the model to estimate forage nutritive value under those grazing conditions.

To further test the performance of the model on previously unseen data, we explored the use of the spectral sensor in a previously unseen field. Within this field, we completed spatial scanning of nine, 3.9 × 3.9 m subplots. The global positioning system location was recorded during the spectral scan of each subplot, which was translated to universal transverse Mercator coordinates for better visualization. After completing the scan, two samples of each subplot were obtained as described previously. These samples were subjected to bench chemistry analysis for DM, ADF, NDF, and CP. The spectral scan data were then used with the previously established random forest regressions for in-field spectral readings with cloud cover to estimate DM, CP, NDF, and ADF of each subplot. The observed forage nutritive values were compared to the predicted values by calculating the RMSPE.

Results and Discussion

A total of 648 samples were collected over the 6-mo period. Two samples were lost due to contamination, resulting in 646 samples being used for analysis. Of those 646 samples, 3 samples only yielded enough ground forage to conduct DM, NDF, and ADF analyses, and these 3 samples were omitted in the CP analysis due to missing values. There were few missing values for spectral data, with 24 of the 646 samples having no in-field reading due to sensor malfunction. All samples had an in-lab spectral reading. Samples were collected under a range of cloud cover percentages, temperatures, and humidity. Time of day was consistent week to week with fields being sampled in a consistent order and beginning at dawn. The EMM of the in-field versus in-lab spectral readings show that the scanning conditions of each spectral system significantly influenced the resulting spectral readings (Table 2). These differences can be attributed to the lighting conditions of each scanning location.

Table 2.

The influence of spectral sensing conditions on spectral readings

Spectral wavelength (nm) In-field EMM1
(nW/cm2)		 In-lab EMM1
(nW/cm2)		 P-value2
410 674 (26.4) 0.008 (25.9) <0.001
435 615 (24.3) 0.919 (23.9) <0.001
460 891 (35.3) 5.9 (34.7) <0.001
485 623 (24.5) 5.18 (24.0) <0.001
510 598 (23.7) 1.11 (23.2) <0.001
535 672 (26.8) 0.278 (26.2) <0.001
560 658 (26.1) 0.284 (25.6) <0.001
585 754 (29.9) 0.125 (29.3) <0.001
610 362 (14.9) 0.676 (14.6) <0.001
645 634 (25.2) 0.122 (24.7) <0.001
680 1,450 (56.4) 0.604 (55.2) <0.001
705 681 (28.6) 0.000 (28.1) <0.001
730 644 (27.0) 0.336 (26.5) <0.001
760 558 (23.9) 1.07 (23.4) <0.001
810 2,620 (103) 0.305 (101) <0.001
860 3,250 (126) 0.148 (123) <0.001
900 3,790 (146) 0.193 (143) <0.001
940 3,850 (146) 1.02 (143) <0.001
Open in a new tab

1EMM of each spectral wavelength for the in-field spectral readings and the in-lab spectral readings with SEM in parentheses.

2 P-value estimating the effect of the spectral sensing system (in-field versus in-lab) on spectral readings.

Time-series chemical composition of fields

Table 3 provides descriptive statistics for the forage nutritive value observations. When analyzed in time-series, the DM, CP, and ADF/NDF ratio of fields, as determined by bench chemistry, behaved as expected for the duration of the trial (Supplementary Figures 1–3). Histograms depicting the distribution of each chemical component can be found in Supplementary Figure 4. The early season typically has a higher DM concentration than mid-season, with DM concentration increasing again as the late season begins (Palacio et al., 2008). These trends were observed on all twelve fields, however, more dramatic increases and decreases in DM concentrations were noted on fields that were classified as poor quality. More drastic changes were also present on fields postmowing, where new growth led to a steady increase in DM concentration (Bryant et al., 2016). Trends in CP concentrations decreased during the initial sampling weeks, May through June, before increasing again throughout the remaining sampling period of July through October. These trends align well with the documented pattern for CP concentrations over the growing season for cool season grasses, which suggests that CP concentrations should be decreasing throughout the early summer and increase into the late season (Machado et al., 2005; Lawrence et al., 2006). The ADF/NDF ratio slowly decreased from May through August, followed by a more rapid increase throughout September and October. This trend is also consistent with expectations for cool season grasses, with the increasing ADF/NDF ratio reflecting maturity of these forages (Demanet et al., 2015; Karabulut and Çomaklı, 2023). Although these forages were subjected to various grazing conditions, the nutritive value concentrations are still influenced by the time of year, however, the rates of maturity may be slower than forages with unrestricted growth (Trlica, 1992).

Table 3.

Descriptive statistics of forage nutritive value observations

Nutritive Value Minimum Maximum Mean Standard Deviation
DM % 10.3 46.3 23.4 6.09
ADF % 25.8 52.9 40.5 4.05
NDF % 31.2 73.2 55.3 6.38
CP % 6.10 36.3 18.9 5.13
Open in a new tab

DM, dry matter; ADF, acid detergent fiber; NDF, neutral detergent fiber; CP, crude protein.

Evaluation of the spectral sensor in predicting forage composition

The in-field and in-lab random forest regressions for DM resulted in a RMSPE of 22.0% and 22.2%, respectively. The evaluation of the residuals showed minimal notable systematic error (Table 4), and the plotted predicted versus observed DM values showed minimal difference in performance between the two models (Figure 2).

Table 4.

Model performance of spectral predictions

Forage nutritive value Data RMSPE1 (%) Residual error (%) Mean bias
(% MSE2)		 Slope bias
(% MSE2)		 RSR3
DM In-Field 22.0 98.8 0.420 0.795 0.840
In-Field, Cloud Cover 21.8 98.5 0.308 1.16 0.836
In-Lab 22.2 98.8 1.12 0.054 0.847
ADF In-Field 9.95 99.1 0.066 0.830 0.941
In-Field, Cloud Cover 9.88 94.8 0.722 4.45 0.952
In-Lab 9.61 98.5 1.35 0.119 0.941
NDF In-Field 10.2 99.0 0.744 0.230 0.931
In-Field, Cloud Cover 10.1 99.6 0.255 0.169 0.925
In-Lab 12.0 95.6 4.28 0.153 0.976
CP In-Field 22.7 99.1 0.009 0.881 0.855
In-Field, Cloud Cover 21.9 99.7 0.185 0.108 0.823
In-Lab 24.4 96.5 1.64 1.85 0.875
Open in a new tab

1Root Mean Squared Prediction Error of each chemical component as predicted on the independent evaluation data.

2Mean Squared Error.

3Root mean standard deviation ratio.

DM, dry matter; ADF, acid detergent fiber; NDF, neutral detergent fiber; CP, crude protein.

Figure 2.

Figure 2.

Open in a new tab

Plot of predicted versus observed DM values for the in-field and in-lab spectral models.

Although in-field and in-lab performance were similar, both RMSPE were higher than expected, and the high error could be due to a variety of factors, including the variability of species composition in the fields, or the variability in DM concentrations across the sampling period. Within individual plant species, spectroscopy has been used to determine DM with a RMSPE as low as 0.08% or 1.8% (Clark et al., 2003; Islam et al., 2018). Because our models predict values from mixed-grass pastures which contain a variety of species, it is not surprising that the performance is not as robust as these previous works.

The DM observed through bench chemistry also had a large variability over the growing season, with observed values ranging from 10.3% to 46.3% and an average value of 23.4% (Table 3, Supplementary Figure 4). As moisture level decreases (increased DM), photosynthesis decreases as well. Moisture levels are directly correlated with photosynthesis, however the effects are not always immediately visible (Schneider and Childers, 1941). If wilting and color change are not visible until a few days postdrought, the spectral system may have limited capacity to respond to these rapid shifts in DM concentrations. Because the chemical composition changes more rapidly than the external plant color, it is likely that DM predicting using only spectral data may be challenging.

The limited predictive performance for DM could also be associated with the wavelengths monitored by the spectral sensor. Previous literature evaluating the most important scanning wavelengths for DM determination has shown that a number of these wavelengths are within our current scanning capabilities; however, there are also a number of important wavelengths in the near infrared (NIR) spectrum that are outside of these capabilities (Biewer et al., 2009; Smith et al., 2020). That said, Biewer et al. (2009) showed adequate DM predictions with a reduced scanning range of 630 nm to 1,000 nm. Given the scanning range of our spectral system, from 410 nm to 940 nm, it should be possible to achieve similar DM predictions; however, it may be more challenging without the additional NIR wavelengths.

It is important to note that although the RMSPE of DM predictions is higher than expected, the goal of this study was to evaluate the spectral sensor as a management support tool to facilitate real-time forage nutritive value predictions, not as a replacement of bench chemistry precision. Although the RMSPE for our spectral sensing system was higher than observed in previous studies, it may be adequate to support management decisions via directionality of changes relative to periodic analyses via traditional bench chemistry. Furthermore, the RSR values for the DM in-field and in-lab models were 0.840 and 0.847, respectively (Table 4). Although these RSR values are unfavorable, as ideal values are closer to zero and a value greater than 0.70 is generally regarded as unsatisfactory (Khosravi et al., 2018), it is logical considering the relatively high error rates for our DM models. The larger range of observed DM values does, however, appear to contribute to overall model performance. With observed DM values exhibiting a 36% difference between the minimum and maximum values, this larger range likely improves the ability of the models to generalize over a range of DM values. Given these findings and the goals of the study, the data suggest that larger datasets based on multiple years of data and a broader geographical representation of local conditions (i.e., across Virginia or the mid-Atlantic region) are necessary before broader conclusions about the suitability of the system can be drawn.

The in-field and in-lab random forest regressions for ADF resulted in a RMSPE of 9.95% and 9.61%, respectively, showed minimal systematic error when evaluating the residuals (Table 4), and the plotted predicted versus observed ADF values showed minimal performance differences between the two models (Figure 3). With similar model performance observed in-field and in-lab, these RMSPE suggest minimal accuracy loss associated with the easier-to-perform, in-field scanning vs the more controlled laboratory conditions. Observed ADF values ranged from 25.8% to 52.9% with an average value of 40.9% (Table 3; Supplementary Figure 4). A range of observed values was expected due to the variety within both plant type and plant maturity. Cellulose and lignin are the components that makeup ADF, and their presence within the cell wall changes throughout the growth cycle of grasses. The primary cell wall is thinner and composed of up to thirty percent cellulose (O’Neill and York, 2018). As the secondary cell wall develops, however, the structural components thicken and cellulose composes up to eighty percent of the structure (Kumar et al., 2015). As photosynthesis increases, sugar (glucose) availability increases as well, which is required to produce cellulose (Boex-Fontvieille et al., 2014). As a result, cellulose biosynthesis increases, and it can be inferred that this increased cellulose occurs simultaneously with increased chlorophyl from the active photosynthesis (Bollivar, 2006). Lignin, although not as prevalent as cellulose, also increases concentration in the secondary cell wall. While lignin can be present in the primary cell wall if the proper conditions are present, it is often unobserved in young plants (Müse et al., 1997). In the secondary cell wall, however, lignin accounts for up to thirty percent of the structure (Meents et al., 2018). Given that cellulose and lignin can have such variation between forage samples, the observations of ADF concentrations can be quite variable.

Figure 3.

Figure 3.

Open in a new tab

Plot of predicted versus observed ADF values for the in-field and in-lab spectral models.

The RSR values for the in-field and in-lab ADF models were both 0.941, which are higher than those of DM (Table 4). Even though the ADF models had lower error rates than the DM models, the range of observed ADF values measured was also less variable. As a result, the unfavorable RSR values for the ADF models reflect the narrow range of observed values relative to the error rates of these models. Given these observations, the lower RMSPE from the ADF models may actually be the result of a narrow range of observed values. As a result, further expansion of the calibration library is needed to improve performance of this system.

The in-field and in-lab random forest regressions for NDF resulted in a RMSPE of 10.2% and 12.0%, respectively, and, once again, showed limited systematic error (Table 4) and no notable difference in performance between the in-field and in-lab models (Figure 4). Although the in-lab model demonstrates a slightly higher error rate than the in-field model, these RMSPE values are promising for future use of the spectral system given the multiple chemical components contributing to NDF composition. Observed NDF values ranged from 31.2% to 73.2% with an average value of 55.3% (Table 3; Supplementary Figure 4). NDF is composed of cellulose, hemicellulose, and lignin. As discussed previously, both cellulose and lignin increase in concentration in the secondary cell wall as plants grow. Hemicellulose, which is unique to NDF, accounts for up to forty percent of the secondary cell wall (Kumar et al., 2015). Similar to cellulose, hemicellulose biosynthesis relies on the presence of sugars that result from photosynthesis, so increased hemicellulose percentages are linked to increased photosynthesis activity and chlorophyl presence (Zhang et al., 2021). These direct relationships between the chemical constituents in NDF and the physical properties of the forage likely support the relatively low error of the spectral system in estimating NDF concentrations. The in-field and in-lab NDF models resulted in RSR values of 0.931 and 0.976, respectively (Table 4). Similar to the ADF model evaluation, the high RSR values from the NDF models are likely an artifact of the narrower range of observed NDF values in the dataset relative to the RMSPE of the models. These values indicate a need for further testing under a variety of conditions to help improve the generalizability of the models.

Figure 4.

Figure 4.

Open in a new tab

Plot of predicted versus observed NDF values for the in-field and in-lab spectral models.

The in-field and in-lab random forest regressions for CP resulted in a RMSPE of 22.7% and 24.4%, respectively, showed minimal systematic error (Table 4), and no notable difference in performance between the two models (Figure 5). Both of these RMSPE values were higher than expected considering previous successes in protein prediction and is likely due to the high variation among observed values. CP is a measure of the nitrogen concentration present in any given sample as nitrogen is one of the main components of amino acids, which then make up proteins. Nitrogen is also one of the components of chlorophyl, meaning that spectral data should correlate with nitrogen percentage, as it directly influences photosynthesis (Bojović and Marković, 2009). Another study conducted using a tool capable of measuring red (660 nm) and near infrared (840 nm) spectral bands showed that optical chlorophyl measurements and CP are highly correlated (Hughes et al., 2017). Given that this range is within the range of wavelengths measured by our spectral sensor suite, we expected to see a strong capacity to estimate CP concentrations.

Figure 5.

Figure 5.

Open in a new tab

Plot of predicted versus observed CP values for the in-field and in-lab spectral models.

Although the model was able to able to predict CP for the large majority of samples, it has the highest RMSPE compared with our other models estimating forage chemical composition. The observed CP values had an average of 18.9% within a range of 6.10% to 36.3% (Table 3, Supplementary Figure 4). A larger RMSPE was also seen when predicting DM, which had a similarly large range of observed values. The low RMSPE error of ADF and NDF, on the other hand, is also accompanied by a narrower range of observed values. This increased error when predicting CP is thought to be attributed to the large variations of species, harvest stage, and regrowth period (Kirchhof et al., 2010; Krawutschke et al., 2013). Although this error rate is highest among the predicted parameters, the CP models resulted in some of the lower RSR values, which can be attributed to the wider range of observed CP values in the dataset. The in-field and in-lab RSR values for the CP models were 0.855 and 0.875, respectively (Table 4). The RSR values of ADF and NDF were highest, despite their low error rates, because they also demonstrated the narrowest ranges of observed values. The observed CP values, on the other hand, vary around 30% from the lowest to highest observation. Although there is room for improvement in model performance, these RSR values indicate that expanding the datasets used for calibration will be critical in assessing the performance of this tool.

Influence of cloud cover on predictions

Many spectral sensing systems implement a reference card for calibration or to aid in eliminating noisy wavelengths driven by variation in external conditions, as these imperfections in the data can influence spectral readings and their usability (Knox et al., 2012; Ferner et al., 2015; Geipel et al., 2021). Because the goal of this system is to eventually be deployed for in-field, real-time sensing of forage composition, this reference card procedure reflects one limitation to ease-of-use. As such, we explored how omitting a referencing step, and instead accounting for cloud cover variation during collection of our in-field data would influence the accuracy of prediction (Table 4). If the inclusion of cloud cover dramatically improved the prediction accuracy, it would be advisable to explore strategies to augment the sensing system to capture ambient conditions in real-time to ensure better capacity to monitor forage nutritive value during grazing, or to consider an explicit referencing procedure. When compared to the in-field models derived with no representation of ambient conditions, the RMSPE of in-field models including cloud cover had minimal change. Although slight decreases in mean and slope biases were observed for NDF, this was not the case for ADF, CP, and DM, and the observed changes have minimal practical relevance. Generally, the models performed very similarly, however, including cloud cover as a feature should help to account for any variation in measurements attributable to atmospheric conditions. Developing models that account for this environmental variation, Geipel et al. (2021) were able to achieve RMSPE values of 15.2%, 11.7%, and 4.8% for DM, CP, and NDF, respectively. Although our models have larger RMSPE values than those previously achieved when accounting for atmospheric conditions, this is likely a result of the limited variation inherent within the calibration dataset rather than the ability of the system to account for environmental influence. Although minimal change in model performance was achieved when including cloud cover as a feature, the use of these models may be a more robust strategy to ensure atmospheric variation is controlled for. Instead, measures such as the UV index in conjunction with cloud cover may provide a better estimate of the amount of incident energy from atmospheric conditions. In future studies, additional atmospheric variables should be explored to determine if another representation of atmospheric conditions is helpful in improving system accuracy.

Residuals analysis to support future model refinement

Of the twelve fields used in this study, 7 housed cattle only, 3 housed horses, 1 housed a rotation of cattle and sheep, and 1 had no animals (Table 1). The 1 field that was not grazed by any animals was mechanically harvested once during the experimental period. Residuals from each in-field, cloud cover model were analyzed using linear models against date, field, or species to explore directions for future broadening of the calibration set. Plots of predicted values versus observed or residual values are available in Supplementary Figures 5–16. Date, field, and species significantly influenced or tended to influence the majority of model residuals (Table 5).

Table 5.

Analysis of in-field, cloud cover model residuals by date, field, and species

Model Explanatory variable P-value
DM Date 0.047
Field 0.015
Species 0.009
ADF Date 0.034
Field 0.102
Species 0.011
NDF Date 0.060
Field <0.001
Species 0.056
CP Date 0.269
Field <0.001
Species <0.001
Open in a new tab

Analysis of residuals for the in-field, cloud cover models by date, field, and species with average significance at P < 0.05 and a tendency at 0.05 < P < 0.10. DM, dry matter; ADF, acid detergent fiber; NDF, neutral detergent fiber; CP, crude protein.

Date significantly influenced the model residuals for the DM (P = 0.047, Supplementary Figure 5) and ADF (P = 0.034, Supplementary Figure 8) models and tended to influence the NDF (P = 0.060, Supplementary Figure 11) model (Table 5). As discussed above, the forage nutritive value observations from the current data follow expected trends in concentrations throughout the growing season. Although the presence of grazing species may slow the rate of growth (Trlica, 1992), we observed increases and decreases in concentrations for each forage chemical component when exploring the time-series patterns. Consequently, it is logical that the sampling week significantly influences model residuals. Because the forage nutritive value concentrations throughout the growing season are subject to change, this residuals analysis suggests that future calibration data should include a variety of dates to ensure that the models are able to predict across the growing season.

Field significantly influenced the residuals for the DM (P = 0.015, Supplementary Figure 6), NDF (P < 0.001, Supplementary Figure 12), and CP (P < 0.001, Supplementary Figure 15) models (Table 5). The fields utilized in this study had varying prevalence of forage species including Kentucky 31 tall fescue, orchardgrass, and ryegrass. Weed infiltration among pastures also varied. Research has shown that commonly observed species of weeds tend to have different ranges of forage nutritive value concentrations than grasses (Marten et al., 1987; Abaye et al., 2009; Bunton et al., 2020). It can be inferred that these contrasting values were more difficult for the models to predict due to their limited occurrence within the dataset. As a result, the prevalence of weeds and the variability in pasture composition led to variable forage nutritive value observations and predictions. In addition to the pasture composition, the fields utilized in this study were deemed poor, moderate, or good quality (Table 1). Overgrazing on a low forage density field can restrict forage growth and maturity while unrestricted forage will reach the same stage of growth and maturity at a much faster rate (Oelberg, 1956; Mysterud, 2006). Consequently, the different levels of pasture quality likely contributed to the significance of field on the residuals of the DM, NDF, and CP models. It is recommended that various forage species, densities, and management types be used in future investigations of the spectral sensing system to improve model generalizability and improve the robustness of the calibration library.

Species significantly influenced the residuals for the DM (P = 0.009, Supplementary Figure 7), ADF (P = 0.011, Supplementary Figure 10), and CP (P < 0.001, Supplementary Figure 16) models and tended to influence the residuals for the NDF (P = 0.056, Supplementary Figure 13) model (Table 5). In this study, fields were grazed by horses, cattle, a rotation of sheep and cattle, or no animals (Table 1). Grazing species exhibit various levels of selectivity and grazing patterns that can influence the level of consumption as well as the forage species consumed (Hongo and Akimoto, 2003; Cuchillo‐Hilario et al., 2018). Because of these differences, there is considerable variation in forage density and prevalent forage species across pastures grazed by these different species. The field with no animals, which was mechanically harvested once during the grazing season, also exhibited different rates of forage growth and maturity than those actively grazed. The significant influence of species on the majority of model residuals suggests that it is important to collect data under each of these grazing conditions to support development of robust, generalizable models.

Overall, these residuals analyses highlight that in addition to sampling across large areas and across multiple growing seasons, samples to inform spectral sensing for forage nutritive value should be obtained from systems with differences in management approach, species, or grazing strategy to ensure robustness. Previously, it has been suggested that reliable prediction of forage nutritive value estimates requires data from multiple growing seasons to achieve models robust enough to perform well (Inostroza et al., 2016). Our residuals analysis also suggests that data used for model derivation should be obtained from a variety of management contexts to represent different grazing or mowing patterns and different levels of forage maturity. Although NIRS calibration libraries tend to be species-specific, research has shown value in developing calibration libraries that incorporate multiple local species for specific geographical regions (Andueza et al., 2011). As such, future work should focus on developing a more generalizable dataset that includes a variety of forages, experiencing grazing by diverse species or mechanical harvesting.

Practical evaluation of model performance on unseen data

In the evaluation of these models on previously unseen data, the RMSPE of DM, ADF, NDF, and CP were calculated (Table 6) and results were plotted showing the differences between the predicted and observed values (Supplementary Figures 17–20). ADF performed well in both the original models and the evaluation on new, unseen data (RMSPE 6.26%), which can be attributed to the reasons discussed previously regarding the alignment between chemical and physical characteristics of forage. Although the NDF model had a slightly higher error rate than the ADF model during this evaluation, the RMSPE of NDF (11.2%) still showed promise for evaluating unseen samples. Interestingly, although CP had a high RMSPE when evaluated as a part of model performance, the CP predicted on the unseen data had one of the lowest error rates (RMSPE 7.95%). The performance of the DM model on unseen samples, however, was very poor (RMSPE 101%) and is likely attributed to the challenges discussed previously with using this spectral sensing system for DM prediction.

Table 6.

Model performance on unseen data

DM % ADF % NDF % CP %
RMSPE1 101 6.26 11.2 7.95
Open in a new tab

1Root mean squared prediction error (% mean) between observed and predicted values of each chemical component when predicted on unseen field plots. DM, dry matter; ADF, acid detergent fiber; NDF, neutral detergent fiber; CP, crude protein.

Although the numerical error values (RMSPE) support conclusions similar to the evaluation data, the residual plots suggest that much of the strength in performance may be reflective of narrow evaluation data within this small sample size. Because these subplots were tested after the conclusion of the experiment, it is possible that the late season presented more consistent visual characteristics which were more directly associated with values such as nitrogen levels, contributing to the lower error rates when estimating CP. Less variation in both the appearance of the plant as well as the observed values may be responsible for the good performance of these models. Overall, the ability of most models to perform well on unseen data supports the potential of this sensing system for forage nutritive value determination on pasture, pending more comprehensive data collection efforts as noted in the above section.

To evaluate the management implications of relying on the spectral sensing system for evaluating forage nutritive value, we used the NASEM Beef Cattle Nutrient Requirements model (NASEM, 2016) to estimate metabolizable protein (MP) and metabolizable energy (ME) allowable gain based either on consumption of forage represented by the bench chemistry analysis or by the spectral system. For the purpose of this investigation, steers in a growing/finishing stage were selected with an initial body weight of 226 kg. On average, the observed values of DM, ADF, NDF, and CP of the grazed forage resulted in an estimated ME allowable gain of 0.692 kg per day and an MP allowable gain of 0.627 kg per day. Similarly, when allowable gains were modeled using the chemical composition values predicted by the spectral system, the ME allowable gain was 0.661 kg per day and MP allowable gain was 0.608 kg per day. Comparison of the allowable gains predicted by forage composition measured by bench chemistry compared with the sensor system resulted in a 4.48% difference in ME allowable gain and a 3.03% difference in MP allowable gain.

Conclusion

The current models predict fiber (ADF and NDF) with lower RMSPE than the DM and CP models. These differences in model performance may be a result of the range of observed values for each forage nutritive value, as the DM and CP values had favorable RSR compared with the ADF and NDF models. Despite the differences in spectral readings resulting from the different experimental conditions (in-field versus in-lab), model performance of the two systems was similar for each forage nutritive value. Overall, more comprehensive data is needed to refine these models across various species, maturities, and atmospheric conditions; however, the current data suggest that the system may eventually be able to enable fast, easy-to-use, low-cost estimations of forage chemical composition.

Supplementary Data

Supplementary data are available at Journal of Animal Science online.

skaf151_suppl_Supplementary_Figures
skaf151_suppl_supplementary_figures.docx (3.4MB, docx)

Acknowledgments

This work was supported by funds appropriated to the Virginia Tech College of Agriculture and Life Sciences and by the Nutrition, Growth, and Lactation, project award number 2018-67007-28452, the Cyberphysical Systems, project award number 2019-67021-29007, and the National Robotics Initiative, project award number 2021-67021-34769 from the U.S. Department of Agriculture’s National Institute of Food and Agriculture.

Glossary

Abbreviations:

ADF

acid detergent fiber

CP

crude protein

DM

dry matter

NDF

neutral detergent fiber

RMSPE

root mean squared prediction error

RSR

root mean standard deviation ratio

Contributor Information

Ryan K Wright, School of Animal Sciences, Virginia Polytechnic Institute and State University, Blacksburg, VA, USA.

Riley K Thompson, School of Animal Sciences, Virginia Polytechnic Institute and State University, Blacksburg, VA, USA.

Chun-Peng James Chen, School of Animal Sciences, Virginia Polytechnic Institute and State University, Blacksburg, VA, USA.

Robin R White, School of Animal Sciences, Virginia Polytechnic Institute and State University, Blacksburg, VA, USA.

Author Contributions

Ryan K. Wright (Data curation, Formal analysis, Investigation, Methodology, Supervision, Writing—original draft, Writing—review & editing), Riley K. Thompson (Data curation, Investigation), Chun-Peng Chen (Formal analysis), and Robin R. White (Conceptualization, Formal analysis, Funding acquisition, Methodology, Supervision, Writing—review & editing)

Conflict of interest statement. The authors declare no real or perceived conflicts of interest.

Literature Cited

  1. Abaye, A. O., Scaglia Alonso G., Teutsch C. D., and Raines P. M... 2009. The nutritive value of common pasture weeds and their relation to livestock nutrient requirements. Virginia Cooperative Extension. p. 418–150. [Google Scholar]
  2. Acosta, J., Castillo M., and Hodge G... 2020. Comparison of benchtop and handheld near‐infrared spectroscopy devices to determine forage nutritive value. Crop Sci. 60:3410–3422. doi: 10.1002/csc2.20264 [DOI] [Google Scholar]
  3. Andueza, D., Picard F., Jestin M., Andrieu J., and Baumont R... 2011. NIRS prediction of the feed value of temperate forages: efficacy of four calibration strategies. Animal 5:1002–1013. doi: https://doi.org/ 10.1017/S1751731110002697 [DOI] [PubMed] [Google Scholar]
  4. AOAC. 2000. AOAC official method 990.03, protein (crude) in animal feed, combustion method. In: Official methods of analysis of AOAC international revision 1, 2006. 18th ed. Gaithersburg (MD): AOAC International; p. 30–31. Chapter 4. [Google Scholar]
  5. Askari, M. S., McCarthy T., Magee A., and Murphy D. J... 2019. Evaluation of grass quality under different soil management scenarios using remote sensing techniques. Remote Sens. 11:1835. doi: https://doi.org/ 10.3390/rs11151835 [DOI] [Google Scholar]
  6. Beeri, O., Phillips R., Hendrickson J., Frank A. B., and Kronberg S... 2007. Estimating forage quantity and quality using aerial hyperspectral imagery for northern mixed-grass prairie. Remote Sens. Environ. 110:216–225. doi: https://doi.org/ 10.1016/j.rse.2007.02.027 [DOI] [Google Scholar]
  7. Berzaghi, P., Cherney J. H., and Casler M. D... 2021. Prediction performance of portable near infrared reflectance instruments using preprocessed dried, ground forage samples. Comput. Electron. Agric. 182:106013. doi: https://doi.org/ 10.1016/j.compag.2021.106013 [DOI] [Google Scholar]
  8. Biewer, S., Fricke T., and Wachendorf M... 2009. Determination of dry matter yield from legume–grass swards by field spectroscopy. Crop Sci. 49:1927–1936. doi: https://doi.org/ 10.2135/cropsci2008.10.0608 [DOI] [Google Scholar]
  9. Boex-Fontvieille, E., Davanture M., Jossier M., Zivy M., Hodges M., and Tcherkez G... 2014. Photosynthetic activity influences cellulose biosynthesis and phosphorylation of proteins involved therein in Arabidopsis leaves. J. Exp. Bot. 65:4997–5010. doi: https://doi.org/ 10.1093/jxb/eru268 [DOI] [PubMed] [Google Scholar]
  10. Bojović, B., and Marković A... 2009. Correlation between nitrogen and chlorophyll content in wheat (Triticum aestivum L.). Kragujevac J. Sci. 31:69–74. https://www.researchgate.net/publication/266609505_Correlation_between_nitrogen_and_chlorophyll_content_in_wheat_Triticum_aestivum_L [Google Scholar]
  11. Bollivar, D. W. 2006. Recent advances in chlorophyll biosynthesis. Photosynth. Res. 90:173–194. doi: https://doi.org/ 10.1007/pl00022068 [DOI] [PubMed] [Google Scholar]
  12. Bryant, R. H., Kingsbury L., and Edwards G... 2016. Does mowing before grazing increase dry matter intake and milk yield?
  13. Bunton, G., Trower Z., Roberts C., and Bradley K. W... 2020. Seasonal changes in forage nutritive value of common weeds encountered in Missouri pastures. Weed Technol. 34:164–171. doi: https://doi.org/ 10.1017/wet.2019.89 [DOI] [Google Scholar]
  14. Catunda, K. L., Churchill A. C., Power S. A., and Moore B. D... 2022. Near infrared spectroscopy calibration strategies to predict multiple nutritional parameters of pasture species from different functional groups. J. Near Infrared Spectrosc. 30:254–263. doi: https://doi.org/ 10.1177/09670335221114746 [DOI] [Google Scholar]
  15. Cherney, J. H., Digman M. F., and Cherney D... 2021. Handheld NIRS for forage evaluation. Comput. Electron. Agric. 190:106469. doi: https://doi.org/ 10.1016/j.compag.2021.106469 [DOI] [Google Scholar]
  16. Clark, C., McGlone V., Requejo C., White A., and Woolf A... 2003. Dry matter determination in ‘Hass’ avocado by NIR spectroscopy. Postharvest Biol. Technol. 29:301–308. doi: https://doi.org/ 10.1016/s0925-5214(03)00046-2 [DOI] [Google Scholar]
  17. Cuchillo‐Hilario, M., Wrage‐Mönnig N., and Isselstein J... 2018. Forage selectivity by cattle and sheep co‐grazing swards differing in plant species diversity. Grass Forage Sci. 73:320–329. doi: https://doi.org/ 10.1111/gfs.12339 [DOI] [Google Scholar]
  18. Demanet, R., Mora M., Herrera M., Miranda H., and Barea J... 2015. Seasonal variation of the productivity and quality of permanent pastures in Andisols of temperate regions. J. Soil Sci. Plant Nutr. 15:111–128. doi: 10.4067/S0718-95162015005000010 [DOI] [Google Scholar]
  19. Durante, M., Oesterheld M., Piñeiro G., and Vassallo M. M... 2014. Estimating forage quantity and quality under different stress and senescent biomass conditions via spectral reflectance. Int. J. Remote Sens. 35:2963–2981. doi: https://doi.org/ 10.1080/01431161.2014.894658 [DOI] [Google Scholar]
  20. Ferner, J., Linstädter A., Südekum K. -H., and Schmidtlein S... 2015. Spectral indicators of forage quality in West Africa’s tropical savannas. Int. J. Appl. Earth Obs. Geoinf. 41:99–106. doi: https://doi.org/ 10.1016/j.jag.2015.04.019 [DOI] [Google Scholar]
  21. Fregulia, P., Neves A. L. A., Dias R. J. P., and Campos M. M... 2021. A review of rumen parameters in bovines with divergent feed efficiencies: what do these parameters tell us about improving animal productivity and sustainability? Livestock Sci. 254:104761. doi: https://doi.org/ 10.1016/j.livsci.2021.104761 [DOI] [Google Scholar]
  22. Geipel, J., Bakken A. K., Jørgensen M., and Korsaeth A... 2021. Forage yield and quality estimation by means of UAV and hyperspectral imaging. Precis. Agric. 22:1437–1463. doi: https://doi.org/ 10.1007/s11119-021-09790-2 [DOI] [Google Scholar]
  23. Hongo, A., and Akimoto M... 2003. The role of incisors in selective grazing by cattle and horses. J. Agric. Sci. 140:469–477. doi: https://doi.org/ 10.1017/s0021859603003083 [DOI] [Google Scholar]
  24. Horwitz, W., and Latimer G. W... 2000. Official methods of analysis of AOAC International. Gaithersburg: AOAC International. [Google Scholar]
  25. Hu, H., Liu H.-q., Zhang H., Zhu J.-h., Yao X.-g., Zhang X.-b., and Zheng K.-f... 2010. Assessment of chlorophyll content based on image color analysis, comparison with SPAD-502. In: 2010 2nd International Conference on Information Engineering and Computer Science. P. 1–3.
  26. Hughes, M., Lallo C., Mlambo V., and Jennings P... 2017. Calibration and validation of optical chlorophyll‐measuring devices for use in predicting crude protein concentration in tropical grass herbage. Grass Forage Sci. 72:118–136. doi: 10.1111/gfs.12219 [DOI] [Google Scholar]
  27. Ikoyi, A., and Younge B. A... 2020. Influence of forage particle size and residual moisture on near infrared reflectance spectroscopy (NIRS) calibration accuracy for macro-mineral determination. Anim. Feed Sci. Technol. 270:114674. doi: https://doi.org/ 10.1016/j.anifeedsci.2020.114674 [DOI] [Google Scholar]
  28. Inostroza, L., Acuña H., Munoz P., Vásquez C., Ibáñez J., Tapia G., Pino M. T., and Aguilera H... 2016. Using aerial images and canopy spectral reflectance for high‐throughput phenotyping of white clover. Crop Sci. 56:2629–2637. doi: https://doi.org/ 10.2135/cropsci2016.03.0156 [DOI] [Google Scholar]
  29. Islam, M. N., Nielsen G., Stærke S., Kjær A., Jørgensen B., and Edelenbos M... 2018. Noninvasive determination of firmness and dry matter content of stored onion bulbs using shortwave infrared imaging with whole spectra and selected wavelengths. Appl. Spectrosc. 72:1467–1478. doi: https://doi.org/ 10.1177/0003702818792282 [DOI] [PubMed] [Google Scholar]
  30. Karabulut, H., and Çomaklı B... 2023. The seasonal variation of forage’s chemical composition of a semi-arid rangeland according to altitude and aspects. Turkish J. Range Forage Sci. 4:27–37. doi: 10.51801/turkjrfs.1264845 [DOI] [Google Scholar]
  31. Khosravi, K., Mao L., Kisi O., Yaseen Z. M., and Shahid S... 2018. Quantifying hourly suspended sediment load using data mining models: case study of a glacierized Andean catchment in Chile. J. Hydrol. 567:165–179. doi: https://doi.org/ 10.1016/j.jhydrol.2018.10.015 [DOI] [Google Scholar]
  32. Kirchhof, S., Eisner I., Gierus M., and Südekum K. H... 2010. Variation in the contents of crude protein fractions of different forage legumes during the spring growth. Grass Forage Sci. 65:376–382. doi: 10.1111/j.1365-2494.2010.00756.x [DOI] [Google Scholar]
  33. Knox, N. M., Skidmore A. K., Prins H. H., Heitkönig I. M., Slotow R., van der Waal C., and de Boer W. F... 2012. Remote sensing of forage nutrients: Combining ecological and spectral absorption feature data. ISPRS J. Photogramm. Remote Sens. 72:27–35. doi: 10.1016/j.isprsjprs.2012.05.013 [DOI] [Google Scholar]
  34. Krawutschke, M., Kleen J., Weiher N., Loges R., Taube F., and Gierus M... 2013. Changes in crude protein fractions of forage legumes during the spring growth and summer regrowth period. J. Agric. Sci. 151:72–90. doi: https://doi.org/ 10.1017/s002185961200024x [DOI] [Google Scholar]
  35. Kumar, M., Campbell L., and Turner S... 2015. Secondary cell walls: biosynthesis and manipulation. J. Exp. Bot. 67:515–531. doi: https://doi.org/ 10.1093/jxb/erv533 [DOI] [PubMed] [Google Scholar]
  36. Lawrence, L., Hayes S., Allman R., and Rich G... 2006. Seasonal variation in cool season grasses. J. Anim. Sci. 84:331. https://research-ebsco-com.ezproxy.lib.vt.edu/c/w5gwtr/search/details/6f5e4bwpgn?db=a9h [Google Scholar]
  37. Lenth, R. V. 2023. emmeans: Estimated marginal means, aka least-squares means. [Google Scholar]
  38. López-Calderón, M. J., Estrada-Ávalos J., Rodríguez-Moreno V. M., Mauricio-Ruvalcaba J. E., Martínez-Sifuentes A. R., Delgado-Ramírez G., and Miguel-Valle E... 2020. Estimation of total nitrogen content in forage maize (Zea mays L.) Using spectral indices: analysis by random forest. Agriculture 10:451. doi: https://doi.org/ 10.3390/agriculture10100451 [DOI] [Google Scholar]
  39. Machado, C., Morris S., Hodgson J., and Fathalla M... 2005. Seasonal changes of herbage quality within a New Zealand beef cattle finishing pasture. [Google Scholar]
  40. Marten, G., Sheaffer C., and Wyse D... 1987. Forage nutritive value and palatability of perennial weeds. Agron. J. 79:980–986. doi: 10.2134/agronj1987.00021962007900060006x [DOI] [Google Scholar]
  41. Meents, M. J., Watanabe Y., and Samuels A. L... 2018. The cell biology of secondary cell wall biosynthesis. Ann. Bot. (Lond.) 121:1107–1125. doi: https://doi.org/ 10.1093/aob/mcy005 [DOI] [PMC free article] [PubMed] [Google Scholar]
  42. Mesa, A. N., Strager M. P., Grushecky S. T., and Kinder P... 2023. Using unmanned aerial vehicles to evaluate revegetation success on natural gas pipelines. Environ. Manage. 72:671–681. [DOI] [PubMed] [Google Scholar]
  43. Müse, G., Schindler T., Bergfeld R., Ruel K., Jacquet G., Lapierre C., Speth V., and Schopfer P... 1997. Structure and distribution of lignin in primary and secondary cell walls of maize coleoptiles analyzed by chemical and immunological probes. Planta 201:146–159. doi: 10.1007/BF01007699 [DOI] [Google Scholar]
  44. Mysterud, A. 2006. The concept of overgrazing and its role in management of large herbivores. Wildl. Biol. 12:129–141. doi: https://doi.org/ 10.2981/0909-6396(2006)12[129:tcooai]2.0.co;2 [DOI] [Google Scholar]
  45. NAHMS. 2017. National animal health monitoring system beef 2017 report 2: Beef Cow-calf Health and Management Practices in the United States, 2017. [Google Scholar]
  46. NASEM. 2016. National Academies of Sciences, Engineering, and Medicine. Nutrient Requirements of Beef Cattle: Eighth Revised Edition. Washington, DC: The National Academies Press. doi: 10.17226/19014 [DOI] [PubMed] [Google Scholar]
  47. Norman, H. C., Hulm E., Humphries A. W., Hughes S. J., and Vercoe P. E... 2020. Broad near-infrared spectroscopy calibrations can predict the nutritional value of> 100 forage species within the Australian feedbase. Anim. Prod. Sci. 60:1111–1122. doi: https://doi.org/ 10.1071/an19310 [DOI] [Google Scholar]
  48. O’Neill, M. A., and York W. S... 2018. The composition and structure of plant primary cell walls. Annu. Plant Rev. 8:1–54. doi: 10.1002/9781119312994.apr0067 [DOI] [Google Scholar]
  49. Oelberg, K. 1956. Factors affecting the nutritive value of range forage. J. Range Manag. 9:220–225. doi: https://doi.org/ 10.2307/3894056 [DOI] [Google Scholar]
  50. Palacio, S., Milla R., Albuixech J., Pérez‐Rontomé C., Camarero J. J., Maestro M., and Montserrat‐Martí G... 2008. Seasonal variability of dry matter content and its relationship with shoot growth and nonstructural carbohydrates. New Phytol. 180:133–142. doi: https://doi.org/ 10.1111/j.1469-8137.2008.02569.x [DOI] [PubMed] [Google Scholar]
  51. Parrini, S., Staglianò N., Bozzi R., and Argenti G... 2021. Can grassland chemical quality be quantified using transform near-infrared spectroscopy? Animals 12:86. doi: https://doi.org/ 10.3390/ani12010086 [DOI] [PMC free article] [PubMed] [Google Scholar]
  52. Schneider, G. W., and Childers N... 1941. Influence of soil moisture on photosynthesis, respiration, and transpiration of apple leaves. Plant Physiol. 16:565. doi: https://doi.org/ 10.1104/pp.16.3.565 [DOI] [PMC free article] [PubMed] [Google Scholar]
  53. Smith, C., Karunaratne S., Badenhorst P., Cogan N., Spangenberg G., and Smith K... 2020. Machine learning algorithms to predict forage nutritive value of in situ perennial ryegrass plants using hyperspectral canopy reflectance data. Remote Sens. 12:928. doi: https://doi.org/ 10.3390/rs12060928 [DOI] [Google Scholar]
  54. Trlica, M. J. 1992. Grass growth and response to grazing. Colorado State University Cooperative Extension. [Google Scholar]
  55. Van Soest, P., Robertson J. B., and Lewis B. A... 1991. Methods for dietary fiber, neutral detergent fiber, and nonstarch polysaccharides in relation to animal nutrition. J. Dairy Sci. 74:3583–3597. doi: 10.3168/jds.S0022-0302(91)78551-2 [DOI] [PubMed] [Google Scholar]
  56. Wickham, H., and Bryan J... 2023. readxl: Read Excel Files.
  57. Zhang, W., Qin W., and Li H., and Wu A.-m... 2021. Biosynthesis and transport of nucleotide sugars for plant hemicellulose. Front. Plant Sci. 12:28. doi: 10.3389/fpls.2021.723128 [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

skaf151_suppl_Supplementary_Figures
skaf151_suppl_supplementary_figures.docx (3.4MB, docx)

Articles from Journal of Animal Science are provided here courtesy of Oxford University Press

RESOURCES