https://orcid.org/0000-0002-6177-1816
Figures
Abstract
The aim of this study was to develop and validate regression models to predict the chemical composition and ruminal degradation parameters of corn silage by near-infrared spectroscopy (NIR). Ninety-four samples were used to develop and validate the models to predict corn silage composition. A subset of 23 samples was used to develop and validate models to predict ruminal degradation parameters of corn silage. Wet chemistry methods were used to determine the composition values and ruminal degradation parameters of the corn silage samples. The dried and ground samples had their NIR spectra scanned using a poliSPECNIR 900–1700 model NIR sprectrophotometer (ITPhotonics S.r.l, Breganze, IT.). The models were developed using regression by partial least squares (PLS), and the ordered predictor selection (OPS) method was used. In general, the regression models obtained to predict the corn silage composition (P>0.05), except the model for organic matter (OM), adequately estimated the studied properties. It was not possible to develop prediction models for the potentially degradable fraction in the rumen of OM and crude protein and the degradation rate of OM. The regression models that could be obtained to predict the ruminal degradation parameters showed correlation coefficient of calibration between 0.530 and 0.985. The regression models developed to predict CS composition accurately estimated the CS composition, except the model for OM. The NIR has potential to be used by nutritionists as a rapid prediction tool for ruminal degradation parameters in the field.
Citation: Pucetti P, Valadares Filho SdC, Roque JV, da Silva JT, de Oliveira KR, Silva FAS, et al. (2024) Predicting ruminal degradability and chemical composition of corn silage using near-infrared spectroscopy and multivariate regression. PLoS ONE 19(4): e0296447. https://doi.org/10.1371/journal.pone.0296447
Editor: Aziz ur Rahman Muhammad, University of Agriculture Faisalabad, PAKISTAN
Received: June 14, 2023; Accepted: December 13, 2023; Published: April 18, 2024
Copyright: © 2024 Pucetti et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: All relevant data are within the paper and its Supporting information files.
Funding: This work was supported by the National Council of Scientific and Technological Development (CNPq - Grant Number - 465377/2014-9), National Institute of Science and Technology in Animal Science (INCT - Ciência Animal - Grant number - 465377/2014-9), and Coordination of Improvement of Personal Higher Education (CAPES, PROEX: 32002017011P9). The funding agency had no role in the study design, data collection, and analyses, decision to publish, or preparation of the manuscript. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors have declared that no competing interests exist.
Introduction
Corn silage (CS) is an important source of nutrients, particularly energy and digestible fiber, in modern dairy cattle operations worldwide and beef cattle operations in Europe, Latin America, and North America [1–5]. Many factors contribute to the high use of CS by producers, including lower harvesting costs, minimized risks of production, elevated yield per area, and flexibility to harvest corn for forage or grain [6]. However, CS quality is influenced by several factors, such as production technique, type of corn hybrid (e.g., flint corn), and climate factors. Thus, the use of tabular values of CS chemical composition found in the literature may not be appropriate to formulate ruminant diets. However, conventional laboratory analyses to measure dietary ingredient chemical composition are laborious and time-consuming.
In addition to feed chemical composition values, nutritional requirement systems for ruminants use ruminal degradability of nutrients to formulate diets with greater precision [7–9]. For example, measuring the ruminal degradability of dietary crude protein allows for predicting the supply of metabolizable protein, which is the total amount of true protein available for intestinal absorption. Additionally, there are models to predict milk yield that consider the effect of NDF degradability [7].
The in situ nylon bag technique is a widely adopted procedure to characterize the dynamics of degradation of feedstuffs and nutrients in the rumen. However, animal ethics committees have recommended stricter animal use protocols and the reduction or, if possible, the replacement of animals with alternative laboratory methods in research settings [10,11]. In addition, the costs of animals, surgical supplies, feeding, and labor are important limitations to these degradability measurements.
In this context, near-infrared (NIR) spectroscopy is an alternative approach to access feed composition and ruminal degradability that does not require the use of animals or chemical analyses. Furthermore, it requires little or no previous sample preparation. Near-infrared spectroscopy has been routinely used for the nutritional analysis of silage and other livestock feeds in the dairy and beef industries and the procedure is rapid and inexpensive as compared to in situ or in vivo measurements [12]. Less information is available accessing the rate and extent of ruminal degradability of nutrients using NIR approaches.
Some published studies demonstrate the potential of NIR spectroscopy to predict feed analysis [12–14]. However, these studies utilized the full-spectrum regression methodology, which do not incorporate variable selection methods that can prevent the use of irrelevant or redundant variables as well as variables that represent noise, commonly observed in large databases such as those used in NIR spectra studies[15–17].
Thus, we hypothesized that regression models can accurately predict both the composition and ruminal degradability parameters of CS produced in Brazil. Hence, the objectives of this study were to develop and evaluate regression models using the ordered predictor selection (OPS) approach to predict the chemical composition and ruminal degradation parameters of CS by NIR spectra.
Material and methods
All procedures were previously approved by the Animal Ethics and Welfare Committee of the Universidade Federal de Viçosa (#042/2019).
Sample collection and preparation
Overall, 94 CS samples were collected in eight states of Brazil (Bahia, Goiás, Minas Gerais, Mato Grosso, Pernambuco, Paraná, Rio Grande do Sul, and São Paulo) to provide sufficient variation of the CS feeding value to develop and evaluate the models. The samples were sent to the Ruminant Nutrition Laboratory of the Animal Science Department at the Universidade Federal de Viçosa (UFV), Viçosa, Minas Gerais, Brazil for processing and performing typical laboratory analyses. Samples were dried in a forced air oven (55°C) for 72 hours and ground to pass a 2-mm and 1-mm screen (Tecnal, Piracicaba, São Paulo, Brazil).
In situ degradation procedures
In situ degradability measurements were conducted at the Ruminant Nutrition Laboratory of the Animal Science Department at the Universidade Federal de Viçosa, Viçosa, Minas Gerais, Brazil.
The 23 samples were divided into 3 groups and ruminally incubated in three Nellore bulls, with an average BW of 450 ± 13 kg and 24 mo of age in a 3 × 3 Latin square design. Of the three corn silage groups, two contained eight corn silage samples, and one group contained seven corn silage samples. Within each period, each corn silage group was incubated in the rumen of a different bull. As reported by Machado et al. [18], the objective of the Latin square was to assist with the organization of the information that was collected in the field while allowing for measurements of degradation of different feeds without confounding the effect of the animal as well as to control sources of variation and avoid bias without estimating the variability.
The bulls were fed for ad libitum intake a 70:30 (DM basis) corn silage:concentrate (corn, soybean meal, wheat bran, urea, ammonium sulfate, salt, bicarbonate, and mineral) diet. The diet contained 12% crude protein (DM basis). The forage-to-concentrate ratio used was based on providing a balanced diet with diverse feed composition with enough fiber and energy to not compromise microbial growth. The animals were adapted to the experimental diet and conditions for 30 d before the in situ incubations.
Individually identified nylon bags (Sefar Nitex; Sefar AG, Thal, Switzerland; porosity of 50 μm-CA and 8 by 15 cm) were used containing 6.0 g of each sample that was previously ground using a 2-mm screen. Incubation times were 0, 3, 6, 12, 24, 48, 72 and 96 h. The number of bags varied as a function of the incubation time to guarantee enough residual sample after incubation (i.e. more bags per sample were incubated for the longer incubation times relative to the shorter incubation times). In situ bags containing samples were attached to a steel chain (90 × 2 cm) with a weight at the end, thus allowing for complete immersion within the ruminal fluid, below the fiber mat. Bags were placed into the rumen in reverse order of incubation hours such that all bags were removed at the same time for rinsing. After the incubation period, the bags were rinsed in running water followed by washing with cold tap water by hand by the same person. The endpoint for rinsing was when the rinse water was clear after flow through the bags [19]. The 0-h bags were not incubated in the rumen but were rinsed using the same procedures for rinsing as described above. Samples were oven-dried at 55°C for 72 h. Bags were placed in an oven at 105°C for 2 h and weighed. The residue was removed from the nylon bags, ground in a knife mill (Tecnal, Piracicaba, São Paulo, Brazil) with a 1-mm sieve, and placed in a labeled plastic bag.
Analytical methods for reference data
Samples were analyzed according to standard analytical methods of the Brazilian National Institute of Science and Technology in Animal Science [20] for DM and DM at 105°C (method INCT—CA G-003/1), ash (complete combustion in a muffle furnace at 600°C; method INCT-CA M-001/1), CP (Kjeldahl procedure; method INCT-CA N-001/1), EE (Randall extraction; method INCT-CA G-005/1), NDF (using a heat-stable α-amylase, omitting sodium sulfite; method INCT-CA F-001/1), ADF (method INCT–CA F-003/1), and lignin (method INCT-CA F-005/1). The OM concentrations were estimated by the difference between the DM and ash concentrations. The NDF and ADF were corrected for residual N compounds (NDIP = method INCT—CA N-004/1; and ADIP = method INCT—CA N-005/1) and ash (method INCT—CA M002/1 and method INCT—CA M003/1). The iNDF concentrations was evaluated as the residual NDF remaining after 288 h of ruminal in situ incubation according to Casali et al. [21]. Starch concentrations were analyzed following the method described by Zinn [22] and modified by Silva et al. [23]. Sample NFC concentrations were analyzed as suggested by Detmann and Valadares Filho [24] as NFC = OM—CP—apNDF—EE, where apNDF is the NDF corrected for ash and residual N compounds, as described above. The residues from in situ incubations were analyzed for DM, OM, CP, and NDF by the methods described above.
In situ degradation parameters
In situ degradation parameters for DM, OM, and CP were estimated by using the first-order asymptotic model described by Ørskov and McDonald [25]:
where Yt = degraded fraction of DM, OM or CP in time ‘t’, (%); a = readily soluble fraction, (%); b = potentially degradable fraction in the rumen, (%); kd = rate constant for degradation of b, per h; and t = time, h.
The NDF degradation parameters were estimated by using the model proposed by Van Milgen et al. [26]
where RNDFt = non-degraded NDF residue at time “t” (%); b = potentially degradable fraction in the rumen (%); λ = joint fractional rate of latency and degradation (h-1); t = time independent variable (h); I = undegradable fraction (%).
The calculation for the equation used for DM and OM is for the degraded fraction and the calculation for the equation used for NDF is for the non-degraded residue. Therefore, NDF degradation was calculated by difference: Degradation (%) = 100 –residue.
The NDF degradation rate was calculated based on λ, using the properties of the Γ(2) distribution [27]:
where kd = rate constant for degradation of b (% per h); λ = joint fractional rate of latency and degradation (h-1). The parameters "a", "b", kd, λ, and “I” of the in situ incubation models were estimated using the PROC NLIN procedure (version 9.4, SAS Institute Inc., Cary, NC, USA), and assuming the Marquardt algorithm for convergence.
Acquisition of spectral data
Ground CS samples were homogenized, placed in Petri dishes (60 mm diameter), and scanned at 2-nm intervals from 902 to 1,680 nm, using a poliSPECNIR 900–1700 model NIR sprectrophotometer (ITPhotonics S.r.l, Breganze, IT.). Spectra were recorded three times, averaged, and stored as the logarithm of the reciprocal of reflectance (1/R). The PoliDATA software (ITPhotonics S.r.l, Breganze, IT) was used to acquire spectral data.
Regression models
The NIR spectra data were exported as an.xls file using the SensorLogic GmbH software (Software + Sensor Systeme, Norderstedt, Germany) and imported by the Matlab 2019b software (Math Works, Natick, USA). A data matrix with the NIR spectra was built and named X matrix for each evaluated property, which represented the independent variables. The rows of the X matrix corresponded to the samples, and the columns corresponded to the variables (wavelength).
Twenty-seven properties were determined and identified as y vectors (dependent variables). The properties evaluated to estimate the CS composition were DM, DM at 105°C, OM, CP, NDIP, ADIP, EE, NDF, apNDF, iNDF, ADF, apADF, lignin, NFC, and starch. The in situ degradation parameters evaluated to estimate ruminal degradation of CS were the “a” of DM, OM, and CP, “b” and “kd” of DM, OM, CP, and NDF, and “I” of NDF. The y vector had a length equal to the number of rows in the X matrix. For each property, a dataset was prepared.
Post-outlier removal, the dataset was then divided into calibration and prediction sets using the Kennard-Stone algorithm [28]. The Kennard-Stone algorithm is advantageous as it aids in the creation of representative subsets for model development, contributing to enhanced predictive accuracy. Additionally, the Kennard-Stone algorithm is known for its computational efficiency, making it a practical choice for large datasets. The data set for ruminal degradation of CS was not split, because of the limited number of samples evaluated, and the model was evaluated by cross-validation. The y variable was mean-centered for all properties evaluated and different pre-treatments were studied for each X matrix. The pre-treatments tested were mean centering, autoscale, smoothing, first and second derivative, multiplicative scatter correction (MSC), normalize, baseline, and standard normal variate (SNV), and detrend. Combinations of two, three, and four pre-treatments were also evaluated. The best pre-treatment for each property was chosen based on the lowest root mean square error of cross-validation (RMSECV).
Partial least squares regression (PLS) was used to develop the models using variables selected through three approaches of the OPS method: automatic OPS (autoOPS), feedback OPS (feedOPS), and OPS intervals (iOPS) as described by Roque et al. [29]. The OPS algorithms were applied using windows of 10 and increments of 5 variables where 100% of variables were tested and random cross-validation was applied and splits were set at 10% of the X matrix rows. In feedOPS, the convergence criteria were 2% as the minimum difference between two consecutive RMSECV and 10 as the maximum number of loops. In iOPS, when the option to run the selection using feedOPS was used, the convergence criteria were the same as those used in feedOPS. The X matrix was divided into intervals of 10% of its size, limited in at least 50 variables. Additionally, the number of latent variables for OPS (hOPS) was calculated for each interval in iOPS.
The prediction set was used to validate the predictive capacity of the calibration models developed. Each property value from the laboratory typical analyses and the value predicted by the calibration models were compared. The root mean square error of prediction (RMSEP) and cross-validation (RMSECV) and the correlation coefficient between the values measured in the typical laboratory analyses and those predicted by the model (RP) and in the cross-validation (RCV), were used as parameters to verify the predictive capacity of the calibration models.
Also, the CS constituents measured by typical laboratory analyses and those predicted by regression models were compared using the following regression model: y = β0 + β1 × x, where x = predicted values; y = observed values; β0 = intercept of equation; and β1 = slope of equation. Regression was conducted according to the following statistical hypothesis [30]: H0: β0 = 0 and β1 = 1; and Ha: not H0. If the null hypothesis was not rejected, it was concluded that the equations accurately and precisely estimated the CS feeding value. When the regression coefficients β0 and β1 are 0 and 1, respectively, predicted and observed values were considered equivalent.
Estimates were also evaluated using the estimated value of the mean squared error of prediction (MSEP) and its components [31]: MSEP = SB + MaF + MoF = 1/n ∑ i = 1 (xi − yi) 2, SB = (xi − yi) 2, Maf = (sx − sy) 2, Mof = 2sxsy (1 − r), where x = predicted values; y = observed values; MSEP = mean squared error of prediction; SB = squared bias; MaF = component relative to the magnitude of random fluctuation; MoF = component relative to the model of random fluctuation; sX and sY = standard deviations of predicted and observed values, respectively; and r = Pearson’s linear correlation between predicted and observed values. The smallest MSEP indicates the best model in the evaluation. These calculations can indicate whether the model error is associated with the SB, errors related to the high dispersion of data around the mean (MaF), or systematic errors concerning the direction of the curve predicted (MoF).
For all variance and covariance calculations, the total number of observations was used as the denominator [31]. The prediction efficiency was determined by estimating the correlation and concordance coefficient (CCC) or reproducibility index as described by Tedeschi [32]. With the CCC, it is possible to ascertain whether the model is overestimating or underestimating the observed values (the closer to 1, the better the model) in addition to giving an indication of the model’s precision and accuracy. Evaluation analyses were performed using the Model Evaluation System (32) and significance was established at α = 0.05.
Results and discussion
CS composition
The NIR spectra of the CS samples (n = 94) are shown in Fig 1. The descriptive statistics of the size of the calibration and prediction sets, and the pre-treatments performed in the respective datasets for CS constituents are provided in Table 1. A wide variation in CS composition was found in the data set, which is commonly observed in production systems such as Brazil. The variation in concentrations of CP (5.00–9.06% MS), NDF (34.15–57.81% MS), and starch (11–40.53% MS) demonstrate the concerns of using book values for CS to formulate diets.
For all CS constituents, the regression models obtained using pre-treated spectra resulted in lower RMSECV, indicating a better model fit and precision for regression models that underwent pre-treatments (Table 1). Pre-treatments are mathematical tools used to adjust multivariate regression models and correct random and systematic errors [33]. Pre-treatment could, for example, correct disturbances caused by physical phenomena such as dispersions (multiplicative scatter correction), equalize the magnitude of samples through a normalization factor (normalization), and minimize interferences caused by noise (smoothing), among others. Therefore, the use of pre-treatments usually improves the goodness of fit of models.
The autoOPS, feedOPS, and iOPS algorithms were applied simultaneously during model development. Thus, it was possible to select regions of the NIR spectra with more relevant information and more correlated with the evaluated CS constituents (Fig 2). Variable selection is an important step in multivariate regression and has become a fundamental tool in many different research areas. Moreover, the proper choice of variables in the X matrix may improve the goodness of fit of multivariate regression models. Thus, the proper selection of variables can be obtained by the triage of certain regions of the spectrum (wavelength set), which may minimize the error of prediction. As a result, more robust, simpler, and more accurate models may be obtained [34].
NIR spectrum regions selected by OPS algorithms in regression models building to predict the concentrations of (A) dry matter (DM), (B) dry matter at 105°C (DM105), (C) organic matter (OM), (D) crude protein (CP), (E) neutral detergent insoluble protein (NDIP), (F) acid detergent insoluble protein (ADIP), (G) ether extract (EE), (H) non-fibrous carbohydrates (NFC), (I) neutral detergent insoluble fiber (NDF), (J) NDF corrected for ash and protein (apNDF), (K) indigestible NDF (iNDF), (L) acid detergent insoluble fiber (ADF), (M) ADF corrected for ash and protein (FDAcp), (N) lignin and (O) starch of corn silage.
The number of variables selected by the OPS algorithms, the number of latent variables selected in the calibration process, and the performance parameters for all developed models are presented in Table 2. In general, models presented RMSEP values that suggest good fit of them. Those developed to estimate OM and ADIP presented a lower prediction correlation coefficient than the other models, demonstrating a possible fit problem for these models. The graphs with the relative prediction errors for each model are shown in Fig 3. These results indicate that the calibration models presented a good predictive capacity.
The relative error in the prediction of the concentrations of (A) dry matter (DM), (B) dry matter at 105°C (MS105), (C) organic matter (OM), (D) crude protein (CP), (E) neutral detergent insoluble protein (NDIP), (F) acid detergent insoluble protein (ADIP), (G) ether extract (EE), (H) non-fibrous carbohydrates (NFC), (I) neutral detergent insoluble fiber (NDF), (J) NDF corrected for ash and protein (apNDF), (K) indigestible NDF (iNDF), (L) acid detergent insoluble fiber (ADF), (M) ADF corrected for ash and protein (apADF), (N) lignin and (O) starch of corn silage by the regression models.
Zicarelli et al. [14] employed a NIR spectrophotometer to assess undried samples, yielding RP and RMSEP values of 0.9 and 1.9, respectively, for DM. These parameters are greater to those obtained in the current study. This disparity can be attributed to the fact that, in this study, the samples were dried and ground, which required the models to predict the water content, which is no longer present in the samples. The authors also provided model parameters for starch, CP, NDF, ADF, and EE, all of which exhibited lower RP and RMSEP values. The reduced RP reported by these authors suggests that the selection of variables can enhance model accuracy, given that their models were developed using a full-spectrum regression methodology. The lower RMSEP observed in their study can be attributed to the larger sample size they employed.
The results of the independent dataset prediction for each CS constituent are presented in Tables 3–5. Corn silage concentrations of DM, DM at 105°C, EE, CP, ADIP, NDIP, NDF, apNDF, iNDF, ADF, apADF, lignin, NFC, and starch were correctly estimated by the developed models as they did not reject the null hypothesis of intercept and slope equal to 0 and 1 respectively (P > 0.05). Moreover, the MSEP decomposition indicated that prediction errors were mostly (more than 70% of the MSEP) associated with random errors (Mof), suggesting that the prediction errors were not associated with the developed models. Furthermore, the calibration models precisely estimated the concentrations of DM at 105°C, EE, CP, NDIP, NDF, apNDF, iNDF, ADF, apADF, lignin, NFC, and starch of CS, as the CCC was close to 1 (CCC ≥ 0.713); in contrast, models for predicting the concentrations of DM and ADIP presented CCC of 0.665 for DM and 0.546 for ADIP. The calibration model for predicting the OM content of CS had adjustment problems according to the regression analysis (P < 0.05), indicating poor accuracy. In addition, 42.23% of the MSEP of the prediction of OM concentration in CS was associated with the central tendency or bias (SB) and systematic errors (Maf), indicating a model adjustment problem.
Graphs with values measured by conventional laboratory methods and predicted by the regression models are shown in Fig 4.
Measured (x axis) vs. predicted (y axis) values obtained by developed regression models for the concentrations of (A) dry matter (DM), (B) dry matter at 105°C (MS105), (C) organic matter (OM), (D) crude protein (CP), (E) neutral detergent insoluble protein (NDIP), (F) acid detergent insoluble protein (ADIP), (G) ether extract (EE), (H) non-fibrous carbohydrates (NFC), (I) neutral detergent insoluble fiber (NDF), (J) NDF corrected for ash and protein (apNDF), (K) indigestible NDF (iNDF), (L) acid detergent insoluble fiber (ADF), (M) ADF corrected for ash and protein (apADF), (N) lignin and (O) starch of corn silage. (■) Calibration set; (●) Prediction set.
The OM content is obtained by calculation (DM—Ash), thus cumulating the errors of the separate analyses, which may explain the lack of accuracy of the model. Moreover, the chemical constituents of OM can vary greatly across samples.
Ruminal degradability parameters
The descriptive statistics of the size of the calibration set and the pre-treatments performed in the dataset for ruminal degradation parameters of CS are provided in Table 6. A wide variation in ruminal degradation parameters was detected in the study, demonstrating that the ruminal degradability of CS varies significantly. For example, CP ruminal degradation parameters ranged from 57.21 to 78.92 for fraction "a", 14.63 to 33.51 for fraction "b", and 0.03 to 0.08 for rate "kd". Average values are 60 for fraction "a", 24 for fraction "b" and 0.04 for "kd" [7]. This fact demonstrates the difficulty of using tabulated values and the need for faster and more viable analytical methods, which will allow for greater precision in the formulation of diets.
For all ruminal degradability parameters, the regression models obtained using pre-treated spectra resulted in lower values of RMSECV, indicating a better model fit and precision. It was not possible to find suitable models to predict the fraction “b” of OM and CP, and the “kd” of OM. The microbial contamination of residual particles of incubated feeds is an important source of errors in the in situ method, resulting in the underestimation of CP degradability [18,35,36]. Thus, such a source of error can impact the construction of prediction models for CP and OM degradation parameters.
The selected regions are shown in Fig 5. The number of variables selected by the OPS algorithms, the number of latent variables selected in the calibration process, and the performance parameters for all developed models are presented in Table 7. The developed models to predict "a" fraction of DM, OM and CP achieved Rcv values between 0.802 to 0.985 and the model of "a" fraction of CP showed the highest RMSEcv among them. The model to predict the "b" fraction of DM presented a Rcv of 0.776 and RMSEcv of 2.146. The variable selection was not efficient to improve the model to predict the fraction “b” of the NDF, since all the variables of the spectrum were used. The prediction models for the NDF "b" and "I" fractions, and "kd" rate presented Rcv of 0.66, 0.53, and 0.90.
NIR spectrum regions selected by OPS algorithms in regression models building to predict the readily soluble fraction “a” of (A) dry matter (DM), (B) organic matter (OM), (C) crude protein (CP), the potentially degradable fraction in the rumen “b” (D) of DM, (E) of neutral detergent fiber (NDF), (F) NDF undegradable fraction “I”, rate constant for degradation of b “kd” of (G) DM, (H) CP, and (I) NDF of corn silage.
Graphs illustrating the relationship between values measured by the in situ method and predicted by the regression models are shown in Fig 6. The development of more comprehensive regression models would increase their accuracy and precision, but this would require the addition of corn silage samples collected across multiple years and various environments; this would increase the spectral, composition, and degradability diversity of the samples compared to those presented in the current study.
Measured (x axis) vs. predicted (y axis) values obtained by developed regression models for the readily soluble fraction “a” of (A) dry matter (DM), (B) organic matter (OM), (C) crude protein (CP), the potentially degradable fraction in the rumen “b” (D) of DM, (E) of neutral detergent fiber (NDF), (F) NDF undegradable fraction “I”, rate constant for degradation of b “kd” of (G) DM, (H) CP, and (I) NDF of corn silage.
Thus, this study has demonstrated that NIR spectroscopy associated with chemometric methods has the potential to be used in the prediction of more complex variables, facilitating the correct and precise application of these variables in the field. Thus, new and broader research in this area should be encouraged.
Conclusion
The regression models developed to predict CS composition accurately estimated the concentrations of DM, DM at 105°C, CP, ADIP, NDIP, EE, NDF, apNDF, iNDF, ADF, apADF, lignin, NFC, and starch. The models developed to predict the ruminal degradation parameters showed moderate predictive performance and has potential to be used by nutritionists as a rapid prediction tool in the field. Further development of the models using larger and more diverse sample datasets would improve model robustness and accuracy.
Supporting information
S1 File. Near-infrared spectra and laboratory values used to develop regression models for predicting composition of corn silage.
https://doi.org/10.1371/journal.pone.0296447.s001
(XLSX)
S2 File. Near-infrared spectra and ruminal in-situ degradation parameters used to develop models for predicting ruminal degradability of corn silage.
https://doi.org/10.1371/journal.pone.0296447.s002
(XLSX)
References
- 1. Keady TWJ, Kilpatrick DJ, Mayne CS, Gordon FJ. Effects of replacing grass silage with maize silages, differing in maturity, on performance and potential concentrate sparing effect of dairy cows offered two feed value grass silages. Livest Sci. 2008 Dec 1;119(1–3):1–11.
- 2.
Adesogan AT. Challenges of Tropical Silage Production Proceeding. In: 15th International Silage Conference. Madison, Winsconsin.; 2009.
- 3. Queiroz OCM, Ogunade IM, Weinberg Z, Adesogan AT. Silage review: Foodborne pathogens in silage and their mitigation by silage additives. J Dairy Sci. 2018 May;101(5):4132–42. pmid:29685282
- 4. Silva DP, Pedroso AM, Pereira MCS, Bertoldi GP, Watanabe DHM, Melo ACB, et al. Survey of management practices used by brazilian dairy farmers and recommendations provided by 43 dairy cattle nutritionists. Can J Anim Sci. 2019;99(4):890–904.
- 5. Silvestre AM, Millen DD. The 2019 Brazilian survey on nutritional practices provided by feedlot cattle consulting nutritionists. Revista Brasileira de Zootecnia [Internet]. 2021 Jul 5;50(2019). Available from: https://www.rbz.org.br/article/the-2019-brazilian-survey-on-nutritional-practices-provided-by-feedlot-cattle-consulting-nutritionists/.
- 6.
Allen MS, Coors JG, Roth GW. Corn Silage. 2003 [cited 2023 Jan 24]; https://acsess.onlinelibrary.wiley.com/doi/10.2134/agronmonogr42.c12.
- 7.
Nutrient Requirements of Dairy Cattle. 8th Revised Edition. Washington, D.C.: National Academies Press; 2021.
- 8.
NASEM. Nutrient Requirements of Beef Cattle, 8th Revised Edition [Internet]. Washington, D.C.: National Academies Press; 2016. http://www.nap.edu/catalog/19014.
- 9.
Valadares Filho SV, Costa e Silva LF, Gionbelli MP, Rotta PP, Marcondes MI, Chizzotti ML, et al. BR-CORTE—Nutrient Requirements of Zebu and Crossbred Cattle. 3rd ed. Viçosa, MG: Suprema Gráfica Ltda, Viçosa; 2016.
- 10. Hartung T. Thoughts on limitations of animal models. Parkinsonism Relat Disord. 2008 Jul;14:S81–3. pmid:18585949
- 11. Silva BC, Pacheco MVC, Godoi LA, Silva FAS, Zanetti D, Menezes ACB, et al. In situ and in vitro techniques for estimating degradation parameters and digestibility of diets based on maize or sorghum. J Agric Sci. 2020 Mar 22;158(1–2):150–8.
- 12. Thomson AL, Humphries DJ, Rymer C, Archer JE, Grant NW, Reynolds CK. Assessing the accuracy of current near infra-red reflectance spectroscopy analysis for fresh grass-clover mixture silages and development of new equations for this purpose. Anim Feed Sci Technol [Internet]. 2018;239(March):94–106. Available from:
- 13. Zicarelli F, Sarubbi F, Iommelli P, Grossi M, Lotito D, Lombardi P, et al. Nutritional Characterization of Hay Produced in Campania Region: Analysis by the near Infrared Spectroscopy (NIRS) Technology. Animals. 2022 Nov 1;12(21). pmid:36359159
- 14. Zicarelli F, Sarubbi F, Iommelli P, Grossi M, Lotito D, Tudisco R, et al. Nutritional Characteristics of Corn Silage Produced in Campania Region Estimated by Near Infrared Spectroscopy (NIRS). Agronomy. 2023 Mar 1;13(3).
- 15. Goodarzi M, Heyden Y Vander, Funar-Timofei S. Towards better understanding of feature-selection or reduction techniques for Quantitative Structure–Activity Relationship models. Trends in Analytical Chemistry [Internet]. 2013 Jan;42:49–63. Available from: https://linkinghub.elsevier.com/retrieve/pii/S0165993612002804.
- 16. Nadler B, Coifman RR. The prediction error in CLS and PLS: the importance of feature selection prior to multivariate calibration. J Chemom [Internet]. 2005 Feb;19(2):107–18. Available from: https://onlinelibrary.wiley.com/doi/10.1002/cem.915.
- 17. Xiaobo Z, Jiewen Z, Povey MJW, Holmes M, Hanpin M. Variables selection methods in near-infrared spectroscopy. Anal Chim Acta [Internet]. 2010 May;667(1–2):14–32. Available from: https://linkinghub.elsevier.com/retrieve/pii/S0003267010003582. pmid:20441862
- 18. Machado PAS, Valadares Filho SC, Detmann E, Santos SA, D Valadares RF, Ducatti C, et al. Development of equations to estimate microbial contamination in ruminal incubation residues of forage produced under tropical conditions using 15 N as a label. J Anim Sci [Internet]. 2013 [cited 2023 Jan 24];91:3836–46. Available from: https://academic.oup.com/jas/article/91/8/3836/4731422. pmid:23658352
- 19. Zanetti D, Menezes ACB, Silva FAS, Costa e Silva LF, Rotta PP, Detmann E, et al. In situ and in vitro estimation of mineral release from common feedstuffs fed to cattle. J Agric Sci. 2017 Sep 13;155(7):1160–73.
- 20.
Detmann E, Souza MA, Valadares Filho SC, Queiroz AC, Berchielli TT, Saliba EOS, et al. Métodos para análise de alimentos. Visconde do Rio Branco, MG: Suprema; 2012. 214 p.
- 21. Casali AO, Detmann E, de Valadares Filho S C, Pereira JC, Henriques LT, de Freitas SG, et al. Influência do tempo de incubação e do tamanho de partículas sobre os teores de compostos indigestíveis em alimentos e fezes bovinas obtidos por procedimentos in situ. Revista Brasileira de Zootecnia. 2008;37(2):335–42.
- 22. Zinn RA. Influence of Flake Density on the Comparative Feeding Value of a Barley-Corn Blend for Feedlot Cattle. J Anim Sci. 1997;75(4):904–9. pmid:9110200
- 23. Silva de BC, Godoi LA, de Valadares Filho S C, Zanetti D, Benedeti PDB, Detmann E. A suitable enzymatic method for starch quantification in different organic matrices. MethodsX [Internet]. 2019;6:2322–8. Available from: https://linkinghub.elsevier.com/retrieve/pii/S221501611930264X. pmid:31667131
- 24. Detmann E, Valadares Filho SC. On the estimation of non-fibrous carbohydrates in feeds and diets. Arq Bras Med Vet Zootec. 2010;62(4):980–4.
- 25. Ørskov ER, McDonald I. The estimation of protein degradability in the rumen from incubation measurements weighted according to rate of passage. J Agric Sci. 1979;92(1):499–503.
- 26. Van Milgen J, Murphy MR, Berger LL. A Compartmental Model to Analyze Ruminal Digestion. J Dairy Sci. 1991 Aug;74(8):2515–29. pmid:1918531
- 27.
Ellis WC, Matis JH, Hill TM, Murphy MR. Methodology for Estimating Digestion and Passage Kinetics of Forages. In John Wiley & Sons, Ltd; 1994. p. 682–756.
- 28. Kennard RW, Stone LA. Computer Aided Design of Experiments. Technometrics. 1969;11(1):137–48.
- 29. Roque J V., Cardoso W, Peternelli LA, Teófilo RF. Comprehensive new approaches for variable selection using ordered predictors selection. Anal Chim Acta. 2019;1075:57–70. pmid:31196424
- 30. Mayer DG, Stuart MA, Swain AJ. Regression of real-world data on model output: An appropriate overall test of validity. Agric Syst. 1994;45(1):93–104.
- 31. Kobayashi K, Salam MU. Comparing Simulated and Measured Values Using Mean Squared Deviation and its Components. Agron J [Internet]. 2000 Mar;92(2):345–52. Available from: https://onlinelibrary.wiley.com/doi/abs/10.2134/agronj2000.922345x.
- 32. Tedeschi LO. Assessment of the adequacy of mathematical models. Agric Syst [Internet]. 2006 Sep;89(2–3):225–47. Available from: https://linkinghub.elsevier.com/retrieve/pii/S0308521X05002568.
- 33. Pasquini C. Near infrared spectroscopy: Fundamentals, practical aspects and analytical applications. J Braz Chem Soc. 2003;14(2):198–219.
- 34. Oliveira FCC, De Souza ATPC, Dias JA, Dias SCL, Rubim JC. The choice of the spectral region in the use of spectroscopic and chemometric methods. Quim Nova. 2004;27(2):218–25.
- 35. Wulf M, Südekum KH. Effects of chemically treated soybeans and expeller rapeseed meal on in vivo and in situ crude fat and crude protein disappearance from the rumen. Anim Feed Sci Technol. 2005 Feb;118(3–4):215–27.
- 36. Westreicher-Kristen E, Steingass H, Rodehutscord M. In situ ruminal degradation of amino acids and in vitro protein digestibility of undegraded CP of dried distillers’ grains with solubles from European ethanol plants. Animal. 2013;7(12):1901–9. pmid:24237671