The study was conducted in the Holeta, Adaberga, Sululta, Bishoftu and Adama districts of the Oromia Regional State of Ethiopia ( Figure 1 ). The study areas were selected based on the potential availability of different feed markets and farms. All the study areas are characterized by a cool, sub-tropical climate with bimodal rainfall of 1200 – 1700 mm per year. The study areas are a mixed crop-livestock farming systems with the cultivation of a range of cereal and legume crops and livestock species including dairy and beef cattle, small ruminants and backyard poultry. Livestock are fed on home-grown feed resources and agro-industrial byproducts including oil seed meals.

The oilseed cake (n = 142) samples were collected from selected farms, oil factories, wholesalers, and retail shops, and comprised meals from linseed (also known as flaxseed, n=53), noug seed (n=22), cotton seed cake (n=56), groundnut (n= 7), soybean (n=2), sunflower (n=1), and rape seed (n=1). The total number of samples were determined according to Williams (2001), who recommended a minimum of 20 to 30 samples for initial calibration with larger numbers of samples improving the robustness of NIRS prediction equations.

The experiment involved collecting spectra data from intact ‘as received’ (Unground, UGr) and dried and ground samples (Gr), with several scan frequencies (one, two and ten scans per sample; 1s, 2s and 10s) and in triplicate. Thus, there were six treatments comprising UGr-1s, UGr-2s, UGr-10s, Gr-1s, Gr-2s and Gr-10s, and these were analysed as a 2x3 factorial completely randomized design with three replicates.

The samples were transported to the animal nutrition laboratory in the International Livestock Research Institute (ILRI), Addis Ababa. The NIRS spectra of each oilseed cake sample was measured in both ‘intact’ (UGr, i.e. ‘as received’) and Gr samples. The UGr samples were scanned before the samples were dried (forced air oven for 48 hours at 60°C), ground (1 mm screen, Wiley mill), and then scanned again. Both intact and Gr samples were packaged into transparent plastic bags (20 cm width and 30 cm length with 20 µm thickness) for scanning, and subsamples were stored for subsequent reference analyses ( Figure 2 ). The spectra of each Ugr and Gr sample was measured with one, two or ten scans (1s, 2s and 10s) in three replicates. Samples were scanned using a Tellspec Enterprise model handheld NIR spectrometer (Tellspec Enterprise). This NIR spectrometer is a smartphone NIR spectrometer weighing 140 g which measures from 900–1700 nm, with a spectral interval of approx. 3 nm (Crocombe, 2018; Rukundo et al., 2021). Data is collected using a smartphone. The data collection and management software tool designed by Tellspec Enterprise (DC&M2) was downloaded from the Google Play-store and installed on a smartphone with a Bluetooth connection to the instrument for spectra collection. The function of the instrument was checked daily using a standard white reference samples supplied by the manufacturer.

Wet chemistry analyses for each feed attribute in each of the samples was conducted to provide their reference values. Total nitrogen (TN) concentration was determined by the Kjeldahl method with the crude protein (CP) content calculated as N x 6.25 (AOAC, 1995). Neutral detergent fiber (NDF), acid detergent fiber (ADF) and acid detergent lignin (ADL) fractions were analyzed according to Van Soest and Robertson (1985). In vitro organic matter digestibility (IVOMD) was determined as described by Tilley and Terry (1963).

The spectra of the feed samples were exported to the WinISI 3.0 software for chemometric analyses. The spectral data was randomly divided into calibration (n= 95, 2/3) and validation (n=47, 1/3) sets to develop calibration models and then to test their performance (Westad and Marini, 2015; Despal et al., 2020). Calibration equations were developed using modified partial least squares (MPLS) regression after scatter correction using standard normal variate (SNV) and detrend. The mathematical treatment applied was (1, 6, 4, 1), where the first number indicates the order of derivative (the first derivative of log 1/R), and the gap in nm over which the derivative is calculated, the number of data points used in a first smoothing, and the number of nm over which the second smoothing was applied. The chemometrics were calculated with WinISI version 3.0 software, and the calibration and validation statistics included the standard error of calibration (SEC), R²cal,SEP, R²val, and the ratio of the SD of the population to the SEP (RPD).

The calibration and prediction models were developed by using WINISI 3.0 software. The data obtained from the calibration model and validation SEP were subjected to General Linear Model (GLM) using Statistical Analysis System (SAS, version, 9.0) (SAS, 2002). The Duncan multiple range test was employed for the separation of treatment means after checking for homogeneity of variance.

The model used for calibration and validation statistics mean separation was:

Where, Y_ij = Response variable,

The mean, standard deviation (SD) minimum (Min.) and maximum (Max.) values for each feed attribute of the various oilseed cake samples measured by wet chemistry are presented in Table 1 . The TN averaged 49 g/kg and ranged from 24 g/kg DM in cotton seed cake to 88 g/kg in a groundnut cake sample. The IVOMD averaged 657 g/kg DM and ranged from 529 g/kg DM in noug seed cake to 815 g/kg DM in a groundnut cake. The NDF, ADF, and ADL contents also had a wide range among the oilseed cake samples. Finally, the mean lignin content of the oil meals was 100 g/kg DM and ranged from 18 g/kg DM in peanut cake to 156 g/kg DM in noug cake.

Chemometric analysis of the spectra and the b-coefficients after scatter correction, detrend and first derivative math treatment indicated that a number of spectral regions in the measured range (915 - 1647 nm contributed to the modified partial least squares calibration equations ( Figure 3 ). Generally, the most important spectral regions contributing to the calibrations were 1000-1100 and 1400-1600 nm, but there were some differences among the feed attributes. The calibration for TN concentration depended primarily on the regions between 940 – 1080 and 1420-1590 nm. However, calibrations for IVOMD content depended primarily on the regions 1020 – 1090 and 1430-1560 nm, and for NDF content 1040-1080 and 1480-1590 nm. Hence the 920-1000 nm region was important for TN but not for IVOMD or NDF. The most important spectral regions to predict ADF and ADL tended to be similar to those for NDF.

The calibration and prediction equation statistics for the constituents TN and IVOMD of oilseed meals are shown in Table 2 . Both the number of scans (i.e. 1, 2 and 10) and grinding of the samples, and their interactions affected the accuracy of the TN and IVOMD calibrations (p<0.001). However, the number of scans had a much greater effect than the sample form. For TN increasing the number of scans of each sample from 1 to 10 decreased the SEP by 51% (from 6.5 to 3.2 g/kg DM) in Gr samples, but surprisingly had little effect on the SEP in UGr material. However, with measurement of IVOMD there was a different pattern where SEP was reduced by 28% due to increasing the number of scans in both Gr and UGr samples; also the SEP was consistently lower in unground samples. The highest R² for calibration and validation for N and IVOMD were obtained for Gr samples scanned ten times (calibration 0.95 and 0.89, validation 0.91 and 0.81, respectively). The relationships between the reference and predicted values for Gr samples scanned 10x are shown in Figure 4 .

The fiber composition of oilseed cake samples as NDF, ADF, and ADL were accurately predicted ( Table 3 ). The statistics of R²cal, SEC, R²val, and SEP were significantly affected by the interaction between the numbers of scans and grinding of samples (P<0.001). The highest R²cal for NDF and ADL were found for Gr (0.98 and 0.93, respectively) and UGr samples scanned ten times (0.97 and 0.93, respectively), whereas the maximum (0.92 and 0.94) R²val for these parameters were obtained for Gr samples scanned ten times, respectively. Similarly, the best R²cal and R²val for ADF were observed for Gr samples scanned ten times (0.96 and 0.90), respectively. For NDF and ADF increasing the number of scans of each sample from 1 to 10 reduced the SEP in Gr samples by 50% (from 53.8 to 26.9 g/kg DM) and 52% (from 53.1 to 25.3 g/kg DM), respectively, but not in UGr samples. However, when ADL was measured, the SEP was lowered by 67% (from 20.4 to 6.8 g/kg DM) due to increasing the number of scans in Gr samples, but there was little effect in UGr samples (23% reduction from 13.8 to 10.7). The relationships between the reference and predicted values for Gr samples scanned 10x are shown in Figure 5 .

The high variation in the chemical composition of the oilseed meals analyzed is a positive factor for calibration when predicting the composition of different types of feeds (Naes et al., 2002; Arzani et al., 2012). The overall mean of the total nitrogen (TN) content in the current study were in accordance with other studies e.g. the findings of Adugna (2008), who found TN content of oilseed meals ranging from 46.7 to 58.3 g/kg DM, and Fekede et al. (2015), who reported an average TN content of 45.7 to 57.2 g/kg DM g/kg DM for oilseed meals in the central highlands of Ethiopia. Likewise, the observed mean IVOMD was comparable with those reported by the above authors. However, in the current study, the mean ADF content was higher than reported by Fekede et al. (2015). The observed variations in the chemical composition of oilseed cake samples might be associated with the type of the oilseed cake, the method, and efficiency of extracting oil from the seeds, and environmental conditions for crop growth (Adugna, 2008; Fekede et al., 2015). This large variability in chemical composition of reference samples was considered suitable to develop NIRS calibrations (Dardenne et al., 2000).

There is extensive evidence in the literature that NIR spectroscopy is suitable for evaluating the TN content of various types of animal feeds (Andueza et al., 2011; Chen et al., 2015; Rushing et al., 2016). The observed R² values for prediction of TN in the present study were consistent with findings from Rukundo et al. (2021), who reported R²cal (0.96) and R²val (0.90) on mixed feeds using Foss and ASD Quality NIRS instruments, respectively. However, the R²cal value was higher than the reports of Aufrere et al. (1996); Modrono et al. (2017), and Wiedemair et al. (2019) who reported using Foss 6500 (0.92), Phazir (0.91), and Phazir (0.57) spectrophotometers for compound feeds, respectively, and lower than the study of Perez-Marin et al. (2004) who found 0.98 and 0.97 on ground and unground compound feeds, respectively, using Foss 6500 instrument. In the present study, the standard error of calibration (SEC) and standard error of prediction (SEP) for TN were lower than in previous studies reported by Aufrere et al. (1996); Swart et al. (2012), and Modrono et al. (2017). Also the results of the current study showed that as the number of scans increased from one to ten, the SEC decreased by 29% (from 3.1 to 2.2 g/kg DM) and 49% (from 4.7 to 2.4 g/kg DM) on D&Gr and UGr samples, respectively. Likewise, the SEP decreases by 51% (from 6.5 at one scan to 3.2 g/kg DM at ten scans) on Gr samples.

Apart from the TN content, the quality trait most frequently measured by NIR spectroscopy in forage crops is digestibility. The IVOMD calibration model had a greater R²cal (0.89) than those found previously by Rukundo et al. (2021) on mixed feeds and forage with a Foss XDS (0.82), Foss 6500 (0.80), and Tellspec (0.83) NIRS instruments and by Xiccato et al. (2003) who found (0.84) using a Foss 6500 with compound feeds. However, the R²cal and R²val values for the IVOMD model were lower than the values reported by Prasad et al. (2019) on roughage feeds using a Phazir (0.96) and Tellspec (0.91). This could be related to the narrow range of values found in the samples analysed for this parameter, in contrast to that of TN, NDF, ADF and ADL. The SEC value for IVOMD in the present study was in line with the report of Prasad et al. (2019) (26.3 g/kg DM) using a Tellspec handheld NIRS instrument on roughage feeds but lower than the report of Dereje et al. (2010) on natural pasture hay (41.2 g/kg DM) using a Foss 5000, and higher than the findings of Andueza et al. (2001) on alfalfa hay (20.5 g/kg DM) using a Foss 6500, and Prasad et al. (2019) on roughage feeds (20.8 g/kg DM) using a Phazir instrument. Similarly, the SEP for IVOMD prediction was lower than the study carried out by Rukundo et al. (2021) using Tellspec (34.5), Foss XDS (40.0), Foss 6500 (41.7), and ASD Quality (40.8) instruments. The present study indicated that when the number of scans increased from one to ten, the SEC decreased 14% (from 12.4 to 10.7 g/kg DM) and 22% (from 23.3 to 18.2 g/kg DM) on ground and unground samples, respectively. The SEP IVOMD prediction decreased as well by 29% (from 27.9 to 20.1) on UGr samples, and by 28% (from 16.2 to 17 g/kg DM) on Gr samples when the scanning frequency increased from one scan to ten scans.

Many studies showed that NDF and ADF concentrations could be well predicted by NIRS in forage (Chen et al., 2015; Rushing et al., 2016). In this study, the performance of the models for NDF and ADF were successful and the precision of the models was consistent with previous studies (Prasad et al., 2019; Mazabel et al., 2020). The R²cal value for the NDF equation in the current study was higher than those from the study reported by Aufrere et al. (1996); Modrono et al. (2017), and Swart et al. (2012) in compound feed and total mixed ration using Foss 6500 (0.92), Phazir (0.91), and FT-NIR (0.94), respectively. This finding shows that 98% and 97% of the observed variation in NDF concentration in both UGr and Gr oilseed cake samples, respectively, was explained by NIRS scans when samples were scanned ten times. Similarly, the accuracy of the Tellspec handheld NIRS to predict the NDF content of ground oilseed cake samples scanned in locally available plastic bags was excellent and the observed R²val (92%) fitted the regression line, or 92% of the variation observed on dependent variables was explained by independent variables. The SEC for the NDF equation was in accordance with the study conducted by Prasad et al. (2019). Likewise, the R²val value for the NDF prediction was in agreement with the report of Mazabel et al. (2020) and Andueza et al. (2001) on Brachiaria grass (0.92) and alfalfa hay (0.92), respectively. Similarly to other parameters, the SEC and SEP for NDF decreases by 52 and 50% as the scanning frequency increase from one to ten, respectively. Although the precision of the NDF (0.92) and ADF (0.90) models was also lower when compared to ADL (0.94), they were still sufficient to differentiate between high and low NDF and ADF content in oilseed meals.

The findings in the present study revealed that a ‘robust’ calibration was developed for the prediction of ADF content in oilseed cake samples, with higher R²cal (0.96) and R²val (0.90) and lower SEC (32.1 g/kg DM) and SEP (53.1 g/kg DM) values when compared with the results reported by other authors. The precision of the calibration model pertaining to the R²cal value obtained in this study was consistent with Aufrere et al. (1996) on compound feed using Foss 6500 (0.95), Swart et al. (2012) on the total mixed ration using an InfrAlyzer 500 (0.95), and Prasad et al. (2019) on roughage feeds using Tellspec (0.96), but higher than Xiccato et al. (2003) on compound feed using an InfraAlyzer 500 (0.84) instrument. Lyons and Stuth (1992) stated that an R² value of above 0.90 is considered good for agricultural products. Similarly, the R²val value obtained in this study was in accord with the report of Acosta et al. (2020) on grass (0.87) using a Nano NIRS instrument. It was higher than the study of Rukundo et al. (2021) who reported (0.63) R²val for mixed feed and forage using a Tellspec, and lower than Prasad et al. (2019), who obtained 0.97 and 0.96, respectively, using Phazir and Tellspec instruments on roughage feeds. The SEC and SEP for ADF reduced by 46 and 52%, respectively, when the scanning frequency increases from one to ten on Gr samples.

Among the fiber components, lignin is the one most commonly measured by NIR spectroscopy. Lignin calibrations have been reported mostly for grasses and legume species commonly used to produce hay and pasture in subtropical regions (Roberts et al., 2004). The R²cal (0.93) and R²val (0.94) for ADL was higher than the study reported by Goi et al. (2020) on ground (R²cal = 0.76) and intact (R²cal = 0.78) kibbles using a SCiO instrument, and Despal et al. (2020) on roughage feeds (R²val = 0.77) by FT-NIR, and Rukundo et al. (2021) on mixed feeds (R²val = 0.62). The R² of validation for ADL in the current study might be an indication of the high performance of the Tellspec instrument to predict the ADL content of unknown samples, and 94% of the observed data fitted the model. The SEC for ADL was comparable with the study conducted by Despal et al. (2020) on roughage feeds (10.9 g/kg DM) by using a FT-NIR instrument and higher than the study of Goi et al. (2020) who reported 0.30 and 0.29 on ground and intact kibble samples, respectively, using a SCiO NIR spectrometer device. The current result indicates that Tellspec has the potential to measure the ADL concentration of both Gr and UGr oilseed cake samples. The observed SEP value for ADL was higher than Despal et al. (2020) findings for roughage feeds (10.7 g/kg DM) using FT-NIR, but lower than Rukundo et al. (2021) results for mixed feeds (13.58) using the Tellspec instrument.

For most of the variables the ratio of standard deviation to the SEP (RPD) g/kg DM found was higher than 3 for NDF (RPD = 3.17), ADF (RPD = 3.05), and ADL (4.06), while RPD were less than 3 for TN (RPD =2.68) and IVOMD (RPD = 2.30). The RPD values represent the ability of the NIRS model to predict the nutrient content of animal feeds (Williams and Sobering, 1993). An RPD value of more than 2 was categorized as a relevant prediction of NIRS (Baillères et al., 2002). In this study, the developed oilseed meals prediction equation was relevant in predicting the oilseed cake samples composition because the RPD value found for TN, IVOMD, NDF, ADF, and ADL were higher than 2 for Gr samples scanned ten times. However, Lobos et al. (2013) gave a higher category (RPD>2.5) as a valid measurement ability of a NIRS model to predict the constituent of a particular feed. In this case, the IVOMD database produced needs to be improved. The low RPD found in the prediction of IVOMD might be related to the interaction between laboratory measurements and the collected spectra of a particular feed. According to Williams (2004) there are five categories of prediction accuracy based on RPD values, i.e., 1) the RPD<1.5 indicated an unusable; 2) the 1.5<RPD<2.0 categorized as the ability of the prediction to distinguish between high and low values; 3) the 2.0<RPD<2.5 produced an “approximate” quantitative prediction; 4) the 2.5<RPD<3.0 reflected a “good” quantitative prediction; and 5) the RPD >3.0 indicated an “excellent” quantitative prediction.

The discrepancy in R²cal, R²val, SEC, SEP, and RPD for TN, IVOMD, NDF, ADF, and ADL might be associated with instrumental differences, feed samples, feed forms, number of scans, scanning material, environmental conditions of crop growth, and the interaction between feed forms and number of scans. Generally, the current results indicate that increasing the number of scans leads to higher R-square values and lower standard errors of calibration (SEC) and standard errors of prediction (SEP). By scanning multiple times the various traits measured could be predicted with reasonable precision, implying that the composition predicted by NIRS agreed closely with that of chemical analysis for the studied quality components.