Ⅰ. INTRODUCTION
Italian ryegrass (Lolium multiflorum Lam.) is an important annual forage crop in temperate livestock production systems. It is widely cultivated because of its rapid growth, high forage yield, and good nutritional quality. In Korea, Italian ryegrass is commonly grown as a winter forage crop and plays an important role in maintaining forage supply for ruminants during periods of limited pasture availability (Oh et al., 2021; Shin et al., 2012). As the demand for high-quality forage increases, a stable supply of good-quality seed is also important for successful forage production. Seed quality affects germination, seedling establishment, crop growth, and final forage productivity (Reed et al., 2022).
Among the different factors affecting seed quality, moisture content is particularly important during harvesting, processing, and storage. High seed moisture can increase respiration and provide favorable conditions for fungal and bacterial growth, which may reduce seed viability during storage (Martín et al., 2022). Excessive drying may also damage seed tissues and increase their susceptibility to mechanical injury (Bewley and Black, 2013). Seed moisture and storage conditions are therefore closely related to seed deterioration and longevity (Corbineau, 2024). Accurate determination of moisture content is important for maintaining seed quality during postharvest handling and storage.
The oven-drying method is commonly used as a reference method for determining seed moisture content. Although this method provides reliable results, it requires several hours and destroys the analyzed sample. These limitations make it less suitable when a large number of samples need to be analyzed rapidly. NIRS provides an alternative method because it allows rapid and non-destructive measurement with little sample preparation (Manley, 2014). NIRS is based on the interaction of near-infrared radiation with molecular bonds such as O–H, C–H, and N–H, and the resulting spectral information can be related to chemical and physical properties using multivariate analysis (Siesler et al., 2008).
Moisture is suitable for NIRS analysis because water has strong absorption characteristics in the near-infrared region. In particular, O–H bonds produce important absorption bands that can be related to differences in water content among biological materials (Cen and He, 2007; Manley, 2014). NIRS has been used for moisture determination and quality evaluation in different agricultural materials. Fassio and Cozzolino (2004) reported that NIRS could be used for non-destructive prediction of moisture and other chemical components in intact sunflower seeds. Similarly, (Oblath et al., 2016) developed NIRS calibration models for intact Brassicaceae seeds and reported good prediction performance for seed moisture content. These studies show that NIRS can provide useful information for rapid moisture determination in intact seed samples.
However, the performance of a NIRS calibration model depends on the samples used for model development. The calibration population should represent the range and variation expected in samples that will be analyzed in the future. Differences in production year, growing conditions, cultivar, harvest stage, and postharvest management may influence the characteristics of agricultural samples and their spectral information. (Shetty et al., 2012) showed that representative calibration samples covering variation among years, species, and cultivars are important for developing reliable NIRS models. Therefore, evaluation using samples collected from different production years can provide useful information about the stability and practical application of a calibration model.
In the first production year of the present study, a NIRS calibration model was developed using Italian ryegrass seed samples collected in 2025. The model showed high calibration and cross-validation performance, but independent validation was based on only nine additional samples. To further evaluate the method, additional seed samples were analyzed in 2026, including a larger independent validation dataset. In 2026, 360 samples were used for calibration and cross-validation, while 90 samples were kept separately for independent validation. This two-year approach allowed the performance of NIRS to be evaluated using different annual sample populations and provided a stronger assessment of prediction performance using independent samples.
Therefore, the objectives of this study were to develop and evaluate NIRS calibration models for determining the moisture content of Italian ryegrass seeds using samples from two production years, to compare model performance between the 2025 and 2026 datasets, and to evaluate the predictive ability of the 2026 model using a larger independent validation set. The results were also used to examine the importance of including samples from different production years when developing a reliable NIRS calibration model for routine seed moisture evaluation.
Ⅱ. MATERIALS AND METHODS
1. Sample preparation and experimental design
Italian ryegrass (Lolium multiflorum Lam.) seed samples collected during two production years were used in this study. In 2025, a total of 222 seed samples were collected from Pyeongchang, Jangheung, and Gimje, Korea. After collection, the samples were transported to the Grassland and Forage Science Laboratory, Pyeongchang Campus, Seoul National University, Republic of Korea. The samples were manually cleaned to remove plant debris, broken seeds, dust, and other foreign materials before moisture and spectral analyses.
In 2026, an additional 450 Italian ryegrass seed samples were analyzed using the same experimental procedure. The 2026 samples were divided into calibration and independent validation datasets before final model development. Of the 450 samples, 360 were used for calibration and cross-validation, while the remaining 90 samples were kept separately for independent validation. The independent validation samples were not included during calibration model development or model selection. This design allowed the final model to be evaluated using samples that were independent of model development.
2. Reference moisture determination
The reference moisture content of the seed samples was determined using the same oven-drying procedure in both production years. Approximately 10 g of each sample was weighed using an analytical balance and placed in a labeled paper bag. The samples were dried in an oven at 105°C for 4 h. After drying, the samples were immediately transferred to a desiccator and allowed to cool to room temperature before being weighed again. The use of a desiccator helped prevent the reabsorption of atmospheric moisture during cooling.
Moisture content was calculated from the difference between the initial and dried sample weights and expressed as a percentage of the initial sample weight using the following equation:
The moisture values obtained by oven drying were used as reference values for NIRS calibration, cross-validation, and independent validation.
3. Spectral data collection using NIRS
NIR spectra were collected using a SpectraStar XT spectrometer (Unity Scientific, USA) equipped with a circular sample holder. Each seed sample was evenly distributed in the sample holder to provide uniform surface coverage and consistent sample presentation. All samples were measured under the same instrument conditions to reduce spectral variation caused by differences in sample placement.
Spectra were recorded at 1-nm intervals. In 2025, the spectral range used for calibration was 680–2500 nm. The same spectral acquisition procedure was followed for the 2026 samples. The collected spectral data were stored and matched with the corresponding oven-determined moisture values before calibration analysis.
4. Development of the NIRS calibration models
Partial least squares (PLS) regression was used to establish the relationship between the NIR spectra and reference moisture values. Before model development, standard normal variate (SNV) and detrending were applied to reduce unwanted spectral variation caused by light scattering and baseline shifts (Barnes et al., 1989).
Several mathematical treatments were evaluated to determine the most suitable preprocessing condition for moisture prediction. The treatments included 1,4,4; 1,8,8; 1,16,16; 2,4,4; 2,8,8; 2,16,16; 3,4,4; 3,8,8; 3,16,16; 4,4,4; 4,8,8; and 4,16,16. The final treatment for each production year was selected based on the overall calibration and cross-validation performance.
For the 2025 dataset, 17 samples were identified as outliers and removed from the original dataset. Therefore, 205 samples were used for final calibration. The selected model used SNV and detrending with a mathematical treatment of 2,16,16 and nine PLS factors.
For the 2026 dataset, 360 samples were used for calibration and cross-validation, and no samples were rejected from the final calibration. The selected model used SNV and detrending with a mathematical treatment of 1,8,8 and nine PLS factors.
All spectral preprocessing, PLS calibration, and model evaluation were performed using UCal software (version 4, Unity Scientific, USA).
5. Model validation and statistical evaluation
The performance of the calibration models was evaluated using the coefficient of determination (R2), standard error of calibration (SEC), coefficient of determination of cross-validation (R2CrV), standard error of cross-validation (SECrV), and ratio of performance to deviation (RPD). RPD was calculated as the standard deviation of the reference moisture values divided by SECrV.
For the 2025 dataset, the final model was developed using 205 samples and evaluated by cross-validation. In addition, nine seed samples that were not included in calibration were used for independent validation. The moisture contents of these samples were determined by the oven-drying method and predicted using the final NIRS calibration model.
For the 2026 dataset, five-group cross-validation was used during model development. Of the total 450 samples, 90 samples were kept separately and were not used during calibration, preprocessing selection, or model optimization. After the final calibration equation was selected using the remaining 360 samples, the same equation was applied to the 90 independent samples without further modification. Independent validation performance was evaluated using R2, standard error of prediction (SEP), bias-corrected standard error of prediction [SEP(C)], bias, and slope.
Ⅲ. RESULTS AND DISCUSSION
1. Descriptive statistics of seed moisture content
The descriptive statistics of Italian ryegrass seed moisture content for the 2025 and 2026 datasets are presented in Table 1. In 2025, the moisture content of the 222 seed samples ranged from 11.2% to 34.5%, with a mean of 23.3% and a standard deviation of 7.33%. In 2026, the 360 samples used for calibration covered a moisture range of 12.7% to 37.5%, with a mean of 19.158% and a standard deviation of 5.061%. These results show that both datasets contained a wide range of moisture values, although the distribution of moisture differed between the two production years.
Variation in reference values is important when developing NIRS calibration models. The calibration population should cover the range of values expected in unknown samples because a model developed from a narrow reference range may have limited ability to predict samples outside that range. (Williams and Norris, 1987) emphasized the importance of representative calibration samples for NIRS analysis. Similarly, (Shetty et al., 2012) showed that the selection of representative samples covering biological variation is important for developing reliable NIRS calibration models.
The 2025 dataset showed a higher mean moisture content and a larger standard deviation than the 2026 calibration dataset. Although the maximum moisture content was higher in 2026, the lower standard deviation indicates that most of the 2026 reference values were distributed more closely around the mean. Differences in the range and distribution of reference values can influence calibration statistics because model performance depends partly on the variation represented in the calibration population (Shetty et al., 2012; Williams and Norris, 1987).
The difference in moisture distribution between the two years may also reflect normal variation among seed populations collected in different production years. Seed moisture can be affected by conditions during seed development, harvesting, drying, and storage (Corbineau, 2024). However, these factors were not evaluated separately in the present study. Therefore, the specific reason for the difference in moisture distribution between 2025 and 2026 cannot be determined from the present data.
The inclusion of two production years provided a broader basis for evaluating NIRS performance than the use of a single-year dataset. (Shetty et al., 2012) reported that calibration populations containing samples from different years and biological sources can provide useful variation for developing NIRS models. Therefore, the differences observed between the 2025 and 2026 datasets are useful for evaluating the performance and future development of NIRS calibration models for Italian ryegrass seed moisture determination.
2. Spectral characteristics of italian ryegrass seeds
The NIR spectra of Italian ryegrass seed samples collected in 2025 are shown in Fig. 1. The spectra showed a similar general pattern across the measured wavelength range, although differences in absorbance intensity were observed among samples. These differences indicate variation in the physical and chemical characteristics of the seed samples. Because moisture content varied among the samples, part of the spectral variation was expected to be related to differences in seed moisture.
Clear absorption features were observed around 1450 and 1940 nm. These regions are strongly associated with water absorption in the near-infrared region. The absorption near 1450 nm is mainly related to the first overtone of O–H stretching, while the region around 1940 nm is associated with combination vibrations of O–H bonds in water molecules (Osborne et al., 1993; Workman Jr and Weyer, 2007). (Cen and He, 2007) also reported that these water-related absorption regions are important for moisture determination in biological materials.
Spectral variation was also observed outside the main water absorption regions. Seeds contain carbohydrates, proteins, lipids, and other organic compounds containing C–H, N–H, and O–H bonds, which can also contribute to the NIR spectrum (Siesler et al., 2008). Because these absorption features can overlap, PLS regression is useful for relating information from many wavelengths to the reference moisture values rather than depending on a single wavelength.
The spectral characteristics observed in the present study therefore provided useful information for moisture prediction. Previous studies have also shown that NIRS can be used for moisture determination in intact seeds and other agricultural materials because of the strong spectral response of water (Manley, 2014; Toscano et al., 2022). The application of SNV and detrending before calibration further helped reduce unwanted spectral variation caused by scattering and baseline differences (Barnes et al., 1989; Rinnan et al., 2009).
3. Calibration and cross-validation performance of the NIRS models
The calibration and cross-validation statistics for moisture prediction in 2025 and 2026 are presented in Tables 2. In 2025, 205 samples were used for final calibration after removing 17 outliers. The selected model used SNV and detrending with a mathematical treatment of 2,16,16 and nine PLS factors. The model showed high calibration performance, with R2 = 0.996 and SEC = 0.463%. The cross-validation results were also high, with R2CrV = 0.992, SECrV = 0.675%, and RPD = 10.865. The small difference between the calibration and cross-validation results indicates that the model performed consistently within the 2025 dataset. The relationship between the reference and NIRS-predicted moisture values for the 2025 dataset is shown in Fig. 2. The close distribution of the data points around the regression line was consistent with the high calibration performance of the model.
The 2026 model was developed using 360 calibration samples, and no samples were rejected from the final calibration. Among the different mathematical treatments tested, 1,8,8 with nine PLS factors was selected based on the overall cross-validation performance. The model showed R2 = 0.874 and SEC = 1.799%. During five-group cross-validation, R2CrV was 0.838, SECrV was 2.036%, and RPD was 2.49. These results show that the 2026 model maintained a clear relationship between the spectral information and reference moisture values, although the prediction accuracy was lower than that obtained in 2025. The NIR spectra of Italian ryegrass seed samples analyzed in 2026 are shown in Fig. 3. The overall spectral pattern was similar to that observed in 2025, although differences in absorbance intensity were observed among samples.
PLS regression is widely used for NIRS calibration because it can handle a large number of correlated spectral variables and identify the spectral information related to the reference values (Martens and Naes, 1989). Spectral preprocessing is also important because light scattering and baseline differences can affect NIR spectra. (Barnes et al., 1989) and (Rinnan et al., 2009) reported that SNV, detrending, and derivative treatments can reduce unwanted spectral variation and improve the information used for calibration. In the present study, the same SNV and detrending procedures were used in both years, while different mathematical treatments were selected for the two datasets.
Good prediction of seed moisture using NIRS has also been reported in previous studies. Fassio and Cozzolino (2004) developed NIRS models for intact sunflower seeds and reported good prediction of moisture and other seed components. Similarly, (Oblath et al., 2016) developed calibration models using intact Brassicaceae seeds and obtained R2 = 0.93 and RPD = 3.8 for moisture prediction. These studies support the use of NIRS for rapid moisture determination in intact seeds.
The difference in model performance between 2025 and 2026 is an important result of the present study. Although the 2026 calibration dataset contained more samples, its R2 and R2CrV values were lower and its prediction errors were higher than those obtained in 2025. The standard deviation of the reference moisture values was also lower in 2026 (5.061%) than in 2025 (7.33%). The range and distribution of reference values can influence calibration statistics, and a larger number of samples does not necessarily result in better prediction if the calibration population contains different spectral or reference variation (Shetty et al., 2012; Williams and Norris, 1987).
The different mathematical treatments selected in the two years also indicate differences between the annual datasets. The 2025 model selected 2,16,16, whereas the 2026 model selected 1,8,8. In the 2026 treatment screening, 1,8,8 produced R2CrV = 0.838 and SECrV = 2.036%, which provided the best overall cross-validation performance among the treatments evaluated. Differences in the optimal preprocessing treatment can occur when spectral populations differ because preprocessing changes the way scattering, baseline variation, and overlapping spectral information are handled (Rinnan et al., 2009).
The lower performance of the 2026 model should not be attributed to a single environmental or management factor because these factors were not evaluated separately in this study. However, the results show that model performance can change between production-year datasets. (Shetty et al., 2012) reported that representative calibration populations containing variation from different years and biological sources are important for developing reliable NIRS models. Therefore, the results from the two years suggest that including samples representing different production years may be important for developing a more stable calibration model for Italian ryegrass seed moisture determination.
4. Independent validation of the NIRS models
Independent validation provides an additional evaluation of model performance because the samples used for prediction are not included during calibration development (Williams and Norris, 1987; Workman Jr and Weyer, 2007). In 2025, independent validation was performed using nine additional seed samples. The NIRS-predicted values were generally close to the oven-determined moisture values, although the small number of validation samples limited the evaluation of model performance.
A larger independent validation dataset was used in 2026. The final calibration equation developed from 360 samples was applied to 90 independent samples that had not been used during calibration or model selection. The mean reference moisture content of the validation samples was 19.097%, while the mean NIRS-predicted moisture content was 19.147%. Independent validation resulted in R2 = 0.851, SEP = 2.081%, SEP(C) = 2.080%, bias = −0.050%, and a slope of 0.877. The relationship between the reference and NIRS-predicted moisture values for the 2026 independent validation set is shown in Fig. 4.
The R2 value of 0.851 indicates a clear relationship between the reference and predicted moisture values in the independent samples. The very small bias (−0.050%) also shows that the mean prediction was close to the mean reference value. However, the slope of 0.877 was lower than the ideal value of 1.0, and some variation between the reference and predicted values can be seen in the independent validation plot. Therefore, the 2026 model provided useful independent prediction of seed moisture, but some prediction error remained.
The independent validation results are important because they provide a more realistic evaluation of the model using samples that were not involved in model development. The use of 90 independent samples in 2026 provided a stronger validation dataset than the nine samples available in 2025. The results therefore support the potential use of NIRS for rapid moisture prediction in Italian ryegrass seeds while also showing that further expansion of the calibration population may improve prediction accuracy for future samples.
5. Effect of production year on NIRS model performance
The NIRS models showed different calibration and prediction performance between the two production years. The 2025 model showed higher R2 and R2CrV values and lower SEC and SECrV values than the 2026 model. Although more samples were used for model development and validation in 2026, the model performance was lower than that obtained in 2025. This result shows that the number of samples alone does not determine the performance of a NIRS calibration model.
The reference moisture distribution also differed between the two years. The 2025 samples had a mean moisture content of 23.3% and a standard deviation of 7.33%, whereas the 2026 calibration samples had a mean of 19.158% and a standard deviation of 5.061%. The lower variation in the 2026 reference data may partly explain the lower RPD obtained for this dataset because RPD is influenced by both the prediction error and the standard deviation of the reference values (Williams and Sobering, 1993). However, differences in reference distribution alone cannot fully explain the difference in model performance.
Variation between production years has also been reported in previous NIRS studies. (Shetty et al., 2012) evaluated grass samples collected from different years, species, and cultivars and showed that representative sample selection is important for developing reliable calibration models. In another study using grass samples collected over five years, (Shetty et al., 2012) found that year-to-year variation affected NIRS prediction and that adding samples from a new production year to the calibration database improved the ability of the model to predict future samples.
Differences between annual seed populations may result from several factors, including growing conditions, harvest stage, seed composition, drying, and storage conditions. These factors can affect both the reference moisture distribution and spectral characteristics of the samples. However, these factors were not evaluated separately in the present study. Therefore, the difference between the 2025 and 2026 models cannot be attributed to a specific environmental or management factor.
The mathematical treatment selected for the two years was also different. The 2025 model selected 2,16,16 with nine PLS factors, whereas the 2026 model selected 1,8,8 with nine PLS factors. Different preprocessing treatments can be selected when the spectral characteristics of calibration populations differ because derivative and smoothing treatments influence the removal of baseline variation and the extraction of useful spectral information (Barnes et al., 1989; Rinnan et al., 2009).
The difference between the two production years provides useful information for future model development. The high performance obtained in 2025 shows that NIRS can accurately predict Italian ryegrass seed moisture within a well-represented calibration population. The 2026 results, including independent validation with 90 samples, show that useful moisture prediction can also be obtained with a different annual dataset, although prediction accuracy was lower. These findings support the use of samples from different production years when developing a calibration model intended for routine application. A calibration population containing representative variation from several years may improve the ability of the model to predict future Italian ryegrass seed samples (Shetty et al., 2012).
6. Practical implications for seed moisture evaluation
The results of this study show that NIRS has potential as a rapid and non-destructive method for determining moisture content in Italian ryegrass seeds. Compared with oven drying, NIRS requires less analysis time and does not destroy the analyzed sample. These characteristics are useful when a large number of seed samples need to be evaluated during processing, drying, and storage (Manley, 2014).
Accurate moisture determination is important for maintaining seed quality because seed moisture affects storage stability, microbial activity, and seed longevity (Corbineau, 2024). Rapid moisture measurement may therefore help in monitoring seed conditions during postharvest handling and storage. Previous studies have also shown that NIRS can be used for rapid and non-destructive evaluation of moisture and other quality characteristics in agricultural and seed materials (Fassio and Cozzolino, 2004; Oblath et al., 2016).
The two-year results also show that calibration development should be considered a continuing process. A model developed from samples collected in one production year may not fully represent the variation found in samples collected in another year. Previous work with multi-year grass datasets similarly found that adding samples from new years to the calibration database can improve prediction for future samples. Therefore, adding representative samples from future production years may improve the reliability and practical application of the NIRS model developed for Italian ryegrass seed moisture.
Overall, the present results support the use of NIRS as a rapid screening method for Italian ryegrass seed moisture. However, further expansion of the calibration database using samples from additional production years, cultivars, and production conditions would provide a stronger basis for routine application of the model.
Ⅵ. CONCLUSION
This study evaluated NIRS for the rapid and non-destructive determination of moisture content in Italian ryegrass seeds using samples collected over two production years. The 2025 model showed high calibration and cross-validation performance, while the 2026 model showed lower but useful prediction performance. More importantly, the 2026 model was evaluated using 90 independent samples, resulting in R2 = 0.851, SEP = 2.081%, bias = −0.050%, and a slope of 0.877. The small bias and close mean values between the reference and NIRS predictions showed that the model provided reasonable moisture prediction for samples that were not included during calibration.
The difference in model performance between 2025 and 2026 indicates that the characteristics of the calibration population can influence NIRS prediction. Therefore, a model developed using samples from a single production year may not fully represent the variation found in future seed samples. Including samples from different production years may improve the reliability and wider application of the calibration model.
Overall, the results show that NIRS has good potential as a rapid method for determining moisture content in Italian ryegrass seeds. Further expansion of the calibration database using representative samples from additional production years, followed by independent validation, would help develop a more reliable model for routine seed moisture evaluation.
