Promoting Alfalfa (Medicago sativa L.) Silage Fermentation using Carbohydrate Additives and Optimizing NIRS-Based Rapid Evaluation Strategies via Machine Learning Approaches = 탄수화물 첨가에 따른 알팔파 사일리지 발효 특성 향상 및 기계학습 기반 NIRS 신속 평가 전략 최적화

다국어 초록 (Multilingual Abstract)

The production of high-quality alfalfa silage is often constrained by its low concentration of fermentable sugars and high buffering capacity, which frequently necessitates the application of pre-ensiling treatments or additives to ensure satisfactory fermentation. Accordingly, this study was conducted with two primary objectives. Experiment 1 aimed to evaluate the effects of sugar additives on the fermentation characteristics of alfalfa silage prepared at different dry matter (DM) levels. Experiment 2 focused on addressing the limitations of near infrared spectroscopy (NIRS) when applied to untreated samples by systematically assessing multiple machine learning approaches to improve model performance for quality and gas traits of wet, unground alfalfa silage, thereby facilitating the development of a rapid and non-destructive evaluation system.
In Experiment 1, fourth-cut alfalfa was wilted to three dry matter (DM) levels (20.83 %, 42.97 %, and 52.53 %) and ensiled with glucose, sucrose, starch (each at 2 % fresh matter), Lactiplantibacillus plantarum+cellulase (LP+C, 1×106 CFU/g and 100 mg/kg fresh matter), or distilled water as a control. Approximately 400 g of forage mass was packed into polyethylene bags and fermented for 45 days. Experiment 2 utilized 348 alfalfa silage samples to develop NIRS models for fifteen quality traits. The modeling performance of Partial Least Squares Regression (PLSR), Principal Component Regression (PCR), Ridge Regression (RR), and Linear Support Vector Regression (L-SVR) were first evaluated based on a ten-fold cross-validation. Then, three spectral preprocessing methods—Savitzky-Golay Smoothing and Standard Normal Variate (SG+SNV), SG and Multiplicative Scatter Correction (SG+MSC), and SNV with Detrending (SNV+D) were employed to facilitate the modeling process. Besides, Least Absolute Shrinkage and Selection Operator (LASSO) and Binary Particle Swarm Optimization (BPSO) techniques were applied for selecting feature wavelengths to further optimize model performance.
The results of Experiment 1 demonstrated that additive treatments at the LDM and MDM levels significantly reduced silage pH and ammonia-N concentrations while increasing lactic acid (LA) production and DM retention (p < 0.001), indicating improved fermentation quality. These effects were most pronounced in the glucose- and sucrose-treated silages, which also retained higher residual water-soluble carbohydrate (WSC) contents. In contrast, the fermentation-promoting effects of additives were not evident at the HDM level. This was likely due to the substantial improvement in silage quality achieved through wilting alone (p < 0.001), as well as the reduced moisture content limiting additive utilization by lactic acid bacteria. With respect to in vitro gas production, all sugar additives increased total gas and methane production to varying extents (p < 0.001), particularly sucrose and starch, while wilting itself also contributed to higher gas and methane yields (p < 0.05). Although the LP+C treatment exerted minimal effects on fermentation quality, it consistently resulted in the lowest methane production, comparable to the control. Based on the TOPSIS analysis, the glucose and sucrose treatments at the HDM level, along with the glucose treatment at the LDM level, achieved the highest comprehensive scores for overall silage quality.
In Experiment 2, PLSR-based models generally exhibited superior predictive performance, whereas RR achieved the lowest prediction errors. Spectral preprocessing primarily contributed to error reduction rather than improvements in R²CV, with SG+SNV showing the most consistent performance. Furthermore, BPSO outperformed LASSO in wavelength selection, particularly when combined with PLSR and PCR. Overall, the integration of multiple machine learning algorithms substantially enhanced the predictive capability of NIRS models for wet, unground alfalfa silage across chemical composition, fermentation characteristics, and gas production traits. The acceptable predictions were obtained for pH, LA, acetic acid (AA), DM, WSC, neutral detergent fiber (NDF), and relative feed value (RFV).
In conclusion, based on TOPSIS assessment and practical considerations, wilting alfalfa to about 50 % DM before ensiling, with or without sucrose supplementation (2 % FM), is recommended under favorable weather conditions, whereas glucose supplementation (2 % FM) is suggested when wilting is not feasible. Meanwhile, for NIRS-based evaluation of wet, unground alfalfa silage, the PLSR–SG+SNV–BPSO modeling framework demonstrated the best overall predictive performance and is therefore recommended for future rapid assessment of wet forage samples.

국문 초록 (Abstract)

본 연구는 알팔파 사일리지 발효에 대한 탄수화물 첨가제의 영향을 평가하고(실험 1), 머신러닝 기법을 활용하여 알팔파 사일리지 품질의 신속 평가를 위한 근적외선 분광법(NIRS) 예측 모델을 최적화하고자 수행되었다(실험 2). 실험 1에서는 4차 수확 알팔파를 건물 함량(DM) 20.83 %, 42.97 %, 및 52.53 %의 세 수준으로 예건한 후, 포도당, 자당, 전분(각각 생물질 기준 2 %), Lactobacillus plantarum과 셀룰라아제 혼합제(LP+C), 또는 증류수를 처리하였다. 저 및 중간 DM 수준에서 탄수화물 첨가제 처리는 pH와 암모니아태 질소 농도를 유의적으로 감소시키고 젖산 생성량과 DM 보존율을 증가시켰으며(p<0.001), 특히 포도당과 자당 처리구에서 그 효과가 두드러졌다. 반면, 고 DM 수준에서는 예건 처리만으로도 발효 품질이 현저히 개선되어 탄수화물 첨가 효과는 제한적이었다. 탄수화물 첨가제는 메탄 및 총 가스 생성량을 증가시킨 반면, LP+C 처리구는 모든 처리 중 가장 낮은 메탄 생성량을 나타냈다. TOPSIS 분석 결과, 예건과 병행한 포도당 및 자당 처리가 전반적인 사일리지 품질 개선에 있어 최적의 전략으로 평가되었다. 실험 2에서는 총 348점의 시료를 대상으로 PLSR, PCR, RR 및 선형 서포트 벡터 회귀(L-SVR) 모델을 다양한 스펙트럼 전처리 및 파장 선택 기법과 결합하여 구축하였다. 그 결과, PLSR–SG+SNV–BPSO 조합이 가장 우수한 예측 성능을 보였으며, 특히 pH, 젖산, DM, 수용성 탄수화물(WSC), 중성세제섬유(NDF) 및 상대사료가치(RFV)에 대해 높은 예측 정확도를 나타냈다.
종합적으로, 알팔파를 약 50 % DM 수준까지 예건한 후 자당을 처리하거나 단독 예건하는 전략은 사일리지 품질 개선에 효과적인 것으로 판단되며, 최적화된 NIRS 모델은 신선 알팔파 사일리지의 신속하고 비파괴적인 품질 평가를 가능하게 할 것으로 기대된다.

목차 (Table of Contents)

• Abstract I
• Contents IV
• List of Tables VII
• List of Figures VIII
• List of Abbreviations IX
• 1. Introduction 1
• 1.1 Research background 1
• 1.2 Research purpose 3
• 2. Literature review 4
• 2.1 Alfalfa 4
• 2.1.1 Definition 4
• 2.1.2 Cultivation scope and utilization 4
• 2.2 Alfalfa silage 5
• 2.2.1 Principle 5
• 2.2.2 Feeding value 7
• 2.2.3 Challenges in alfalfa silage production 7
• 2.3 Wilting pretreatment for alfalfa silage 8
• 2.4 Carbohydrates supplement for alfalfa silage 10
• 2.4.1 Application of simple sugars 10
• 2.4.2 Application of sugar-rich materials 11
• 2.5 Greenhouse gas emission from livestock industry 12
• 2.5.1 Basic profile 12
• 2.5.2 CH4 emission from animals 13
• 2.5.3 Solutions to mitigate CH4 emission from ruminants 15
• 2.6 Near infrared spectroscopy (NIRS) 16
• 2.6.1 Definition 17
• 2.6.2 Application of NIRS in the feed industry 18
• 2.6.3 General procedure for establishing an NIRS model 19
• 2.6.4 Factors affecting NIRS model performance 20
• 3. Materials and methods 22
• 3.1 Experiment 1 22
• 3.1.1 Silage preparation 22
• 3.1.2 Chemical composition analysis 23
• 3.1.3 Fermentation condition analysis 24
• 3.1.4 Microbiological analysis 25
• 3.1.5 In vitro gas production determination 25
• 3.1.6 Statistical analysis 27
• 3.2 Experiment 2 28
• 3.2.1 NIR spectral acquisition 28
• 3.2.2 Chemometric analysis: Regression algorithms 29
• 3.2.3 Chemometric analysis: Spectral preprocessing 31
• 3.2.4 Chemometric analysis: Wavelength selection 31
• 4. Results and discussion 33
• 4.1 Experiment 1 33
• 4.1.1 Effects of wilting treatment on the chemical composition and microbial population of alfalfa 33
• 4.1.2 Effects of wilting treatment and additives on the nutrients and fermentation quality of alfalfa silage 37
• 4.1.3 Effects of DM level and sugar additives on rumen gas production from alfalfa silage 48
• 4.1.4 TOPSIS analysis 53
• 4.2 Experiment 2 56
• 4.2.1 Spectra investigation 56
• 4.2.2 Comparison of model performance based on regression algorithms 57
• 4.2.3 Effect of spectral preprocessing on model performance 62
• 4.2.4 Effect of wavelength selection on model performance 64
• 4.2.5 Final optimal models 68
• 5. Conclusion 75
• Reference 77
• Abstract in Korean 92
• Acknowledgement 94
• Supplementary Materials 97