@prefix vivo: . @prefix edm: . @prefix ns0: . @prefix dcterms: . @prefix dc: . @prefix skos: . vivo:departmentOrSchool "Land and Food Systems, Faculty of"@en ; edm:dataProvider "DSpace"@en ; ns0:degreeCampus "UBCV"@en ; dcterms:creator "Kennedy, Carol Ann"@en ; dcterms:issued "2009-02-06T22:55:41Z"@en, "1996"@en ; vivo:relatedDegree "Master of Science - MSc"@en ; ns0:degreeGrantor "University of British Columbia"@en ; dcterms:description """Diverse fresh grass and grass silage samples (n = 292) were collected from 138 dairy farms in southwestern British Columbia, Canada. Sources of variation within the forage sample population included different crop years, harvests, grass species, ensiling additives and storage facilities. These samples were used to determine the accuracy of near infrared reflectance spectroscopy (MRS) for predicting the feeding value of undried, unground fresh grass and grass silage. NIRS spectra were collected from intact samples with a NIRSystems 6500 scanning monochromator instrument followed by analyses for dry matter (DM), crude protein (CP), acid detergent fibre (ADF), and fermentation end-products (lactic, acetic, propionic, isobutyric, butyric, isovaleric, and valeric acids) by conventional chemical laboratory procedures. The means and standard deviations of the sample population were 34.7 and 13.0% for DM, 17.1 and 3.5% for CP, and 34.3 and 3.9% for ADF when corrected to a moisture-free basis. The calibration set (n =216) was selected by spectral variation using the "neighbourhood Ff" method and the spectral duplicates were assigned to the validation set (n = 67). Prediction equations were developed with modified partial least square regression and cross validation utilizing different scatter treatments, derivatives and wavelength segments. The calibration R2 and standard errors of cross validation (SECV) for DM, CP, and ADF corrected to a moisture-free basis were 1.00 and 1.15%, 0.95 and 1.05%, and 0.95 and 1.16%, respectively. Standard errors of performance (SEP), means, and coefficients of variability (SEP* 100 ⌯ mean) for the validation set were 0.73, 27.2 and 2.69% for DM, 0.79, 15.7 and 5.03% for CP, and 0.95, 34.5 and 2.75% for ADF. The errors associated with the equations developed for the short chain organic acids were unacceptably high. Prediction equations were also developed on reference values calculated on an "as received" basis. Different procedures for calibration and validation set selection were compared with no one common method producing the lowest error on all constituents. It was concluded that the NIRS prediction equation for DM produced excellent accuracy as indicated by the low SECV and SEP. The prediction equations for CP and ADF had acceptable accuracy for monitoring forage nutrient quality for livestock ration balancing programs. The NIRS method of analysis will provide forage quality information faster and at a reduced cost compared to conventional chemical procedures."""@en ; edm:aggregatedCHO "https://circle.library.ubc.ca/rest/handle/2429/4266?expand=metadata"@en ; dcterms:extent "4345408 bytes"@en ; dc:format "application/pdf"@en ; skos:note "NIRS ANALYSIS OF INTACT GRASS SILAGE AND FRESH GRASS FOR THE PREDICTION OF DRY MATTER, CRUDE PROTEIN AND ACTD DETERGENT FIBRE by CAROL ANN KENNEDY B.Sc.(Agr.), The University of British Columbia, 1974 A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF MASTER OF SCIENCE in THE FACULTY OF GRADUATE STUDIES Department of Animal Science We accept this thesis as conforming to^he requijed^ standard THE UNIVERSITY OF BRITISH COLUMBIA March 1996 © Carol Ann Kennedy, 1996 In presenting this thesis in partial fulfilment of the requirements for an advanced degree at the University of British Columbia, I agree that the Library shall make it freely available for reference and study. I further agree that permission for extensive copying of this thesis for scholarly purposes may be granted by the head of my department or by his or her representatives. It is understood that copying or publication of this thesis for financial gain shall not be allowed without my written permission. Department of The University of British Columbia Vancouver, Canada Date DE-6 (2/88) Abstract Diverse fresh grass and grass silage samples (n = 292) were collected from 138 dairy farms in southwestern British Columbia, Canada. Sources of variation within the forage sample population included different crop years, harvests, grass species, ensiling additives and storage facilities. These samples were used to determine the accuracy of near infrared reflectance spectroscopy (MRS) for predicting the feeding value of undried, unground fresh grass and grass silage. NIRS spectra were collected from intact samples with a NIRSystems 6500 scanning monochromator instrument followed by analyses for dry matter (DM), crude protein (CP), acid detergent fibre (ADF), and fermentation end-products (lactic, acetic, propionic, isobutyric, butyric, isovaleric, and valeric acids) by conventional chemical laboratory procedures. The means and standard deviations of the sample population were 34.7 and 13.0% for D M , 17.1 and 3.5% for CP, and 34.3 and 3.9% for ADF when corrected to a moisture-free basis. The calibration set («=216) was selected by spectral variation using the \"neighbourhood Ff\" method and the spectral duplicates were assigned to the validation set (n-67). Prediction equations were developed with modified partial least square regression and cross validation utilizing different scatter treatments, derivatives and wavelength segments. The calibration R 2 and standard errors of cross validation (SECV) for D M , CP, and ADF corrected to a moisture-free basis were 1.00 and 1.15%, 0.95 and 1.05%, and 0.95 and 1.16%, respectively. Standard errors of performance (SEP), means, and coefficients of variability (SEP* 100 + mean) for the validation set were 0.73, 27.2 and 2.69% for D M , 0.79, 15.7 and 5.03% for CP, and 0.95, 34.5 and 2.75% for ADF. The errors associated with the equations developed for the short chain organic acids ii were unacceptably high. Prediction equations were also developed on reference values calculated on an \"as received\" basis. Different procedures for calibration and validation set selection were compared with no one common method producing the lowest error on all constituents. It was concluded that the NIRS prediction equation for DM produced excellent accuracy as indicated by the low SECV and SEP. The prediction equations for CP and ADF had acceptable accuracy for monitoring forage nutrient quality for livestock ration balancing programs. The NIRS method of analysis will provide forage quality information faster and at a reduced cost compared to conventional chemical procedures. iii Table of Contents Page Abstract ii Table of Contents iv List of Tables vi List of Figures viii Acknowledgement ix 1. General Introduction 1 2. A Discussion of Factors Affecting and Parameters Used to Estimate the Quality of Grass Silage • 5 2.1 Literature Review 5 2.2 The Ensiling Process 5 2.2.1 Factors Affecting Silage Fermentation 8 2.2.1.1 Water-Soluble Carbohydrate Content 8 2.2.1.2 Buffering Capacity 9 2.2.1.3 Nitrogenous Substances 10 2.2.1.4 Dry Matter Content 11 2.2.1.5 Bacterial Content 13 2.2.1.6 The Importance of Oxygen 14 2.2.1.7 Temperature 15 2.2.1.8 Mechanical Pretreatment 16 2.2.1.9 Silage Additives 16 2.2.2 Summary of Dry Matter Losses 22 iv 2.3 The Evaluation of Silage Quality 22 2.3.1 Errors Associated with the Chemical Analysis of Forages 26 2.4 Near Infrared Reflectance Spectroscopy (NIRS) and Its Application to Forage Analysis 27 2.4.1 Brief History of NIRS 27 2.4.2 Principles of NIRS 29 2.4.3 Sampling and Sample Presentation 31 2.4.4 NIRS Reference Methods 33 2.4.5 NIRS Statistical Methods 37 2.4.5.1 Structuring of the Sample Population 37 2.4.5.2 Calibration 40 2.4.5.3 Regression Methods 42 2.4.5.4 Accuracy of NIRS 44 2.4.6 New Applications in Forage Analysis Using NIRS 48 2.5 Summary 48 3. NIRS Analysis of Intact Grass Silage and Fresh Grass for the Prediction of Dry Matter, Crude Protein, and Acid Detergent Fibre 51 3.1 Introduction and Objective 51 3.2 Materials and Methods 52 3.3 Results and Discussion 60 3.4 Conclusions 79 4. Summary 81 Bibliography 83 v List of Tables Table Description Page 2.1 Silo Type and Recommended Dry Matter Contents 12 2.2 Effect of Lactic Acid Bacteria (LAB) and Cell-Wall Degrading Enzymes on the Composition of Grass-Legume Silage 20 2.3 Expected Dry Matter Losses During Forage Harvest, Storage, and Feeding 22 2.4 Fermentation Products (Dry Matter Basis) and Silage Quality 24 2.5 Nitrogenous Components of Three Different Silages 25 2.6 Examples of Coefficients of Variation of Chemical and NIR Analyses on Ground Forages Between Laboratories 27 2.7 Calibration Statistics for Dried Ground (DG) and Undried, Unground (UD/UG) Haylage and Pasture Grass 32 2.8 Precision of 64 Analyses of a Ground Sealed Soybean Meal Standard Over a 75 Day Period 47 3.1 Number of Fresh Grass and Grass Silage Samples by Crop Year and Cut Number 54 3.2 Grass Species Mix and Combinations as Seeded in the Sample Population 54 3.3 Ensiling Additives Utilized in the Grass Silage Population 55 3.4 Grass Silage Storage Facilities 55 3.5 Fresh Grass and Grass Silage Samples Reference Method Statistics 61 3.6 Fresh Grass and Grass Silage Reference Method Statistics for Dry Matter, Crude Protein (DM), and ADF (DM) on the Calibration and Validation Sets 62 3.7 Calibration Equation Statistics Before and After Outlier Removal 64 3.8 Calibration and Validation Statistics for the Best Performing Fresh Grass and Grass Silage Equations Developed on Calibration and Validation Sets Selected by the \"Neighbourhood FT' Method 65 vi 3.9 Performance Statistics for Prediction of Lactic, Acetic, and Propionic Acid in Fresh Grass and Grass Silage 70 3.10 Best \"As Received\" Equation and Validation Set Statistics for Crude Protein and ADF in Intact Fresh Grass and Grass Silage 71 3.11 Ratio of the Standard Error of Performance to the Standard Deviation (RPD) and the Coefficient of Variability (CV) 72 3.12 Comparison of Performance Statistics for DM, CP (DM), and ADF (DM) Predictions in Dry, Ground Forage Versus Intact Forage 74 3.13 Comparison of the Accuracy of NIRS in Analyzing Undried, Unground Fresh Grass and Grass Silage on the Three Calibration and Validation Set Selection Methods 77 vii List of Figures Figure Description Page 2.1 Major Phases Before, During, and After Ensiling when Various Plant, Microbial and Chemical Processes are Most Active 6 3.2 Fresh Grass and Grass Silage Population when Viewed in S Y M M E T R Y 60 3.3 Linear Regression Plot of the Reference Method and Predicted Dry Matter of Fresh Grass and Grass Silage 66 3.4 Linear Regression Plot of the Reference Method and Predicted Crude Protein Corrected to a Dry Matter Basis CP (DM) of Fresh Grass and Grass Silage 68 3.5 Linear Regression Plot of the Reference Method and Predicted Acid Detergent Fibre Corrected to a Dry Matter Basis ADF (DM) of Fresh Grass and Grass Silage 69 viii Acknowledgements I would like to thank my graduate supervisor Dr. J. A. Shelford for his advice and direction and Dr. L. J. Fisher of Agriculture and Agri-Food Canada, Agassiz, and Dr. P. C. Williams of the Canadian Grain Commission in Winnipeg for their invaluable technical and professional guidance in the creation of this thesis. Their council and suggestions upon reading this thesis were immensely helpful and appreciated. I am also grateful to Dr. M. Tait and Dr. A. Bomke at UBC for reading and commenting on this thesis. Mary Lou Swift and Pro Form Feeds Inc. were also instrumental in making this project possible. Without their encouragement and support, this would not have been possible. The assistance of the lab staff at Pro Form Feeds Inc. is gratefully acknowledged for their laboratory analyses and understanding. I am also indebted to the Science Council of British Columbia and the B.C. Ministry of Agriculture, Fisheries and Food for providing me with financial support for my studies at UBC. The staff at U B C who helped arrange and analyse the samples for this project are also thanked for their time and efforts. I would also like to thank my parents, Denis and June Murray, for their generous offer of accommodations and support on those many days and nights I stayed in their home for the duration of classes. Finally, I would like to thank my husband, Patrick Kennedy, for his patience and caring throughout the duration of the research project and during the times I was seldom at home. Thank you to everyone. ix Chapter 1 General Introduction For many dairy farmers, preserving and storing a nutritionally suitable winter feed supply in adequate amounts is an essential part of livestock production. In the Lower Fraser Valley of the South Coastal Region of British Columbia, approximately 13,800 hectares of land are managed for silage and hay production for the dairy industry. Of this area, only 2,080 hectares or 15% is used for grazing. From the rest of this area, roughly 552,000 tonnes of grass and 220,000 tonnes of corn silage on a fresh weight basis are harvested annually. Of the grass portion, 4 to 6 cuttings are harvested per season with around 20% harvested as green chop or hay and 80% stored as grass silage. The grasses grown in this area often include mixtures of species that incorporate orchardgrass (Dactylis glomerata), perennial ryegrass (Lolium perenne), tall fescue (Festuca arundinacea) and/or white or red clover of different varieties and ratios. The major varieties of orchardgrass include Benchmark and Pro-file, perennial ryegrasses include Barlano (a diploid variety) and Fantoom (a tetraploid variety), tall fescues include Barcel and Fuego, and the clovers include California Ladino and Pacific DC. Approximately 50% of the feed offered to dairy cows is locally grown forages in the form of grass hay, corn silage and grass silage and of these, as much as 75% may be grass. The average dry matter (DM) intake of a lactating cow is 20 to 25 kg per cow per day including both grain and forage with approximately half of this amount as forage. The nutritive value of a forage crop is mainly affected by the stage of maturity rather than by fertilization, species, and climatic conditions. To make silage, forage is normally wilted in the 1 field to a D M content between 30 to 50% and in the case of large round bale silage, to as high as 75% D M which normally requires between 1 to 3 days depending on weather conditions. In the Lower Fraser Valley, at least 50% of all grass silage is ensiled in bunker silos at less than 35% D M . As the moisture content decreases, plant respiration decreases. Plant respiration results in the loss of soluble carbohydrates due to the complete oxidation of hexose sugar to C 0 2 and water with the loss of energy as heat. Rewetting of the crop by dew or rain reactivates enzyme activity and prolongs respiration resulting in increased D M losses. Ensiling grass offers the opportunity to preserve high quality feed with a minimum of harvesting losses when timely harvesting is practiced. Any grass crop that has adequate water-soluble carbohydrates (WSC) and adequate moisture content possesses the potential of being ensiled for livestock feed. Grass silage can be made in a variety of facilities including oxygen limiting silos, tower silos, bunker silos, pit silos, and bags. The ensiling process is not simple and the amount of storage loss that occurs varies widely with storage facility, species of grass, maturity, cut, moisture content, wilting conditions, weather, length of chop, degree of packing, ratio of protein to energy, additives, WSC, buffering capacity, soil and/or manure contamination, and the presence of fungi. Average D M losses in silage production are 14 to 24% of total D M with about half of this loss occurring during storage. These losses typically cause an increase in neutral detergent fibre (NDF) concentration of 3 to 12% (31). The ensiling process, itself, can greatly affect the quality of the feed. If ensiled properly, the ensiling process maintains the crop D M and allows the storage of the maximum nutrients in the crop with minimal loss. Forages, together with manufactured supplementary feeds, make up the ration for dairy cows. Feed manufacturers are involved in formulating grain supplements for dairy farms. It is 2 desirable that forage nutrient content be taken into consideration in order to arrive at the most cost effective ration formulation. Currently, laboratory wet chemistry methods are employed in the determination of forage quality, but the methods used are expensive, labour intensive, time-consuming, and the chemicals used are not environmentally friendly. They also, do not always accurately predict the feeding value or feed intake (13). This is a major limitation in formulating proper feeding programs for animals. A frequently measured parameter of silage is pH but, unfortunately, it gives no indication of the route taken during the fermentation. High levels of butyric acid and ammonia-nitrogen indicate an undesirable fermentation and it is generally agreed that less than 0.5% butyric acid and an ammonia-nitrogen level of less than 10% of the total nitrogen is desired (13). As knowledge of ruminant nutrition increases, so does interest in improving the quality of forages, through researching the effects of forage additives, and investigating different methodologies that can provide forage quality information. Determining the nutritive value of silage quickly, accurately and inexpensively is a desirable goal. Near infrared reflectance spectroscopy (NIRS) is a relatively new analytical application technique recently used more frequently for forage analysis and prediction of nutrient value for livestock. In its 24 year history, its usefulness has been demonstrated in many industries including agriculture, food, pharmaceutical, textile, and the petroleum industries (36). NIRS is a low-cost method of analysis that uses light to predict the nutrient composition of forages and it can satisfy the increasing need for speed in the analysis of livestock feed. It also has the advantage of being environmentally friendly and does not require chemical alteration of the samples. NIRS has the potential to predict the concentration of nutrient constituents that have been measured by a reference method, such as Kjeldahl nitrogen for crude protein. These analytical 3 values are correlated to the absorption spectrum of a group of samples. However, since NIRS calibration is dependent on chemical analyses, it would be no more accurate than the reference methods used in the calibration development. Currently, NIRS calibration equations have been developed and used extensively in the U.S. for the prediction of nutrient content of dried, ground forages but few studies have been conducted with undried, unground forages. Being able to analyse fresh forage reduces the time interval for analysis and eliminates the need for drying and grinding the sample prior to conventional wet chemistry or NIRS analysis. With greater understanding of factors influencing feed intake by animals and the NIRS prediction of forage quality, NIRS can become a more useful tool in the animal feed industry (36). The objective of this study was to determine the accuracy of NIRS calibration equations developed from undried, unground grass silage and fresh grass for the prediction of DM, CP, and acid detergent fibre (ADF). These calibration equations were developed on fresh grass and grass silage produced from grass grown in southwestern British Columbia. They will be used as a starting set or base to build future equations in future growing years. In the following sections, the ensiling process and its effect on grass and NIRS as an analytical technique will be reviewed. 4 Chapter 2 A Discussion of Factors Affecting and Parameters Used to Estimate the Quality of Grass Silage 2.1 Literature Review 2.2 The Ensiling Process Silage is defined as the feedstuff resulting from the preservation of green forage crops by anaerobic fermentation. Ensiling involves the conversion of WSC to lactic acid, which reduces the pH to a level sufficient to inhibit any further biological activity in the ensiled material when it is maintained under anaerobic conditions (7). There are two main steps in the ensiling process: the aerobic stage and the anaerobic stage. Since the goal of ensiling is to preserve the material with minimum nutrient loss, it is important to limit the nutrient consuming aerobic stage. The major phases of ensiling and when each process is most active are summarized in Figure 2.1 (20,31). The first phase or pre-seal takes place when oxygen is present in the forage up to the point of packing in the silo. The oxygen is consumed through the process of respiration by plant enzymes and microorganisms utilizing the plants' WSC (sugars) resulting in the production of water, C 0 2 and heat. The heat generated during respiration can be sufficient to damage protein in poorly compacted silage when the length of time of respiration is extended (13). Good quality silage tends to have a short respiration period and is achieved by chopping the silage to a short length, packing it thoroughly in the silo or bags, and sealing the storage facility effectively and quickly to exclude air (31). 5 Pre-Seal Active Fermentation Stable Phase Feed Out Plant Respiration Proteolysis Enzyme Hydrolysis of Carbohydrates Forage Cell Lysis Yeasts Molds Acetic Acid Bacteria Bacilli Lactic Acid Bacteria Clostridia Maillard Reactions Acid Hydrolysis of Hemicellulose > - > - - > . — > v > . v V v — > - > . — > . - > — > - v — V Figure 2.1. Major Phases Before, During, and After Ensiling when Various Plant, Microbial and Chemical Processes are Most Activea a From (20,31) The second stage is the anaerobic stage which includes the active fermentation and stable phase as indicated in Figure 2.1. It begins after the available oxygen is depleted through plant respiration and aerobic bacteria cease to function. Anaerobic bacteria then begin to multiply rapidly and the fermentation process begins. Ideally the lactobacilli species will predominate and will produce lactic acid and acetic acid from the plant material. The acids produced will lower the pH of the silage. Fermentation completely ceases after 3 to 4 weeks when the pH becomes so low (pH < 4.0) that all microbial growth is inhibited (31). If the ensiling conditions do not encourage the growth of lactic acid producing bacteria, then clostridial type microorganisms, listeria, coliform and yeasts will grow. If these organisms become established during ensiling, the resulting silage can cause a drop in D M intake in animals, increase the possibility of reproductive problems, and cause significant losses in animal productivity 6 such as milk production (13). Clostridial type microorganisms utilize lactic acid and WSC as energy sources and degrade protein to amines and ammonia. Ammonia can also be produced through the reduction of nitrate and nitrite by Clostridia. In addition, clostridial microorganisms produce butyric acid, C0 2 , and results in an increase in pH. The quality of the resulting silage is greatly reduced with this type of fermentation and results in a loss of nitrogen and a silage of low palatability and nutrient content (31). In addition to lactobacilli and clostridial microorganisms, silages also contain yeasts, molds, and coliform and propionic acid producing bacteria. Many fermentation products are formed in addition to lactic, acetic, and butyric acids and many do not contribute to preservation. At time of feeding, a second aerobic stage begins when the silage is exposed to air such as at the exposed face of a storage facility and/or upon the opening of an ensiling structure. As indicated in Figure 2.1, aerobic bacteria begin to grow and this causes the temperatures to rise. Molds and yeasts often appear and the silage starts to undergo aerobic deterioration. Aerobic deterioration involves the conversion of fermentation acids, proteins, and WSC to C0 2 , water and heat. The pre-harvest management and ensiling technique seem to be the major influence on the rate of deterioration (20,44). This results in greater DM loss, nutrient degradation, the formation of toxic products and, because fermentation acids are destroyed, an increase in pH. Silages in which fermentation was limited by additives such as minerals acids, with no specific antimicrobial properties and low DM silages with a high buffering capacity such as legume/grass mixtures, are most susceptible to aerobic deterioration as they are most likely to become contaminated by molds and yeasts. Silages most resistant to aerobic deterioration are those ensiled quickly, packed adequately and exposed to oxygen for a minimum length of time during ensiling (44). 7 2.2.1 Factors Affecting Silage Fermentation The main factors affecting silage fermentation are interrelated and can be affected by every phase of the silage making process. The factors include WSC content, buffering capacity, amount of nitrogenous substances, DM content, bacterial count, oxygen content, and mechanical pretreatment. 2.2.1.1 Water-Soluble Carbohydrate Content Aerobic microorganisms utilize WSC as their main source of energy for growth. The main WSC or sugars present in plants include glucose, fructose, sucrose and fructosans (44). The principal storage polysaccharide in perennial temperate grasses is fructan, while the main reserve polysaccharide in legumes is starch. Since starch is insoluble in cold water, it is not included in the WSC fraction. There is only limited fermentation of plant carbohydrates such as starch, cellulose and hemicellulose since they are not available as a fermentable substrate to most lactic acid bacteria. Fermentation can be limited by a low WSC content, as in very young grass, which can result in a pH too high to achieve safe preservation. Normally a minimum of 6 to 12% WSC on a DM basis is required for proper silage formation. Grasses have been reported to contain 6 to 23% WSC with orchard grass varieties containing WSC at the lower end of the range. Water-soluble carbohydrate content in grasses is influenced by the species of grass, stage of maturity, growing conditions, management conditions (such as wilting), moisture availability for growth, time of day, and fertilization. Thus, the timing of forage harvesting for ensiling is often a compromise between increasing yield and decreasing quality. For example, in the Lower Fraser Valley of B.C., orchardgrass/ryegrass/clover mixtures are common and if cutting intervals are longer than 35 days, nutrient quality can decline rapidly. Generally, WSC content is higher during April to June for the 8 first and second cuttings than later in the year. It has been suggested that this may be due to the high leaf-to-stem ratio, and also to climate, light intensity, increasing day length, and temperature effects. High rates of nitrogen application as chemical fertilizer or manure shortly before ensilage can influence the nitrate concentration in forages. High soil nitrogen levels produce high forage nitrate levels which are undesirable and are generally associated with lower levels of WSC and an increase in silage pH. At levels of 0.11 to 0.20 % nitrates in forage fresh weight, nitrate is reduced to nitrite, a radical that inhibits the undesirable formation of butyric acid. Some of the nitrates in forages are ultimately degraded to ammonia which tends to raise the pH of the silage and result in nitrogen losses (44). Nitrogen oxides are also formed when nitrates are degraded by microbial metabolism. Nitrogen dioxide (N02), often called \"silo gas\", results when nitrogen monoxide contacts air and it is highly toxic to humans and animals in concentrations greater than 25 ppm. Other gases formed are nitrogen tetroxide, nitric oxide and carbon dioxide and these gases are also produced during the first few days after silo filling. 2.2.1.2 Buffering Capacity (Organic Acids) The buffering capacity of a forage is defined as the degree to which forage material resists changes in pH and it influences the ease with which the forage can be ensiled (17). A high buffering capacity resists a reduction in pH which is necessary for good preservation. Therefore, more acid must be produced to reduce the pH to desired levels. The organic acids (malic, succinic, malonic and glyceric) in forages are mainly responsible for buffering capacity. During the ensiling process, these organic acids are degraded by lactic acid bacteria and replaced by fermentation acids (lactate, formate, and acetate), ethanol, 2,3-butane diol, 9 and acetoin with stronger buffering properties that cause the buffering capacity of the forage to increase two to four fold. When compared to forages with a low buffering capacity, these highly buffered forages have been estimated to require twice the amount of lactic acid to produce a good fermentation that results in a low pH. Therefore, if a forage has a high buffering capacity and an insufficient amount of WSC, insufficient lactic acid will be produced and an undesirable type of fermentation will occur. An undesirable fermentation also results in a loss of dry matter in the form of C0 2 . The ratio of the final products is dependent on the pH at the conclusion of the ensiling process, with lactate as the predominant product at a pH of 4.0 and larger proportions of formate and acetate with little lactate at pH greater than 5.0 (44). The organic acid content of grasses accounts for 2 to 6% of the DM whereas in legumes it is higher, from 6 to 8% of the DM. Studies have indicated that legumes are more highly buffered than are grasses due mainly to their higher organic acid content, but also to their high content of proteins and salts (17). This disadvantage in terms of ensiling characteristics can be overcome by wilting. Silage prepared from wilted forage will have a lower buffering capacity than the corresponding unwilted material. This can be explained by the fact that during wilting under aerobic conditions, the organic acids will not be transformed into stronger acids but respired to C0 2 and water (44). Another factor influencing the buffering capacity is maturity. The buffering capacity of ryegrass has been shown to decrease with maturity (44). 2.2.1.3 Nitrogenous Substances Between 75 and 90% of the total nitrogen in fresh forage are in the form of protein-nitrogen. The remainder consists of amino acids, amines, amides (glutamine and asparagine), 10 chlorophyll, ureides, nucleotides and low molecular-weight peptides (44). Nitrate will also be present and is dependent on the weather and the amount and the timing of application of nitrogen fertilizer applied to the crop. High soil nitrogen levels can predispose the silage to a clostridial type of fermentation by: 1) lowering WSC therefore a low pH becomes difficult to achieve, 2) increasing nitrates which are converted to ammonia-nitrogen resulting in an increase in silage pH, and 3) increasing the buffering capacity which slows the rate of silage pH decline (13). Once a forage is cut, proteolysis occurs which may be encouraged by wilting. Moisture appears to be the most important factor in proteolysis and during a rapid wilt, there is little evidence of protein breakdown. Nitrogen stability is reached at DM contents of 40 to 45%. Proteolysis continues during ensiling and decreases as hydrogen ion activity increases. The nitrogenous degradation products are discussed more fully in a later section. 2.2.1.4 Dry Matter Content Moisture is essential for the proliferation of desirable microorganisms but, an excess (>80%) will encourage the growth of undesirable organisms and a butyric acid fermentation may occur. In general, the higher the DM content of the grass, the lower the bacterial activity (44). To achieve a higher DM content with minimum proteolysis and WSC loss, the grass is wilted quickly after harvesting. Wilting reduces the production of effluent that results from plant cell lysis after active fermentation begins. The amount of effluent that leaves the silo is primarily dependent on the DM content of the crop and the type of silo. For example, the DM content at which no effluent leaves bunker silos ranges between 29 and 33% (20,31) and 35 and 40% for top unloading tower silos lower than 30 feet high (20). In tower silos more than 30 feet high, seepage and loss of nutrients due to pressure can occur at DM contents less than 40% (17). The effluent 11 contains valuable amounts of highly digestible nutrients such as sugars, fermentation products, soluble protein, nonprotein nitrogen, and minerals. The easiest way to reduce effluent is by prewilting the crop to at least 30% DM (44). Most of the effluent flow from a silo occurs within the first week of the ensiling period (17,31). Wilting is also beneficial for ensiling because lactic acid bacteria are more tolerant of higher DM contents than are the undesirable clostridial organisms. Forages with less than 30% DM are undesirable since clostridial growth may not be inhibited even when the pH drops to 4.0. The recommended DM content for grass stored in different types of silos are given in Table 2.1. Table 2.1: Silo Type and Recommended Dry Matter Contents8 a From (2,32) At DM levels in excess of 55%, mold growth and excessive heating is more likely to occur unless an oxygen limiting silo is used or the forage is finely chopped and well packed. Another exception is large round bale silage that can also be fermented successfully to a high quality product starting with a high DM content providing the exclusion of air is maintained. Silo Type Horizontal (bunker) Concrete Tower Silo Bags (plastic tubes) Round Bales (plastic wrapped) Oxygen Limiting Silos Recommended Dry Matter Contents 30 - 40 % 35 -40% 30-40% 40 - >50 % 50 - 60% 12 2.2.1.5 Bacterial Content The number of lactic acid bacteria present on fresh forage crops is low (44). These bacteria multiply during the fermentation stage and the most desirable fermentation will occur when lactic acid producing bacteria predominate. The fermentation of hexoses by these bacteria is either homolactic or heterolactic. The homolactic bacteria ferment hexoses by the glycolytic pathway to produce pyruvate followed by dehydrogenation to yield lactate. One mole of glucose or fructose is fermented to two moles of lactate with no loss of DM. The heterolactic bacteria ferment glucose to yield one mole each of lactate, ethanol, and C0 2 . This results in up to a 24% loss of dry matter through the evolution of C0 2. If fructose is fermented by heterolactic bacteria, then every three moles will yield one mole each of lactate, acetate, and C0 2 and two moles of mannitol. This results in a 4.8% loss of DM but little lactic acid production (31). The homolactic fermentation is more efficient in terms of lactic acid production than is heterolactic fermentation and a heterolactic fermentation will be even less efficient in terms of lactic acid production if fructose is more abundant than glucose, as is the case with grass. It has been found that a pH of 4.0 in silage can be reached with approximately 100 g of lactic acid per kg of DM depending on the buffering properties of the material (44). Badly preserved silage results when the fermentation has been dominated by enterobacteria or Clostridia or both. Enterobacteria (eg. E. coli) may contaminate the forage if manure application is made shortly before harvest. Enterobacteria are normally regarded as acetic acid producers and Clostridia are regarded as butyric acid producers. Both acetic and butyric acids provide no energy for rumen microbial growth because they are nonfermentable (10). However, they can be absorbed directly from the gastrointestinal tract by the animal and used as a source of energy. 13 Clostridial bacteria utilize lactate and sugars to form pyruvate followed by butyrate production. It is thought that the pathway from lactate is the major route and two moles of lactate yields one mole of butyrate. This would mean there is a close correlation between the amount of lactate utilized and the amount of butyrate formed (44). Badly preserved silages are characterized by a low lactic acid to butyric and/or acetic acid ratio and have ammonia nitrogen concentrations in excess of 10% of total nitrogen. An excess of ammonia is harmful to cows and can result in high blood urea nitrogen. Blood urea nitrogen greater than 20 mg/100 ml has been linked to potential problems with cow reproductive performance. Silages with high levels of ammonia may also contain amines that have been shown to be toxic. Feeding badly preserved silage containing compounds produced by clostridial fermentation can disrupt rumen ecology resulting in reduced feed intake and milk production (20). 2.2.1.6 The Importance of Oxygen The most important factor in ensuring good ensiling procedures and optimal fermentation is to exclude oxygen from the storage facility. Oxygen exerts its influence from the time the crop is cut until it is eaten. Exposure to oxygen can lead to the delay of lactic acid formation, the proliferation of clostridial bacteria and the promotion of protein breakdown. The presence of oxygen will also lead to an increase in temperature of the ensiling forage which can cause the silage to become heat damaged. Air will also delay lysis, the breakdown of the plant cells and the release of cell contents. This then causes a delay in the onset of rapid fermentation since the nutrients in the cell contents are not immediately available to the microorganisms. 14 Exposure to oxygen during the filling of the silo will encourage the growth of fungi that cause instability and susceptibility to aerobic deterioration when the silage is being fed. In most instances, small amounts of fiingi will not cause problems, but, under certain circumstances, these organisms (eg. Aspergillus) can produce mycotoxins that can cause severe diarrhea, loss of appetite, abortion, muscular tremors, pneumonia, allergic reactions and in severe cases, death. The microorganisms in the rumen can also be affected. The forage may be exposed to air if the silage is not covered with plastic or sealed in some manner. Forage that is not chopped finely enough or is inadequately packed may also result in excessive air in silage. 2.2.1.7 Temperature In the ensiling process, the growth of microorganisms results in a rise in temperature and high temperatures can affect the rate of microbial metabolism. Higher temperatures (> 30° C) encourage the growth of undesirable Clostridia resulting in increased butyric acid and ammonia formation (44). Clostridial growth can be inhibited by raising the DM content to greater than 30% and rapidly lowering the silage pH (31). At DM contents greater than 30%, the growth of lactic acid producing bacteria has been shown to increase at temperatures between 15 and 25° C while inhibiting the growth of undesirable clostridial species (31). Temperatures in excess of 40° C, in the presence of oxygen, will result in a chemical reaction between WSC and protein. This reaction produces heat damaged silage (Maillard products) characterized by a brown colour with a tobacco or caramelized odour. The protein bound in this reaction is largely indigestible to the rumen microorganisms and to the animal. These compounds increase the silage concentrations of ADF and acid detergent insoluble nitrogen (ADIN) and are most common in silage with a high DM content (31). 15 2.2.1.8 Mechanical Pretreatment The cutting, chopping and bruising of the forage before ensiling improves the potential for making good silage. The forage is first cut, followed by chopping to the desired length. This causes some cell breakage which favours bacterial growth and facilitates packing for air exclusion. Anaerobic conditions can be established quicker in chopped forages and temperatures do not rise as high in forages not chopped. Shorter chop lengths result in lower silage pH, reduced protein degradation and ammonia content, reduced butyric acid concentrations, and increased lactic acid concentrations (44). 2.2.1.9 Silage Additives Silage additives encompass a wide range of products and are designed to improve silage quality (7). They are used to reduce DM and nutrient losses, particularly the loss of protein, improve fermentation in the silo, reduce deterioration after the silo is opened and improve the nutritional value of the silage with the aim of increasing silage DM intake and animal performance. Good silage can be made without additives or preservatives if proper ensiling techniques are used. However, anaerobic conditions and the attainment of high DM levels may be difficult to achieve. Additives are most useful where low D M low WSC or high buffering capacity materials are ensiled or when storage facilities are not well sealed. They will not make \"poor\" silage \"good.\" The silage additive industry is expanding and the range of products is changing continuously. Specific silage additives and preservatives are classified on the basis of their mode of action and those used in grass silage making are separated into three types: 16 A. Fermentation Inhibitors Fermentation inhibitors reduce the pH to less than 4.5 thus inhibiting microbial activity. I) Mineral Acids and Organic Acids These include mineral acids such as sulfuric and hydrochloric acid and organic acids such as formic, acetic, lactic, benzoic, and propionic acids. Mineral acids and benzoic, and formic acid-treated silages are not commonly used in Canada and are, therefore, are not reviewed in this study. In Canada, propionic and acetic acid are the most commonly used fermentation inhibitors. They reduce bacterial growth and have high anti-fungal properties. At recommended application rates, they reduce ensiling losses and improve aerobic stability. They are most practical in cases where DM is high and where aerobic stability is important. Lactic acid, as an additive, inhibits endospore-forming bacteria but does not affect fungi because many of these organisms utilize it as a carbon source. There are few reports of lactic acid use as a silage additive due to its high cost (44). Forages that benefit the most from the addition of acids are those high in moisture, inadequate in WSC for optimal fermentation, and have a high buffering capacity, such as legumes, ii) Non-Acidic Fermentation Inhibitors (Sterilants) Non-acidic inhibitors reduce bacterial activity mainly by partially sterilizing the crop. Formaldehyde, sulfur dioxide, and sodium metabisulfite are some of the fermentation inhibitors that have been used to aid ensiling by inhibiting the microflora in general. B. Fermentation Stimulants Fermentation stimulants, including enzymes and microbial inoculants, encourage rather than suppress the silage fermentation process. 17 Enzyme addition, alone or in combination with microbial inoculants, is a relatively new method of improving the nutritive value of ensiled forages. Cellulolytic enzymes such as cellulases, hemicellulases, xylanases, cellobiases, and pectinases break down structural carbohydrates (cellulose) to glucose and other individual sugars providing bacteria with an energy source and achieving a better fermentation. This can result in an increase in digestibility, an increase in silage DM intake and a significant increase in milk yield (5,39). Amylolytic enzymes break down starch, the principal form of storage carbohydrate in legumes, and have improved silage quality under specific conditions (22). Responses to enzyme additions have been inconsistent. Forage species was the most consistent factor determining whether enzyme additions reduced the concentration of fibre. Grasses were twice as likely to respond than were legumes possibly because the fiber structure in grasses may be more susceptible to enzyme attack than is that of legumes (5,22). Pure enzymes are currently too expensive for practical application, however, there are a number of less expensive commercial enzymes or fermentive breakdown products available. In recent years, interest has been shown in the microbiology of silages and the potential for enhanced digestibility of the silage (16). The aim of inoculation is to supply large numbers of lactic acid-producing bacteria to the crop. They convert WSC to lactic acid more rapidly and more efficiently than the indigenous population of bacteria, thereby inhibiting the activity of spoilage organisms and producing more lactic acid from a limited supply of WSC (39). This also reduces plant protein breakdown (16). Microbial inoculants are now widely used in silage additives and are the most commonly used silage additives in Canada and the U.S. They are considered safer than chemical additives, lower in cost, leave no residue problem, and are non-corrosive to farm machinery. 18 Often there are low numbers of lactobacilli on fresh plants and inoculation with lactobacilli encourages the development of their population, which in turn causes a rapid drop in pH. Microbial inoculation is not beneficial if there is already a large population of lactic acid bacteria on the forage, if the bacterial strains in the inoculum are not compatible to the forage or ensiling conditions, or if the amount added is too low (22). The ideal bacteria used in an ensiling inoculant must produce only lactic acid from glucose, fructose, sucrose and fructosans as an end product of fermentation, have a high growth rate, be able to dominate other organisms in the silage, be acid tolerant down to a pH of 4.0 or less, work at temperatures from 4° up to 50° C, grow in high dry matter situations and not degrade organic acids or proteins. It should not act on organic acids as these are responsible for the preservation and stability of the silage. Examples of suitable bacteria include Lactobacillus plantarum, Lactobacillus acidophilus, and Streptococcus faecalis which all perform well given the availability of substrate. The streptococci that are used in many silage additives are added because of their ability to reduce the pH rapidly to a level where the lactobacilli are more active (16). Performance of the lactic acid bacteria is limited by several factors with WSC supply as the major factor. If the WSC levels are inadequate, then the inoculant will have little chance of ensuring preservation of the silage. This suggests crops with high WSC may benefit from inoculation. Lactic acid bacteria are also limited by their inability to ferment cellulose or carbohydrates larger than oligosaccharides. Other limiting factors include some strains that are not as competitive as others and are liable to be eliminated by other more vigorous strains or, the temperature at which they grow may not be ideal. Also, certain strains of lactic acid bacteria grow best on the crop from which they were isolated. 19 The addition of enzymes to microbial inoculants is becoming more common to enhance the breakdown of forage cell walls during fermentation. Table 2.2 summarizes the effects of lactic acid bacteria and cell-wall degrading enzymes on the fermentation of grass/legume silage. The combination of additives significantly reduces the pH and NDF and increased lactic acid content. Also, the combination of lactic acid bacteria and cell-wall degrading enzymes can result in a desirable increase in the digestibility of the forage. Table 2.2. Effect of Lactic Acid Bacteria (LAB) and Cell-Wall Degrading Enzymes on the Composition of Grass-Legume Silage\" Treatment Cell-wall LAB + cell-degrading wall degrading Treated Untreated LAB enzvmes enzymes pH 4.45 4.04 4.17 4.02 Lactic acid, % of DM 4.25 5.95 5.26 6.22 % NDF b reduction 0 0 1.8 3.5 % ADF b reduction 0 NS° NSC 4.0 a From (13) b Percent of reduction from pre-ensiled fibre composition c Not significant at P > 0.10 C. Nutrient Additives Molasses, sugars, cereal grains, beet pulp, whey, ammonia, and mineral additives can be added to grass forage to improve the nutritional value of silage. The impact of nutrient additives was not reviewed for this study. 20 Effects of Silage Additives The effects of microbial inoculants have been extremely variable due to the many different factors involved such as application rate, type of forage, combinations of forage, storage structures, ensiling conditions (such as DM content and chemical composition) and specific properties and purposes of the inoculants. It has been found that enzyme additions frequently improve DM digestibility and occasionally increase fibre digestibility (21,22). Improvements in animal performance are closely linked with improvements in digestibility. This suggests that increases in digestibility may be the key factor in explaining why microbial inoculants improve animal performance (21). The addition of enzymes, with or without microbial inoculation, was the most effective silage treatment for improvement of milk yield. Twenty-one of the 22 comparisons were with wilted or minimum wilted forage, but responses were also positive with unwilted material. However, more studies are needed to determine the mode of action of the enzymes on cell wall structures and to produce enzyme and inoculant mixtures appropriate for different crops and environmental conditions (13). Silage additives can affect all crop and fermentation variables and can cause interactions between silage quality and animal performance. More research is required on the actual benefits and animal performance improvements resulting from forages ensiled with additives/inoculants along with descriptions of the conditions under which success, failure, and no apparent change occurs. There is also a need for comparisons of the effectiveness of different types of additives applied to the same forage and to different cultivars. However, animal performance is not the only measure of inoculant success. Improved aerobic stability and/or improved chances of successful fermentation of forages difficult to ensile would also produce benefits on the farm. 21 2.2.2 Summary of Dry Matter Losses Estimated DM losses associated with harvesting, storing and feeding grass silage at different moisture contents are summarized in Table 2.3. The table considers DM losses only and losses in nutrient quality are not considered, although, variations in quality could be considerable depending on the many factors discussed in this thesis. Table 2.3. Expected Dry Matter Losses During Forage Harvest, Storage, and Feeding3 Hay/Silage Crop Harvest Storage Feeding Total Drv Matter % % % % % <30 % 2.0 21.2 11.0 34.2 31 -40% 5.0 10.1 11.0 26.1 over 40 % 11.5 8.2 11.0 30.7 \"From (14) Harvest losses include respiration losses, rain damage, and mechanical damage. Mechanical damage normally causes a greater loss of leaves which contain higher concentrations of nutrients than in the stems. As shown in Table 2.3, DM losses during harvest increase with increasing DM content and storage losses decrease with increasing DM content. Feeding losses stayed constant throughout the range of DM contents studied. Losses associated with feeding include feed refusal, feed wastage, loss of DM due to poor aerobic stability, and handling losses from the silo to the feed area (14). 2.3 The Evaluation of Silage Quality The nutritional quality of a forage does not improve with ensiling but it does change. Even if the original material was of good quality, the material may have ensiled poorly resulting in 22 undesirable qualities. Since grasses are ensiled in large amounts, it is extremely important that a representative sample of silage be obtained for quality determinations using good sampling techniques. It is of little value to analyse a silage sample if it does not represent the silage being fed. It is suggested that grab samples be taken and mixed during silo filling and again at the end of the ensiling period during feedout. The sample container must then be sealed to avoid moisture loss and frozen until it is submitted to a laboratory. Samples deteriorate quickly in warm conditions. Volatile constituents evaporate and digestible carbohydrates disappear as a consequence of aerobic activity (8). Sampling and analysis of the silage is recommended throughout the silo so that revisions of feeding programs can be made as the quality of the ensiled forage varies. Silage quality is a function of both forage DM intake and digestibility. The traditional methods used to estimate forage quality are through forage analysis and the common tests include D M CP, and fibre (ADF and/or NDF). Digestibility is determined by in vivo, in situ, and in vitro methods. There are, however, errors associated with sampling, drying, grinding, subsampling, and analysis. These errors accumulate during the whole process. Differences also exists between methods and laboratories. The results of forage analysis can only be as good as the sampling, handling, and analytical procedures used. The CP in silage is largely degradable in the rumen. If the silage has heated too much, the digestibility of protein and DM may be greatly reduced. If heat damage is suspected, as evidenced by a charred appearance and/or a tobacco-like odour, an ADESf analysis may be helpful. In general, if the ADIN content of the silage is above 15% of the total nitrogen in the forage, it is indicative of excessive heating (14). 23 Normally, forage analysis laboratories do not determine the amounts of acids or ammonia-nitrogen in silage, however, the information can be very useful in interpretations of silage quality. The greater the ratio of lactic acid to butyric acid, the higher the ensiling efficiency (7). A description of silage quality based on the different amounts of fermentation products is provided in Table 2.4. Table 2.4. Fermentation Products (Dry Matter Basis) and Silage Quality8 Silage Quality Item Good Intermediate Poor pH of silages with under 65% moisture under 4.8 under 5.2 over 5.2 pH of silages with over 65% moisture under 4.2 under 4.5 over 4.8 % Lactic acid 3-14 variable variable % Butyric acid under 0.2 0.2-0.5 over 0.5 Proportion of total acids % Lactic over 60 40-60 under 40 Acetic under 25 25-40 over 40 Butyric under 5 5-10 over 10 Ammonia N (% of total N) under 10 10-16 over 16 ADIN (% of total N) under 15 15-30 over 30 From (2,13,31) As indicated in Table 2.4, a low pH is often desirable and indicates that a lactic acid type fermentation has occurred. However, pH is not always a good indicator of quality as in the case of large round bale silage. Large round bale silage will typically contain greater than 50% DM and often results in a satisfactorily ensiled product with a pH of -6.0. Good round bale silage can be fermented with little lactic acid or butyric acid production, due to limited availability of moisture for the fermentation process. i Chemically, fresh grass and ensiled grass differ greatly in only two ways. The first, is that grass silage contains less WSC than fresh forage since they are utilized by bacteria in the 24 production of organic acids in the ensiling process. Secondly, in fresh forage, 75 to 90% of the nitrogen is in the form of protein with the remainder occurring as amino acids, amides, amines, chlorophyll, nucleotides, nitrates and low molecular-weight peptides (44). After ensiling, significant amounts of the proteins (25 to 85% of total nitrogen) are degraded to nonprotein nitrogen compounds such as amino acids, amines, and ammonia (20). The amount of proteolysis depends on the forage species, pH, time in the silo, temperature, and moisture content. The different levels of nitrogenous components that can result from different fermentations are illustrated in Table 2.5. Table 2.5. Nitrogenous Components of Three Different Silages8 Nitrogen Components Lactate Acetate/Butyrate Chemically- Restricted (% of Total Nitrogen) Silage Silage Silage Protein-N 30.8 26.0 81.4 Non-protein-N 62.9 74.0 18.6 Free-amino-N 36.9 27.8 3.5 Amlde-N trace 0.1 0.2 Amine-N 9.6 11.4 0.7 Ammonia-N 12.1 29.2 0.8 Nitrate-N 4.9 0.1 1.0 Other-N 3.2 4.2 8.0 From (40) Non-protein nitrogen degradation products are not as nutritionally useful as intact proteins. Amines, in particular, may cause adverse reactions and a lower feed intake in animals. Free rumen amino acids are used rapidly by the rumen microorganisms and not by the cow itself. As a result, lactating cows may require additional rumen-undegradable protein for maximum milk production with a diet already high in crude protein. 25 Under conditions that favour clostridial organisms, Clostridium botulinum can grow which, when fed, can cause botulism.. It was suggested that the presence of these organisms may be due to contamination of the fresh forage with manure (44). Ruminants are less susceptible to this disease because the ruminal microbes rapidly inactivate the toxin responsible. The majority of hazardous substances (eg. amines) found in silage are associated with poor quality silage or where ensiling conditions are unsatisfactory. 2.3.1 Errors Associated with the Chemical Analyses of Forages Since fibre is not well defined chemically, different analytical methods produce different results and the residues left at the end of an extraction are only representative of the conditions employed in that particular procedure. Any small change in the method or extraction conditions will yield a different fibre recovery. This characteristic, for example, results in coefficients of variation for ADF on the same sample between different laboratories that are high and average approximately 3.96% for wet chemistry and 3.55% when predicted by NIRS as shown in Table 2.6. Neutral detergent fibre analyses produce average coefficients of variation even higher at 6.40% from chemical methods and 5.37% from MRS predictions. In contrast, DM determinations typically produce a coefficient of variation of 0.94% for standard oven-drying techniques and 1.82% by NIRS prediction. Crude protein analysis when performed by different methodologies produce coefficients of variation of 3.35% and 4.67% for wet chemistry and NIRS, respectively. The wet chemistry CP coefficient of variation is high due to the differences in procedures and instrumentation utilized in this procedure. Combustion Nitrogen Analysers typically recover greater amounts of nonprotein nitrogen than Kjeldahl methods thus resulting in greater differences between laboratories. 26 These coefficients of variation were derived from the National Forage Testing Association Check Sample Program quarterly reports for 1994. The forage check samples were dried and ground prior to wet chemistry analysis and MRS predictions. Table 2.6 does not take into account variation associated with sampling or grinding. It also suggests that MRS is capable of better precision than some of the methods used for its calibration. Table 2.6. Examples of Coefficients of Variation of Chemical and NIRS Analyses on Ground Forages Between Laboratories3 Chemical Analysis MRS Predictions Dry Matter 0.94 1.82 Crude Protein 3.35 4.67 ADF 3.96 3.55 NDF 6.40 5.37 a Derived from the National Forage Testing Association Check Sample Program quarterly reports for 1994. 2.4 Near Infrared Reflectance Spectroscopy (NIRS) and its Application to Forage Analysis In the following section, the principles of MRS will be briefly summarized together with sample selection, reference methods, statistical methods, calibration methods, and a discussion of the accuracy of MRS. 2.4.1 Brief History of NIRS The analysis of forages by traditional wet chemistry methods is expensive and time consuming. An instrument capable of analyzing the physical and chemical composition of a sample in minutes, with little or no sample preparation would improve analysis significantly. In 1976 the original MRS paper related to forages was published (24). This paper described the development 27 and testing of prediction equations for protein, NDF, ADF, lignin, in vitro dry matter disappearance (IVDMD), in vivo digestibility, voluntary DM intake, and intake of digestible energy for forage evaluation. These workers showed that forages could be analyzed for quality constituents by MRS. Based on their results, instrument manufacturers developed M R discrete filter instruments capable of forage analysis. Scanning monochromator instruments capable of collecting spectral data every two nanometres (nm) were developed later and provided additional evidence that MRS could provide rapid and accurate analysis of forage quality. In 1978, the U.S. Department of Agriculture MRS Forage Network was formed to coordinate research, develop software, and prepare the technology for the private sector. The project was later expanded to include cereal grains, soybeans, and other agricultural products. Since forage and manufactured feed are more variable in physical and chemical composition than grains and oilseeds, scanning monochromators, scanning filter instruments, or fixed filter instruments with more than ten filters were found to be necessary to produce acceptable accuracy (37). Research before 1986 involved improvement of software and instrument design, calibrations for new applications and constituents, and feasibility studies. More recent advances have included instrument standardization, calibration techniques, monitoring of instrument performance and interpretation of the spectra. Currently, an estimated 70 to 80% of dairy cows in the American Midwest are fed rations whose nutrient content is predicted with an MRS instrument. The U. S. GIPS A (Grain Inspection, Packaging and Storage Administration) use MRS to predict moisture and protein for commodity trading decisions. Since 1975, the Canadian Grain Commission has utilized MRS analysis in all grain shipping terminals for the rapid analysis of wheat for international commerce. European and North American food processing industries and agricultural companies use the technology to 28 monitor the fat, protein, fibre, and moisture composition of products ranging from candy to hamburgers. 2.4.2 Principles of NIRS The near infrared region of the electromagnetic spectrum extends from 750 nm to 2500 nm. The spectral region from 800 to 1900 nm is referred to as the overtone region, whereas 1900 to 2500 nm represents the combination band region. Fundamental absorbances or wavelengths at which energy is absorbed, mainly occur in the mid-infrared region between 3000 and 15000 nm. The absorptions associated with excitations of molecules to higher energy levels are referred to as overtones and since fewer molecules reach these higher levels, the higher (second, third, etc.) overtone bands are generally much weaker in intensity than the fundamentals. Absorbers in the combination area are combinations of fundamental absorbances and their overtones. Absorption in the 400 to 750 nm region contains information on visible parameters such as colour. The principle of MRS is that the major components of a sample have near infrared absorption properties which can be used to differentiate one component from the others. Calibration of the instrument using samples of known composition enable the technique to become quantitative. Calibration involves the establishment of a mathematical relationship between absorption at a specific wavelength and a reference set of sample materials of known composition. Effective calibration depends on many factors including assembly of samples containing all sources of expected variation, consistent sample preparation, accurate reference data, a mathematical model for their relationship and consistency throughout the whole process (23,36,37,42). Monochromatic light produced by a tungsten-halogen lamp in a M R instrument irradiates the sample material and causes transitions from the \"ground\" energy level to higher energy levels. 29 These cause absorbances of irradiated energy in the mid-infrared region, together with associated overtones and combinations. The absorbances are detected by the instrument in reflectance or transmittance mode, amplified, and computed into optical density signals, which in turn are displayed as composition data, after the instrument has been calibrated. Absorption in the NIR region is primarily due to carbon-hydrogen (C-H), nitrogen-hydrogen (N-H), and oxygen-hydrogen (O-H) bonds with carbonyl carbon-to-oxygen (C=0) double bonds and carbon-to-carbon (C-C) bonds also showing importance. Signals are recorded as log (1/R), where R represents reflectance, at each wavelength point (usually every two nanometres). A higher log (1/R) value means that more radiation has been absorbed (less reflected) at that wavelength (42). NIRS data from scanning monochromators produce a spectrum composed of many overlapping absorption bands between 400 and 2500 nm. These bands are defined by location, height, and width. Due to overlapping and the fact that stronger bands may mask other bands, the individual characteristics of most bands cannot be accurately defined in log (1/R) form, and derivatization can be utilized to separate the bands. Derivatization also counteracts baseline drift. Water is a strong absorber of energy at 1425 and 1940 nm and the differences in moisture are apparent in the spectra. The water content in products such as fresh grass and silages has a large effect on the state of other constituents in the sample and can affect the accuracy of determination of these constituents when present in lesser concentrations. Hydrogen bonding and changes in hydration can lead to interactions among and between constituents. Protein absorbs M R energy more weakly than water and spectral differences due to high and low levels of CP are not as apparent in the spectra (42). 30 2.4.3 Sampling and Sample Presentation Efficient sampling is essential to ensure a representative sample of the material to be analyzed. If a sample is not representative, the laboratory analysis values will not correspond with the actual nutrient value of the forage. Presentation of the sample to the instrument is also important. Traditional NIRS prediction equations are based on dried, ground samples packed and presented in small, round (3.75 to 5.0 cm diameter), ring cups fitted with quartz windows. Some researchers have reported results from undried, ground forages also presented in small, round cups and in rectangular cells (1,6,27,28,29,30). Their results have indicated a potential to predict the composition of high moisture materials. Abrams et al. (1) were first to show promising calibrations on undried, chopped silages for DM, insoluble nitrogen and total nitrogen on samples taken from a broad range of ensiling conditions. Standard errors of calibration on an \"as received\" basis and multiple coefficient of determination (R2) were reported as 1.80 and 0.98 for DM and 0.070 and 0.98 for total nitrogen, respectively. However, it was concluded that the results were more accurate if the samples were scanned after drying and grinding. Steps that might be taken for improving accuracy in future studies include a better method of reducing the sample to a more uniform particle size before scanning, a larger presentation cell, and calibrations based on silage samples collected from numerous farms. Recently, the natural products cell (Perstorp Analytical / NIRSystems) has been introduced. It is a larger cup that allows more heterogenous samples to be presented in their undried, unground form. This cup, along with multiple sample repacks and spectra averaging, reduces variation among samples and enables larger volumes of sample to be scanned. Researchers are now reporting calibrations with unground forages utilizing these improvements (33,37,38). 31 Prediction results obtained from DM, CP, and ADF calibrations developed from fresh, undried, unground haylage and pasture grass compared to calibrations derived from the same samples scanned after drying and grinding are illustrated in Table 2.7 (33). In all cases, the standard errors of cross validation (SECV), an estimation of prediction accuracy, were higher for undried, unground material than on dried, ground materials. Lower calibration SECVs indicate a more accurate calibration. Also, the R 2 values for CP and ADF were higher for the dried, ground samples than on the intact samples. However, the R 2 value for the DM calibration was higher for the undried, unground forage mainly due to the greater range in DM content in the sample set. A conclusion reached from this work was that the prediction equations developed from undried, unground forages were accurate enough for routine use. The accuracy of these prediction equations is discussed in greater detail later in this thesis on pages 71 to 74. Table 2.7. Calibration Statistics for Dried, Ground (DG) and Undried, Unground (UD/UG) Haylage and Pasture Grass3 Haylage Pasture Grass DG UD/UG DG UD/UG Constituent SDb SECVC R 2 SD SECV R 2 SD SECV R 2 SD SECV R2 % % % % % % % % Dry Matter 1.36 0.76 0.84 12.73 1.20 0.99 3.34 0.58 0.68 8.12 0.73 0.99 CP(DM) 3.37 0.71 0.97 3.37 1.16 0.91 7.73 1.41 0.97 7.73 1.93 0.95 ADF(DM) 2.12 1.63 0.90 2.12 2.12 0.83 7.75 1.96 0.95 7.75 2.49 0.86 a From (33) b Standard deviation of the reference values in the calibration set 0 Standard error of cross validation: an estimation of prediction accuracy 32 2.4.4 NIRS Reference Methods Errors are defined as differences between the measured values and the true or correct values (42). The errors associated with NIRS are largely dependent on the reference method and sampling error (19,25,37,42). One of the problems with consumptive chemical analyses is that when the results of analysis of replicate samples do not agree, it is sometimes difficult to determine if there was an experimental error during analysis or an actual difference between the samples (4). Outside of carelessness, all errors are of two types: systematic and random. Systematic error can be minimized through careful review of all steps in the procedure but random errors are not easily explained. The reference methods used for analysis of samples for calibration, validation, and monitoring must be accurate or the calibration will not be effective. Also, reference analysis must be performed on the same sample as the one from which absorption information has been collected because variations in the sample can greatly influence results from the primary analytical method and affect calibration development. The sample must, therefore, be well mixed. Many of the current laboratory methods used for calibrating NIRS instruments for forage analysis are poorly-defined chemically (eg. fibre) and relating them to spectroscopic data can be difficult (36). Precision or reproducibility in wet chemistry analysis can limit the performance of NIRS. Precision or repeatability of an NIR test is the agreement between successive readings of a sample by the instrument after loading the cell (42). Bias is defined as the mean of the differences between the actual measured reference values and the predicted values. Moisture: The most common reference method for measuring moisture in forages is oven drying. This method drives off volatile compounds such as alcohols, acids, esters, amines and ammonia present 33 in the sample as well as moisture. However, even after oven drying, NIRS can often detect a trace of moisture remaining in a sample. Apparently calibrations developed from oven moisture methods contain analytical error and are not as accurate as expected. The NIRS calibration error for moisture increases with increasing concentrations of volatile compounds (36). NIRS correlates more closely with the toluene distillation method of moisture determination than to oven drying methods. Toluene distillation gives grass silage DM values up to 11% higher than the oven drying method but the method is not normally carried out as a routine procedure in analytical laboratories due to the tedious nature of the determination and the small amount of sample used. The difference between DM estimated by oven drying and toluene distillation is greatest in silages of high pH, high volatile fatty acid concentrations, low lactic acid content, and high amounts of ammonia nitrogen (12). Ideally, the toluene method of moisture determination would be the most accurate method for calibration due to its specificity for water, however, accurate DM calibrations can be developed from forced air drying methods. Knowledge of the calibration samples and whether these samples are subject to losses of volatiles during drying is important in understanding and correcting for errors associated with the calibration (43). The moisture bands are among the most prominent features in forage spectra and occur at 1150, 1390, 1425, and 1940 nm with the most prominent, broad bands at 1425 and 1940 nm (36,42). These broad bands consist of multiple overlapping bands containing information on more than one hydrogen-bonded subspecies. Variations in hydrogen bonding molecular subspecies can broaden and shift peak positions (26,36). This is critical for the analysis of intact fresh grass and grass silage since water is the most important single constituent. Water peak positions shift with changes in the relative proportions of the individual bands making up the composite bands, the concentration of major constituents in solution in the sample such as the organic acids in grass 34 silage samples, water content itself (26), and with changes in temperature (36). This implies that different wavelengths may be necessary for determination of water in samples varying widely in moisture content and samples varying widely in temperature at the time of analyses (42). Protein: The nitrogen content of forages varies considerably. Nitrogen content is one of the most successful measurements made by NIRS and the main reason for this is the superior precision of the Kjeldahl test relative to some other reference tests. However, it does not differentiate between true protein and non-protein nitrogen (eg. N0 3, N02). Nitrogen content is multiplied by 6.25 to convert nitrogen to protein. By this method, total nitrogen content is used to estimate CP. NIRS can measure =N-H bonds (peptide bonds), amines and other nitrogenous substances present as well as protein and may be measuring several different components in the sample. The relationship between MRS and the reference method depends on the agreement between absorbances detected by the instrument, and the total nitrogen content (37). Crude protein content is more precisely determined in fresh grass than in silage due to protein degradation products, losses of volatiles, or sampling difficulties that occur with wet silage (23). The most important wavelengths for measurement of protein by MRS include the carbonyl of the primary amide at 2054 nm, and the 2168 to 2180 nm band consisting of N-H, C-H, C-N, and C=0 absorptions (23,36,37). The N-H absorption bands at 1500 to 1530 nm, 1690, 1734, and 1974 nm can also be useful. Amino acids and proteins which contain aromatic rings exhibit aromatic C-H absorptions in the 1640 to 1680 nm region. Variation in the nitrogen content of forages is shown by a change in shape of the log (1/R) spectrum in the region from 2000 to 2200 nm (37). High nitrogen content leads to a \"bulging\" or flattening of the region while low nitrogen leads to a sharper peak at 2100 nm (23). 35 Fibre: NIRS instruments have been successfully calibrated to measure Van Soest detergent components (11,41). Since cellulose is a major portion of the fibre in forages, calibration equations often involve the prominent cellulose absorption peaks at 2272 and 2330 nm. Lignin consists of phenylpropanoid residues, so wavelengths in the regions of 1680 nm and 2200 to 2300 nm are normally present in the calibration equation (37). The lignin content of forage is the chemical compound most inversely correlated to digestibility but its measurement by wet chemistry has proved difficult. The absorption at 2266 nm is common to both lignin and cellulose while the band at 1666 nm absorbs for lignin, but not cellulose and could be used to differentiate between cellulose and lignin in plant tissues (23). The wavelength at 2266 nm has been found to be associated with lignocellulose (23). Acid detergent fibre is defined as the fraction of a feedstuff remaining after extraction with acid detergent solution when analyzed by the Van Soest system of detergent analysis (11). This procedure divides the carbohydrates in plant constituents into highly available and poorly available (undegradable) components to ruminants. It is often used as a measure of cellulose and lignin and is most influenced by stage of growth. As plants grow, there is a greater need for structural tissues, and cellulose, hemicellulose, and lignin all increase with age. The ADF residue contains mainly cellulose and lignin but also contains some minerals, tannins, cutin, and pectin. Total cell walls, which also includes hemicellulose, is measured by the NDF test. Carbohydrates: In 1987, cell wall carbohydrates and starch in alfalfa were predicted by NIRS (3). Because of the complex nature and mixtures of these substances in the cell wall, due to species, plant maturity, and methods of forage storage, these prediction equations do not predict as accurately as those for nitrogen analysis (36). 36 NIRS can be used to determine WSC in fresh grass to determine if the sugar status of a grass crop is sufficient to allow natural lactic fermentation to preserve the silage without using acid additives. It has also been used by plant breeders to screen extremely large numbers of different plant varieties at various stages of plant maturity. Plants of different maturity show significant differences in NIR spectra and, therefore, NIRS spectra can be used to measure plant age as well as composition (23). Minerals: NIRS measures absorption by molecular bonds. Mineral forms of calcium and phosphorus have no NIR absorption bands. However, organic complexes such as the phosphorus in phytate, and chelates could be detected by NIRS. Some ionic substances do have absorbances in the NIR region including sulphates, carbonates, and phosphates. Calibrations for mineral content suggest that NIRS absorption bands for organic complexes associated with minerals might exist, although direct evidence is lacking (36). More research is required to determine the relationships between the organic complexes in crops and individual mineral concentrations. 2.4.5 NIRS Statistical Methods 2.4.5.1 Structuring of the Sample Population Accurate NIRS calibration development is dependent on the collection of representative samples that contain all sources of the population variance. Ideally, calibration sets should cover both the constituent and spectral range uniformly, as well as being composed of a uniformly distributed set of sample types (37,42,45). If the constituent distribution is uneven, the calibration equation will become biased toward the unevenness in the constituent distribution (42,45). 37 Often a large number of samples from a population is available but obtaining laboratory reference values by wet chemistry methods for all samples would be expensive. Therefore, different methods have been developed that can be used to define a population and select samples for a calibration. The first method employs principal components to condense the information in the spectrum to a small number of principal components or scores. A principal component score is the amount of change in the principal component pattern. Each of these scores is a linear combination of the original measurements and contains information from the entire spectra without loss of spectral information (25). Principal component analysis (PCA) identifies patterns in a group of spectra that contribute the most to the variation among the spectra. The first principal component should account for most of the variance and corresponds to the most important differences among the spectra, often particle size. Subsequent principal components are determined on the residual variance. The second component often corresponds to the largest absorbance band changes, such as moisture. Additional components include variation in other constituents, and interactions among constituents (25,37). Software is available to rank the spectra according to their Mahalanobis distance from the average spectra of the sample set, using principal component scores. The Mahalanobis distance is a measure of the squared multidimensional distance of a spectra from the mean of all samples and all the wavelengths in the calibration model are used. The value is independent of the concentration of any of the constituents and it removes redundant information so it will not be counted twice. The Mahalanobis distances, standardized by dividing them by their average value, are called \"global H\" values (36). The average score is then calculated and the global \"H\" is determined from each sample relative to the average sample point. This procedure also defines the spectral boundaries of a population of spectra. Samples more than three standard deviations away from the mean are 38 considered to be spectral outliers (global \"H\" is greater than 3.0) and have the largest values. The spectra can then be sorted from the smallest to the largest global \"H\". Samples with extreme spectra have a large amount of influence during calibration development. The nature of the difference is not necessarily clear and it may not have a strong effect on predictability of composition. A sample with a large global \"H\" may have a high or low constituent value, an unusual combination of constituent values, a different particle size distribution, or an excessive amount of an extraneous constituent (19). Verification of the extreme spectra (ie. to determine if the extremes are true extremes) can be accomplished by repeating the MRS scans. If they are still found to be extreme, examination of the samples to determine the cause of the differences, followed by evaluation to determine if they are a valid extension of the population is essential. If they belong with the population, they will broaden the range of samples that can be predicted. However, if they do not belong with the sample population, they will affect the accuracy of the calibration (37,42). It is important to determine whether such samples are likely to occur in future populations. Calibration samples can be chosen by selecting every nth sample from the ordered set of samples with the remainder being retained in a validation (prediction) set. A validation or prediction set is a set of samples independent from the calibration set that should contain as much variation as the calibration set. Validation refers to the comparison of the results of the MRS analysis relative to the reference data, and is used to determine the accuracy of the prediction equations. A second method of calibration sample selection also utilizes spectral variation and principal component scores. The Mahalanobis distance is computed between the scores of two samples and is called the \"neighbourhood H ' value. The neighbourhood of a sample is the space near a sample 39 and is defined as all sets of scores with neighbourhood \"H\" distances to that sample less than a specified cut-off value (34,37). The procedure assumes that only one sample is required to represent all samples in a neighbourhood. The spectrally similar samples not chosen for the calibration set can then be used as a validation set. However, the validation set may not include all variation of the calibration set if there are no spectral \"siblings\" at the extreme ends of the range. A third method of calibration sample set selection involves the ordering of the samples by constituent value after the samples have been analyzed by the reference method. The samples are ordered by constituent value from low to high and values represented by multiple samples can be removed to produce a rectangular boxcar distribution. Every nth sample can then be selected for a calibration set and the remainder can be used as a validation set. This produces calibration and validation sets with similar statistics (mean and standard deviation) but spectral differences may occur between calibration and validation sample sets. A fourth method involves random selection of samples without regard for spectral variation or constituent value. There is no guarantee that the calibration and validation sets will contain the same amount of variation as the population sampled. Calibration set selection for the present work was performed utilizing the first three methods and the results were compared. It must be borne in mind that in the \"real world\", all samples submitted to a laboratory must be analyzed. Calibrations should be monitored continually, and updated periodically. 2.4.5.2 Calibration Calibration is the term used to describe the relationship between NIR absorption data and laboratory reference method information. To develop the calibration equations, different 40 mathematical pretreatments can be employed to reduce interferences from variables such as particle size. Chemical information in a spectrum is present at specific wavelengths while physical properties, such as particle size, affect the whole spectrum. Particle size variations cause changes in radiation scatter which causes baseline shifts. Large particles result in higher absorption than smaller particles (42). Different mathematical procedures for particle size correction can be used if the particle size variance in the spectra is larger than the variance due to chemical information. Detrending removes the linear and quadratic curvature of each spectrum and accounts for baseline shift, but does not affect the particle size variance. Standard normal variate (SNV) transformations standardize the spectra by scaling each spectrum to have a standard deviation of 1.0. If a multiplicative effect is present, then strong absorbers such as water show more change with particle size than the weak absorbers. Normal multiplicative scatter correction can be applied to shift each spectrum up or down to look most like the average spectrum in the file. These scatter corrections for particle size may or may not improve the accuracy of the NIRS equations (37). Derivatives can also be utilized to separate overlapping peaks, emphasize small absorption peaks, and to correct for baseline variations. Derivatives are characterized by a \"segment\" and a \"gap\". The segment refers to the wavelength range in nm averaged together and it has the effect of \"averaging\" out the random noise in the spectral data (smoothing). The gap refers to the distance between wavelength point averages being subtracted. A first derivative utilizes two wavelengths and is calculated as log (1/R)A - log (1/R)B separated by the specified gap. A second derivative uses three wavelength points and is calculated as log (1/R)A - 2 * log (1/R)B + log (1/R)C. The size of the segment and gap should be optimized during calibration development. No single mathematical treatment is ideal for all applications. 41 The number of samples required to develop a calibration that proves to be robust in routine use for forages is unlikely to be less than 100, except for constituents such as moisture or protein. In the case of fresh grass and grass silage, a population of at least 200 is recommended for calibration. Again, it is emphasized that the population should contain samples representing the full range of variance due to composition, stage of maturity, growing conditions, species, storage (ie. ensiling) conditions, and other sources of variance. A calibration based on 200 samples will not perform better than one based on 100 samples if there is no extension of the variance (42). 2.4.5.3 Regression Methods The regression method used in this thesis was modified partial least squares (PLS). Modified PLS uses all wavelengths in the spectrum and includes laboratory reference method values to develop the equation. The resulting equation effectively uses all of the data or wavelengths because each of the new variables involves all of the original measurements. Each wavelength has a corresponding coefficient and is weighted depending on its correlation strength. Coefficient size is directly related to the importance of that data point or wavelength in explaining the variance in the reference value. The NIR residuals are standardized (divided by the standard deviations of the residuals at a wavelength after each iteration). Cross validation then minimizes overfitting of the equation. In cross validation, the calibration samples are divided into groups. One group is reserved for validation and the remaining groups are used for calibration. This is repeated for the whole sample set and results in an estimate of the average prediction error of the sample set representing the population. 42 PLS regression is recommended for complex materials where unique spectral features are not readily identified or absorption bands are susceptible to matrix effects. PLS is sensitive to un-modelled sources of spectral variation and therefore, requires a larger calibration set (36). The occurrence of outliers is a common occurrence in NIRS work and the question is whether to include them or eliminate them from the calculations. An outlier is a sample that does not fit the calibration either spectrally or due to its reference value. A sample can be a t-statistic outlier if it's predicted result differs from the true result by at least three times the standard error of performance (SEP). A global \"H\" statistic outlier is a spectral outlier. Global \"H\" outliers have atypical spectra outside the boundaries relative to the calibration set but may still predict reasonably well even if they are out of the scope of the regression model. If outliers are free from prediction error, more samples similar to them should be included in the calibration set. The critical \"T\" outlier indicates that the sample analyzed has extreme analytical values relative to the calibration population and should be added to the calibration model (34). Some workers consider it acceptable to eliminate outliers from the calibration. This often results in a favourable decrease in the standard error of cross validation (SECV) but may also eliminate some of the variance in the calibration. The development of calibrations with all sources of variance represented and with all samples carefully prepared in the same way reduces the likelihood of outliers. There is no set rule for which scatter treatment, derivative, regression treatment, mathematical treatment, wavelength segment, outlier elimination, etc. is best for a calibration (25). Trial and error along with experience are the best guides. What is important is that the final prediction equation produces accurate results on unknown samples. Modern software and computers enable optimizing mathematical treatments and wavelength ranges quickly. A wide 43 range in mathematical treatments, scatter corrections and wavelength range optimization can be completed in a reasonable time depending on the regression method utilized. Once calibration equations are developed, they are retained in the instrument's computer. They are not usually printed because they are too cumbersome. For example, PLS equations involve weighted regression coefficients at every wavelength used in the calibration. If the wavelength segment from 800 to 2500 nm is used in a calibration, 850 wavelengths or data points are used in the equation. Coefficient size is directly related to the importance of that data point (wavelength) in explaining the variation in the reference value. Typically, the wavelength segment, derivative, mathematical treatment, regression method, and the regression statistics for the calibration and the validation sets are reported to indicate the settings used in the development of the equations and their associated error. It is also important to note that equations developed on a specific MRS instrument cannot be transferred to another MRS instrument unless the two are spectrally matched. Spectrally matched instruments produce identical optical data. Once similar instruments are matched, calibration constants developed on the first instrument should work on a second instrument, sometimes after a slope and bias adjustment. 2.4.5.4 Accuracy of NIRS Accuracy is defined as closeness to the true or correct value (40). The accuracy of MRS is defined as the agreement between the MRS predicted values and the laboratory reference values and is determined on the basis of samples not included in the calibration set. The accepted method of expressing accuracy in chemical analyses is to compare the analysis results with known samples. The most common way of expressing accuracy in M R analysis is by calculating the standard deviation of the differences between M R predicted values and the reference method values. 44 The accuracy of NIRS equations generally falls into three categories. The most accurate procedures measure single chemical entities such as crude protein and moisture with high precision, in most agricultural products. Slightly lower accuracy is afforded by laboratory procedures that do not measure a single chemical entity, such as fibre in forages. These methods involve several steps such as filtering, washing and drying residues, so the procedural errors are larger than those of nitrogen determinations. The lowest accuracy occurs when measuring constituents that occur in small quantities or that do not directly absorb NIR energy, such as minerals (36). Correlations for these constituents usually rely on secondary correlations that are associated with the parameter of interest. Included in this category are parameters such as digestibility and voluntary intake. Confidence in NIRS analysis requires an understanding of sampling errors and laboratory errors. One method of determining the error is to calculate the standard error of the laboratory (SEL). The SEL is the standard error of the difference between the blind duplicate measurements of the reference method values. SEL = / ^2 (L t - L 2 ) 2 where Lj = the first laboratory reference value n L 2 = the second laboratory reference value n = the number of samples In NIRS, accuracy is determined by the standard error of performance (SEP) or the standard error of cross validation (SECV) of NIRS results, compared to those of the laboratory reference method. In both cases, the lower the error, the more accurate the calibration. The SEP is defined as the standard deviation of the differences between the reference values and the NIR predicted values and it contains the combined error associated with the laboratory reference method, the NIR tests, sampling error, sample preparation error, and residual error (42). 45 where L = the laboratory reference value M = NJR predicted value n = the number of samples The SEP contains a systematic component known as the bias which is the difference between the mean laboratory measurement and the mean of the M R prediction. BIAS= L - M The SECV (standard error of cross-validation) statistic is an estimate of prediction accuracy from the calibration samples. It is obtained by predicting samples removed from the calibration set. Different samples are removed each time until all samples not having been included in the calibration have been predicted. The SEP may be as high as twice the magnitude of the SEL in successful M R calibrations. Ideally, the SEP or SECV and bias should be as low as possible and the lower it is, the better the calibration is performing. To determine if the predicted value is accurate enough for a constituent, the SECV for the equation is compared to the sampling error in the population the sample came from. The SECV should be lower than the sampling error in an accurate equation (36). The sampling error of silage is determined by collecting multiple samples from the face of a silo (eg. bunker silo) and analyzing each one for the constituent of interest. The mean, standard deviation and coefficient of variation can then be determined. The criteria for equation selection are a low SECV and a high R2 (multiple coefficient of determination) along with a low SEP and bias, a high r2 (coefficient of determination for linear regression), and a slope close to 1.0 for M R analysis of the validation set. The error between M R analysis of silages and the reference method is considered acceptable if it is less than one third of SEP = / E f L - M ) 2 n 46 the standard deviation of the reference set (42). The most important factors in gaining confidence in the NIR method are the accuracy, and day to day reproducibility of the analyses. In forages, the influence of growing year and its representation in the calibration set can be pronounced. When prediction equations prepared from samples collected during one calendar year were used to analyse samples collected during another year, correlation coefficients were generally lower and standard errors of analysis higher than when prediction equations were used within the same year's samples (6). Calibration equation upgrade by adding a limited number of samples from the new sample population to an existing calibration set is one procedure used to improve accuracy. Precision and stability of NIR analysis is often better than many reference methods. Over a 75 day period, a ground sealed soybean standard gave the values shown in Table 2.8. The greatest sources of error in any calibration are due to reference method error, non-homogeneity of the sample, nonrepresentative sampling in the calibration set, and poor sample preparation. Table 2.8. Precision of 64 Analyses of a Ground Sealed Soybean Meal Standard over a 75 Day Period a Dry Matter % Crude Protein % Crude Fibre % 4.62 5.14 4.88 0.09 1.844 Crude Fat % Minimum Maximum 90.59 90.96 90.78 0.06 0.066 51.73 52.59 52.16 0.14 0.268 2.27 2.40 2.34 0.02 0.855 Mean SDb CV° a Pro Form Feeds, Inc. 1995 b Standard deviation c Coefficient of variation = (standard deviation + mean) * 100 47 2.4.6 New Applications in Forage Analysis Using NIRS Currently, the analysis of fresh, unground forage material is under intense investigation and the accuracy of these calibrations is being evaluated. Nitrogen content is expected to be less precisely determined in silages than in hays due to protein degradation products, losses of volatiles, and sampling difficulties with fresh silages. At present, there are few reports of the use of NIRS to measure protein quality or protein degradation products. There is some promising work in the area of soluble and insoluble nitrogen in silages and soluble nitrogen and ammonia in intact, wet silages (23). ADAS (Agricultural Development and Advisory Service) in Great Britain has recently developed a method for predicting lactic acid and individual volatile fatty acids in fresh silage via NIR. This is one of the techniques that promises to improve forage analysis in the future (9). 2.5 Summary The ensiling of a forage is a complex fermentation process that produces many end products and quality variations within the silo. Factors affecting the efficiency of ensiling include species, maturity at harvest, DM content, WSC content, buffering capacity, climate, fertilization rate, bacterial content, pretreatment, silo air exclusion, speed of ensiling, storage facilities, additives, management practices, and others. Traditional chemical analysis of grass silage for quality determination is time-consuming, expensive, and the results are not necessarily well correlated to animal performance. Both the turn-around time and the cost of analysis can limit the number of samples analyzed from a silo as well as the practical usefulness of the data. Once the results are obtained, they may no longer relate to the forage being fed. NIRS offers the potential of fast, accurate silage analysis that can measure 48 many constituents within minutes, without the use of environmentally unfriendly chemicals. This will enable and encourage forage analysis at all stages of the ensiling process from when best to harvest, to the final, ensiled end product. It will also be possible to analyse feed-out silage samples from the silos more often, so that variations in quality can be identified and rations adjusted according to these changes. Although a sample's spectrum carries information about it's physical and chemical composition, MRS analysis cannot measure every constituent in the sample. While it is possible to calibrate and predict many of the nutritional variables in grass silage by MRS that have been addressed in this chapter, if it is not possible to determine them accurately by chemical methods, then calibration and prediction by MRS will not make the results more accurate. There is a need to develop suitable laboratory methods for determination of quality parameters in grass and grass silage in their fresh state and to secure their acceptance by the industry as reliable indications of nutritional value. In order to produce a successful calibration equation, the following steps must be strictly followed (42): 1. The calibration samples must adequately represent all variation in the population to be analyzed. 2. The laboratory reference values must be accurate. 3. A suitable regression technique must be chosen to extract the pertinent spectral information. 4. The selected equations must be validated with an independent, representative set of samples. Once calibration is completed, all of the error associated with the sampling, preparation, and reference methods will be included in the prediction error. However, the advantages of MRS predictions far outweigh the disadvantages. MRS prediction does not require hazardous chemicals or extensive sample preparation, which makes the analysis much safer and faster. The estimated 49 time for the simultaneous prediction of DM, CP, and ADF on an undried, unground silage sample is less than five minutes per sample. In addition to speed, the calibration for silage DM, CP, and ADF on undried, unground silage will reduce analytical error associated with the oven drying technique and provide forage quality predictions that are currently restricted by time per test and cost. Since the precision (reproducibility) of NTR analysis is often superior to that of the reference methods, MRS has the potential for improving the overall accuracy of analysis. The application of the MRS technique to feed analysis has only a 24 year history. In these 24 years we have learned about the absorption spectrum and its properties but still lack the ability to use the MRS absorption information directly without using a reference method. This will continue to be a goal for research in the nature. The objective of this thesis is to evaluate the accuracy and sources of error of MRS calibrations for the prediction of D M CP, and ADF on locally grown, intact, fresh grass and grass silage. 50 Chapter 3 NIRS Analysis of Intact Grass Silage and Fresh Grass for the Prediction of Dry Matter. Crude Protein and Acid Detergent Fibre 3.1 Introduction and Objective Forages are an important feed source for lactating dairy cows. Milk producers require forage nutritional information so that grain supplements can be proposed that balance the dairy cattle's diet at least cost. In order to optimize milk production and ensure herd health, silage analyses must be performed a number of times while feeding from a storage facility to compensate for variations in quality and DM within the silos. The cost of forage analysis on multiple samples from the same silos by traditional wet chemistry methods can be limiting. Near infrared reflectance spectroscopy is a rapid, non-destructive method of analysis that utilizes light instead of chemicals to analyse forage samples for nutritional information. The spectral data recorded by an NIR instrument contains information on the total physical and chemical properties of a sample (36). In 1988, the Association of Official Analytical Chemists (AOAC) approved the MRS method for the prediction of ADF and CP in dried, ground forages. Now, the evaluation of forages by MRS has become a major analytical method due to the increased demand for rapid analysis. The calibration equations currently used for forage in many of these laboratories, were derived from forages collected from different areas of the continent and from around the world. These equations are called \"global calibrations\" and they contain all the variance associated with the broad-based population. Calibration error increases as a population expands to include additional variance from new groups of samples. However, when samples containing variance not 51 represented in the calibration set are predicted, the results are not always accurate. Usually, a well defined local calibration is more accurate than a global calibration (36). The demand for same day forage analysis is increasing. Much of the literature on NIRS analysis of forages addresses dried, ground samples but very little research has addressed the analyses of undried, unground, high moisture samples (eg. fresh grass and silages). In addition to speed of analysis, the driving force for analyzing fresh samples are the errors caused by oven-drying samples to determine DM content (37). Other advantages of direct analysis include increased speed of analysis due to the elimination of drying and sample grinding. Conventional wet chemistry methods may require three to five days to complete DM, CP and ADF determinations on grass silage. Classical NIRS methodology requires at least two days for sample drying and preparation, while the analysis of undried, unground grass silages can be performed in minutes. Furthermore, the form in which it is analyzed is the same form as that fed to the dairy cows. The prediction of intact fresh grass and grass silage by MRS to determine the chemical composition and nutritional value quickly is an attractive prospect as an aid in formulating total diets for dairy cows. The objective of this study was to evaluate the accuracy and sources of error of MRS calibrations for the prediction of D M CP, and ADF on locally grown, intact, fresh grass and grass silages. 3.2 Materials and Methods Forage Samples Fresh grass and ensiled grass samples were collected from 138 dairy farms located in the Lower Fraser Valley in the southwestern corner of British Columbia, Canada. Multiple grab samples from a storage facility were collected and combined in plastic bags and brought to the 52 laboratory where they were mixed and presented in quadruplicate in a 4.5 x 5.7 x 21 cm coarse reflectance sample cell in their intact (undried, unground) form to a NTRSystems 6500 instrument (Perstorp Analytical, Silver Spring, MD) equipped with a sample transport device. The coarse reflectance sample cell holds approximately 40 to 80 grams of fresh forage which ensures that a representative sample is scanned from multiple points within the cell. The samples were not cut or changed in any way once they were received in the laboratory so that moisture losses were minimized. The samples were coarse, uneven in size, shape and composition and some of the samples were a mixture of leaves, stems and included different forage species. Spectral information was collected from 400 to 2500 nm at two nm intervals in the reflectance mode. Four separate loadings and readings were performed for each sample and the four spectra were averaged for each sample. The samples were chosen to include as much variance as possible in the growing area, and included two crop years - 1993 and 1994. The 1993 crop year was characterized by a warm, dry growing season while the 1994 crop year was cooler and wet. Five samples from the 1992 crop year were also included in the sample set. These samples had been ensiled for more than one year. Different harvests or cuts from each crop year were also included. These samples covered a range of plant maturities and included high quality as well as poor quality forages. Cutting dates were variable depending on weather conditions. Poor quality grass silages included poorly fermented samples, samples high in ADIN and fibre, and samples low in protein and DM. The sample set included 41 fresh grass samples and 251 ensiled samples to give a total of 292 samples. The characteristics of samples collected over the crop years and the number of samples collected from each harvest are listed in Table 3.1. Due to weather conditions, the majority of the 53 first cut harvested in the area is ensiled. After the first cutting, more of the harvested material is stored as hay resulting in less later-cut silages available, as indicated by their numbers on Table 3.1. Table 3.1. Number of Fresh Grass and Grass Silage Samples by Crop Year and Cut Number Crop Year Cut#l Cut #2 Cut #3 Cut #4 Cut #5 Cut #6 Total 1992 2 1 1 0 1 0 5 1993 51 39 23 14 13 0 140 1994 82 27 10 6 19 3 147 Total 135 67 34 20 33 3 292 Different grass species and mixtures were sampled and are listed in Table 3.2. Orchardgrass (Dactylis glomerata) and perennial ryegrass (Lolium perenne) combinations are common in the area, and the majority of the samples reflected this, but species separation was not attempted. Pure species of orchardgrass, ryegrass and tall fescue (Festuca arundinaced) were sampled as well. Also included were 17 first cut ryegrass varieties from a research variety trial conducted in 1994. These samples were scanned in the fresh, unwilted form, ensiled in research silos for 34 days, and 13 of the samples were scanned again after ensiling. Table 3.2. Grass Species Mix and Combinations as Seeded in the Sample Population Species # of Samples Orchardgrass / Perennial Ryegrass Mix 210 Perennial Ryegrass 34 Orchardgrass 33 Tall Fescue 3 Orchardgrass / Tall Fescue Grass Mix 6 Orchardgrass / Clover Mix 4 Miscellaneous 2 Total 292 54 Different additives were used on the farms to aid the ensiling process and the 251 ensiled samples included different additives as indicated in Table 3.3. Approximately 51.8% of the silage samples were ensiled without the aid of an ensiling additive. Table 3.3. Ensiling Additives Utilized in the Grass Silage Population # of Samples % of Samples No Additives 130 51.8 Enzymes and/or Microbial Inoculants 103 41.0 Acid (propionic) 18 7.2 Total 251 100.0 Storage facilities for the grass silages were varied (Table 3.4) with the majority of farmers utilizing bunker silos. Samples were obtained from other types of storage facilities as well, such as \"ag bags\", oxygen limiting silos, tower silos, research silos and round bales. Ag bags are long (approximately 10 metres) heavy, plastic tubes that are packed with silage. Research silos are 36 cm long by 10 cm in diameter PVC pipe fitted with rubber gaskets at each end that are used to simulate ensiling conditions in farm silos. Table 3.4. Grass Silage Storage Facilities # of Samples Fresh Grass 41 Bunker Silos 168 Ag Bags 26 Oxygen Limiting Silos 18 Tower Silos 16 Research Silos (34 days) 13 Round Bales (plastic wrapped) 9 Stack with Plastic Covering 1 Total 292 55 Additional sources of variation included in this sample set were assorted chop lengths (from 1.5 to 20 cm), various types of fertilizers, application rates and soil types, different personnel preparing the samples for analysis and performing the wet chemistries, and different sample temperatures at the time of scanning. All of these sources of variation were included in the calibration equations and contributed to the final error. Chemical Analysis After spectra were collected, the fresh forages were manually mixed and split into three subsamples. One portion was dried for 24 hours at 105° C until a constant weight was achieved for DM values. A second portion was dried at 60° C in a forced-air oven followed by grinding in a Wiley Mill to pass a 1-mm screen and was then used for chemical analyses. Crude protein (nitrogen x 6.25) was determined using the Kjeldahl procedure and ADF was determined according to the method of Van Soest (41). Residual moisture was removed from the ground samples by drying at 105° C in a forced-air oven until constant weight was achieved. All samples were analyzed in duplicate and the averaged results reported on a moisture-free basis. The third portion was frozen in plastic bags for volatile fatty acid and lactic acid determinations. The methods utilized to determine volatile fatty acids and lactic acid were developed by the Department of Animal Science at the University of Alberta and the procedures were as follows: Volatile fatty acids in 94 silages harvested in the 1994 crop year were determined after 5 g of unground, previously frozen silage were shaken for five hours with 35 mL of deionized water, 5 mL of internal standard, and 2 mL of 25% H 3P0 4 to bring the pH down to less than 4.0, in a 50 mL Teflon lined screw-capped tube. The internal standard was composed of 0.03 M isocaproic acid to determine recovery during the procedure. After shaking, the samples stood overnight and an aliquot of the supernatant was analyzed by gas chromatography using a 30 m x 0.25 mm inside 56 diameter Stabilwax-DA column. Percentages of acetic, propionic, isobutyric, butyric, isovaleric and valeric acid on a weight basis were determined by this method. It is recognized that volatiles also include alcohols and ammonia-nitrogen; however, they were not measured in this study. The lactic acid procedure was developed by Mamer and Gibbs (18) and modified by the Department of Animal Science at the University of Alberta. Lactic acid was determined on 96 forages harvested in the 1994 crop year utilizing 5 g of unground, previously frozen silage shaken for 16 hours with 40 mL of methanol containing 80 mg of malonic acid as an internal standard in a 50 mL screw-capped tube. After extraction, 0.5 mL of silage methanol extract was transferred to a 12 x 100 mm teflon lined screw-capped tube and 200 uL of 3 M NaOH was added. The samples were mixed and evaporated to dryness under nitrogen at 70° C to convert the lactic acid to a salt form. The samples were then esterified with 1 mL of 3 M methanol and HC1 (1.2 mL HC1 in 250 mL methanol) after mixing to dissolve the solids, followed by heating at 100° C for 25 minutes. The acidic solution was allowed to cool. Once the precipitate had settled, the supernatant plus 0.5 mL of methanol and 50 uL base (3 M NaOH in methanol) was transferred to a gas chromatography vial for a final solution pH of 4.0 to 5.0. Standards of lactic acid (2 mg/mL methanol) and malonic acid (2 mg/mL methanol) were carried through the same derivatizing procedure as the samples of silage extract to determine response factors. The samples were then analyzed by gas chromatography using a 30 m x 0.25 mm inside diameter Stabilwax-DA column. All steps and procedures in the sample preparation and wet chemistries were maintained as consistent as possible for both the calibration and validation set for the development of the calibration equations. 57 Statistical Analysis There are different approaches to developing calibration equations and each has similarities and differences. There is no one \"correct\" procedure (19,23,25,42). This study incorporated different scatter treatments, derivatives, mathematical treatments, wavelength segments, outlier elimination passes, and different calibration and validation set selection methods to generate calibration equations for each constituent. For this thesis, the data were analyzed utilizing established statistical procedures used in the ISI Software (InfraSoft International, Port Matilda, PA, USA) which was specifically developed for users of NIRSystems' instruments for agricultural products. The CENTER option within the ISI program establishes population boundaries with a maximum standardized \"H\" distance of 3.0 (35). The spectra were ranked according to their distance from the average spectra using principal component scores. The distances were standardized by dividing them by their average value, and called \"global FT values. This procedure defines the spectral boundaries of a population of spectra and identifies extreme spectra beyond a critical distance of three standard deviations away from the mean spectra. Examination of the spectral data revealed nine global \"FT outliers that were identified as being spectrally different from the rest of the population. Two of the outliers were heat damaged due to excessive heating during fermentation, three samples were products of a clostridial type of fermentation that resulted in high amounts of volatiles, three samples were excessively mature and dry, and the remaining sample did not have any readily identifiable differences. All nine outliers were orchardgrass/ryegrass combinations from different farms but were from varying crop years, cuts, and were ensiled in differing storage facilities with and without additives. Samples with extreme spectra have a large amount of influence during calibration. After careful consideration, 58 these nine global \"H\" outliers were eliminated from the set, leaving 283 samples for experimental work. The SYMMETRY option within the ISI Software provides a Chi-square evaluation of the position of each spectra in relation to each other in a spectral file (36). The graphic display plots each sample spectra in three-dimensional space within a cube. Spectra close to the average spectrum are near the middle of the plot and more extreme spectra are further away from the middle. Once a sample file is plotted, samples with global \"H\" values less than the critical value are marked with a \"+\", and those with global \"H\" values greater than the critical value are marked with a \"* \". When the sample population was viewed in the SYMMETRY option within the ISI Software and plotted with the first three principal component scores, there appeared to be a second population that included the fresh, unwilted ryegrass variety trial samples and their associated research silo samples, as shown in Figure 3.2. These samples were clustered together because they contain less variation than the rest of the population. These ryegrass samples were more similar to each other than to the rest of the population due to their low DM content ranging from 14 to 18%. Some of the samples that appeared between the two populations were fresh, wilted grass samples. The remainder of the fresh samples overlapped the main population of fermented grass. Samples with spectra close to the average spectra were near the middle of the scatter plot while spectra with extreme features were near the outer edge of the plot. The nine outliers are marked with a \"*\". 59 Figure 3.2. Fresh Grass and Grass Silage Population when Viewed in SYMMETRY. 3.3 Results and Discussion Chemical Composition of Fresh Grass and Grass Silage Samples The set of 283 samples varied widely in chemical composition and was characterized by the statistics listed in Table 3.5. The large DM range from 13.9 to 72.3 % resulted in a standard deviation of 13.04%. The range for CP corrected to a DM basis, CP(DM), was 9.5 to 27.8% and the range for ADF corrected to a DM basis, ADF(DM), was 22.6 to 46.6% which provided reasonably large standard deviations of 3.5 and 3.9%, respectively. All 283 samples were treated as one, broad-based, multi-species population for the purpose of this chapter. 60 Table 3.5. Fresh Grass and Grass Silage Samples Reference Method Statistics Constituent n a Mean SDb Range Dry Matter 283 34.66 % 13.04 13.9 -72.3 CP (DM)C 283 17.11 3.51 9.5 -27.8 CP (as rec)d 283 6.02 2.71 1.6 - 14.6 ADF (DM) 283 34.33 3.90 22.6 -46.6 ADF (as rec) 283 11.84 4.58 4.2 -26.7 Lactic Acid 96 1.37 0.90 0.11 -4.11 Acetic Acid 94 0.50 0.38 0.13 - 1.20 Propionic Acid 94 0.040 0.061 0.001 - 0.097 Isobutyric Acid 94 0.009 0.018 0.001 - 0.080 Butyric Acid 94 0.072 0.168 0.001 - 0.770 Isovaleric Acid 94 0.010 0.027 0.001 -0.170 Valeric Acid 94 0.007 0.018 0.001 -0.130 a Number of samples b Standard deviation c Dry matter corrected values d Value expressed at the dry matter content of the forage when received in the laboratory. Short-Chain Organic Acids Lactic acid and acetic acid were the predominant organic acids found in the samples. Propionic, isobutyric, butyric, isovaleric, and valeric acids were present in much lower quantities; often in amounts below the detection limit of 0.001% of the reference method. The chemical composition statistics for the measured volatiles are summarized in Table 3.5. A total of 96 samples were analyzed for lactic acid and 94 samples for the volatiles (acetic, propionic, isobutyric, butyric, isovaleric, and valeric acids). Total volatiles were not estimated. Chemical analyses of the short-chain fatty acids were conducted on extracts of the previously frozen, undried, unground samples. This presented a problem for obtaining a representative sample for the reference method. This was indicated by poor precision between 61 duplicates. If the sample could have been mixed adequately before sampling, the agreement between duplicates may have improved. Population Structuring The calibration set was selected utilizing the \"neighbourhood H\" selection method which is based on the assumption that only one sample is required to represent all samples in a neighbourhood. Distances between all spectra are computed and the spectrum with the most close neighbours (spectrally similar samples) is retained in the calibration set. The close neighbours comprise the validation set. This procedure is repeated with all the remaining spectra until no spectra are closer than the critical distance. The critical distance between spectra can be varied to keep a desired number of samples in the calibration set. Complex materials, such as grass silage, contain unique spectral features that are not readily identified and absorption bands that are susceptible to matrix effects (30). Due to this complexity, the desired minimum number of samples to develop a calibration of this nature was 200. With a neighbourhood \"H\" of 0.25, 216 spectrally different samples were selected for the calibration set and the remaining 67 spectral duplicates were assigned to the validation set. The reference method statistics for both the calibration and validation sets are displayed in Table 3.6. Table 3.6. Fresh Grass and Grass Silage Reference Method Statistics for Dry Matter, Crude Protein(DM), and ADF(DM) on the Calibration and Validation Sets. Calibration Set Validation Set Constituent na Mean SDb Range n\" Mean SDb Range (%) (%) 216 36.95 13.04 14.6-72.3 67 27.31 10.05 13.9 - 58.0 216 17.57 3.67 9.5 -27.8 67 15.66 2.48 11.2 -22.2 216 34.22 4.14 22.6-46.6 67 34.70 3.00 28.8 -43.2 a Number of samples b Standard deviation Dry Matter CP (DM) ADF (DM) 62 Prediction equations were developed from the calibration samples utilizing modified PLS regression with four cross validation groups and using the various scatter treatment options available in the ISI program for particle size correction: no scatter correction, SNV (standard normal variate) and detrend, SNV, detrend, and normal multiplicative scatter correction. For each constituent, the scatter correction treatment producing the lowest error was used to continue with equation development. Different derivatives, mathematical treatments, wavelength segments and outlier elimination passes were then applied to optimize the equations. The three wavelength segments tested included 800 to 2498 nm, 1106 to 2498 nm, and 800 to 1850 nm using four or eight nm between wavelengths. A minimum of 40 equations were developed for each constituent. Outlier elimination passes were utilized due to the difficulty of obtaining good agreement between the spectral information from the intact sample and the laboratory reference values. A sample is a t-statistic outlier (high residual) if its predicted result differs from the reference method result by at least three times the SEP. Some samples are outliers due to errors either in the reference method analysis or in the spectra. Unless outliers are verified to be free from errors, it is considered safer to eliminate them (36). If outliers are free from error (ie: no mistake during spectra collection, reference method analysis, population assignment, etc.), more samples like them should be included in the calibration set, or the outliers should be split off to form a separate calibration. Elimination passes during regression analysis eliminated samples with high residuals between laboratory reference values and predicted values, and calibration equations were developed without their presence. The calibration equation statistics that were obtained before and after outlier removal are summarized in Table 3.7. In all cases, the elimination of t-statistic outliers increased the coefficient 63 of determination (R2) and decreased the SECV. The SECV is defined as an estimation of prediction accuracy from the calibration samples and is expressed in the same units as the constituent is measured. All developed calibration equations were then used to predict the nutrient content of the validation set. The results from these predictions were used to evaluate the accuracy of the calibration equations. All samples in the validation set were independent of the calibration set. Since the equations were developed with outlier elimination, the t-statistic outliers from the validation sets were also eliminated. The eliminated samples included different sources of variations with no one source consistently present. Table 3.7. Calibration Equation Statistics Before and After Outlier Removal All Samples After Outlier Removal Constituent n R 2 SECV 3 n R 2 SECV % % Dry Matter 216 0.99 1.58 204 1.00 1.15 CP (DM) 216 0.94 1.19 212 0.95 1.05 CP (as rec) 216 0.98 0.45 206 0.99 0.34 ADF (DM) 216 0.93 1.38 210 0.95 1.16 ADF (as rec) 216 0.98 0.76 206 0.99 0.61 Lactic Acid 68 0.48 0.74 64 0.62 0.50 Acetic Acid 67 0.78 0.26 64 0.91 0.19 Propionic Acid 67 0.60 0.05 61 0.84 0.017 3 Standard error of cross validation: an estimation of prediction accuracy from the calibration samples. The factors considered in the final equation selection were a low SECV and a high R2 in the calibration set; and a low SEP, a high r2, a low bias, and a slope close to 1.0 when the validation set was predicted. Another important consideration in equation selection was to retain as many samples in the calibration and validation sets with a minimum of outlier elimination. The 64 elimination of too many samples will remove important sources of variation needed for the accurate prediction of unknown samples. This becomes apparent when the equation with the lowest SECV does not perform well and produces a large SEP on an independent test set. The spectra data pretreatment methods and the statistics on the calibration equations considered to have performed the most accurately on the validation set are summarized in Table 3.8. In the case of all three constituents, the use of a scatter correction pretreatment improved the predictive ability of the calibrations by improving the SEP over calibrations with no scatter correction pretreatment. Also, these equations did not have the lowest SECVs, however, they performed the best when applied to the independent validation set. Table 3.8. Calibration and Validation Statistics for the Best Performing Fresh Grass and Grass Silage Equations Developed on Calibration and Validation Sets Selected by the \"Neighbourhood H\" Method. Constituent Math\" Scatterb Calibration Set Validation Set nc R 2 SECVd ne SEPf r2 Bias Slope (%) (%) (%) Dry Matter 1,4,4 SNV&Detrend 204 0.995 1.15 66 0.73 0.99 -0.02 1.00 CP(DM) 1,4,4 Detrend 212 0.948 1.05 66 0.79 0.89 -0.04 0.99 ADF(DM) 2,5,5 Detrend 210 0.951 1.16 61 0.95 0.89 -0.08 1.00 a Mathematical treatment as derivative, segment gap, and segment length for smoothing b Different mathematical procedures for particle size variation removal 0 Number of calibration samples retained for equation development d Standard error of cross validation (four groups) on samples selected for the calibration set e Number of samples retained in the validation set f Standard error of performance (SEP) is the standard deviation of the differences between the reference method values and the predicted values. 65 Major Constituent Results Dry Matter The size of the SEP value is influenced by the analytical method and what the technique actually measures. In the case of DM, the oven drying reference method drives off the volatiles as well as water, resulting in under-estimation in DM predictions in samples containing high amounts of volatiles (36). This increases the SEP values for DM and continues to introduce error when other constituent analysis values are converted to a moisture-free basis. Y = 1.00X + 0.04 Reference Method Dry Matter (%) Figure 3.3. Linear Regression Plot of the Reference Method and Predicted Dry Matter of Fresh Grass and Grass Silage. 66 The calibration and validation statistics for the DM equation are presented in Table 3.8. The R 2 and r2 on the calibration and validation set, respectively, are very high and are associated with a low SEP and bias. The linear regression correlation plot of the reference method and NIRS predictions is shown in Figure 3.3. The sample with the highest residual was a sample with a difference of -1.90% dry matter between the reference method and the predicted results. Crude Protein The reference method for CP measures elemental nitrogen while NIRS predicts N-H bonds in protein and peptide molecules. This discrepancy means that the two methods are measuring different components in the sample which causes SEP values for CP to be elevated. This increase in error is apparent in the SEP value of 0.79 as indicated for CP(DM) in Table 3.8. Reeves et al. (27) reported a SEP value of 0.61 on dried, ground silages with both a R 2 and r2 of 0.99. The coefficients of determination (R2 and r2) on the undried, unground samples used in this study were 0.95 and 0.89, respectively. The large, irregular particle sizes in the samples may also contribute to this error. The linear regression plot of the reference method and NIRS predictions is shown in Figure 3.4. The sample with the largest protein residual was a sample with a difference between the reference method and the predicted results of 2.22% CP. This was the same sample with the highest DM residual. Of the 66 samples in the validation set, 52 samples had residual values less than 1.0% CP. 67 Y = 0.99X + 0.09 Reference Method Crude Protein (DM) (%) Figure 3.4. Linear Regression Plot of the Reference Method and Predicted Crude Protein Corrected to a Dry Matter Basis CP(DM) of Fresh Grass and Grass Silage. Acid Detergent Fibre The ADF reference method measures many compounds that contribute to the ADF value. These compounds absorb NIR energy differently and result in an increase in SEP values for ADF. When the ADF(DM) errors in Table 3.8 are examined, the SECV and SEP are both higher than those from the D M and CP(DM) calibrations. The errors are, however, lower than those reported by Reeves et al. (27) for undried, ground silage and on dried ground silage. The R 2 and r2 are the same as those for CP(DM) and may be due to the coarse nature of the sample. 68 Y - 1 .OOX - 0.19 Reference Method ADF(DM) (%) Figure 3.5. Linear Regression Plot of the Reference Method and Predicted Acid Detergent Fibre Corrected to a Dry Matter Basis ADF(DM) of Fresh Grass and Grass Silage. The linear regression correlation plot of the ADF reference method and NIRS predictions is shown in Figure 3.5. The sample with the highest residual was again, the same sample as for DM and CP with a difference between the reference method and the predicted results of -2.40% ADF(DM). Since the same sample produced a large residual for all three equations, the portion of the sample analyzed by the reference method may not have been exactly the same as the portion scanned. This may be a result of particle size variations and irregular cell packing. 69 In a successful NIR calibration, the SEP is often twice the magnitude of the standard error of the reference method (SEL) for a particular method (23). The SEL for DM, CP, and ADF were approximately 0.35, 0.35, and 0.57, respectively, while the SEP values of 0.73, 0.79 and 0.95, as reported in Table 3.8, indicate successful calibrations. Short Chain Organic Acids Calibration statistics for lactic, acetic, and propionic acid are shown in Table 3.9. The prediction equations developed for isobutyric, butyric, isovaleric, and valeric acids were unsuccessful partially due to the small amounts present in the samples, the small range associated with the sample set, and the high error associated with the reference method sampling procedure. High t-statistic outliers eliminated from the calibration set contained concentrations of volatile fatty acids at the high end of the range. There were few samples representing the mid range. This resulted in a lower standard deviation and R 2 than was desirable for a successful calibration. Table 3.9. Performance Statistics for Prediction of Lactic, Acetic, and Propionic Acid in Fresh Grass and Grass Silages. Calibration Set Validation Set Constituent Math3 Scatterb nc R 2 SECVd ne SEPf r2 Bias Slope Lactic Acid Acetic Acid Propionic Acid 1.4.4 3.5.5 3,5,5 SNV&Detrend SNV&Detrend None 64 64 61 0.62 0.91 0.84 % 0.50 0.194 0.017 24 23 22 % 0.48 0.14 0.02 0.52 0.73 0.63 0.14 1.24 0.05 1.13 0.00 1.12 3 Mathematical treatment as derivative, segment gap in nm, and segment length in nm for smoothing b Different mathematical procedures for particle size variation removal 0 Number of calibration samples retained for equation development d Standard error of cross validation (four groups) on samples selected for the calibration set. e Number of samples retained in the validation set f Standard error of performance (SEP) is the standard deviation of the differences between the reference method values and the predicted values. 70 Table 3.10. Best \"As Received\" Equation and Validation Set Statistics for Crude Protein and ADF in Intact Fresh Grass and Grass Silage. Calibration Set Validation Set Constituent Math\" Scatterb nc R 2 SECVd ne SEPf r2 Bias Slope % % % CP(asrec) 1,4,4 SNV&Detrend 206 0.989 0.342 67 0.27 0.98 -0.06 1.02 ADF(asrec) 1,4,4 SNV&Detrend 206 0.988 0.608 65 0.39 0.99 -0.10 1.01 a Mathematical treatment as derivative, segment gap, and segment length for smoothing b Different mathematical procedures for particle size variation removal 0 Number of calibration samples retained for equation development d Standard error of cross validation (four groups) on samples selected for the calibration set. e Number of samples retained in the validation set f Standard error of performance (SEP) is the standard deviation of the differences between the reference method values and the predicted values. Equations Generated from \"As Received\" Reference Values Prediction equations were developed from the sample population on \"as received\" (as rec) values for CP (as rec) and ADF (as rec). The best prediction equation results are summarized in Table 3.10. An equation is considered acceptable if the RPD (standard deviation divided by the SEP) is greater than three, and ideally the ratio should be ten or higher (42). This comparison is referred to as the ratio of the standard error of performance to the standard deviation, and all five equations from the major constituents met this criterion, as shown in Table 3.11. The DM equation had a high RPD of 13.79. The RPD values for the \"as received\" equations indicated improvement in accuracy over the moisture-free equations. The RPD of the CP(DM) prediction equation was 3.08 and rose to 7.11 when the equation was developed from \"as received\" laboratory values. The ADF(DM) prediction equation resulted in a RPD of 3.07 and increased to 9.95 for the \"as received\" equation. The low RPD values for lactic, acetic, and propionic acid indicate that the 71 equations were not performing accurately because the SEP was not appreciably lower than the standard deviation. When the SEP value is similar to the standard deviation of the test set, the MRS is not predicting any better than the standard deviation of the original set. Table 3.11. Ratio of the Standard Error of Performance to the Standard Deviation (RPD) and the Coefficient of Variability (CV) Constituent SEP3 SDb Mean RPDC CV d Dry Matter 0.73 10.07 27.18 13.79 2.69 CP (DM) 0.79 2.43 15.72 3.08 5.03 CP (as rec) 0.27 1.92 4.37 7.11 6.18 ADF (DM) 0.95 2.92 34.53 3.07 2.75 ADF (as rec) 0.39 3.88 9.49 9.95 4.11 Lactic Acid 0.48 0.66 1.39 1.37 34.53 Acetic Acid 0.14 0.25 0.45 1.79 31.11 Propionic Acid 0.02 0.03 0.02 1.50 100 CP(as rec) back corrected6 0.75 2.43 15.98 3.24 4.69 ADF(as rec) back corrected 1.14 2.67 35.10 2.34 3.25 3 Standard error of prediction of the validation set b Standard deviation of the reference method results in the validation set 0 Standard deviation standard error of prediction (SEP) of the validation set d Coefficient of variability = (SEP * 100) + mean of validation set 6 Back corrected = predicted (as rec) results calculated to a DM basis and compared to the reference method results. When the RPD values are compared, the CP (as rec) and ADF (as rec) prediction equations appear to be more accurate than the CP(DM) and ADF(DM) prediction equations. However, when the CP (as rec) and ADF (as rec) prediction values were converted to a moisture-free basis using the reference method DM values, and these results compared to the error associated with the CP(DM) and ADF(DM) prediction calibrations, the advantage was not so pronounced. The SEP for CP(as rec) rose from 0.27 to 0.75 and the bias increased to 0.24, but, the r2 increased from 0.87 to 0.91 and only ten samples out of 67 predictions had CP residuals greater than 1.00% compared 72 to 14 in the CP(DM) calibration. The SEP and bias for ADF (as rec) increased to 1.14 and 0.29, respectively, when back corrected to a DM basis. This may indicate that the prediction equations developed from \"as received\" reference values produced a larger error when the results were converted to a moisture free basis. The coefficients of variability for the CP(DM) and ADF(DM) prediction equations are lower than for CP (as rec) and ADF (as rec) prediction equations which indicates the equations developed from constituents corrected to a DM basis, may be more accurate. More research is needed to determine the effect of predictions on a fresh versus dry basis on the different constituents in forages. The moisture content is critical. Comparison of the Accuracy of Intact Forage and Dry. Ground Forage Calibrations A comparison of performance statistics summarized from published papers on forages from three publications (24,27,33) is summarized in Table 3.12. In most cases, the error is higher and the RPD is lower for the prediction of CP and ADF in undried, unground forage when compared to dried, ground forage. For calibrations where SECV is reported without a SEP, the ratio of the SECV to standard deviation of the original data is displayed. The calibration equations developed for this thesis indicate that DM and ADF(DM) in particular can be predicted with similar or improved accuracy to prediction equations developed from dry, ground forage. The prediction of CP appeared to be less accurate in undried, unground silage and this may have been a result of the high moisture content, reference method, and the high sampling error associated with wet silage. High moisture content causes shifts in the absorption bands depending on the compounds in solution and the amount of moisture in the sample (26) and 88% of the samples in this study had moisture contents greater than 50%. Also, error is introduced into the 73 Table 3.12. Comparison of Performance Statistics for DM, CP(DM), and ADF(DM) Predictions in Dry, Ground Forage Versus Intact Forage. Dry, Ground Forage Calibration Set Constituent R 2 SECV Validation Set r2 SEP\" Slope SDC RPDd Forage Reference % % % DM 0.84 0.76 1.36 1.79 Haylage (33) DM 0.68 0.58 3.34 5.76 Pasture Grass (33) CP(DM) 0.98 1.07 5.98 5.59 Mixed Silages (24) CP(DM) 0.99 0.99 0.61 1.00 2.7 4.43 Mixed Silages (27) CP(DM) 0.97 0.71 3.37 4.75 Haylage (33) CP(DM) 0.97 1.41 7.73 5.48 Pasture Grass (33) ADF(DM) 0.92 2.5 5.60 2.24 Mixed Silages (24) ADF(DM) 0.96 0.96 1.48 1.00 5.4 3.65 Mixed Silages (27) ADF (DM) 0.90 1.63 2.12 1.30 Haylage (33) ADF(DM) 0.95 1.96 7.75 3.95 Pasture Grass (33) Intact Silage Calibration Set Validation Set Constituent R 2 SECV r2 SEP Slope SD RPD Forage Reference % % % DM 0.96 0.96 2.22 1.01 12.3 5.54 Mixed Silages (27) DM 0.99 1.20 12.73 10.61 Haylage (33) DM 0.99 0.73 8.12 11.12 Pasture Grass (33) DM 1.00 1.15 0.99 0.73 1.00 10.1 13.84 Grass Silage this thesis CP(DM) 0.97 0.96 1.09 0.99 2.8 2.57 Mixed Silages (27) CP(DM) 0.91 1.16 3.37 2.91 Haylage (33) CP(DM) 0.95 1.93 7.73 4.01 Pasture Grass (33) CP(DM) 0.95 1.05 0.89 0.79 0.99 2.43 3.08 Grass Silage this thesis ADF(DM) 0.86 0.83 3.13 1.04 5.5 1.76 Mixed Silages (27) ADF(DM) 0.83 2.12 2.12 1.00 Haylage (33) ADF(DM) 0.86 2.49 7.75 3.11 Pasture Grass (33) ADF(DM) 0.95 1.16 0.89 0.95 1.00 2.92 3.07 Grass Silage this thesis a Standard error of cross validation b Standard error of performance is the standard deviation of the differences between the reference method values and the predicted values. 0 Standard deviation d Standard deviation of the reference values from the validation set + SEP or the standard deviation of the reference values from the calibration set SECV 74 calibration by drying the samples prior to analysis. The spectra were collected from the intact silage and the reference data was determined on dried, ground forage after the ammonia and volatile fatty acids had volatilized. The Effect of Calibration Set Selection Method The effect of different combinations of samples in the calibration and validation set was compared by selecting samples by three methods. The first method involved ordering the sample set by global \"FT' to establish the spectral boundaries of the population, followed by the selection of every fifth sample for the validation set. The second method entailed ordering the samples from lowest to highest constituent value after reference method analysis and selecting every fifth sample for the validation set. Both methods resulted in calibration and validation sets containing 226 and 57 samples respectively. The third method of calibration set selection was by \"neighbourhood FT'. A neighbourhood \"EC\" of 0.25 was chosen to ensure at least 200 samples were selected for the calibration set. This selection method resulted in 216 samples being selected for calibration development; each sample representing a unique spectral neighbourhood. The remaining 67 samples were assigned to the validation set. This final method of calibration set selection allows for the inclusion of additional samples with different spectral features for future equation expansion. Prediction equations were developed from the calibration sample sets for each constituent utilizing different scatter correction treatments, wavelength segments, mathematical treatments and outlier elimination passes as described under Structuring of the Population and Calibration in Chapter 2. Fifty equations for each constituent were developed and all the equations were then used to predict the validation set. Equations developed for the prediction of isobutyric, butyric, 75 isovaleric and valeric acids were not successful and are not reported. Insufficient samples representing the middle and high range of these constituents resulted in an unacceptable relationship that was heavily influenced by the highest sample value retained in the regression. The majority of the samples contained measured amounts less than or close to the detection limits of the analysis method of these acids. The t-statistic outliers eliminated during equation development were often the samples containing the highest amounts of volatiles. Addition of more samples in the mid to high ranges may improve the accuracy of the equations in the future. The comparison of the best performing equations between calibration method selections are reported in Table 3.13. Crude protein and ADF prediction equations developed on \"as received\" results are also given. In all cases, the validation set r2 was equal to or lower than the calibration set R2. Based on the given statistics, the equations for DM developed from the Global \"FT and constituent value method of calibration set selection appeared to have produced the most accurate prediction equations when the SECV is compared. However, the SEP for the neighbourhood \"H\" method was the lowest of the three methods and the number of samples retained in the validation set was the highest. The elimination of fewer samples in the validation set indicates that more samples fit the developed equation. The prediction equations for CP(DM) were more varied. The neighbourhood \"H\" method eliminated the fewest outliers from the calibration and validation sets but had a higher error than the constituents value method, as indicated by the SECV and SEP. The constituent value method of calibration set selection produced an equation for ADF(DM) with the lowest SECV, however, the neighbourhood \"H\" method produced a higher SECV with a much lower SEP of 0.95 and a slope of 1.00. 76 C O C O « CD > P H w C/a T 3 O O X Ul 3 O '53 fc 0) o n \"3 U > o c / a CD C/a > C/3 c CD 3 t i c o O CD c/a O > O W c / a CD C/a \"3 > c / a ed X ) o 5 c / a \"ed O > u c / a G CD 3 co G O U O N as O N 00 O0 O N O N 00 O N O N cs i n CO CO V O b b b b b © b b co O N CN m O N O N co 00 cs 0 b b © b b b b b 0 0 m O N O N O N i n O N as O N cs vo O N 00 b © b © b © b m O co V O V O 0 m O N cs 0 b b b b 0 \" 0 CS n O N O N O N V O V O 00 00 O N 00 © b © © b © © 0 O N O co co 0 CO i n 0 i n O N cs 0 — 1 ^ b b © b b m O CN CS 00 cs O N CS CO cs V O 00 vo V O p R PL, -13 * * 9 8 8 u u < < j o •a o o CD CO G O CD CO G O .\"2 > CD X 32 ,0 * CD •£3 J D \"a3 CO \"H. £ CO > CO > 3 - O c u t5 *e3 CO CD f CO c t t c u > c u c4 O -a CU O H s c / a g a ' > > O H 3 o o CO U CO CU CO CU u. O H CD CD t-i 1 CD l -(D CO CD t c u c / a c u 3 > •s 3 O U O H > *e3 -a > O CD G O 'I ed > \"3 l-H +-» o c u O H CO >> -o o J H CO CO CD CD CD CO CD u CD VH CD ^ CO CD Id O \"o, 3 -O ~3 o CD I \" o o u. 3 o £P \"53 G O O X I l-H 3 O S H -•s « V ' O H /-I CO ed G O UH *3 o CD X ! >> o ed l-H 3 o o ed G O o CD l-H O H I o CD 3 CO ed CD B . G CD X H-» T 3 3 CO CD Ul T 3 O X S \"co Ji o O H *H b £ cd n CO u te S • G c te -3 ed C G CD ed X +^ H 5 / 3 CD X H-» o G O • > CD ed T 3 G ed CD X H CD O G cd CD O H C+H O Ui O fc CD T3 u ed T3 u c / a CO CD 3 \"3 > CD u X ca CD BP CD + J CD -o O ed T 3 C cd G '53 +-> o U O H CD a o T3 CD > T 3 CD CD O CD u CO < T 3 CD G — £ CD +e ' C D - K CD H-> O fc 77 It is not always obvious which equation will predict most accurately. For example, the SECVs for CP(as rec) were similar for all three calibration sample set selection methods, but, the SEP was the lowest without producing any outliers in the validation set with the neighbourhood \"H\" method. The ADF(as rec) prediction equation developed from samples selected with the neighbourhood \"FT method, had a larger SECV than the other methods, however, the SEP was the lowest with the fewest outliers eliminated from the validation set. The SECV is a measure of accuracy on the calibration set, but does not imply how the equation would perform on samples independent of the calibration set. The selection of a calibration and validation sample sets by global \"FT produced the best lactic acid and propionic equations from this sample set when compared to the other two methods of calibration set selection. Both equations resulted in better R2, r2, and SEP values. The acetic acid calibration performed slightly better when the calibration and validation sets were chosen by the neighbourhood \"FT method. Calibration equations were also developed on calibration sample sets selected with a wider neighbourhood \"FT value of 0.60. All generated prediction equations for all constituents had a higher error than the equations developed utilizing a neighbourhood \"H\" value of 0.25. The number of spectrally different samples selected for the calibration set dropped to 120. Prediction equations were also generated after the 31 fresh ryegrass and research silo samples were eliminated from the population. The elimination of these samples did not result in prediction equations with improved accuracy. 78 3.4 Conclusions This study established that the accuracy associated with NIRS prediction equations developed from intact fresh grass and grass silages for DM, CP(DM), and ADF(DM) determinations is similar or slightly lower than for calibration equations developed from dried ground samples when sample data is collected and analyzed using the methods described in this thesis. With the addition of more variation from future samples of fresh grass and grass silages, including more samples with undesirable fermentation end products, these equations will be a sound base to build future, expandable, stable (robust) calibrations that will accurately predict nutrient content in fresh grass and grass silage samples from the local agricultural area. Equations developed from \"as received\" reference values appeared to be more accurate than equations developed on DM corrected values for the prediction of nutritional content of intact fresh grass and grass silages, however, more studies are required using different forage calibration sets to determine if this is a valid observation. This study involved the collection of absorption information from undried, unground forages, but the wet chemistry methods for CP and ADF were still performed on dried, ground samples. Future research needs to address methods that incorporate sampling procedures, particle size reduction of fresh samples, and techniques that enable the provision of reference data that can be used on fresh samples that are better correlated with NIRS absorption bands. This may reduce analysis errors further by eliminating the need for drying and grinding before analysis of the reference samples. The error associated with equation development for volatile fermentation end-products could be reduced if the prediction equations were developed from a sample population with a more even compositional distribution over the measured range. The end products of a desirable type of fermentation are different in composition than from an undesirable type of fermentation and it is 79 difficult to obtain samples representing the mid range of these constituents. Future research in this area may be more successful if samples containing high amount of volatile fatty acids are mixed with samples containing low amounts of volatiles to produce samples containing volatiles in the mid range. In retrospect, accuracy could have been improved by grinding the samples in liquid nitrogen prior to analysis to ensure a representative subsample for analysis. The method of calibration and validation set selection can produce equations with variable accuracies within the same population sample set. The best calibration and validation set selection method appears to be dependent on the individual sample constituents and sample set, and therefore, must be assessed for each individual population. The prediction equations display acceptable accuracy for limited applications and their usefulness will largely depend on the accuracy required. Farmers need to know the nutrient content of their forages so balanced diets can be proposed and these prediction equations can do the job quickly. If speed of analysis is more important then absolute accuracy, then these prediction equations can provide a good estimate of DM, CP, and ADF in fresh grass and grass silages within minutes rather than days. This will enable farmers to monitor the forage quality in their silos more frequently which will aid in effective total diet formulations at considerable savings in time and cost. 80 Chapter 4 Summary In summary, this thesis describes the accuracy of NIRS prediction equations that can be obtained from fresh pasture grass and grass silage using this particular combination of sampling procedure, scanning and packing method, reference analyses methods, calibration set selection method, data treatment and regression analysis. No single mathematical treatment, wavelength segment, or scatter correction treatment consistently produced the best performing prediction equation. The method of calibration set selection also produced variable results when compared within the same sample set. The DM prediction equation error may be marginally high due to spectral alterations caused by the presence of water. Studies by Reeves (26) have shown shifts in the absorption bands depending on the compounds in solution and the amount of moisture in the sample and 88% of the samples in this study had moisture contents greater than 50%. The error may possibly be lowered for the DM equation by using the Karl Fischer method which involves direct titration of water with a water-specific titrant, or the toluene distillation method for true moisture determination as the reference method so that volatile acids and alcohols are not lost as in oven drying (15,43). However, the accuracy of the DM prediction equation is sufficient for many applications as it stands. The CP and ADF calibration equations showed considerable promise for most ration balancing applications. One way to reduce the error may be to reduce the chop length of the fresh hay and silage samples to an approximate maximum length of two cm. Some of the samples had visible air pockets when packed in the natural products cell, which indicated variations in packing 81 densities. The sample presented to the instrument may not have been representative of the sample submitted for reference method analysis. This is an on-going problem with the establishment of calibration equations. In the future, other directions that could be taken with this sample set, could include segregation of the different plant species, separation of the ensiled samples from the fresh grass, and/or segregation of poorly fermented samples into separate calibration equations. Some studies have indicated an increase in accuracy when the forage species were separated. From a practical standpoint, this may not be useful because samples are seldom accurately identified by species when sent to the laboratory for nutrient analysis. Pure species from research studies only could be predicted with calibration equations developed from pure species. The prediction equations developed for this thesis were derived from locally grown forages and contain only the variance found in this agricultural area. They were developed from fresh, intact samples that were analyzed immediately upon receipt. The prediction equations reflect the true quality of the forage being fed, with minimum degradation or nutrient quality changes before analysis. These prediction equations will be used as a base for future calibration equations that will be utilized to predict forage nutrient value for dairy farmers in the local agricultural area. The dairy farmers will then be able to react quickly to variations in the forage DM content and nutritional constituents as it is fed from the silo and, thus, benefit from information obtained from the forages they are currently feeding. 82 Bibliography 1. Abrams, S. M., J. S. Shenk, and H. W. Harpster. 1988. Potential of near infrared reflectance spectroscopy for analysis of silage composition. J. Dairy Sci. 71:1955-1959. 2. Alberta Agriculture. 1988. Silage manual. Print Media Branch, Edmonton, Alberta. AGDEX 120/52-2. 3. Albrecht, K. A., G. C. Marten, J. L. Halgerson, and W. F. Wedin. 1987. Analyses of cell-wall carbohydrates and starch in alfalfa by near infrared reflectance spectroscopy. Crop Sci. 27:586-588. 4. Barton, II, F. E. 1991. New Methods for the structural and compositional analysis of cell walls for quality determinations. Animal Feed Sci. and Tech. 32:1-11. 5. Berger, L .L., G. C. Fahey Jr., L.D. Bourquin, and E. C. Titgemeyer. 1994. Modification of forage quality after harvest. Proc. Natl. Conf. on Forage Quality, Evaluation and Utilization. Univ. of NE. Lincoln, NE. Editor-in-Chief: G. C. Fahey, Jr. 922-966. 6. Blosser, T. H , J. B. Reeves, III, and J. Bond. 1988. Factors affecting analysis of the chemical composition of tall fescue with near infrared reflectance spectroscopy. J. Dairy Sci. 71:398-408. 7. Bolsen, K. and J. I. Heidker. 1985. Silage additives USA. Chalcombe Publications, Manhattan, Kansas, USA. 8. Crawshaw, R. 1992. Evaluation of silage. Forward with Grass Into Europe: Proc. of the British Grassland Society Winter Meeting. 97-108. 9. Givens, D. I. 1993. Evaluating energy and protein in grass and grass silage. Grass Farmer, Autumn 1993, No. 45. British Grassland Society, 26-27. 10. Givens, D. I., A. R. Moss, and A. H. Adamson. 1993. The digestibility and energy value of badly preserved grass silages. Animal Feed Sci. and Tech. 42:97-107. 11. Goering, H. K. and P. J. Van Soest. 1970. Forage fiber analysis (apparatus, reagents, procedures, and some applications). Agricultural Handbook No. 379, ARS, USD A, U.S. Government Printing Office, Washington, DC. 12. Haigh, P. M. and J. R Hopkins. 1977. Relationship between oven and toluene dry matter in grass silage. J. Sci. Food Agric. 28: 477-480. 13. Harrison, J. H., R. Blauwiekel, and M. R. Stokes. 1994. Symposium: Utilization of grass silage. J. Dairy Sci. 77:3209-3235. 83 14. Harrison, J. H. and S. Fransen. 1991. Silage management in North America. Field Guide for Hay and Silage Management in North America. National Feed Ingredients Association. 33-67. 15. Kaiser, A. G., R. J. Mailer, and M. M. Vonarx. 1995. A comparison of Karl Fischer titration with alternative methods for the analysis of silage dry matter content. J. Sci. Food Agric. 69:51-59. 16. Leger, D. A. and S. K. Ho. 1992. Proceedings of the International Roundtable on Animal Feed Biotechnology Research and Scientific Regulation. Agriculture Canada Research Branch. 17. McDonald, P. 1981. The biochemistry of silage. John Wiley & Sons, Ltd. Chichester. 18. Mamer, O. A., and B. F. Gibbs. 1973. Simplified gas chromatography of trimethylsilyl esters of C x through C 5 fatty acids in serum and urine. Clinical Chemistry. Vol. 19, No.9: 1006-1009. 19. Mark, H. 1991. Calibration practice. Pages 85-146 in Principles and Practice of Spectroscopic Calibration. J. D. Winefordner ed. John Wiley & Sons, Inc. NY. 20. Muck, R. E. 1993. Ensiling and its effects on crop quality. Silage Production from Seed to Animal. Proc. of the Natl. Silage Production Conf. Syracuse, NY. 21. Muck, R. E. 1993. The role of silage additives in making high quality silage. Silage Production from Seed to Animal. Proc. of the Natl. Silage Production Conf. Syracuse, NY. 22. Muck, R. E. and K. K. Bolsen. 1991. Silage preservation and silage additive products. Field Guide for Hay and Silage Management in North America. National Feed Ingredients Association. 105-126. 23. Murray, I. Sward measurement handbook. Chapter 14: Forage Analysis by Near Infra-red Spectroscopy. The British Grassland Society (in press). 24. Norris, K. H., R. F. Barnes, J. E. Moore, and J. S. Shenk. 1976. Predicting forage quality by near infrared reflectance spectroscopy. J. Anim. Sci. 43:889-897. 25. Osborne, B. G., T. Fearn, and P. H. Hindle. 1993. Practical NTR spectroscopy with applications in food and beverage analysis. John Wiley & Sons, New York, NY. 26. Reeves, J. B., III. 1993. Influence of water on the near infrared spectra of model compounds. J. of AOAC International. Vol. 76, No. 4. 84 27. Reeves, J. B., Ill, and T. H. Blosser. 1991. Near infrared spectroscopic analysis of undried silages as influenced by sample grind, presentation method, and spectral region. J. Dairy Sci. 74:882-895. 28. Reeves, J. B., Ill, T. H. Blosser, A. T. Balde, B. P. Glenn, and J. Vandersall. 1991. Near infrared spectroscopic analysis of forages samples digested in situ (nylon bag). J. Dairy Sci. 74:2664-2673. 29. Reeves, J. B., Ill, T. H. Blosser, and V. F. Colenbrander. 1989. Near infrared reflectance spectroscopy for analyzing undried silage. J. Dairy Sci. 72:79-88. 30. Reeves, J. B., Ill, T. H. Blosser, and V. F. Colenbrander. 1990. Analysis of silage composition by near infrared reflectance spectroscopy. SPEE Vol. 1379 Optics in Agriculture. 31. Rotz, C. A. and R. E. Muck. 1994. Changes in forage quality during harvest and storage. Proc. Natl. Conf. on Forage Quality, Evaluation, and Utilization. Univ. of NE, Lincoln, NE. Editor-in-Chief: G.C. Fahey, Jr. 828-868. 32. Savoie, P., D. Tremblay, and J. Wauthy. 1993. Novel harvesting equipment for silage. Silage Production from Seed to Animal. Proc. Natl. Silage Production Conf. Syracuse, NY. 46-56. 33. Shenk, J. S. 1993. Analysis of undried, unground forage with a visible-near-infrared monochromator. Proc. of the XVII International Grassland Congress, pp 592-593. 34. Shenk, J. S. and M. O. Westerhaus. 1991. Population definition, sample selection, and calibration procedures for near infrared reflectance spectroscopy. Crop Sci. 31:469-474. 35. Shenk, J. S., and M. O. Westerhaus. 1991. Population structuring of near infrared spectra and modified partial least squares regression. Crop Sci. 31:1548-1555. 36. Shenk, J. S., and M. O. Westerhaus. 1993. Monograph: Analysis of agriculture and food products by near infrared reflectance spectroscopy. InfraSoft International, Port Matilda, PA, USA. 37. Shenk, J. S., and M. O. Westerhaus. 1994. The application of near infrared reflectance spectroscopy (NIRS) to forage analysis. In: G.C. Fahey, Jr. (Ed) Forage Quality, Evaluation, and Utilization, pp 406-449. ASA, Inc. Madison WI. 38. Sinnaeve, G., P. Dardenne, R. Agneessens, and R. Biston. 1994. The use of near infrared spectroscopy for the analysis of fresh grass silage. J. Near Infrared Spectrosc. 2, 79-84. 85 39. Steen, R. W. J. 1990. Recent advances in the use of silage additives for dairy cattle. Pages 87-101 in C. S. Mayne, ed. Management Issues for the Grassland Farmer in the 1990's. Occasional Symposium of the British Grassland Society, No. 25. Malvern, Worchestershire. 40. Thomas, C. and R. C. Rae. 1988. Concentrate supplementation of silage for dairy cows. Pages 327-354 in P. C. Garnsworthy, ed. Nutrition and Lactation in the Dairy Cow. Butterworths, London, UK. 41. Van Soest, P. J., J. B. Robertson, and B. A. Lewis. 1991. Symposium: Carbohydrate methodology, metabolism, and nutritional implications in dairy cattle: Methods for dietary fiber, NDF, and nonstarch polysaccharides in relation to animal nutrition. J. Dairy Sci. 74:3583-3597. 42. Williams, P. C. 1987. Variables affecting near-infrared reflectance spectroscopic analysis. Pages 143-167 in P. Williams and K. Norris ed. Near-Infrared Technology in the Agricultural and Food Industries. American Assoc. of Cereal Chemists. Saint Paul, MN. 43. Windham, W. R., J. A. Robertson, and R. G. Leffler. 1987. A comparison of methods for moisture determination of forages for near infrared reflectance spectroscopy calibration and validation. Crop Sci. 27:777-783. 44. Woolford, M. K. 1984. The silage fermentation. Microbiology Series. (Ed. Al. Laskin and R. I. Mateles.) Marcel Dekker, Inc. NY. and Basel. 45. Workman, J. and H. Mark. 1991. Statistics in spectroscopy. Academic Press, Inc. NY 86 "@en ; edm:hasType "Thesis/Dissertation"@en ; vivo:dateIssued "1996-05"@en ; edm:isShownAt "10.14288/1.0087077"@en ; dcterms:language "eng"@en ; ns0:degreeDiscipline "Animal Science"@en ; edm:provider "Vancouver : University of British Columbia Library"@en ; dcterms:publisher "University of British Columbia"@en ; dcterms:rights "For non-commercial purposes only, such as research, private study and education. Additional conditions apply, see Terms of Use https://open.library.ubc.ca/terms_of_use."@en ; ns0:scholarLevel "Graduate"@en ; dcterms:title "NIRS analysis of intact grass silage and fresh grass for the prediction of dry matter, crude protein and acid detergent fibre"@en ; dcterms:type "Text"@en ; ns0:identifierURI "http://hdl.handle.net/2429/4266"@en .