Advertisement

Herbage dry matter intake estimation of grazing dairy cows based on animal, behavioral, environmental, and feed variables

Open AccessPublished:January 31, 2019DOI:https://doi.org/10.3168/jds.2018-14834

      ABSTRACT

      Information about the individual herbage DMI (HDMI) of grazing dairy cows is important for an efficient use of pasture herbage as an animal feed with a range of benefits. Estimating HDMI, with its multifaceted influencing variables, is difficult but may be attempted using animal, performance, behavior, and feed variables. In our study, 2 types of approaches were explored: 1 for HDMI estimation under a global approach (GA), where all variables measured in the 4 underlying experiments were used for model development, and 1 for HDMI estimation in an approach without information about the amount of supplements fed in the barn (WSB). The accuracy of these models was assessed. The underlying data set was developed from 4 experiments with 52 GA and 50 WSB variables and one hundred thirty 7-d measurements. The experiments differed in pasture size, herbage allowance, pregrazing herbage mass, supplements fed in the barn, and sward composition. In all the experiments, cow behavioral characteristics were recorded using the RumiWatch system (Itin and Hoch GmbH, Liestal, Switzerland). Herbage intake was estimated by applying the n-alkane method. Finally, HDMI estimation models with a minimal relative prediction error of 11.1% for use under GA and 13.2% for use under WSB were developed. The variables retained for the GA model with the highest accuracy, determined through various selection steps, were herbage crude protein, chopped whole-plant corn silage intake in the barn, protein supplement or concentrate intake in the barn, body weight, milk yield, milk protein, milk lactose, lactation number, postgrazing herbage mass, and bite rate performed at pasture. Instead of the omitted amounts of feed intake in the barn and, due to the statistical procedure for model reduction, the unconsidered variables postgrazing herbage mass and bite rate performed at pasture, the WSB model with the highest accuracy retained additional variables. The additional variables were total eating chews performed at pasture and in the barn, total eating time performed at pasture, number of total prehension bites, number of prehension bites performed at pasture, and herbage ash concentration. Even though behavioral characteristics alone did not allow a sufficiently accurate individual HDMI estimation, their inclusion under WSB improved estimation accuracy and represented the most valid variables for the HDMI estimation under WSB. Under GA, the inclusion of behavioral characteristics in the HDMI estimation models did not reduce the root mean squared prediction error. Finally, further adaptation, as well as validation on a more comprehensive data set and the inclusion of variables excluded in this study such as body condition score or gestation, should be considered in the development of HDMI estimation models.

      Key words

      INTRODUCTION

      In addition to a better consumer image (
      • Getter K.L.
      • Behe B.K.
      • Howard P.H.
      • Conner D.S.
      • Spaniolo L.M.
      Increasing demand for pasture-based dairy: What attributes and images do consumers want?.
      ), improved animal welfare (
      • Arnott G.
      • Ferris C.
      • O'Connell N.
      Review: Welfare of dairy cows in continuously housed and pasture-based production systems.
      ), and higher product quality (
      • O'Callaghan T.F.
      • Hennessy D.
      • McAuliffe S.
      • Kilcawley K.N.
      • O'Donovan M.
      • Dillon P.
      • Ross R.P.
      • Stanton C.
      Effect of pasture versus indoor feeding systems on raw milk composition and quality over an entire lactation.
      ), grazing dairy cows offer potential ecological (
      • Guyader J.
      • Janzen H.
      • Kroebel R.
      • Beauchemin K.
      Forage use to improve environmental sustainability of ruminant production.
      ) and economic (
      • Holshof G.
      • Evers A.
      • De Haan M.
      • Galama P.
      Grazing and difficult circumstances: Economic benefits depend on milk price and grazing efficiency.
      ) benefits. However, recent estimates have shown that only 49 to 52% of lactating dairy cows in European countries have access to pasture, and the numbers have been declining in recent years (
      • Van den Pol A.
      • de Vliegher A.
      • Hennessy D.
      • Isselstein J.
      • Peyraud J.
      The future of grazing.
      ). Although pasture herbage has widely been identified as the cheapest source of nutrients for dairy cows (
      • Peyraud J.L.
      • Delaby L.
      • Garnsworthy P.
      • Wiseman J.
      Ideal concentrate feeds for grazing dairy cows—Responses to supplementation in interaction with grazing management and grass quality.
      ), the decline in grazing may be the result of larger herds, fragmentation or lack of land, development of automatic milking systems, and farmer expectations regarding productivity in a pasture-based system, as mentioned by
      • Kristensen T.
      • Madsen M.L.
      • Noe E.
      The use of grazing in intensive dairy production and assessment of farmers' attitude towards grazing.
      . Information about individual herbage DMI (HDMI) allows estimating the nutrient supply from pastures and determining an adapted optional supplementation in the barn that may improve efficiency and lead to higher acceptance in practice. Furthermore, information about intake compared with production can be used to assess the nutrient and energy efficiency of individual grazing dairy cows and may enable selection for this trait. Several methods for HDMI estimation exist. These include measuring herbage mass or sward surface height before and after grazing, back calculation from the energy requirements of the cows and the energy density of the offered herbage, near-infrared reflectance spectroscopy, and marker techniques (
      • Decruyenaere V.
      • Buldgen A.
      • Stilmant D.
      Factors affecting intake by grazing ruminants and related quantification methods: A review.
      ). However, these techniques are expensive, time consuming, and impracticable at the farm level and yield only group- or herd-mean HDMI estimations. Behavioral characteristics may also be considered in estimating the individual animal HDMI (
      • Andriamandroso A.L.H.
      • Bindelle J.
      • Mercatoris B.
      • Lebeau F.
      A review on the use of sensors to monitor cattle jaw movements and behavior when grazing.
      ).
      • Oudshoorn F.W.
      • Cornou C.
      • Hellwing A.L.F.
      • Hansen H.H.
      • Munksgaard L.
      • Lund P.
      • Kristensen T.
      Estimation of grass intake on pasture for dairy cows using tightly and loosely mounted di-and tri-axial accelerometers combined with bite count.
      calculated grass intake using regressions based on grazing time and animal individual bite frequency. Various methods based on jaw switches, pressure sensors, microphones, accelerometers, and electromyography have been developed to record behavioral characteristics automatically. Among these, pressure sensors and microphones can detect jaw movements with high accuracy (
      • Andriamandroso A.L.H.
      • Bindelle J.
      • Mercatoris B.
      • Lebeau F.
      A review on the use of sensors to monitor cattle jaw movements and behavior when grazing.
      ). Because the RumiWatch system (RWS; Itin and Hoch GmbH, Liestal, Switzerland), based on a pressure sensor and a triaxial accelerometer, showed reliable accuracy in detecting eating and rumination behavior (
      • Ruuska S.
      • Kajava S.
      • Mughal M.
      • Zehner N.
      • Mononen J.
      Validation of a pressure sensor-based system for measuring eating, rumination and drinking behaviour of dairy cattle.
      ;
      • Rombach M.
      • Münger A.
      • Niederhauser J.
      • Südekum K.-H.
      • Schori F.
      Evaluation and validation of an automatic jaw movement recorder (RumiWatch) on unsupplemented or supplemented dairy cows on pasture.
      ), it was chosen for the present investigation. However, eating chews or prehension bites alone seem insufficient to estimate HDMI, and they explain only a minor part of HDMI variation. Therefore, in addition to behavioral characteristics,
      • Decruyenaere V.
      • Buldgen A.
      • Stilmant D.
      Factors affecting intake by grazing ruminants and related quantification methods: A review.
      mentioned animal and feed characteristics that might have an influence on the HDMI of grazing ruminants.
      • Timmer B.
      • Zom R.L.G.
      • Holshof G.
      • Spithoven M.
      • Van Reenen C.G.
      The application of behavior sensors and sward measurement to support grazing management.
      showed a reliable estimation of HDMI using behavioral as well as animal- and herbage-related variables in the HDMI estimation model. In our study, we built models for HDMI estimation based on 2 approaches using the RWS, which can differentiate between mastication chews and prehension bites. Data from 4 experiments were used. The data contained differences in farm management, pasture size, grazing duration, herbage allowance, pregrazing herbage mass, supplements fed in the barn, and sward composition. The main objective of our study was to develop models for HDMI estimation based on the merged data set of the 4 experiments mentioned above. First, a global approach (GA) was explored, where all measured variables were used for the model development. Then, a second approach was investigated without the information about the supplemented amounts of forage or concentrate in the barn (referred to as WSB). Finally, we studied the effect of behavioral characteristics recorded with the RWS on the accuracy of the HDMI estimation under both approaches.

      MATERIALS AND METHODS

      Experimental Design, Animals, and Housing

      Three grazing experiments were conducted at the Agroscope experimental farm in Posieux, Switzerland, and 1 was conducted at the Ferme-Ecole in Sorens, Switzerland, from 2014 to 2016 (Table 1). All experimental procedures were in accordance with Swiss guidelines for animal welfare and were approved by the Animal Care Committee of the Canton Fribourg, Switzerland (no. 2014_38_FR, 2014_51_FR, 2015_11_FR, and 2015_22_FR). Before selection of the experimental cows, all cows passed a medical check. In general, each period consisted of 2 wk of adaptation to adjust pasture-accustomed cows to the various feeding treatments and measuring devices followed by 1 wk of data collection. Consequently, the experiments lasted 21 to 63 d. Between 18 and 28 Holstein and Red Holstein cows were used in the experiments. The experimental cows were distributed equally across the experimental groups based on BW, milk yield, DIM, and lactation number. At the beginning of the experiments, the cows had an average BW of 601 ± 58.7 kg, were 155 ± 64.5 DIM, were in lactation 2.5 ± 1.75 on average, and produced 24.2 ± 5.18 kg of milk/d (±SD).
      Table 1Basic data of the 4 experiments
      ItemExperiment 1Experiment 2Experiment 3Experiment 4
      PeriodJul. 28–Sep. 27, 2014May 18–Jun. 7, 2015Aug. 31–Sep. 20, 2015Jun. 27–Jul. 17, 2016
      Data sets in database (no.)54282424
      Experimental designCrossover blockBalanced blockBalanced blockBalanced block
      Treatment × period3 × 32 × 12 × 12 × 1
      Animals (no.)18282424
      Pasture size
      Experiments 1 and 2: all experimental cows grazed together in the same paddocks; experiments 3 and 4: the 2 experimental groups grazed on 2 paddocks.
      (ha)
      0.300.900.41, 0.060.21, 0.17
      Herbage allowance (kg of DM/d)21.426.621.918.1, 14.9
      Supplement
      NSP = not supplemented in the barn; corn = supplemented in the barn with chopped whole-plant corn silage; corn protein = supplemented in the barn with chopped whole-plant corn silage and with protein concentrate; concentrate = supplemented in the barn with concentrate (UFA 275; UFA AG, Herzogenbuchsee, Switzerland).
      NSP, corn, corn proteinNSP, concentrateNSPNSP, corn
      Farm managementConventionalOrganicConventionalConventional
      Primiparous cows (%)30.064.341.725.0
      Cow genetics
      CH = Swiss Holstein; NZ = New Zealand Holstein.
      CHCH, NZCHCH
      Outdoor temperature (°C)14.8 (min. 10.0, max. 19.9)
      min. = minimum; max. = maximum.
      20.2 (min. 13.7, max. 26.7)12.9 (min. 6.8, max. 19.2)15.1 (min. 10.1, max.19.7)
      Daily precipitation (mm)6 (min. 0.1, max. 11)0.5 (min. 0.1, max. 2.7)9.6 (min. 0.1, max. 30.9)12.1 (min. 0.2, max. 15.3)
      ReferenceRombach et al. (2018)Schori et al. (unpublished data)Rombach et al. (unpublished data)Menzi et al. (unpublished data)
      1 Experiments 1 and 2: all experimental cows grazed together in the same paddocks; experiments 3 and 4: the 2 experimental groups grazed on 2 paddocks.
      2 NSP = not supplemented in the barn; corn = supplemented in the barn with chopped whole-plant corn silage; corn protein = supplemented in the barn with chopped whole-plant corn silage and with protein concentrate; concentrate = supplemented in the barn with concentrate (UFA 275; UFA AG, Herzogenbuchsee, Switzerland).
      3 CH = Swiss Holstein; NZ = New Zealand Holstein.
      4 min. = minimum; max. = maximum.
      In all 4 experiments, the paddocks were grazed rotationally, and the stocking periods lasted 1 to 3 d. All cows grazed day and night between 16 and 19 h/d depending on the experiment. In the meantime, the cows were kept in freestall barns and were supplemented if intended as part of the treatment structure. In addition, they were milked, milk was sampled, and alkane marker capsules were administered. On average, the pastures (permanent and sown) comprised grasses (mainly Lolium perenne, Poa pratensis, and Lolium multiflorum; 58–84% of the fresh herbage biomass), legumes (Trifolium repens and Trifolium pratense; 6–26% of the fresh herbage biomass), and forbs (mainly Taraxacum officinale, Plantago lanceolata, and Rumex acetosa; 2–29% of the fresh herbage biomass).
      The first experiment (EX1) was organized as a crossover block design with 3 treatments and 3 periods (Table 1). All experimental cows in EX1 grazed as a single group in the same paddocks. Two out of 3 experimental groups were supplemented in the barn at weighing troughs (Insentec B.V., Marknesse, the Netherlands) with either 10 kg DM/d of chopped whole-plant corn silage or with 8.2 kg DM/d of chopped whole-plant corn silage mixed with 1.8 kg DM/d of protein concentrate (60% expeller soybean meal, 25% corn gluten, 10% potato protein, and 5% dried sugar beet pulp). The third group was not supplemented. Each cow served as its own control and underwent all 3 treatments. The second (EX2), third (EX3), and fourth (EX4) experiments were performed as balanced block designs including 2 treatments and 1 period. In EX2, 1 group was supplemented with an average of 3 kg DM/d of concentrate (UFA 275; UFA AG, Herzogenbuchsee, Switzerland) through an automatic concentrate feeder in the barn. The other group received no concentrate. All experimental cows in EX2 grazed as a single group in the same paddocks. In EX3, one group of cows grazed on swards with a high herbage mass of 2,288 kg of DM/ha, and the other group grazed on swards with a low herbage mass of 589 kg of DM/ha. The cows in both groups had a similar herbage allowance of approximately 22 kg of DM/d per cow and were not supplemented in the barn. In EX3 and EX4, the cows from each group grazed on different paddocks. One experimental group in EX4 was supplemented in the barn at the weighing troughs (Insentec B.V.) with 4 kg DM/d of chopped whole-plant corn silage. The other group received no supplementation. The chemical composition of the herbage and the supplements fed during the experimental periods are shown in Table 2, Table 3. The cows had free access to drinking water during all experiments. The ambient outdoor temperature and the amount of rainfall during the experiments were recorded daily at the meteorological station in Grangeneuve (Meteo-Schweiz, Station Posieux, Switzerland).
      Table 2Chemical composition of the herbage grazed during the 4 experiments
      ItemExperiment 1Experiment 2Experiment 3
      Low = group that grazed on swards with a sward height of 10 click units and an herbage mass of 589 kg of DM/ha; High = group that grazed on swards with a sward height of 18 click units and an herbage mass of 2,288 kg of DM/ha.
      Experiment 4
      NSP = not supplemented in the barn; SP = supplemented in the barn with 4 kg of DM/d of chopped whole-plant corn silage.
      Period 1Period 2Period 3MeanSDLowHighNSPSP
      MeanSDMeanSDMeanSDMeanSDMeanSDMeanSDMeanSD
      DM (g/kg of wet weight)14925.715228.115726.621826.917430.216722.020041.220141.6
      Analyzed nutrient composition (g/kg of DM)
       OM8878.68917.28898.19007.19024.19023.19082.19074.9
       CP16915.42159.62126.815822.824026.41849.91798.416514.8
       ADFom
      ADF corrected for residual ash.
      26118.720611.11915.722127.618813.521810.923913.324518.5
       NDFom
      NDF corrected for residual ash.
      41871.633419.230118.140555.432828.235917.742831.544141.2
       Crude fiber18719.818018.819221.120926.016712.91958.821811.922911.9
      Calculated energy and APDE
      Absorbable protein in the small intestine when rumen-fermentable energy is limiting microbial protein synthesis in the rumen.
      content
      According to Agroscope (2015).
      (per kg of DM)
       NEL (MJ)6.00.226.50.076.50.086.10.336.60.146.30.076.20.096.10.15
       ADPE (g)1034.51152.31151.81016.91184.01071.61062.01033.7
      Analyzed n-alkane content (mg/kg of DM)
       HC32
      Dotriacontane (C32H66).
      5.70.715.40.494.80.223.40.706.70.596.30.514.60.355.60.31
       HC33
      Tritriacontane (C33H68).
      72.45.3566.84.8165.24.1147.16.4887.49.9783.78.6877.96.9393.48.16
      1 Low = group that grazed on swards with a sward height of 10 click units and an herbage mass of 589 kg of DM/ha; High = group that grazed on swards with a sward height of 18 click units and an herbage mass of 2,288 kg of DM/ha.
      2 NSP = not supplemented in the barn; SP = supplemented in the barn with 4 kg of DM/d of chopped whole-plant corn silage.
      3 ADF corrected for residual ash.
      4 NDF corrected for residual ash.
      5 Absorbable protein in the small intestine when rumen-fermentable energy is limiting microbial protein synthesis in the rumen.
      6 According to
      • Agroscope
      Fütterungsempfehlungen und Nährwerttabellen für Wiederkäuer (Feeding recommendations and nutrient tables for ruminants).
      .
      7 Dotriacontane (C32H66).
      8 Tritriacontane (C33H68).
      Table 3Average chemical composition of the supplements fed during the experiments
      ItemExperiment 1
      CS = chopped whole-plant corn silage; Protein = protein concentrate consisting of (as-fed basis) 60% expeller soybean meal, 25% corn gluten, 10% potato protein, and 5% dried sugar beet pulp.
      Experiment 2Experiment 4
      CSProteinCS + proteinConcentrate
      Commercial concentrate (UFA 275; UFA AG, Herzogenbuchsee, Switzerland).
      CS
      MeanSDMeanSDMeanSDMeanSDMeanSD
      DM (g/kg of wet weight)39833.888223.948532.08854053.4
      Analyzed nutrient composition (g/kg of DM)
       OM9710.59430.29660.49469751.0
       CP725.75626.61605.9115893.2
       ADFom
      ADF corrected for residual ash.
      19429.3763.017324.67320610.5
       NDFom
      NDF corrected for residual ash.
      35149.731634.834541.019538925.0
       Crude fiber16323.8340.114019.54716422.1
      Calculated energy and APDE
      Absorbable protein in the small intestine when rumen-fermentable energy is limiting microbial protein synthesis in the rumen.
      content
      According to Agroscope (2015).
      (per kg of DM)
       NEL (MJ)6.90.247.57.20.207.07.00.17
       APDE (g)703.742951113.0785740.82
      Analyzed n-alkane content (mg/kg of DM)
       HC32
      Dotriacontane (C32H66).
      1.10.211.31.221.10.21ND
      Not detected.
      0.80.03
       HC33
      Tritriacontane (C33H68).
      9.71.430.40.278.01.231.710.40.82
      1 CS = chopped whole-plant corn silage; Protein = protein concentrate consisting of (as-fed basis) 60% expeller soybean meal, 25% corn gluten, 10% potato protein, and 5% dried sugar beet pulp.
      2 Commercial concentrate (UFA 275; UFA AG, Herzogenbuchsee, Switzerland).
      3 ADF corrected for residual ash.
      4 NDF corrected for residual ash.
      5 Absorbable protein in the small intestine when rumen-fermentable energy is limiting microbial protein synthesis in the rumen.
      6 According to
      • Agroscope
      Fütterungsempfehlungen und Nährwerttabellen für Wiederkäuer (Feeding recommendations and nutrient tables for ruminants).
      .
      7 Dotriacontane (C32H66).
      8 Not detected.
      9 Tritriacontane (C33H68).

      Data Recording and Sample Collection

      The sward height was measured daily in all experiments using an electronic rising plate meter (Jenquip, Feilding, New Zealand; 1 click unit = 0.5 cm; EX1, EX3, and EX4) or a C-Dax Pasture Meter (C-Dax Ltd., Palmerston North, New Zealand; EX2). Herbage mass was measured before and after grazing each paddock by cutting 2 strips (1 m wide × 9.4 ± 3.02 (SD) m long; >6.7 click units or 3.35-cm stubble height) with a sickle-bar mower (Rekord 38, Bucher Landtechnik AG, Niederweningen, Switzerland). The herbage mass of the cut strips and the analyzed herbage DM content were used to calculate the herbage mass.
      The RWS, consisting of the RumiWatch Halter (RWH; Itin and Hoch GmbH) and the RumiWatch Converter 0.7.3.31 (Itin and Hoch GmbH), was used in all experiments to record and evaluate the behavioral characteristics of the experimental cows. Further details about the RWS, the accuracy and definition of the recorded behavioral characteristics, and the setup and handling of the RWH have been described previously (
      • Rombach M.
      • Münger A.
      • Niederhauser J.
      • Südekum K.-H.
      • Schori F.
      Evaluation and validation of an automatic jaw movement recorder (RumiWatch) on unsupplemented or supplemented dairy cows on pasture.
      ). To accustom the cows to the measuring system, the RWH were attached to the cows 4 d before the start of each measuring week and left on throughout the week. In EX1, 15 cows were equipped with version 3.0 of the RWH and 3 cows were equipped with version 6.0. These differ in materials, adjustability, and wearing comfort. In EX2, EX3, and EX4, all cows were equipped with RWH version 6.0. The recorded raw data were read through the interface software RumiWatch Manager (version 2.1.0.0; Itin and Hoch GmbH) and processed using the evaluation software RumiWatch Converter (version 0.7.3.31). Compared with the converters used in
      • Rombach M.
      • Münger A.
      • Niederhauser J.
      • Südekum K.-H.
      • Schori F.
      Evaluation and validation of an automatic jaw movement recorder (RumiWatch) on unsupplemented or supplemented dairy cows on pasture.
      , RumiWatch Converter 0.7.3.31 allows a differentiation of other bites, mastication chews with the head up, mastication chews with the head down, prehension bites, and time spent masticating and eating.
      Milk yield was measured twice daily during the milkings in the milking parlor (EX1, EX3, EX4: Fullwood, Arnold Bertschy AG, Guschelmuth, Switzerland; EX2: MidiLine, DeLaval AG, Sursee, Switzerland) with a Pulsameter (EX1, EX3, EX4: LMS GmbH Stützerbach, Ilmenau, Germany) or an MM15 (EX2: DeLaval AG). Milk composition was measured for EX1, EX2, and EX4 on d 2 and 5 and for EX3 on d 2, 4, and 6 during the measuring week. Aliquots of subsamples from the morning and evening milkings were pooled and preserved in 1 sample tube containing a Broad-Spectrum Microtab II (Gerber Instruments AG, Effretikon, Switzerland) and stored at 8°C for subsequent analysis of milk fat, protein, lactose, and casein content. After each milking, BW was measured with an animal weighing system (EX1, EX3, EX4: Ga5010, Insentec B.V.; EX2: W-2000, DeLaval AG).
      Individual HDMI was estimated using the n-alkane double indicator method (
      • Mayes R.W.
      • Lamb C.S.
      • Colgrove P.M.
      The use of dosed and herbage n-alkanes as markers for the determination of herbage intake.
      ). Six days before each measuring week until the next-to-last day of the measuring weeks, cows were dosed twice daily with 1 gelatine capsule (HGK-17-60 sl; Capsula GmbH, Ratingen, Germany) containing 0.5 g of dotriacontane (C32H66, HC32; Minakem Beuvry Production S.A.S., Beuvry la Forêt, France) as the external alkane marker on a carrier of 4.5 g of dried fruit pomace. During 7 d, once per day after the morning milking, feces of each cow were spot-sampled indoors to determine the content of alkanes in the feces. Samples were taken from spontaneous defecations or with mild stimuli, pooled for each cow and measuring week, and stored at −20°C for further analysis. To determine the content of alkanes in the herbage eaten by the cows, herbage collection was carried out 7 d, in the morning and afternoon. By following the experimental cows one after another and observing their grazing selection, small samples of the most likely grazed herbage were cut with a battery grass shearer (Gardena; Husqvarna Schweiz AG, Mägenwil, Switzerland). The herbage sampling started 24 h before the feces sampling and ended 24 h earlier. These samples were chopped and stored at −20°C for further analysis.

      Laboratory Analysis

      Milk samples were analyzed using Fourier-transform mid-infrared spectrometry (Combi-Foss FT +; Foss, Hillerød, Denmark) to determine fat, protein, casein, and lactose content. Fluorescence flow cytometry (Fossomatic FC200; Foss) was used to count the number of somatic cells in the milk samples.
      The herbage and supplement samples, except the protein supplement in EX1 and concentrate supplement in EX2, were stored at −20°C until they were lyophilized (Delta 1-24 LSC; Christ, Osterode, Germany). Subsequently, all samples, including protein and concentrate supplements, were milled through a 1.0-mm screen (Brabender mill with titanium blades; Brabender GmbH & Co. KG, Duisburg, Germany), dried for 3 h at 105°C to determine DM, and finally incinerated at 550°C until a stable mass was reached to determine the ash content (
      • AOAC International
      Official Methods of Analysis.
      ; method 942.05). Mineral residues in the ash were dissolved by nitric acid and analyzed for Ca, P, Na, Mg, and K with inductively coupled plasma optical emission spectrometry (ICP-OES Optima 7300 DV; PerkinElmer, Waltham, MA) based on
      • ISO (International Organization for Standardization)
      Animal feeding stuffs—Determination of calcium, sodium, phosphorus, magnesium, potassium, iron, zinc, copper, manganese, cobalt, molybdenum, arsenic, lead and cadmium by ICP-AES. Method 27085.
      ; method 27085). The contents of the HC32 and tritriacontane (C33H68) were analyzed as described by
      • Thanner S.
      • Dohme-Meier F.
      • Görs S.
      • Metges C.
      • Bruckmaier R.
      • Schori F.
      The energy expenditure of 2 Holstein cow strains in an organic grazing system.
      . The N content of herbage and supplement samples was analyzed using the Dumas method (
      • ISO (International Organization for Standardization)
      Food products—Determination of the total nitrogen content by combustion according to the Dumas principle and calculation of the crude protein content. Method 16634-1.
      ; method 16634-1) on a C/N analyzer (Trumac CNS; Leco Instruments, St. Joseph, MI); the results were multiplied by 6.25 to obtain the CP content. The contents of ADF (
      • AOAC International
      Official Methods of Analysis.
      ; method 973.18), NDF (
      • AOAC International
      Official Methods of Analysis.
      ; method 2002.4), and crude fiber (
      • AOAC International
      Official Methods of Analysis.
      ; method 978.10) for the herbage and supplement samples were analyzed with Gerhardt Fibertherm (Gerhardt GmbH & Co. KG, Königswinter, Germany). The NDF and ADF contents were separately determined (parallel). For NDF analysis, heat-stable amylase and sodium sulfite were added. A correction for the residual ash obtained after 2 h of incineration at 550°C was made for ADF corrected for residual ash and NDF corrected for residual ash.

      Calculations and Data Analysis

      The NEL content of herbage was calculated from chemical composition according to
      • Agroscope
      Fütterungsempfehlungen und Nährwerttabellen für Wiederkäuer (Feeding recommendations and nutrient tables for ruminants).
      . For chopped whole-plant corn silage, NEL content was calculated according to
      • Agroscope
      Fütterungsempfehlungen und Nährwerttabellen für Wiederkäuer (Feeding recommendations and nutrient tables for ruminants).
      . Herbage intake calculation was based on equations proposed by
      • Mayes R.W.
      • Lamb C.S.
      • Colgrove P.M.
      The use of dosed and herbage n-alkanes as markers for the determination of herbage intake.
      . Equation 1 was used to calculate the daily HDMI of every single experimental cow in the 4 experiments:
      HDMI=F33F32×{[A32+(P×P32)+(CN×CN32)+(CR×CR32)](P×P33+CN×CN33+CR×CR33)}H33F33F32×H32,
      [1]


      where HDMI represents the daily HDMI (kg); F33, H33, P33, CN33, and CR33 are the concentrations of tritriacontane (mg/kg of DM) in feces, herbage, protein supplement, concentrate, and chopped whole-plant corn silage, respectively; F32, H32, P32, CN32, and CR32 are the concentrations of HC32 (mg/kg of DM) in feces, herbage, protein supplement, concentrate, and chopped whole-plant corn silage consumed, respectively; P, CN, and CR are the amounts (kg of DM/d) of consumed protein supplement, concentrate, and chopped whole-plant corn silage, respectively; and A32 is the daily dose of HC32 (mg/d) administered via the alkane capsules.
      The weekly averages of pasture, herbage, intake, animal, milk, and behavioral data used for the development of the different HDMI estimation models are shown in Table 4. The HDMI estimated by the n-alkane double indicator method was used as the reference herbage intake on pasture for the development and validation of 4 approaches. First, a GA was explored, where all variables in the data set were used for the model development. Afterward, a WSB approach was investigated without information about the amount of the supplements fed in the barn; the variables whole-plant corn silage, protein supplement, and concentrate intake were removed from the data set for the model development. Finally, to study the benefits of behavioral characteristics as predictors for HDMI estimation, the behavioral characteristics recorded with the RWS were removed from the data set of the GA and the WSB. These approaches without RWS variables are hereafter referred to as GAwRW and WSBwRW, respectively.
      Table 4Mean (n = 109) and range of the pasture, intake, animal, milk, and behavioral characteristics used for the herbage DMI estimation models
      ItemMeanMinimumMaximumSD of mean
      Pasture variables
       Postgrazing herbage mass (kg of DM/ha)22263554143.4
       Pregrazing herbage mass (kg of DM/ha)1,2065892,333628.5
       Residence time on pasture (h/d)1815191.2
       Herbage allowance (kg of DM/cow per d)23.611.138.99.28
      Herbage variables
       CP (g/kg of DM)18715824027.1
       Ash (g/kg of DM)102921229.1
      Intake variables (kg/d)
       Herbage DMI12.44.720.42.93
       Protein or concentrate intake
      Consumed in the barn.
      0.80.04.01.18
       Corn silage intake
      Consumed in the barn.
      Chopped whole-plant corn silage.
      3.70.07.93.08
      Animal variables
       BW (kg)61042871958.3
       Lactation number2.71.09.01.92
      Milk yield and content
       Milk yield (kg/d)23.314.038.04.56
       Fat (%)4.12.75.60.57
       Protein (%)3.32.43.90.28
       Lactose (%)4.64.05.20.21
      Daily behavioral characteristics
       Total eating time (min/d)61344174257.4
       Prehension bites (no./d)30,16511,78441,3466,578.4
       Total eating chews (no./d)44,02731,66854,1744,495.2
       Bite rate (total eating bites/min)7262803.5
       Bite mass (DMI/prehension bites)0.540.271.600.216
      Daily behavioral characteristics performed at pasture
       Total eating time (min/d)54835569162.4
       Prehension bites (no./d)28,75711,03740,3046,664.4
       Total eating chews (no./d)40,00426,22548,7104,842.2
       Bite rate (total eating bites/min)7362813.7
       Bite mass (herbage DMI/prehension bites)0.470.261.040.136
       Head down (min/d)667179956118.9
      1 Consumed in the barn.
      2 Chopped whole-plant corn silage.
      Univariate and bivariate graphics and descriptive statistics were used to provide an overview of the data set. Linear dependencies between regression variables were detected by Pearson correlation coefficients near or equal to 1 and prevented by the exclusion of redundant variables.
      A preliminary set of predictive linear models was based on the combined results of various statistical approaches (principal component analysis; partial least squares; forward, backward, and sequential selection; and best subset regression) using Systat 13 (version 13.0; Systat Software, Chicago, IL) and R (
      • R Core Team
      R: A Language and Environment for Statistical Computing.
      ) packages Rcmdr (
      • Fox J.
      The R Commander: A basic statistics graphical user interface to R.
      ,
      • Fox J.
      Using the R Commander: A Point-and-Click Interface for R.
      ;
      • Fox J.
      • Bouchet-Valat M.
      Rcmdr: R Commander. R package version 2.3-2.
      ), leaps (
      • Lumley T.
      • Miller A.
      Leaps: Regression subset selection. R package version 3.0.
      ), and rms (
      • Harrell F.E.
      rms: Regression Modeling Strategies. R package version 5.1-0.
      ).
      Spearman correlation coefficients facilitated the identification of monotonous relations between the reference variable (HDMI) and the continuous or ordinal regressors. Principal component analysis using the Pearson and Spearman correlation matrices, partial least squares regression (including dummy variables for categorical factors), linear models (forward, backward, and sequential selection), canonical correlation, and finally best subset regression modeling was applied to identify variables that (from a statistical point of view) could be deleted from the set of regressors, leaving a list of potentially important predictor variables. The complete set of variables as well as the statistical findings were assessed, and variables were included or excluded based on the statistical findings, pertinent knowledge, and experience. The resulting combined set of 25 variables was still too large. Therefore, model reduction by the best subset regression approach was applied, and models with fewer than 14 variables were kept for the final validation. As the sample size was too small to keep an independent validation set, the bootstrap validation method proposed by
      • Harrell F.E.
      Regression Modeling Strategies: With Applications to Linear Models, Logistic and Ordinal Regression, and Survival Analysis.
      and implemented in the R function rms:validate.ols (
      • Harrell F.E.
      rms: Regression Modeling Strategies. R package version 5.1-0.
      ) was used. This function combines modeling, model reduction, and bootstrap validation. The root mean squared prediction error (RMSPE) calculation was based on the results of 5,000 bootstrap samples (
      • Efron B.
      • Hastie T.
      The jackknife and the bootstrap and bootstrap confidence intervals.
      ) to identify the optimal predictive model while still preventing overfitting. Moreover, the bootstrap validation method was enhanced by including the optimism bias to prevent an underestimation of the RMSPE. Equation 2 shows the multiple linear regression structure of the HDMI estimation models:
      y=μ+(V1×C1)+(V2×C2)++(Vn×Cn),
      [2]


      where y represents the average daily HDMI (kg/cow) over 1 wk; µ is the model mean; and V1, V2, …, Vn are the explanatory variables with the corresponding coefficients C1, C2, …, Cn.

      RESULTS

      A data set of 130 measurements taken over 7 d consisting of 52 variables for GA and 50 variables for WSB (without protein supplement or concentrate and chopped whole-plant corn silage intake) was the database for the development and validation of the HDMI estimation models. For the development and validation of the WSBwRW and GAwRW HDMI estimation models, the data sets consisted of 42 variables for GAwRW (without 10 behavioral characteristics) and 40 variables for WSBwRW (without 10 behavioral characteristics and without protein supplement or concentrate and chopped whole-plant corn silage intake), and 130 measurements taken over 7 d were used. Due to technical difficulties with the RWH, in particular liquid leakage in the pressure sensor tubes, 21 (16%) of the records could not be correctly evaluated over the whole measuring week and were therefore not used for further model development and validation if behavioral characteristics were used in the HDMI estimation models. Thus, the final data set for GA and WSB model development consisted of 109 measurements.

      HDMI Estimation Under a GA

      The model with the lowest RMSPE for HDMI estimation under GA without overfitting (eliminating the nonsignificant variables) was model GA7 (Table 5). In this model, according to the β value (used to classify the variables and their effect on the HDMI estimation models), intake of chopped whole-plant corn silage [P < 0.001; standardized coefficient (β) = −0.646] showed the greatest effect on the target variable (HDMI), followed by milk yield (P < 0.001, β = 0.570), milk protein content (P < 0.001, β = 0.410), intake of protein supplement or concentrate (P < 0.001, β = −0.393), herbage CP (P < 0.001, β = −0.265), lactation number (P = 0.002, β = −0.191), milk lactose content (P < 0.001, β = −0.187), postgrazing herbage mass (P < 0.001, β = −0.180), BW (P = 0.021, β = 0.170), and bite rate performed at pasture (P = 0.046, β = −0.101).
      Table 5Evaluated linear model GA7 for herbage DMI estimation for use under a global approach
      ItemCoefficientSEβ
      Standardized coefficient; helps classify the variables according to their effect on the corresponding herbage DMI estimation model.
      95% CIP-value
      LowerUpper
      Model mean μ11.276.1820−0.995423.53930.071
      Protein or concentrate intake (kg of DM/d)−1.090.153−0.393−1.3944−0.7865<0.001
      Corn silage intake (kg of DM/d)−0.640.047−0.646−0.7369−0.5497<0.001
      Milk lactose (%)−2.520.740−0.187−3.9876−1.0497<0.001
      Lactation number−0.290.091−0.191−0.46686−0.10550.002
      Herbage CP (g/kg of DM)−0.030.006−0.265−0.0388−0.0150<0.001
      Postgrazing herbage mass (kg of DM/ha)−0.0040.0010−0.180−0.0056−0.0015<0.001
      Bite rate
      Performed at pasture.
      (total eating bites/min)
      −0.080.038−0.101−0.1514−0.00150.046
      BW (kg)0.0080.00340.1700.00120.01480.021
      Milk protein (%)4.240.5950.4103.05995.4208<0.001
      Milk yield (kg/d)0.350.0370.5700.27860.4272<0.001
      1 Standardized coefficient; helps classify the variables according to their effect on the corresponding herbage DMI estimation model.
      2 Performed at pasture.

      HDMI Estimation Model Under an Approach Without Knowledge of the Supplements Fed

      The model with the lowest RMSPE for HDMI estimation under WSB without overfitting (eliminating the nonsignificant variables) was model WSB8 (Table 6). In this model, according to the β value, number of prehension bites performed at pasture (P < 0.001, β = 2.475) showed the greatest effect on HDMI, followed by number of prehension bites (P = 0.002, β = −1.994), total time spent eating on pasture (P < 0.001, β = 0.636), number of total eating chews (P < 0.001, β = −0.488), BW (P = 0.001, β = 0.265), milk lactose content (P < 0.001, β = −0.236), herbage ash (P = 0.004, β = 0.236), milk yield (P = 0.004, β = 0.216), herbage CP (P = 0.003, β = −0.209), lactation number (P = 0.011, β = −0.199), and milk protein content (P = 0.014, β = 0.172).
      Table 6Evaluated linear model WSB8 for herbage DMI estimation for use under the approach without knowledge of the supplements fed
      ItemCoefficientSEβ
      Standardized coefficient; helps classify the variables according to their effect on the corresponding herbage DMI estimation model.
      95% CIP-value
      LowerUpper
      Model mean μ−1.017.7300−16.348814.33870.897
      Milk lactose (%)−3.180.868−0.236−4.9019−1.4573<0.001
      Lactation number−0.300.115−0.199−0.5278−0.07010.011
      Herbage CP (g/kg of DM)−0.020.007−0.209−0.03530.00700.003
      Total eating chews
      Performed at pasture and in the barn.
      (no./d)
      −0.00030.00007−0.488−0.0005−0.0002<0.001
      Total eating time
      Performed at pasture.
      (min/d)
      0.030.0050.6360.01860.039<0.001
      Prehension bites
      Performed at pasture and in the barn.
      (no./d)
      −0.00090.00027−1.994−0.0014−0.00030.002
      Prehension bites
      Performed at pasture.
      (no./d)
      0.0010.000272.4750.00050.0016<0.001
      Milk protein (%)1.780.7120.1720.36833.19300.014
      Herbage ash (g/kg of DM)0.100.0340.2360.03310.16670.004
      Milk yield (kg/d)0.130.0450.2160.04420.22360.004
      BW (kg)0.010.0040.2650.00520.01970.001
      1 Standardized coefficient; helps classify the variables according to their effect on the corresponding herbage DMI estimation model.
      2 Performed at pasture and in the barn.
      3 Performed at pasture.

      Validation of the HDMI Estimation Models

      The models suggested for GA (GA1–GA7) showed an estimation accuracy [100% − relative prediction error (RPE)] of 86.9% (GA1), increasing by an average of 0.3% for every variable added to 88.9% for model GA7. Model GA7 explained 79% of the HDMI variation, with an RMSPE of 1.38 kg of DM/animal per day and an RPE of 11.1% (Table 7).
      Table 7Structure and accuracy of the evaluated herbage DMI estimation models
      GA = global approach; WSB = approach without the information about the supplemented amounts of forage or concentrate in the barn.
      ItemGAWSB
      1234567891012345678910
      Pasture variables
       Postgrazing herbage mass (kg of DM/ha)xxxxxxx
      Herbage variables (g/kg of DM)
       CPxxxxxxxxxxxxx
       Ashxxxxxxx
      Intake variables (kg/d)
       Chopped whole-plant corn silage intake
      Consumed in the barn.
      xxxxxxxxxx
       Protein or concentrate intake
      Consumed in the barn.
      xxxxxxxxxx
      Animal variables
       BW (kg)xxxxxxxxxxxxxxxxxx
       Lactation numberxxxxxxxxxxxx
      Milk yield and content
       Milk yield (kg/d)xxxxxxxxxxxxxx
       Fat (%)xx
       Protein (%)xxxxxxxxxxxxx
       Lactose (%)xxxxxxxxxxxxxxx
      Daily behavioral characteristics
       Total eating chews (no./d)xxxxxxxxxxxx
       Total eating chews
      Performed at pasture.
      (no./d)
      xxx
       Bite rate (total eating bites/min)x
       Bite rate
      Performed at pasture.
      (total eating bites/min)
      xxxx
       Prehension bites (no./d)xxxxxxxx
       Prehension bites
      Performed at pasture.
      (no./d)
      xxxxxxxxxx
       Total eating time
      Performed at pasture.
      (min/d)
      xxxxxxxxxx
       Head down time
      Performed at pasture.
      (min/d)
      x
      Accuracy of evaluated models
       No.130130130130130130109109109109109109109109109109109109109106
       Multiple R0.850.860.860.870.880.880.900.900.910.910.790.810.820.830.830.840.840.850.860.86
       Squared multiple R (R2)0.720.740.750.760.770.770.810.810.820.830.620.650.670.680.690.700.710.730.740.74
       Adjusted squared multiple R
      Related to the number of variables and observations.
      0.710.730.730.740.750.750.790.790.800.800.600.630.650.660.660.670.680.700.710.71
       Standard error of the estimate1.581.541.521.491.471.461.311.311.281.261.801.721.691.661.651.631.601.561.541.54
       RMSPE
      Root mean squared prediction error.
      (bootstrapping)
      1.621.591.571.551.531.521.381.391.361.341.851.781.751.731.731.711.681.641.651.65
       RPE
      Relative prediction error.
      13.112.812.712.512.312.311.111.211.010.814.914.414.114.014.013.813.513.213.313.3
      1 GA = global approach; WSB = approach without the information about the supplemented amounts of forage or concentrate in the barn.
      2 Consumed in the barn.
      3 Performed at pasture.
      4 Related to the number of variables and observations.
      5 Root mean squared prediction error.
      6 Relative prediction error.
      The models suggested for WSB (WSB1–WSB8) showed an estimation accuracy of 85.1% (WSB1), increasing by an average of 0.2% for every variable added to 86.8% for model WSB8. Furthermore, model WSB8 explained 70% of the HDMI variation, with an RMSPE of 1.64 kg of DM/animal per day and an RPE of 13.2% (Table 7).
      The models suggested for GAwRW (GAwRW1–GAwRW7; detailed data not shown) showed an estimation accuracy of 86.9% (GAwRW1), increasing by an average of 0.2% for every variable added to 88.3% for model GAwRW7. Furthermore, the model for GAwRW7 explained 77% of the HDMI variation, with an RMSPE of 1.5 kg of DM/animal per day and an RPE of 11.7%.
      The models suggested for the approaches without behavioral characteristics (WSBwRW1–WSBwRW8; detailed data not shown) exhibited an estimation accuracy of 81.9% (WSBwRW1), decreasing by an average of 0.03% for every variable added to 81.7% for model WSBwRW8. Furthermore, the model WSBwRW8 explained 45% of the HDMI variation, with an RMSPE of 2.3 kg of DM/animal per day and an RPE of 18.3%.
      Figure 1 presents the RMSPE of the HDMI estimation models with and without behavioral characteristics (RWS variables). Models without information about supplementation and behavioral characteristics exhibit an RMSPE between 2.2 and 2.3 kg of DM/animal per day depending on the number of predictors in the model. Inclusion of behavioral characteristics reduced the error term by about 0.5 kg of DM/animal per day. If amounts of supplements fed in the barn were available for HDMI estimation, the error term decreased again about 0.3 kg of DM/animal per day to end at 1.3 to 1.6 kg of DM/animal per day. In this case, inclusion of behavioral information did not additionally reduce the RMSPE.
      Figure thumbnail gr1
      Figure 1Comparison between the root mean squared prediction error (RMSPE) of the herbage DMI estimation models with or without behavioral characteristics in the data set for model development. GA = global approach; WSB = approach without the information about the supplemented amounts of forage or concentrate in the barn; GAwRW = GA without RumiWatch (Itin and Hoch GmbH, Liestal, Switzerland) system variables; WSBwRW = WSB without RumiWatch system variables.

      DISCUSSION

      The range of the measured variables in the 4 experiments that constituted the data set of our study allowed for the development and validation of different models for HDMI estimation under GA (GA1–GA10), WSB (WSB1–WSB10), GAwRW, and WSBwRW.

      HDMI Estimation Using the n-Alkane Method

      Herbage DMI estimated with the n-alkane method was used as the reference for model development and validation. In earlier investigations with barn-fed dairy cows, compared with the weighed intake, DMI estimation using the n-alkane method showed a low mean deviation of 0.05 kg, and the estimated and weighed DMI were highly correlated (R2 = 0.93) for pooled samples (
      • Berry N.R.
      • Scheeder M.R.
      • Sutter F.
      • Kröber T.F.
      • Kreuzer M.
      The accuracy of intake estimation based on the use of alkane controlled-release capsules and faeces grab sampling in cows.
      ). In conditions that were similar to those in our study,
      • Kaufmann L.D.
      • Münger A.
      • Rérat M.
      • Junghans P.
      • Görs S.
      • Metges C.C.
      • Dohme-Meier F.
      Energy expenditure of grazing cows and cows fed grass indoors as determined by the 13C bicarbonate dilution technique using an automatic blood sampling system.
      found a mean deviation between estimated and weighed DMI of 0.2 kg for cows fed in the barn with fresh herbage. No exact reference for the daily HDMI for individual animals exists under grazing conditions; thus, a genuine validation of the n-alkane method is not possible. Nevertheless, the n-alkane method is commonly suggested as one of the best available techniques for individual HDMI estimation of grazing animals (
      • Smit H.J.
      • Taweel H.Z.
      • Tas B.M.
      • Tamminga S.
      • Elgersma A.
      Comparison of techniques for estimating herbage intake of grazing dairy cows.
      ;
      • Decruyenaere V.
      • Buldgen A.
      • Stilmant D.
      Factors affecting intake by grazing ruminants and related quantification methods: A review.
      ;
      • Pérez-Ramírez E.
      • Peyraud J.L.
      • Delagarde R.
      N-alkanes v. ytterbium/faecal index as two methods for estimating herbage intake of dairy cows fed on diets differing in the herbage:maize silage ratio and feeding level.
      ). However, there are weak points that may impede an accurate HDMI estimation.
      • Lippke H.
      Estimation of forage intake by ruminants on pasture.
      mentioned the diurnal variation of the n-alkane concentration in feces as one possible source of variation. To prevent variations, we administered HC32 on an apple pomace carrier twice daily. A further challenge was the collection of a representative sample of the herbage eaten by the grazing dairy cows during the day, especially in multispecies swards. Grazing cows can select for certain plant groups and graze on layers that may contain different alkane concentrations (
      • Dove H.
      • Mayes R.W.
      Using n-alkanes and other plant wax components to estimate intake, digestibility and diet composition of grazing/browsing sheep and goats.
      ;
      • Heublein C.
      • Südekum K.-H.
      • Gill F.
      • Dohme-Meier F.
      • Schori F.
      Using plant wax markers to estimate the diet composition of grazing Holstein dairy cows.
      ). To minimize this risk, we took herbage samples twice daily by imitating the feed selection of almost each experimental cow on pasture and pooled these. To prevent varying alkane dosing, as mentioned by
      • Smit H.J.
      • Taweel H.Z.
      • Tas B.M.
      • Tamminga S.
      • Elgersma A.
      Comparison of techniques for estimating herbage intake of grazing dairy cows.
      , HC32 was not applied on feeds or concentrates; instead, capsules containing the exact intended amount of HC32 were administered twice daily with a bolus gun.
      Comparing our HDMI results generated with the n-alkane method [11.9 ± 3.31 (SD) kg of HDMI], the reference method, with those obtained from estimation equations for grazing dairy cows proposed by
      • Faverdin P.
      • Delagarde R.
      • Delaby L.
      • Meschy F.
      Alimentation des vaches laitieres.
      using our data set [13.3 ± 3.72 (SD) kg of HDMI], an R2 of 0.77 (y = 1.178 + 0.962x; standard error of estimate = 1.76) was obtained (Figure 2). This suggests an acceptable correlation and general accordance of our reference method for HDMI, as no genuine or much better method exists.
      Figure thumbnail gr2
      Figure 2Comparison of our herbage DMI results generated with the n-alkane method and those obtained from estimation equations proposed by
      • Faverdin P.
      • Delagarde R.
      • Delaby L.
      • Meschy F.
      Alimentation des vaches laitieres.
      ; n = 30).

      Variables Used for HDMI Estimation Models

      Behavioral characteristics are useful for HDMI estimation in many cases, although the comparison of RMSPE of the GA and GAwRW models showed no improvement in accuracy of the HDMI estimation. The importance of behavioral characteristics for HDMI estimation is shown under an approach without knowledge of the supplements fed because total eating chews, total eating time performed at pasture, prehension bites performed on pasture, and total prehension bites represent the most important variables in the WSB models. Compared with the WSBwRW models, WSB models had a lower RMSPE on average by 0.58 kg/cow per day. Clearly, behavioral characteristics seem to improve the accuracy of HDMI estimation if information about the amount of supplements fed in the barn to pasture is lacking.
      • Chacon E.
      • Stobbs T.H.
      • Sandland R.L.
      Estimation of herbage consumption by grazing cattle using measurements of eating behaviour.
      already concluded that herbage intake of grazing cattle can be estimated with reasonable precision using eating bites and bite size based on esophageal fistula samples. The challenge would be to know the bite size with intact animals. Later,
      • Halachmi I.
      • Meir Y.B.
      • Miron J.
      • Maltz E.
      Feeding behavior improves prediction of dairy cow voluntary feed intake but cannot serve as the sole indicator.
      investigated the effect of behavioral characteristics in DMI models for intact, TMR-fed cows housed in open, no-stall cowsheds. They found an improvement of 1.3 kg/cow per day on RMSPE, even greater than that in our study, when behavioral characteristics were incorporated in a DMI estimation model. Additional benefits of including behavioral characteristics in an intake estimation model might be the detection of sick or injured cows as well as cows in heat, as short-term feeding behavior is modified in characteristic ways in such cases (
      • González L.A.
      • Tolkamp B.J.
      • Coffey M.P.
      • Ferret A.
      • Kyriazakis I.
      Changes in feeding behavior as possible indicators for the automatic monitoring of health disorders in dairy cows.
      ).
      • Andriamandroso A.L.H.
      • Bindelle J.
      • Mercatoris B.
      • Lebeau F.
      A review on the use of sensors to monitor cattle jaw movements and behavior when grazing.
      advocated the potential use of behavioral characteristics for HDMI estimation, as the bite is the elementary and indivisible unit of the whole grazing process. This is also seen in the present study, as according to β the variables prehension bites performed at pasture and total prehension bites were the 2 most important variables for the WSB models. An explanation for opposite effects of prehension bites performed at pasture and total prehension bites on HDMI might be the effect of supplements fed in the barn on HDMI (substitution of herbage); this also applies to the contrasting effects of total eating chews and total eating time at pasture.
      Although behavioral characteristics represent the most important variables in the WSB model, a large variation in bite mass exists among animals, grazing, pasture, and feeding managements. Stated values range between 0.33 and 0.74 g of DM/bite (
      • Barrett P.D.
      • Laidlaw A.S.
      • Mayne C.S.
      • Christie H.
      Pattern of herbage intake rate and bite dimensions of rotationally grazed dairy cows as sward height declines.
      ;
      • Penning P.D.
      • Rutter S.M.
      Ingestive behaviour.
      ). This leads to an insufficient correlation between behavioral characteristics and HDMI. According to
      • Hellwing A.L.F.
      • Lund P.
      • Weisbjerg M.R.
      • Oudshoorn F.W.
      • Munksgaard L.
      • Kristensen T.
      Comparison of methods for estimating herbage intake in grazing dairy cows.
      , grazing activity displays a close correlation with the predicted intake, yet further characteristics may have an influence on herbage intake. Therefore, in addition to behavioral data, animal, feed, grazing, and environmental variables must be included to reduce the estimation error of HDMI estimation models.
      In our study, postgrazing herbage mass (222 kg of DM/ha; minimum: 63 kg of DM/ha; maximum: 554 kg of DM/ha) had a negative effect (−0.004 kg of DMI/d per kg of DM per ha) on HDMI. Contradicting results were found by
      • O'Neill B.F.
      • Lewis E.
      • O'Donovan M.
      • Shalloo L.
      • Galvin N.
      • Mulligan F.
      • Boland T.
      • Delagarde R.
      Predicting grass dry matter intake, milk yield and milk fat and protein yield of spring calving grazing dairy cows during the grazing season.
      with no or a positive effect of postgrazing sward height up to 0.99 kg of DMI/cm per day. As postgrazing sward height or postgrazing herbage mass are a function of pregrazing herbage mass and daily herbage allowance, differences in these factors as well as their interaction might have been partly the source of the contrasting results. Also, herbage CP content showed a negative effect on HDMI. This contradicts the findings of
      • Timmer B.
      • Zom R.L.G.
      • Holshof G.
      • Spithoven M.
      • Van Reenen C.G.
      The application of behavior sensors and sward measurement to support grazing management.
      , who found a positive effect of milk urea content, which is associated with the intake of either CP or ruminally degraded CP. Results similar to those of
      • Timmer B.
      • Zom R.L.G.
      • Holshof G.
      • Spithoven M.
      • Van Reenen C.G.
      The application of behavior sensors and sward measurement to support grazing management.
      were found for milk protein content; thus, a higher HDMI is correlated with a higher milk protein content.
      • Timmer B.
      • Zom R.L.G.
      • Holshof G.
      • Spithoven M.
      • Van Reenen C.G.
      The application of behavior sensors and sward measurement to support grazing management.
      likewise found an effect of 3.79 kg of DMI/% of milk protein per day. The partially contradicting and partially similar results of our models and those of
      • Timmer B.
      • Zom R.L.G.
      • Holshof G.
      • Spithoven M.
      • Van Reenen C.G.
      The application of behavior sensors and sward measurement to support grazing management.
      might be caused by the interactions between the variables used, as discussed by

      Gruber, L., F. J. Schwarz, D. Erdin, B. Fischer, H. Spiekers, H. Steingass, U. Meyer, A. Chassot, T. Jilg, A. Omermaier, and T. Gruggenberg. 2005. Vorhersage der Futteraufnahme von Milchkühen Datenbasis von 10 Forschungs- und Universitätsinstituten Deutschlands, Österreichs and der Schweiz. 116. VDLUFA-Kongress, Rostock, Germany.

      .
      Amounts of concentrate and protein supplement were merged into a single variable because the substitution rates are similar and lower than those for a forage supplement (chopped whole-plant corn silage;
      • Delagarde R.
      • O'Donovan M.
      Modelling of herbage intake and milk production by grazing dairy cows.
      ). The feeding of protein and concentrate supplements was associated with a decrease of HDMI in the GA model.
      • O'Neill B.F.
      • Lewis E.
      • O'Donovan M.
      • Shalloo L.
      • Galvin N.
      • Mulligan F.
      • Boland T.
      • Delagarde R.
      Predicting grass dry matter intake, milk yield and milk fat and protein yield of spring calving grazing dairy cows during the grazing season.
      observed a substitution rate of 0.58 to 0.71 kg of HDMI/kg of concentrate compared with our results (0.63 kg of HDMI/kg of concentrate). An even smaller reduction of 0.36 kg of HDMI/kg of concentrate was found for early-lactation dairy cows (
      • McEvoy M.
      • Delaby L.
      • Kennedy E.
      • Boland T.
      • O'Donovan M.
      Early lactation dairy cows: Development of equations to predict intake and milk performance at grazing.
      ). The greater substitution rate in our studies could have been the result of a generous herbage allowance (
      • Penno J.
      • Macdonald K.
      • Holmes C.W.
      • Davis S.
      • Wilson G.
      • Brookes I.M.
      • Thom E.
      Responses to supplementation by dairy cows given low pasture allowances in different seasons 1. Pasture intake and substitution.
      ) in most of our studies. Thus, previous studies showed a substitution rate of zero for a high grazing pressure up to 0.6 to 0.8 for a low grazing pressure (
      • Stockdale C.
      Levels of pasture substitution when concentrates are fed to grazing dairy cows in northern Victoria.
      ;
      • Peyraud J.L.
      • Delaby L.
      • Garnsworthy P.
      • Wiseman J.
      Ideal concentrate feeds for grazing dairy cows—Responses to supplementation in interaction with grazing management and grass quality.
      ). Besides the generous herbage allowance, an increased amount of supplemented feed (
      • Penno J.
      • Macdonald K.
      • Holmes C.W.
      • Davis S.
      • Wilson G.
      • Brookes I.M.
      • Thom E.
      Responses to supplementation by dairy cows given low pasture allowances in different seasons 1. Pasture intake and substitution.
      ) or lactation stage could have had an influence; thus, cows in early lactation showed a lower substitution rate compared with cows in mid or late lactation (
      • Penno J.
      • Macdonald K.
      • Holmes C.W.
      • Davis S.
      • Wilson G.
      • Brookes I.M.
      • Thom E.
      Responses to supplementation by dairy cows given low pasture allowances in different seasons 1. Pasture intake and substitution.
      ). Furthermore, the substitution rate is greater when the quality of pasture allows cows to attain greater DMI from pasture alone (
      • Penno J.
      • Macdonald K.
      • Holmes C.W.
      • Davis S.
      • Wilson G.
      • Brookes I.M.
      • Thom E.
      Responses to supplementation by dairy cows given low pasture allowances in different seasons 1. Pasture intake and substitution.
      ).
      Other reasons for the high substitution rate might be that the interaction between variables used in the models affected the coefficients, as discussed by

      Gruber, L., F. J. Schwarz, D. Erdin, B. Fischer, H. Spiekers, H. Steingass, U. Meyer, A. Chassot, T. Jilg, A. Omermaier, and T. Gruggenberg. 2005. Vorhersage der Futteraufnahme von Milchkühen Datenbasis von 10 Forschungs- und Universitätsinstituten Deutschlands, Österreichs and der Schweiz. 116. VDLUFA-Kongress, Rostock, Germany.

      , and maybe some bias due to different conditions between experiments. For example, the average of the calculated substitution rate between supplemented and not-supplemented pairs within experiments [0.63 ± 0.015 (SD) kg of HDMI/kg of concentrate or protein supplement] was lower compared with the coefficient of the explanatory variable protein and concentrate supplements (1.09 kg of DM/d) in GA7. The calculated substitution rate is similar to the findings of
      • Stockdale C.
      Levels of pasture substitution when concentrates are fed to grazing dairy cows in northern Victoria.
      and
      • Peyraud J.L.
      • Delaby L.
      • Garnsworthy P.
      • Wiseman J.
      Ideal concentrate feeds for grazing dairy cows—Responses to supplementation in interaction with grazing management and grass quality.
      . The decrease in HDMI caused by the supplementation of chopped whole-plant corn silage was smaller compared with that found for concentrate and protein supplementation. However, the coefficients (0.64 kg of DM/d) in this case are comparable with the calculated substitution rate [0.47 ± 0.015 (SD) kg of HDMI/kg of chopped whole-plant corn silage supplement]. Usually, substitution rates resulting from supplementing forage to pasture are greater than those from supplementing concentrates due to the greater forage fill (unité d'encombrement;
      • Faverdin P.
      • Delagarde R.
      • Delaby L.
      • Meschy F.
      Alimentation des vaches laitieres.
      ) value representing rumen fill (
      • Delagarde R.
      • O'Donovan M.
      Modelling of herbage intake and milk production by grazing dairy cows.
      ). Different experiments were used to calculate substitution rates. Varying environmental, feed, and animal characteristics may have therefore caused inconsistent substitution rates between concentrates and forage even though the individual substitution rates were in an expected range.
      As a measure for body size and an indicator for rumen size, BW was positively correlated with HDMI in our study (i.e., an increase of 0.8 to 1 kg/100 kg of BW for our models). Similar increases of 0.95 and 1.3 kg/100 kg of BW were measured by
      • Delagarde R.
      • O'Donovan M.
      Modelling of herbage intake and milk production by grazing dairy cows.
      and

      Gruber, L., F. J. Schwarz, D. Erdin, B. Fischer, H. Spiekers, H. Steingass, U. Meyer, A. Chassot, T. Jilg, A. Omermaier, and T. Gruggenberg. 2005. Vorhersage der Futteraufnahme von Milchkühen Datenbasis von 10 Forschungs- und Universitätsinstituten Deutschlands, Österreichs and der Schweiz. 116. VDLUFA-Kongress, Rostock, Germany.

      . A considerably greater increase in grass DMI, 2.0 kg/100 kg of BW (
      • McEvoy M.
      • Delaby L.
      • Kennedy E.
      • Boland T.
      • O'Donovan M.
      Early lactation dairy cows: Development of equations to predict intake and milk performance at grazing.
      ), may be related to the early lactation stage of their cows compared with our study.

      Gruber, L., F. J. Schwarz, D. Erdin, B. Fischer, H. Spiekers, H. Steingass, U. Meyer, A. Chassot, T. Jilg, A. Omermaier, and T. Gruggenberg. 2005. Vorhersage der Futteraufnahme von Milchkühen Datenbasis von 10 Forschungs- und Universitätsinstituten Deutschlands, Österreichs and der Schweiz. 116. VDLUFA-Kongress, Rostock, Germany.

      showed a greater increase in DMI per 100 kg of BW for cows in early lactation compared with cows in mid or late lactation.
      Besides the physical feed intake capacity of the cow, proxied by BW, total DMI seems to be one driving force for milk production and vice versa, a conclusion that has been drawn in several other studies (
      • Delagarde R.
      • O'Donovan M.
      Modelling of herbage intake and milk production by grazing dairy cows.
      ;

      Gruber, L., F. J. Schwarz, D. Erdin, B. Fischer, H. Spiekers, H. Steingass, U. Meyer, A. Chassot, T. Jilg, A. Omermaier, and T. Gruggenberg. 2005. Vorhersage der Futteraufnahme von Milchkühen Datenbasis von 10 Forschungs- und Universitätsinstituten Deutschlands, Österreichs and der Schweiz. 116. VDLUFA-Kongress, Rostock, Germany.

      ;
      • McEvoy M.
      • Delaby L.
      • Kennedy E.
      • Boland T.
      • O'Donovan M.
      Early lactation dairy cows: Development of equations to predict intake and milk performance at grazing.
      ). The coefficient of the explanatory variable milk yield was between 0.13 (WSB8) and 0.35 (GA7) kg of HDMI/kg of milk per day. A reason for differences in the coefficients for milk yield between GA7 and WSB8 might be the aforementioned interaction between the variables used in the models.

      Precision of the HDMI Estimation Models

      The models GA7 and WSB8 showed the highest accuracy in estimating HDMI, with an R2 of 0.81 and 0.73 and an RPE of 11.1 and 13.2%, respectively. The explained proportion of variation of the HDMI and the accuracy of these models are similar to those of other investigations. The models of
      • McEvoy M.
      • Delaby L.
      • Kennedy E.
      • Boland T.
      • O'Donovan M.
      Early lactation dairy cows: Development of equations to predict intake and milk performance at grazing.
      accounted for 79% of the HDMI variation for early-lactating dairy cows, and those of
      • Timmer B.
      • Zom R.L.G.
      • Holshof G.
      • Spithoven M.
      • Van Reenen C.G.
      The application of behavior sensors and sward measurement to support grazing management.
      accounted for 84% of the HDMI variation.
      • Delagarde R.
      • O'Donovan M.
      Modelling of herbage intake and milk production by grazing dairy cows.
      obtained an RPE between 10 and 25% for the HDMI of grazing dairy cows. Likewise,
      • Keady T.
      • Mayne C.
      • Kilpatrick D.
      An evaluation of five models commonly used to predict food intake of lactating dairy cattle.
      found an RPE for grass silage DMI between 10 and 20% for dairy cows fed indoors. Overall, the RPE of all of our models, developed for use under GA (RPE = 10.8–13.1%) as well as under WSB (RPE = 13.3–14.9%), showed an acceptable prediction of HDMI using detectable variables.
      • Fuentes-Pila J.
      • DeLorenzo M.
      • Beede D.
      • Staples C.
      • Holter J.
      Evaluation of equations based on animal factors to predict intake of lactating Holstein cows.
      suggested that an RPE value lower than 10% indicates a satisfactory prediction of the DMI, whereas an RPE between 10 and 20% indicates a relatively good or acceptable prediction, and an RPE greater than 20% indicates an unsatisfactory prediction. According to
      • Delagarde R.
      • Valk H.
      • Mayne C.S.
      • Rook A.J.
      • González-Rodríguez A.
      • Baratte C.
      • Faverdin P.
      • Peyraud J.L.
      GrazeIn: A model of herbage intake and milk production for grazing dairy cows. 3. Simulations and external validation of the model.
      , a threshold for good intake estimation models for practical use should be a mean prediction error of 10%. Nevertheless, this threshold is hard to achieve, particularly for grazing dairy cows, as the available reference methods for development and validation of the HDMI estimation models are indirect methods. Furthermore,
      • Galyean M.
      Predicting forage intake in extensive grazing systems.
      argued that daily intake of cattle is naturally variable, and

      Gruber, L., F. J. Schwarz, D. Erdin, B. Fischer, H. Spiekers, H. Steingass, U. Meyer, A. Chassot, T. Jilg, A. Omermaier, and T. Gruggenberg. 2005. Vorhersage der Futteraufnahme von Milchkühen Datenbasis von 10 Forschungs- und Universitätsinstituten Deutschlands, Österreichs and der Schweiz. 116. VDLUFA-Kongress, Rostock, Germany.

      stated that a biologically determined natural scattering of about 10% might also impede the development of reliable intake estimation models.

      CONCLUSIONS

      With the HDMI estimation models developed in this study, HDMI could be estimated with an RPE between 11.1% (GA7) and 13.2% (WSB8) at best. When the amount of supplements fed was unknown, the inclusion of behavioral characteristics in the model showed a clear added value regarding the accuracy of the individual HDMI estimation. On the other hand, the inclusion of behavioral characteristics in the HDMI estimation model when amount of supplements fed was known did not significantly reduce the RMSPE. Finally, a larger data set (in terms of number of animals and herds and covering more breeds, different grazing systems, and management patterns, including, for example, varying supplementation) will be needed for further development and validation of a robust HDMI estimation model.

      REFERENCES

        • Agroscope
        Fütterungsempfehlungen und Nährwerttabellen für Wiederkäuer (Feeding recommendations and nutrient tables for ruminants).
        • Agroscope
        Fütterungsempfehlungen und Nährwerttabellen für Wiederkäuer (Feeding recommendations and nutrient tables for ruminants).
        • Andriamandroso A.L.H.
        • Bindelle J.
        • Mercatoris B.
        • Lebeau F.
        A review on the use of sensors to monitor cattle jaw movements and behavior when grazing.
        Biotechnol. Agron. Soc. Environ. 2016; 20: 273-286
        • AOAC International
        Official Methods of Analysis.
        16th ed. AOAC International, Arlington, VA1995
        • Arnott G.
        • Ferris C.
        • O'Connell N.
        Review: Welfare of dairy cows in continuously housed and pasture-based production systems.
        Animal. 2017; 11 (27364762): 261-273
        • Barrett P.D.
        • Laidlaw A.S.
        • Mayne C.S.
        • Christie H.
        Pattern of herbage intake rate and bite dimensions of rotationally grazed dairy cows as sward height declines.
        Grass Forage Sci. 2001; 56: 362-373
        • Berry N.R.
        • Scheeder M.R.
        • Sutter F.
        • Kröber T.F.
        • Kreuzer M.
        The accuracy of intake estimation based on the use of alkane controlled-release capsules and faeces grab sampling in cows.
        Ann. Zootech. 2000; 49: 3-13
        • Chacon E.
        • Stobbs T.H.
        • Sandland R.L.
        Estimation of herbage consumption by grazing cattle using measurements of eating behaviour.
        Grass Forage Sci. 1976; 31: 81-87
        • Decruyenaere V.
        • Buldgen A.
        • Stilmant D.
        Factors affecting intake by grazing ruminants and related quantification methods: A review.
        Biotechnol. Agron. Soc. Environ. 2009; 13: 559-573
        • Delagarde R.
        • O'Donovan M.
        Modelling of herbage intake and milk production by grazing dairy cows.
        in: Murphy J.J. Utilisation of Grazed Grass in Temperate Animal Systems: Proc. Satellite Workshop XXth International Grassland Congress, Cork, Ireland. Wageningen Academic Publishers, Wageningen, the Netherlands2005: 89-104
        • Delagarde R.
        • Valk H.
        • Mayne C.S.
        • Rook A.J.
        • González-Rodríguez A.
        • Baratte C.
        • Faverdin P.
        • Peyraud J.L.
        GrazeIn: A model of herbage intake and milk production for grazing dairy cows. 3. Simulations and external validation of the model.
        Grass Forage Sci. 2011; 66: 61-77
        • Dove H.
        • Mayes R.W.
        Using n-alkanes and other plant wax components to estimate intake, digestibility and diet composition of grazing/browsing sheep and goats.
        Small Rumin. Res. 2005; 59: 123-139
        • Efron B.
        • Hastie T.
        The jackknife and the bootstrap and bootstrap confidence intervals.
        in: Efron B. Hastie T. Computer Age Statistical Inference. Cambridge University Press, Cambridge, UK2016: 155-204
        • Faverdin P.
        • Delagarde R.
        • Delaby L.
        • Meschy F.
        Alimentation des vaches laitieres.
        in: Alimentation des Bovins, Ovins et Caprins: Besoins des Animaux-Valeur des Aliments. Editions Quae, Versailles Cedex, France2007: 23-57
        • Fox J.
        The R Commander: A basic statistics graphical user interface to R.
        J. Stat. Softw. 2005; 14: 1-42
        • Fox J.
        Using the R Commander: A Point-and-Click Interface for R.
        CRC Press, Boca Raton, FL2017
        • Fox J.
        • Bouchet-Valat M.
        Rcmdr: R Commander. R package version 2.3-2.
        https://CRAN.R-project.org/package=rms
        Date: 2017
        Date accessed: July 7, 2016
        • Fuentes-Pila J.
        • DeLorenzo M.
        • Beede D.
        • Staples C.
        • Holter J.
        Evaluation of equations based on animal factors to predict intake of lactating Holstein cows.
        J. Dairy Sci. 1996; 79 (8899522): 1562-1571
        • Galyean M.
        Predicting forage intake in extensive grazing systems.
        J. Anim. Sci. 2016; 94: 26-43
        • Getter K.L.
        • Behe B.K.
        • Howard P.H.
        • Conner D.S.
        • Spaniolo L.M.
        Increasing demand for pasture-based dairy: What attributes and images do consumers want?.
        in: Freyer B. Bingen J. Re-Thinking Organic Food and Farming in a Changing World. Springer, Berlin, Germany2015: 125-140
        • González L.A.
        • Tolkamp B.J.
        • Coffey M.P.
        • Ferret A.
        • Kyriazakis I.
        Changes in feeding behavior as possible indicators for the automatic monitoring of health disorders in dairy cows.
        J. Dairy Sci. 2008; 91 (18292258): 1017-1028
      1. Gruber, L., F. J. Schwarz, D. Erdin, B. Fischer, H. Spiekers, H. Steingass, U. Meyer, A. Chassot, T. Jilg, A. Omermaier, and T. Gruggenberg. 2005. Vorhersage der Futteraufnahme von Milchkühen Datenbasis von 10 Forschungs- und Universitätsinstituten Deutschlands, Österreichs and der Schweiz. 116. VDLUFA-Kongress, Rostock, Germany.

        • Guyader J.
        • Janzen H.
        • Kroebel R.
        • Beauchemin K.
        Forage use to improve environmental sustainability of ruminant production.
        J. Anim. Sci. 2016; 94 (27695772): 3147-3158
        • Halachmi I.
        • Meir Y.B.
        • Miron J.
        • Maltz E.
        Feeding behavior improves prediction of dairy cow voluntary feed intake but cannot serve as the sole indicator.
        Animal. 2016; 10 (26387522): 1501-1506
        • Harrell F.E.
        Regression Modeling Strategies: With Applications to Linear Models, Logistic and Ordinal Regression, and Survival Analysis.
        2nd ed. Springer International, Cham, Switzerland2015
        • Harrell F.E.
        rms: Regression Modeling Strategies. R package version 5.1-0.
        https://CRAN.R-project.org/package=rms
        Date: 2017
        Date accessed: July 7, 2016
        • Hellwing A.L.F.
        • Lund P.
        • Weisbjerg M.R.
        • Oudshoorn F.W.
        • Munksgaard L.
        • Kristensen T.
        Comparison of methods for estimating herbage intake in grazing dairy cows.
        Livest. Sci. 2015; 176: 61-74
        • Heublein C.
        • Südekum K.-H.
        • Gill F.
        • Dohme-Meier F.
        • Schori F.
        Using plant wax markers to estimate the diet composition of grazing Holstein dairy cows.
        J. Dairy Sci. 2017; 100 (27939548): 1019-1036
        • Holshof G.
        • Evers A.
        • De Haan M.
        • Galama P.
        Grazing and difficult circumstances: Economic benefits depend on milk price and grazing efficiency.
        in: van den Pol-van Dasselaar A. van den Aarts H.F.M. De Vliegher A. Elgersma A. Reheul D. Reijneveld J.A. Verloop J. Hopkins A. Proc. EGF, Grassland Science in Europe, Trondheim, Norway. Wageningen Academic Publishers, Wageningen, the Netherlands2015: 236-238
        • ISO (International Organization for Standardization)
        Food products—Determination of the total nitrogen content by combustion according to the Dumas principle and calculation of the crude protein content. Method 16634-1.
        ISO, Geneva, Switzerland2008
        • ISO (International Organization for Standardization)
        Animal feeding stuffs—Determination of calcium, sodium, phosphorus, magnesium, potassium, iron, zinc, copper, manganese, cobalt, molybdenum, arsenic, lead and cadmium by ICP-AES. Method 27085.
        ISO, Geneva, Switzerland2009
        • Kaufmann L.D.
        • Münger A.
        • Rérat M.
        • Junghans P.
        • Görs S.
        • Metges C.C.
        • Dohme-Meier F.
        Energy expenditure of grazing cows and cows fed grass indoors as determined by the 13C bicarbonate dilution technique using an automatic blood sampling system.
        J. Dairy Sci. 2011; 94 (21426990): 1989-2000
        • Keady T.
        • Mayne C.
        • Kilpatrick D.
        An evaluation of five models commonly used to predict food intake of lactating dairy cattle.
        Livest. Prod. Sci. 2004; 89: 129-138
        • Kristensen T.
        • Madsen M.L.
        • Noe E.
        The use of grazing in intensive dairy production and assessment of farmers' attitude towards grazing.
        in: Schnyder H. Isselstein J. Taube F. Auerswald K. Schellberg J. Wachendorf M. Herrmann A. Gierus M. Wrage N. Hopkins A. Proc. EGF, Grassland Science in Europe, Kiel, Germany. Mecke Druck und Verlag, Duderstadt, Germany2010: 964-966
        • Lippke H.
        Estimation of forage intake by ruminants on pasture.
        Crop Sci. 2002; 42: 869-872
        • Lumley T.
        • Miller A.
        Leaps: Regression subset selection. R package version 3.0.
        https://CRAN.R-project.org/package=leaps
        Date: 2017
        Date accessed: January 10, 2017
        • Mayes R.W.
        • Lamb C.S.
        • Colgrove P.M.
        The use of dosed and herbage n-alkanes as markers for the determination of herbage intake.
        J. Agric. Sci. 1986; 107: 161-170
        • McEvoy M.
        • Delaby L.
        • Kennedy E.
        • Boland T.
        • O'Donovan M.
        Early lactation dairy cows: Development of equations to predict intake and milk performance at grazing.
        Livest. Sci. 2009; 122: 214-221
        • O'Callaghan T.F.
        • Hennessy D.
        • McAuliffe S.
        • Kilcawley K.N.
        • O'Donovan M.
        • Dillon P.
        • Ross R.P.
        • Stanton C.
        Effect of pasture versus indoor feeding systems on raw milk composition and quality over an entire lactation.
        J. Dairy Sci. 2016; 99 (27720161): 9424-9440
        • O'Neill B.F.
        • Lewis E.
        • O'Donovan M.
        • Shalloo L.
        • Galvin N.
        • Mulligan F.
        • Boland T.
        • Delagarde R.
        Predicting grass dry matter intake, milk yield and milk fat and protein yield of spring calving grazing dairy cows during the grazing season.
        Animal. 2013; 7 (23570842): 1379-1389
        • Oudshoorn F.W.
        • Cornou C.
        • Hellwing A.L.F.
        • Hansen H.H.
        • Munksgaard L.
        • Lund P.
        • Kristensen T.
        Estimation of grass intake on pasture for dairy cows using tightly and loosely mounted di-and tri-axial accelerometers combined with bite count.
        Comput. Electron. Agric. 2013; 99: 227-235
        • Penning P.D.
        • Rutter S.M.
        Ingestive behaviour.
        in: Penning P.D. Herbage Intake Handbook. British Grassland Society, Maidenhead, UK2004: 1-191
        • Penno J.
        • Macdonald K.
        • Holmes C.W.
        • Davis S.
        • Wilson G.
        • Brookes I.M.
        • Thom E.
        Responses to supplementation by dairy cows given low pasture allowances in different seasons 1. Pasture intake and substitution.
        Anim. Sci. 2006; 82: 661-670
        • Pérez-Ramírez E.
        • Peyraud J.L.
        • Delagarde R.
        N-alkanes v. ytterbium/faecal index as two methods for estimating herbage intake of dairy cows fed on diets differing in the herbage:maize silage ratio and feeding level.
        Animal. 2012; 6 (22436181): 232-244
        • Peyraud J.L.
        • Delaby L.
        • Garnsworthy P.
        • Wiseman J.
        Ideal concentrate feeds for grazing dairy cows—Responses to supplementation in interaction with grazing management and grass quality.
        in: Garnsworthy P. Wiseman J. Recent Advances in Animal Nutrition. Nottingham University Press, Nottingham, UK2001: 203-220
        • R Core Team
        R: A Language and Environment for Statistical Computing.
        R Foundation for Statistical Computing, Vienna, Austria2016
        • Rombach M.
        • Münger A.
        • Niederhauser J.
        • Südekum K.-H.
        • Schori F.
        Evaluation and validation of an automatic jaw movement recorder (RumiWatch) on unsupplemented or supplemented dairy cows on pasture.
        J. Dairy Sci. 2018; 101 (29290426): 2463-2475
        • Ruuska S.
        • Kajava S.
        • Mughal M.
        • Zehner N.
        • Mononen J.
        Validation of a pressure sensor-based system for measuring eating, rumination and drinking behaviour of dairy cattle.
        Appl. Anim. Behav. Sci. 2016; 174: 19-23
        • Smit H.J.
        • Taweel H.Z.
        • Tas B.M.
        • Tamminga S.
        • Elgersma A.
        Comparison of techniques for estimating herbage intake of grazing dairy cows.
        J. Dairy Sci. 2005; 88 (15829676): 1827-1836
        • Stockdale C.
        Levels of pasture substitution when concentrates are fed to grazing dairy cows in northern Victoria.
        Anim. Prod. Sci. 2000; 40: 913-921
        • Thanner S.
        • Dohme-Meier F.
        • Görs S.
        • Metges C.
        • Bruckmaier R.
        • Schori F.
        The energy expenditure of 2 Holstein cow strains in an organic grazing system.
        J. Dairy Sci. 2014; 97 (24630659): 2789-2799
        • Timmer B.
        • Zom R.L.G.
        • Holshof G.
        • Spithoven M.
        • Van Reenen C.G.
        The application of behavior sensors and sward measurement to support grazing management.
        in: Höglind M. Bakken A.K. Hovstad K.A. Kallioniemi E. Riley H. Steinshamn H. Østrem L. Proc. EGF, Grassland Science in Europe, Trondheim, Norway. Wageningen Academic Publishers, Wageningen, the Netherlands2016: 80-82
        • Van den Pol A.
        • de Vliegher A.
        • Hennessy D.
        • Isselstein J.
        • Peyraud J.
        The future of grazing.
        https://www.europeangrassland.org/
        Date: 2015
        Date accessed: May 10, 2017