Prediction of sensory quality of cheese during ripening from chemical and spectroscopy measurements

Accepted Manuscript Prediction of sensory quality of cheese during ripening from chemical and spectroscopy measurements Hilde Kraggerud, Tormod Næs, Roger K. Abrahamsen PII:

S0958-6946(13)00185-4

DOI:

10.1016/j.idairyj.2013.07.008

Reference:

INDA 3554

To appear in:

International Dairy Journal

Received Date: 3 October 2011 Revised Date:

8 July 2013

Accepted Date: 9 July 2013

Please cite this article as: Kraggerud, H., Næs, T., Abrahamsen, R.K., Prediction of sensory quality of cheese during ripening from chemical and spectroscopy measurements, International Dairy Journal (2013), doi: 10.1016/j.idairyj.2013.07.008. This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.

ACCEPTED MANUSCRIPT

1

Prediction of sensory quality of cheese during ripening from chemical and

2

spectroscopy measurements

3

5 6 Hilde Kraggerud a,b*, Tormod Næsa, Roger K. Abrahamsena

SC

7 8

M AN U

9 10 11 12

a

13

and Food Science, N-1432 Ås, Norway

14

b

15

Norway.

Norwegian University of Life Sciences, Department of Chemistry, Biotechnology

TE D

TINE R&D Center, Box 8034, N-4068 Stavanger, Norway, 3 Nofima, N-1430 Ås,

16

EP

17 18

* Corresponding author. Tel.: +47 95230454

20

E-mail address: [email protected] (H. Kraggerud)

22

AC C

19

21

RI PT

4

23

1

ACCEPTED MANUSCRIPT

24 25

___________________________________________________________________

26

Abstract

27 Norvegia cheese samples (459 in total) were analysed during maturation using

RI PT

28

a broad range of analytical methods. From eight to 24 and 40 weeks there was a

30

highly systematic development in chemical and sensory attributes. Chemical data

31

and Fourier transform infrared (FTIR) measurements gave almost equivalent

32

validation results in prediction of sensory data. Fluorescence spectroscopy and near

33

infrared spectroscopy, both carried out on the surface of cheese, gave slightly less

34

valid results than FTIR. Using a combination of spectra from all instruments gave

35

higher correlations than spectra from instruments taken one by one. Replacement of

36

sensory measurements would hardly be possible, but chemical analysis can be

37

replaced by spectroscopy to a large extent. Sensory characteristics at a greater age

38

(40 weeks) were not very well forecasted from early measurements on cheese (eight

39

weeks), which could have been useful for prediction of quality development during

40

cheese maturation.

41

___________________________________________________________________

AC C

EP

TE D

M AN U

SC

29

2

ACCEPTED MANUSCRIPT

42 43

1.

Introduction

44 45

In quality control, predicting product quality as early as possible in the production process is of major interest for all products. For a product such as cheese

47

with a long ripening period before the product is ready for sale, it is especially

48

interesting to be able to predict the product quality at an early stage, months before

49

the actual consumption of the food. The ultimate goal is to be able to predict the

50

sensory quality – which is the nearest one could come to consumers’ experience of

51

the product – with non-destructive rapid methods.

SC

M AN U

52

RI PT

46

The sensory characteristics of cheese are very complex and are, among other factors, a result of the ripening process of the particular cheese variety. Initial milk

54

quality and composition and the particular cheese making process used result in a

55

defined chemical composition of the fresh cheese. Biochemical processes such as

56

proteolysis, glycolysis and lipolysis are also of importance. Factors that affect the

57

quality of cheese have been extensively reviewed (Collins, McSweeney, & Wilkinson,

58

2003; Delahunty & Drake, 2004; Fox & Cogan, 2004; McSweeney & Sousa, 2000).

59

The complex task of analysing cheese to determine composition and the

60

monitoring ripening by chemical and instrumental approaches have been reviewed

61

by Subramanian and Rodriguez-Saona (2010). Many attempts have been made to

62

predict chemical or sensory properties of cheese by spectroscopic methods. A

63

number of these have examined relatively few samples, often with large variation in

64

the sample set. Many studies are based on classifications and detailed examination

65

of electromagnetic spectra, whereas fewer are based on comparisons with reference

AC C

EP

TE D

53

3

ACCEPTED MANUSCRIPT

measurements. Often samples have considerable differences in the major chemical

67

components known to affect the ripening process (Beresford, Fitzsimons, Brennan, &

68

Cogan, 2001; Gilles & Lawrence, 1973; Johnson & Law, 1999). For use in industrial

69

quality control, the methods must be able to show small deviations from normal

70

composition.

71

RI PT

66

Characterisation of cheese during ripening using spectroscopic methods has been reviewed by several authors (Karoui & De Baerdemaeker, 2007; Karoui,

73

Mazerolles, & Dufour, 2003). Fourier Transform Infrared (FTIR) spectroscopy has

74

become more or less a standard method for rapid and accurate determination of

75

major chemical components in the dairy industry. The focus has been on analysis of

76

macromolecules in cheese. More recently, characterisation of ripening has also been

77

the subject of research (Subramanian, Alvarez, Harper, & Rodriguez-Saona, 2011).

78

Near infrared (NIR) spectroscopy has also been used for different components

M AN U

SC

72

and sensory attributes of cheese, reviewed by Subramanian and Rodriguez-Saona

80

(2010). Fluorescence spectroscopy has been widely used for prediction of light-

81

induced oxidation with reference measurements including sensory analysis, reviewed

82

by Andersen and Mortensen (2008). But fluorescence spectroscopy has also been

83

regarded as a promising tool for exploring texture development during ripening.

EP

TE D

79

Still, there is a challenge to establish valid methods of analysis for predicting

85

cheese quality for practical use in the dairy industry. Variation between production

86

runs normally is relatively small, but there is a need to detect the occurrence of

87

product deviations. Our approach attempted to simulate commercial production

88

conditions and a variety of situations. A large number of cheese samples (153) were

AC C

84

4

ACCEPTED MANUSCRIPT

analysed at three points during maturation (8, 24 and 40 weeks) making 459 samples

90

altogether. Variations in major chemical components were limited. Production was

91

made during all seasons of the year, with sampling and analysis at 20 points during a

92

period of 18 months. A broad range of chemical, sensory and spectroscopic methods

93

was applied in order to be able to compare a number of possible methods.

RI PT

89

94

The objectives of this study were to examine the development in sensory and

chemical parameters during cheese ripening, and to assess the feasibility of different

96

spectroscopic techniques (NIR, FTIR, fluorescence) for the monitoring of sensory

97

attributes. Relationships between chemical and sensory measurements, and the

98

possibility of early prediction of sensory quality development during ripening of

99

cheese, were also examined. Using realistic conditions with small variations in major components and a high number of samples was essential in this work.

101 2.

Materials and methods

TE D

102 103 104

2.1. Cheese sample selection

EP

105 106

M AN U

100

SC

95

A Dutch-type, semi-hard, rindless cheese variety, Norvegia, with approximately 46% fat in dry matter, was evaluated. During the first 3-4 weeks after

108

cheesemaking, all cheeses underwent ripening at set temperatures – varying from

109

plant to plant – and adjusted to acquire the desired quality. After 3-4 weeks from

110

cheese making, the cheese was matured at ≤ 4 °C.

111

AC C

107

Cheese included in this study came from three coordinated experiments.

5

ACCEPTED MANUSCRIPT

Experiment 1 varied raw milk quality and age in our large scale pilot plant (120 kg

113

cheese per batch). In experiment 2, samples from commercial production were

114

collected on the first day and manipulated with respect to ripening temperatures

115

(Kraggerud, Skeie, Høy, Røkke, & Abrahamsen, 2008). In experiment 3, random

116

samples were collected from commercial production at six different cheesemaking

117

plants. Each sample consisted of four vacuum packaged 5 kg cheeses, three of them

118

used for analysis at different ages. The samples were produced during one year, and

119

there were 12 samplings with four weeks between each. Samples were analysed

120

every four weeks for a period of 48 weeks, within an 18-month timeframe. A

121

production and sampling scheme is shown in Table 1.

122 123

2.2.

Sampling procedures

124

M AN U

SC

RI PT

112

Sampling of cheese for analysis was made in accordance with International

126

Dairy Federation Standard 50C (IDF, 1995). Spectroscopy, pH measurements and

127

sensory analyses were carried out within a few days at each sampling point. Fat, dry

128

matter and salt were only determined at one ripening point, after eight weeks, as they

129

are considered not to change during ripening. Other chemical analyses (not all

130

analyses were carried out at all sampling points) were carried out on grated samples

131

frozen at ≤-20 °C until analysis.

AC C

EP

TE D

125

6

ACCEPTED MANUSCRIPT

132 133

2.3.

Spectroscopy

134 2.3.1. NIR spectroscopy

RI PT

135

An NIRSystems 6500 scanning instrument was used for measuring NIR

137

spectra (FOSS NIRSystems Inc., Silver Spring, MD, USA). The samples were

138

scanned in a reflectance mode using a fibre-optic probe directly in contact with the

139

polyethylene covered surface of cheese samples. The instrument was operated in

140

the spectral region 400–1100 nm and measured in 2 nm steps. The ceramic plate of

141

the instrument was used as a reference for calibration. Sample temperature was

142

between 4 and 8 °C. Each cheese sample was scanned three times at different spots

143

on the cheese surface, the average absorbance spectrum was calculated

144

automatically and used for all further data analysis. Spectra were treated with

145

multiplicative scatter correction (Martens & Næs, 1989), a transformation method

146

used to compensate for additive and multiplicative effects in spectral data.

147 2.3.2. FTIR spectroscopy

EP

148

TE D

M AN U

SC

136

Pre-processing of samples was carried out according to standard methods,

150

using solvents and equipment delivered by the instrument supplier (Foss, Hillerød,

151

Denmark). Two parallel samples of grated cheese were dissolved in a proportion of

152

1/10 in LOSsolver Cheese solvent (Foss) tempered in a water bath at 43 °C, and

153

using a LOSmixer (Foss) with a cooling device, regulated at 45 °C, to homogenise

154

the mixture for 1 min. Antifoam-Y30 (Foss) was used to prevent foaming. Sample

AC C

149

7

ACCEPTED MANUSCRIPT

155

bottles were then placed in the water bath at 43 °C for 14 min to remove air bubbles.

156

FTIR transmittance spectra were the recorded using a Milkoscan FT120 spectrophotometer (Foss). The instrument was equipped with a fat homogenisation

158

module, a flow cell of path length 50 µm and an automated sampler. Spectra were

159

recorded in the range of 974 to 5000 cm-1 with a resolution of 3.858 cm-1. Duplicate

160

spectra of each sample were recorded on each of two parallel samples, and the

161

average of these four spectra used for further analysis. Transmittance values were

162

converted to absorbance values by the transformation A = log (1/T) where A is

163

absorbance, and T transmittance. The spectral regions retained for data analysis

164

were 974-1593 cm-1 and 1700-2986cm-1, because of heavy interference from water

165

in the region 1600-1700 cm-1. The samples from eight spectra (analysed in sequence

166

on the same day) were found to be outliers, and were omitted from the dataset,

167

leaving 451 samples with FTIR spectra for further modelling.

M AN U

SC

RI PT

157

169 170

TE D

168 2.3.3. Fluorescence spectroscopy

Fluorescence was measured on the surface of cheese, taking the average signal over a surface of approximately 20 cm2. The excitation light was generated by

172

a 300 W xenon light source and a 10-nm bandwidth interference filter (both from

173

Oriel Corporation, Stratford, CT, USA). An optical cut-off filter (CVI Laser Optics &

174

Melles Griot, Albuquerque, NM, USA) for filtering the excitation signal and a charge

175

coupled device camera (TEA/CCD-512-TKBM1; Princeton Instruments Inc., Trenton,

176

NJ, USA) with spectrograph for measuring the emission spectra (Acton SP-150;

177

Acton Research Corp., Acton, MA, USA). Emission spectra in the region 287-796 nm

AC C

EP

171

8

ACCEPTED MANUSCRIPT

were measured for each of the excitation wavelengths, 280, 325, 382 and 450 nm.

179

Exposure time was 1 s for all spectroscopic measurements. Two parallel

180

measurements per sample were averaged in further analysis, then pre-processed by

181

normalisation, according to the method described by Bertrand and Scotter (1992).

RI PT

178

182 183

2.4.

Chemical analyses

185

SC

184

Dry matter was determined after eight weeks according to International Dairy Federation (IDF) Standard 4A (IDF, 1982), salt determined according to IDF

187

Standard 88A (IDF, 1988) and fat according to IDF Standard 5B (IDF, 1986). The pH

188

was measured with a calibrated pH meter, Radiometer PHM 210 with electrode

189

GK2401c (Nerliens Mezansky AS, Oslo, Norway) on 25 g of grated cheese with 4-5

190

mL of deionised water added. For determination of total nitrogen, IDF Standard 20B

191

(IDF, 1993) was used with a Kjeltech 2400 Kjeldahl apparatus and auto digestor

192

(Foss Tecator, Höganäs, Sweden). Amino nitrogen (AN) and soluble nitrogen (SN)

193

were analysed according to Mogensen (1948). Amino nitrogen was analysed using a

194

formol titration method comprising (i) neutralisation of cheese solution with 0.25 M

195

NaOH with phenolphthalein as indicator and (ii) addition of 35% formaldehyde in the

196

neutralised mixture, and re-titration with 0.1 M NaOH. Soluble nitrogen was

197

determined using a Kjeldahl apparatus (Foss Tecator) after precipitation with 1 M HCl

198

and filtration.

200

TE D

EP

AC C

199

M AN U

186

Organic acids were analysed with HPLC as described by Skeie, Lindberg, and

Narvhus (2001). Perkin Elmer Series 200 equipment with a UV spectrophotometric

9

ACCEPTED MANUSCRIPT

detector was used (Perkin Elmer, Norwalk, CT, USA). The following components

202

were determined (g kg-1): pyruvate, succinic acid, and lactic acid. Volatile compounds

203

were separated from samples at 90 °C in a gas-tight bottle, and the headspace

204

analysed using a Perkin Elmer Autosystem XL gas chromatography (GC) unit with

205

headspace sampler HS40, headspace injector HS-101, column 183.2 cm 0.2 %

206

Carbowax 1500 80/100 Carbopack C with a flame ionisation detector (FID; Perkin

207

Elmer). The carrier gas was nitrogen and the detector gas hydrogen and air. The

208

following components were determined (in relative units): acetaldehyde, ethanol,

209

acetone, i-propanol, n-propanol, 2-butanone, diacetyl, 2-butanol and acetoin.

210

Volatiles were extracted from the samples in ether and the ether phase analysed by

211

GC in a packed glass column with an FID, Perkin Elmer Autosystem GC, column

212

Supelco 2 m, 0.635 cm OD, 2 mm ID glass column, with Supelco GP 10% SP

213

1000/1% H3PO4 on 100/120 Chromosorb WAW, and FID (Perkin Elmer). The carrier

214

gas was nitrogen and the detector gas hydrogen and air. The components

215

determined (mmol kg-1) were: acetoin, acetic acid, propionic acid and butyric acid.

216

Fatty acid composition was analysed by GC (Kraggerud et al., 2008). Analysis of free

217

amino acids was carried out according to the procedure described in Skeie, Feten,

218

Almøy, Østlie, and Isaksson (2006).

220 221 222 223

SC

M AN U

TE D

EP

AC C

219

RI PT

201

2.5. Sensory analysis

2.5.1. Quality scoring

Quality scoring was carried out using a modified method (TINE Dairy

10

ACCEPTED MANUSCRIPT

Cooperative, Oslo, Norway) of the IDF method for sensory evaluation, Part IV (IDF,

225

1997). At least three authorised expert assessors were used per session. To be

226

authorised as an assessor in TINE, a candidate has to participate in at least 20-30

227

assessment sessions in the course of about one year. In the authorisation test,

228

congruence and correlation with reference assessors, and the candidate’s

229

reproducibility and repeatability are the most important criteria used. Appearance,

230

consistency and odour/flavour as well as overall quality score were assessed using a

231

numerical interval scale from 1 to 5, with 0.5 point intervals. The scale is defined

232

according to IDF 99C, Part 1 clause 9.2 (IDF, 1997) as follows: 5 = conformity with

233

the pre-established sensory specification (PS), 4 = minimal deviation from PS, 3 =

234

noticeable deviation from PS, 2 = considerable deviation from PS, 1 = very

235

considerable deviation from PS. The score 0, which was defined in IDF 99C, (unfit for

236

human consumption) was not used. When an assessor scored <4 points for a

237

sample, a description of the deviation was given, using predefined nomenclature of

238

defect terms.

SC

M AN U

TE D

239

RI PT

224

Blocks of cheese (5 kg each) at a temperature of 14 ± 2 °C were presented to the assessors in random order and placed in a row on a long table. Before sensory

241

assessment, each cheese block was cut in two. Sensory assessment was carried out

242

using a standard cheese slicer for cutting. Each session started with calibration of the

243

assessment with two different cheeses as reference samples. The assessors

244

received no information about the samples and were not allowed to communicate

245

during the following assessment. There was one replicate of each cheese sample,

246

owing to the large number of samples.

AC C

EP

240

11

ACCEPTED MANUSCRIPT

247 248

2.5.2. Descriptive sensory analysis

249

The vocabulary of sensory attributes used in this experiment is in accordance with International Organisation for Standardisation (ISO) 5492 (ISO, 1992). The

251

terms to be included or excluded are evaluated for each data set, depending on the

252

needs of the specific experiments. The attributes have earlier been described in more

253

detail (Kraggerud et al., 2008). An interval scale from 1 to 9 points was used for each

254

sensory attribute, as defined in ISO 4121 (ISO, 2003). An internal laboratory panel

255

was used comprising six assessors, well-trained on cheese assessment for several

256

years. Before each session, a calibration session was carried out in order to obtain

257

an agreement on the use of attributes and scales, using two different cheeses as test

258

samples. All assessments were carried out in individual booths at the sensory

259

laboratory, designed in compliance with international standards for test rooms, ISO

260

8589 (ISO, 1988). Cheese samples were held at 14 ± 2 °C during the assessment.

261

The order of assessment of the samples was randomised for each assessor. Each

262

sample was assessed as one replicate, owing to the large number of samples. The

263

scores of each sample were averaged over the assessors.

264

266 267

2.6.

Analysis of data

AC C

265

EP

TE D

M AN U

SC

RI PT

250

Basic statistics and analysis of variance (ANOVA) were carried out using

268

software package Minitab® 15.1. (MINITAB Inc., State College, PA, USA).

269

Multivariate analysis was carried out using Unscrambler version 10.1. (CAMO

12

ACCEPTED MANUSCRIPT

Software AS, Oslo, Norway), with several standard methods such as principal

271

components analysis (PCA) and partial least squares regression (PLS) (Martens &

272

Næs, 1989). Validation of models and choice of optimal number of components were

273

carried out by cross-validation with uncertainty testing (Martens & Martens, 2000).

274

Correlation coefficient, R, and root mean square error of prediction (RMSECV) were

275

used to evaluate the goodness of fit of models. The values from validation were used

276

(Martens & Næs, 1989). Standard error of cross-validation (SECV), frequently used

277

by other authors, is essentially the same parameter as RMSECV.

SC

Standard error, serror, was calculated for each sensory variable as the standard

M AN U

278

RI PT

270

279

error of the mean (s/√n) averaged over all samples assessed, using Mintab software

280

(Minitab Inc.). Ratio performance deviation (RPD; Williams & Sobering, 1993) was

281

calculated according to equation (1):

282 RPD = StDev(ref)/RMSECV

(1)

TE D

283 284

where StDev(ref) is the standard deviation of the reference method results. Range

286

error ratio (RER; Williams, 1987) was also used for evaluating model performance,

287

and calculated as follows:

289 290 291

AC C

288

EP

285

Range = Max value – Min value

(2)

RER = Range/RMSECV

(3)

292

13

ACCEPTED MANUSCRIPT

293

3.

Results and discussion

294 295

3.1.

Changes in sensory and chemical characteristics during cheese maturation

297

RI PT

296 Different analyses relevant for the characterisation of cheese maturation were carried out at three points: 8, 24 and 40 weeks of age. In addition, some analyses

299

were carried out during processing and on fresh cheese. Tables 2, 6, and 7 show

300

basic statistics for measured sensory and chemical parameters at 8, 24 and 40

301

weeks. The variation in major chemical components was relatively narrow (in

302

parenthesis: range, standard deviation in g 100 g-1): salt (0.45, 0.10), fat (3.90, 0.68),

303

protein (0.59, 0.10) and dry matter (4.00, 0.73). The samples analysed should

304

therefore be appropriate for the purpose of testing methods for use in ordinary

305

production where there will be interest in detecting minor variations. Fig. 1 shows the

306

developments in sensory attributes during maturation. A systematic increase or

307

decrease in almost all measurements during the maturation period investigated was

308

observed. ANOVA was carried out with sample number and age as factors and the

309

results are presented in Table 2. Highly significant differences (P<0.001) in both

310

factors were found for almost all variables. Consistency variables, such as firmness

311

chewing, grainy, cohesion and elasticity decreased, whereas solubility and pasty, as

312

well as flavour variables such as flavour intensity, pungent and bitter typically

313

increased during maturation.

315

M AN U

TE D

EP

AC C

314

SC

298

In Fig. 2, relative development in various chemical measurements at 8, 24,

and 40 weeks age are illustrated. ANOVA with age and sample as factors, showed

14

ACCEPTED MANUSCRIPT

that there were no significant differences for the variable propionic acid (which is not

317

a typical component in Norvegia, and hence varied between 0.0 and 0.2 mmol kg-1).

318

All the other variables had significant age differences with P<0.001, except total N

319

(Kjeldahl) which had P=0.088. Some of the components increased, and others

320

decreased during 8-40 weeks of ripening. A few components were practically

321

constant, such as lactic acid and acetic acid. During ripening, the soluble N and

322

amino acid (amino N) containing fractions gradually increased as protein was

323

degraded during ripening. Free amino acids, illustrated in Fig. 3, all demonstrated a

324

near linear increase with age. Cross-correlation between all amino acids, and also

325

with amino N, were significant, almost all of them very high (Table 3). This can

326

indicate that the pattern of protein degradation was relatively similar during

327

maturation for these samples even though they originated from five different

328

production plants. These findings, in general, correspond well with what was

329

expected, from cheese ripening biochemistry theory (McSweeney, 2004).

TE D

M AN U

SC

RI PT

316

330 331

3.2.

Regression modelling and validation

333

EP

332

Predicting sensory variables, with their relatively low precision compared with the total variation range, is not an easy task. For calibration of models for use in

335

industry, it would probably be best to calibrate on a set of data with larger variation,

336

then use the calibration for prediction on subsequent samples.

337 338

AC C

334

There are a number of different validation parameters that can be used,

among them, validated RMSECV is considered to be relevant. RMSECV can be

15

ACCEPTED MANUSCRIPT

compared with the precision of the reference method, and has the same unit as the

340

reference values, in contrast to regression coefficient, which will be related to the

341

actual sample space. The other validation measures in this article, RPD and RER,

342

are also very dependent of the total sample space, but were chosen as they have

343

been used by other authors. When it comes to comparison of results presented by

344

various authors’ interpretation is always difficult, because number of samples,

345

breadth of range, standard error of the reference method and other factors affect all

346

these validation parameters. Furthermore, there have been very few experiments

347

with high numbers of samples, which could also be an important factor influencing

348

the values of the validation parameters.

SC

M AN U

349

RI PT

339

Cross-validation was carried out in this case. Explained variance with full cross-validation (“leave one out”) was compared with segmented cross-validation

351

with 20 random segments (23 samples per segment). An example of explained

352

variance results for PLS1 regression of amino nitrogen from FTIR spectra is

353

illustrated in Fig. 4. Calibration lines for all samples (459 in total before outlier

354

deletion) are of course identical. Explained validation variance was shown to be

355

almost identical; it is impossible to separate the two lines from each other as the

356

average difference was only 0.2%.

EP

Outliers are also an important challenge to be addressed. The FTIR spectra of

AC C

357

TE D

350

358

eight samples, analysed in sequence, were found to be outliers. Results from

359

modelling amino N from FTIR with and without these outliers in the dataset are also

360

illustrated in Fig. 4. The explained variance was slightly higher when removing the

361

eight outliers, both in calibration and validation, where validation variance was 74.5%

16

ACCEPTED MANUSCRIPT

362

versus 71.5%. The actual eight outliers were excluded from our further modelling

363

based on FTIR. With lower number of samples, outliers will be even more important

364

to handle. The number of factors in the chosen model can be of importance for the

RI PT

365

results. In a PLS model, components are extracted in such a way that the first

367

factor/principal component (PC) explains the largest amount of variation, followed by

368

the second factor, etc. At a certain point, the variation modelled by any new PC is

369

mostly noise. The optimal number of factors – modelling useful information, but

370

avoiding overfitting – is determined with the help of the residual validation variances.

371

If less than 3% of the residual variance is explained in the successor factor, the

372

optimum number of factors is chosen in the Unscrambler software. In our case (Fig.

373

4), the optimal number of factors were five for the 459-sample, and six for the 451-

374

sample segmented cross-validation models. It is also necessary to view the actual

375

explained variance curve to see whether the chosen optimal number of factors is

376

proper, as there can in some cases be temporary drops in the curve. When validation

377

residual variance is minimal, RMSECV is as well, and the model with an optimal

378

number of components will have the lowest expected prediction error.

M AN U

TE D

EP

379

SC

366

With PLS regression, it is possible to model dependent variables one by one (PLS1), or many Y-variables at the same time (PLS2). Table 4 compares PLS2

381

results with PLS1 results for two of the sensory attributes. The same applies to

382

Tables 5 and 6, for two variables in each table. PLS1 compared with PLS2 had only

383

minor differences in RMSECV and correlation for most of the models. Some models

384

were better with PLS1, as would be expected when focusing on modelling only one

AC C

380

17

ACCEPTED MANUSCRIPT

385

variable. As the differences were small, PLS2 was chosen for groups of Y-variables

386

for the main results and discussion, as a large number of Y-variables were modelled.

387 3.3.

Prediction of sensory attributes

RI PT

388 389 390

Table 4 compares validation results from PLS2 modelling of all sensory variables modelled from several chemical variables and from FTIR, NIR and

392

fluorescence spectroscopy of the same samples. RMSECV was found to be only

393

slightly higher than the standard error of the reference method for many of the

394

sensory variables. Chemical variables and FTIR gave the best, and almost the same,

395

prediction results for many sensory variables. Correlation coefficients were relatively

396

low, but, as shown in Fig. 5, this can be caused by noisy reference data as the

397

correlations seem clear. Fluorescence and NIR spectroscopy also showed

398

reasonable results, but generally not as high as FTIR. This might be due to both

399

these methods being carried out on the surface of cheese, which can be a problem if

400

the cheese is not homogeneous. FTIR was carried out on ground and pre-processed

401

cheese samples, and involved a lot more labour for each analysis.

EP

TE D

M AN U

SC

391

Publications comparing FTIR and sensory variables of cheese are few.

403

Kocaoglu-Vurma, Eliardi, Drake, Rodriguez-Saona, and Harper (2009) analysed 15

404

samples from three manufacturers of Swiss cheese with FTIR and sensory

405

descriptive analysis. They found correlations between 0.69 and 0.96 when predicting

406

sensory variables from FTIR using PLS. Fagan et al. (2007b) concluded that mid-

407

infrared spectroscopy has the potential to predict age, SN, and several sensory

AC C

402

18

ACCEPTED MANUSCRIPT

texture attributes of cheddar cheese – with R up to 0.8 for different sensory attributes

409

– in the same area as the models found here. The same group used FTIR for

410

processed cheese and found RER values between 5 and 10 for sensory descriptive

411

attributes, indicating that the models had good practical utility (Fagan et al., 2007a).

412

In the data presented here – even with the relatively narrow range of data – RER

413

values were in the range 5-10 for almost all sensory variables, and some variables

414

showed RER >10, which indicates a high utility value (Williams, 1987). Subramanian,

415

Harper, and Rodriguez-Saona (2009) developed an extraction method which, in

416

combination with measurements with FTIR, showed promising results based on

417

clustering of samples. The experiment consisted of 15 Cheddar samples, with no

418

reference data except for quality classification, and no key validation data, which

419

makes comparison with the present results difficult.

SC

M AN U

420

RI PT

408

Lerma-García, Gori, Cerretani, Simó-Alfonso, and Caboni (2010) were able to classify Pecorino cheese according to both ripening time and production technique

422

from FTIR results. It should be noted that differences between individual cheeses

423

seem relatively large, and that reference analysis measurements were not reported.

424

Martín-del-Campo, Picque, Cosio-Ramirez, and Corrieu (2007) used attenuated total

425

reflectance (ATR)-FTIR to predict ripening date of Camembert cheese, with a

426

precision of ±1 day. FTIR has also been used to study Crescenza cheese during 20

427

days, and it was possible to define the maximum shelf-life of this fresh cheese

428

(Cattaneo, Giardina, Sinelli, Riva, & Giangiacomo, 2005).

430

EP

AC C

429

TE D

421

NIR analysis has been used for assessment of sensory properties of a semi-

hard cheese variety (Sorensen & Jepsen, 1998). Thirty-two batches of cheese were

19

ACCEPTED MANUSCRIPT

made in that study, some with the addition of undesired bacteria, covering a wide

432

range of pH and moisture content, and measured four times during ripening from five

433

to 11 weeks. Squared correlation coefficients of the best predicted sensory attributes

434

were 0.7–0.8, higher than our R2 from spectroscopy. This can be due to wider ranges

435

within the dataset which tend to generate higher correlation coefficients. There was

436

no other parameter directly comparable with our validation parameters. Twenty

437

Emmental cheeses were evaluated by NIR and sensory panel and the method was

438

found to be feasible for prediction of some sensory attributes (Karoui et al., 2006).

439

Eight attributes obtained R2 higher than 0.5. RER was between 5 and 7 for most

440

attributes, whereas one was 10 and one 3.5. RPD was mostly between 1.5 and 2. All

441

these results were quite comparable with our results from different spectroscopic

442

measurements of sensory attributes. Experimental Cheddar cheeses were measured

443

by NIR and sensory evaluation and models were evaluated to be sufficiently accurate

444

for industrial use (Downey et al., 2005). Calculated RPD and RER were comparable

445

with our results from spectroscopic data. Blazquez et al. (2006) modelled sensory

446

and texture parameters with NIR in experimental processed cheese with fairly high

447

variation between samples, and obtained RER values between 9 and 12.

448

Changes in cheese during ripening have also been described with

EP

TE D

M AN U

SC

RI PT

431

fluorescence spectra, some of them with sensory, chemical and rheological

450

measurements as reference. Dufour,Devaux, Fortier, and Herbert (2001) found

451

squared correlations between fluorescence and sensory texture attributes of 0.22-

452

0.69, and for pastiness 0.89, in soft cheese. This observation corresponds well with

453

our findings of pasty being one of the best predicted sensory variables by all our

AC C

449

20

ACCEPTED MANUSCRIPT

spectroscopic measurements (Table 4), as also reported by Lebecque, Laguet,

455

Deveaux, and Dufour (2001) who obtained R2 below 0.5 in prediction of sensory

456

attributes from fluorescence tryptophan spectra, and somewhat higher using

457

fluorescence vitamin A spectra.

RI PT

454

458 459

3.4.

Prediction of chemical parameters from spectroscopic measurements

461

SC

460

Major components of cheese, such as fat, protein and water/dry matter, are commonly measured with spectroscopic methods such as FTIR and NIR

463

(Adamopoulos, Goula, & Petropakis, 2001; Chen, Kocaoglu-Vurma, Harper, &

464

Rodriguez-Saona, 2009; González-Martín, Gonzáles-Pérez, Hernándes-Hierro, &

465

Gonzáles-Cabrera, 2008; Karoui et al., 2006; Wittrup & Nørgaard, 1998). They will

466

not be discussed here, as the variation of the major components was limited in our

467

dataset. In Table 6, validation results from the prediction of different minor chemical

468

attributes are given. In the PLS2 model, the highest values of correlation, RPD and

469

RER, were obtained for amino N and soluble N measured with FTIR, and with mixed

470

spectra. This indicates a high capacity of prediction of important ripening

471

characteristics. Amino acids (Table 7) also showed high correlation coefficients, most

472

in the area 0.8-0.9. All have 1.5
473

they can be useful (Williams & Sobering, 1993). Modelling with PLS1 one variable at

474

a time instead of PLS2 might be useful, for example lactic acid (Table 6) where the

475

correlation coefficient was raised from 0.70 (PLS2) to 0.88 (PLS1) for the models

476

from mixed spectra. This might also be typical for other variables which we have not

AC C

EP

TE D

M AN U

462

21

ACCEPTED MANUSCRIPT

477

tested, as prediction of chemical variables were not the main issue. Other authors have made attempts to measure chemical variables in addition

479

to the major components in cheese. Skeie et al. (2006) used NIR to predict selected

480

amino acids in Norvegia and Präst cheese. They achieved very high correlations

481

(>0.9 for many variables) compared with an HPLC/standard method. Our data

482

showed higher RMSECV values, the best fit (from mixed spectra) only slightly higher.

483

The correlations in our dataset were lower, but all of them were higher than 0.74

484

(mixed spectra). In the data from Skeie et al. (2006) the optimum number of PLS

485

components was much higher. Correlation coefficients for amino acids were in our

486

experiment lower for NIR than for the other spectroscopy methods. NIR

487

measurements from 400-1100 nm were used in our experiment, whereas Skeie et al.

488

(2006) used the region from 780-2500 nm. This indicates that the higher part of the

489

NIR spectrum could be used for higher prediction ability. Another difference in the

490

NIR measurements was that scanning was made directly on the cheese surface (a

491

non-destructive method) whereas Skeie et al. (2006) made measurements on grated

492

cheese, which might result in lower variation in the sample between reference

493

measurement and spectroscopy as the sample is then mixed and is more uniform. It

494

is known that the variation inside one cheese can be of importance. On the other

495

hand, non-destructive measurements would be a great advantage, and there is

496

reason to believe from other results that it is possible to improve the method to obtain

497

acceptable results also with limited variation in the samples.

499

SC

M AN U

TE D

EP

AC C

498

RI PT

478

Karoui et al. (2006) measured non-protein nitrogen, total nitrogen, and SN by

NIR. RPD values were comparable with values obtained in our experiment but, also

22

ACCEPTED MANUSCRIPT

500

in this case, correlation coefficients in the present experiments were lower; the much

501

greater number of samples could explain this observation. Koca, Rodriguez-Saona, Harper, and Alvarez (2007) applied FTIR for

503

monitoring of short-chain free fatty acids (FFA) in Swiss cheese. The range of FFA

504

was much higher than in our experiment. SECV, re-calculated to mmol per 100 g,

505

was quite comparable for acetic acid measured with FTIR (see our results expressed

506

as RMSECV values in Table 6). FTIR was used for monitoring of amino acids,

507

organic acids and ripening changes in water-soluble fractions of 12 Cheddar samples

508

at five points during a ripening period of 7-73 days to predict amino acids and organic

509

acids such as lactic, formic and oxalic acid, with correlation coefficients of 0.89 and

510

higher (Subramanian et al., 2011). RER, based on range were, in their case,

511

between 10 and 20 in most components which is higher than in our experimental

512

data. We had trouble comparing our amino acid range and RMSECV values with the

513

SECV quoted in Table 1 of Subramanian et al. (2011), but given the concentrations

514

reported in Table 3 of the same paper, the results seem more comparable with our

515

data. It is again difficult to compare validation parameters directly as our dataset is

516

much more comprehensive when it comes to number of samples.

517

519 520

3.5.

Early prediction of sensory quality

AC C

518

EP

TE D

M AN U

SC

RI PT

502

In Table 5, validation results are given from prediction of sensory attributes at

521

40 weeks of age from different measurements at eight weeks (number of samples

522

153, as there is only one age group). Correlation coefficients were generally lower,

23

ACCEPTED MANUSCRIPT

but RMSECV results were comparable, even a little better for some attributes

524

compared with predictions on the eight-week samples (Table 4). This is consistent

525

with what was reported in section 3.1, development during cheese ripening is on

526

average linear. But the individual differences per sample are more difficult to predict,

527

making correlations lower, which might also be caused by noise in the sensory

528

measurements. The potential with respect to early prediction of sensory quality is

529

therefore uncertain.

SC

RI PT

523

530 4.

Conclusions

M AN U

531 532

An extensive assembly of 459 samples of Norvegia, a Dutch-type semi-hard

534

cheese, was analysed using a number of chemical, chromatographic, sensory and

535

spectroscopic methods during maturation. From eight to 24 and 40 weeks of age

536

there was a highly systematic development in chemical and sensory parameters.

537

Sensory attributes pasty, grainy, solubility, cohesion, firmness on chewing,

TE D

533

flavour intensity, pungent and bitter showed the highest correlations with chemical

539

variables modelled by PLS2. Noise in the sensory data seems to be the cause of

540

relatively low correlation coefficients. The sensory attributes with the best validation

541

results provided slightly lower correlations with FTIR measurements than with

542

chemical analysis. Fluorescence spectroscopy and spectroscopy using NIR between

543

400 nm and 1100 nm, carried out on the surface of the cheese, gave slightly less

544

valid results than FTIR for measurement of sensory variables in this experiment.

545

Using a combination of spectra from all instruments gave a higher correlation than

AC C

EP

538

24

ACCEPTED MANUSCRIPT

546 547

spectra from instruments taken separately. Results from a mixture of spectroscopic measurements, and also FTIR, NIR and fluorescence alone, were promising in order to replace the more onerous

549

sensory measurements, but there is still some way to go with the calibrations.

550

Chemical analysis, on the other hand, can be replaced to a large extent by

551

spectroscopy, with accurate results. The advantages will be fast testing which makes

552

direct utilisation of the results in production possible. The spectroscopic methods are

553

also usually labour-saving, both compared with chemical and sensory methods. The

554

benefit will be the possibility to analyse more samples in order to cover variability

555

within and between cheese batches. Sensory characteristics at a higher age of the

556

cheese were, however, not very well forecasted by early measurements on cheese,

557

whether sensory, chemical or spectroscopic.

M AN U

SC

RI PT

548

558 References

TE D

559 560 561

Adamopoulos, K. G., Goula, A. M., & Petropakis, H. J. (2001). Quality control during processing of feta cheese - NIR application. Journal of Food Composition and

563

Analysis, 14, 431-440.

Andersen, C. M., & Mortensen, G. (2008). Fluorescence spectroscopy: a rapid tool

AC C

564

EP

562

565

for analyzing dairy products. Journal of Agricultural and Food Chemistry, 56,

566

720-729.

567 568

Beresford, T. P., Fitzsimons, N. A., Brennan, N. L., & Cogan, T. M. (2001). Recent advances in cheese microbiology. International Dairy Journal, 11, 259-274.

25

ACCEPTED MANUSCRIPT

569 570

Bertrand, D., & Scotter, C.N.G. (1992). Application of multivariate analyses to NIR spectra of gelatinized starch. Applied Spectroscopy, 46, 1420-1425. Blazquez, C., Downey, G., O'Callaghan, D., Howard, V., Delahunty, C., Sheehan, E.,

572

et al. (2006). Modelling of sensory and instrumental texture parameters in

573

processed cheese by near infrared reflectance spectroscopy. Journal of Dairy

574

Research, 73, 58-69.

RI PT

571

Cattaneo, T. M. P., Giardina, C., Sinelli, N., Riva, M., & Giangiacomo, R. (2005).

576

Application of FT-NIR and FT-IR spectroscopy to study the shelf-life of

577

Crescenza cheese. International Dairy Journal, 15, 693-700.

M AN U

578

SC

575

Chen, G., Kocaoglu-Vurma, N. A., Harper, W. J., & Rodriguez-Saona, L. E. (2009).

579

Application of infrared microspectroscopy and multivariate analysis for

580

monitoring the effect of adjunct cultures during Swiss cheese ripening. Journal

581

of Dairy Science, 92, 3575-3584.

Collins, Y. F., McSweeney, P. L. H., & Wilkinson, M. G. (2003). Lipolysis and free

583

fatty acid catabolism in cheese: a review of current knowledge. International

584

Dairy Journal, 13, 841-866.

Delahunty, C. M., & Drake, M. A. (2004). Sensory character of cheese and its

EP

585

TE D

582

evaluation. In P.F. Fox, P. L. H. McSweeney, T. M. Cogan, & T. P. Guinee

587

(Eds.), Cheese: Chemistry, physics and microbiology, Vol. 1: General aspects

588

(3rd edn.; pp. 455-487). London, UK: Elsevier Academic Press.

589

AC C

586

Downey, G., Sheehan, E., Delahunty, C., Callaghan, D., Guinee, T., & Howard, V.

590

(2005). Prediction of maturity and sensory attributes of Cheddar cheese using

591

near-infrared spectroscopy. International Dairy Journal, 15, 701-709.

26

ACCEPTED MANUSCRIPT

592

Dufour, E., Devaux, M. F., Fortier, P., & Herbert, S. (2001). Delineation of the

593

structure of soft cheeses at the molecular level by fluorescence spectroscopy-

594

relationship with texture. International Dairy Journal, 11, 465-473. Fagan, C. C., Everard, C., O'Donnell, C. P., Downey, G., Sheehan, E. M., Delahunty,

596

C. M., et al. (2007a). Evaluating Mid-infrared Spectroscopy as a New Technique

597

for Predicting Sensory Texture Attributes of Processed Cheese. Journal of Dairy

598

Science, 90, 1122-1132.

SC

599

RI PT

595

Fagan, C. C., O'Donnell, C. P., O'Callaghan, D. J., Downey, G., Sheehan, E. M., Delahunty, C. M., et al. (2007b). Application of mid-infrared spectroscopy to the

601

prediction of maturity and sensory texture attributes of cheddar cheese. Journal

602

of Food Science, 72, E130-137.

M AN U

600

Fox, P. F., & Cogan, T. M. (2004). Factors that affect the quality of cheese. In P.F.

604

Fox, P. L. H. McSweeney, T. M. Cogan, & T. P. Guinee (Eds.), Cheese:

605

Chemistry, physics and microbiology , Vol. 1: General aspects (3rd edn.; pp.

606

583-608). London, UK: Elsevier Academic Press.

607

TE D

603

Gilles, J., & Lawrence, R. C. (1973). The assessment of Cheddar cheese quality by compositional analysis. New Zealand Journal of Dairy Science and Technology,

609

8, 148-151.

González-Martín, I., González-Pérez, C., Hernández-Hierro, J. M., & González-

AC C

610

EP

608

611

Cabrera, J. M. (2008). Use of NIRS technology with a remote reflectance fibre-

612

optic probe for predicting major components in cheese. Talanta, 75, 351-355.

613 614

IDF. (1982). IDF 4A Cheese - Determination of dry matter. Brussels, Belgium: International Dairy Federation.

27

ACCEPTED MANUSCRIPT

615

IDF. (1986). IDF 5B Cheese and processed cheese products. Determination of fat

616

content - gravimetric method (reference method). Brussels, Belgium:

617

International Dairy Federation. IDF. (1988). IDF 88A Cheese and processed cheese products. Determination of

619

chloride content - potentiometric titration method. Brussels, Belgium:

620

International Dairy Federation.

623 624 625 626

SC

622

IDF. (1993). IDF 20B Milk - Determination of nitrogen content. Brussels, Belgium: International Dairy Federation.

IDF. (1995). IDF 50C Milk and milk products - Guidance on sampling. Brussels,

M AN U

621

RI PT

618

Belgium: International Dairy Federation.

IDF. (1997). IDF 99C Sensory evaluation of dairy products by scoring. Reference method. Brussels, Belgium: International Dairy Federation. ISO. (1988). ISO 8589 Sensory analysis - general guidance for the design of test

628

rooms. Geneva, Switzerland: International Organisation for Standardisation.

629 630

TE D

627

ISO. (1992). ISO 5492 sensory analysis - vocabulary. Geneva, Switzerland: International Organisation for Standardisation. ISO. (2003). ISO 4121 sensory analysis - Guidelines for the use of quantitative

632

response scales. Geneva, Switzerland: International Organisation for

633

Standardisation.

AC C

634

EP

631

Johnson, M., & Law, B. (1999). The origins, development and basic operations of

635

cheesemaking technology. In B. Law (Ed.), Technology of cheesemaking (pp.

636

1-32). Sheffield, England: Sheffield Academic Press Ltd.

637

Karoui, R., & De Baerdemaeker, J. (2007). A review of the analytical methods

28

ACCEPTED MANUSCRIPT

638

coupled with chemometric tools for the determination of the quality and identity

639

of dairy products. Food Chemistry, 102, 621-640. Karouri, R., Mazerolles, G., & Dufour, E. (2003). Spectroscopic techniques coupled

641

with chemometric tools for structure and texture determinations in dairy

642

products. International Dairy Journal, 13, 607-620.

RI PT

640

Karoui, R., Mouazen, A. M., Dufour, É., Pillonel, L., Schaller, E., De Baerdemaeker,

644

J., et al. (2006). Chemical characterisation of European Emmental cheeses by

645

near infrared spectroscopy using chemometric tools. International Dairy Journal,

646

16, 1211-1217.

M AN U

647

SC

643

Koca, N., Rodriguez-Saona, L. E., Harper, W. J., & Alvarez, V. B. (2007). Application

648

of Fourier transform infrared spectroscopy for monitoring short-chain free fatty

649

acids in Swiss cheese. Journal of Dairy Science, 90, 3596-3603.

650

Kocaoglu-Vurma, N. A., Eliardi, A., Drake, M. A., Rodriguez-Saona, L. E., & Harper, W. J. (2009). Rapid profiling of Swiss cheese by attenuated total reflectance

652

(ATR) infrared spectroscopy and descriptive sensory analysis. Journal of Food

653

Science, 74, S232-239.

Kraggerud, H., Skeie, S., Høy, M., Røkke, L., & Abrahamsen, R. K. (2008). Season

EP

654

TE D

651

and ripening temperature influence fatty acid composition and sensory

656

properties of semi-hard cheese during cheese maturation. International Dairy

657

Journal, 18, 801-810.

658

AC C

655

Lebecque, A., Laguet, A., Deveaux, M. T., & Dufour, É. (2001). Delineation of the

659

texture of Salers cheese by sensory analysis and physical methods. Lait, 81,

660

609-624.

29

ACCEPTED MANUSCRIPT

661

Lerma-García, M. J., Gori, A., Cerretani, L., Simó-Alfonso, E. F., & Caboni, M. F. (2010). Classification of Pecorino cheeses produced in Italy according to their

663

ripening time and manufacturing technique using Fourier transform infrared

664

spectroscopy. Journal of Dairy Science, 93, 4490-4496.

665

RI PT

662

Martens, H., & Martens, M. (2000). Modified Jack-knife estimation of parameter

uncertainty in bilinear modelling by partial least squares regression (PLSR).

667

Food Quality and Preference, 11, 5-16.

SC

666

Martens, H., & Næs, T. (1989). Multivariate Calibration. Chichester, UK: Wiley.

669

Martín-del-Campo, S. T., Picque, D., Cosío-Ramírez, R., & Corrieu, G. (2007).

M AN U

668

670

Evaluation of chemical parameters in soft mold-ripened cheese during ripening

671

by mid-infrared spectroscopy. Journal of Dairy Science, 90, 3018-3027.

672

McSweeney, P. L., & Sousa, M. J. (2000). Biochemical pathways for the production

673

of flavour compounds in cheeses during ripening: A review. Lait, 80, 293-324. McSweeney, P. L. H. (2004). Biochemistry of cheese ripening: Introduction and

TE D

674 675

overview. In P.F. Fox (Ed.), Cheese: Chemistry, physics and microbiology, Vol.

676

1: General aspects (3rd edn; pp. 347-360). London, UK: Academic Press. Mogensen, M. T. S. (1948). Bestemmelse af ostens proteolytiske spaltningsgrad med

678

særligt henblik på formoltitreringen. Malmö, Sweden: Statens Mejeriförsök. Skeie, S., Lindberg, C., & Narvhus, J. (2001). Development of amino acids and

AC C

679

EP

677

680

organic acids in Norvegia, influence of milk treatment and adjunct Lactobacillus.

681

International Dairy Journal, 11, 399-411.

682 683

Skeie, S., Feten, G., Almøy, T., Østlie, H., & Isaksson, T. (2006). The use of near infrared spectroscopy to predict selected free amino acids during cheese

30

ACCEPTED MANUSCRIPT

684 685

ripening. International Dairy Journal, 16, 236-242. Sorensen, L. K., & Jepsen, R. (1998). Assessment of sensory pcroperties of cheese by near-infrared spectroscopy. International Dairy Journal, 8, 863-871.

687

Subramanian, A., Alvarez, V. B., Harper, W. J., & Rodriguez-Saona, L. E. (2011).

RI PT

686

Monitoring amino acids, organic acids, and ripening changes in Cheddar

689

cheese using Fourier-transform infrared spectroscopy. International Dairy

690

Journal, 21, 434-440.

SC

688

Subramanian, A., Harper, W. J., & Rodriguez-Saona, L. E. (2009). Cheddar cheese

692

classification based on flavor quality using a novel extraction method and

693

Fourier transform infrared spectroscopy. Journal of Dairy Science, 92, 87-94.

694

M AN U

691

Subramanian, A., & Rodriguez-Saona, L. (2010). Chemical and instrumental

695

approaches to cheese analysis. In S. Taylor (Ed.), Advances in Food and

696

Nutrition Research (pp. 167-213). London, UK: Elsevier. Williams, P. C. (1987). Implementation of near-infrared technology. In P. C. Williams

TE D

697

& K. Norris (Eds.), Near-infrared technology in the agricultural and food

699

industries (pp. 143-167). St. Paul, MN, USA: American Association of Cereal

700

Chemists International.

701

EP

698

Williams, P. C., & Sobering, D. C. (1993). Comparison of commercial near infrared transmittance and reflectance instruments for analysis of whole grains and

703

seeds. Journal of Near Infrared Spectroscopy, 1, 25-32.

704

AC C

702

Wittrup, C., & Nørgaard, L. (1998). Rapid near infrared spectroscopic screening of

705

chemical parameters in semi-hard cheese using chemometrics. Journal of Dairy

706

Science, 81, 1803-1809.

31

ACCEPTED MANUSCRIPT

1

Figure legends

2 Fig. 1. Average values for sensory attributes at each stage of maturity: white eight

4

weeks, grey 24 weeks, and black 40 weeks. Values are adjusted to % within each

5

variable to fit onto the same scale.

RI PT

3

6

Fig. 2. Average values for chemical analysis results at each stage of maturity: white

8

eight weeks, grey 24 weeks, and black 40 weeks. Values are adjusted to % within

9

each variable in order to fit onto the same scale. X axis labels are 1: pH, 2: total

M AN U

SC

7

10

nitrogen (N), 3: soluble-N, 4: amino-N, 5: acetaldehyde, 6: ethanol, 7: acetone, 8: i-

11

propanol, 9: n-propanol, 10: 2-butanone, 11: diacetyl, 12: 2-butanol, 13: acetoin, 14:

12

acetic acid, 15: propionic acid, 16: butyric acid, 17: pyruvate, 18: succinic acid, 19:

13

lactic acid.

TE D

14

Fig. 3. Average values for amino acids measured by HPLC at each stage of maturity:

16

white, eight weeks; grey, 24 weeks; black, 40 weeks. Values are adjusted to % within

17

each variable to fit into the same scale.

18

EP

15

Fig. 4. Partial least squares (PLS) 1 regression validation: explained variance in Y for

20

20 PLS1 factors (equal to ‘number PC’, 0-20 – on the x-axis) with variables Fourier

21

transform infra-red as X, amino nitrogen (AN) as Y. Plot shows: calibration variance

22

for all samples (–··–); sample set with eight outlier samples omitted (——); validation

23

variance for full cross-validation (- - - -); segmented cross-validation with 20 random

AC C

19

1

ACCEPTED MANUSCRIPT

24

segments (–·–·); outlier reduced data set (- - - -).

25 Fig. 5. Partial least squares 1 regression plot with measured value (reference Y) vs.

27

predicted Y for the dependent variable sensory descriptive attribute ‘firmness

28

chewing’. Independent variable sets were: a) chemical variables, R=0.78; b) Fourier

29

transform infra-red spectra, R=0.77; c) near infra-red spectra, R=0.67; d)

30

fluorescence spectra, R=0.76.

AC C

EP

TE D

M AN U

SC

RI PT

26

2

ACCEPTED MANUSCRIPT

Table 1 Overview of sampling and analysis throughout the experimental period.

0

0

Sum 32 36 12 12 12 12 12 13

40 44 14 12 12 12 12 14 12 16 12 16 12 13 24 30 24 25 36 42

48 52 56 60 64 68 72 76 80 84

12 12 12 36

RI PT

Production E1+E3 Production E2 Sampling 8 weeks Sampling 24 weeks Sampling 40 weeks Samples analysed

Weeks from start 0 4 8 12 16 20 24 28 16 13 14 12 12 12 12 12 12 16 12 13 12 14 12 16

14 12 12 12 12 14 13 12 14 12 12 12 12 12 14 39 24 26 24 26 12 12 12 14

a

SC

Parameter

a

153 153 153 153 459

AC C

EP

TE D

M AN U

Production of samples in Experiment 1, 2 and 3 (E1, E2, & E3) are distributed. Number of samples given for each sampling point throughout the experimental period.

ACCEPTED MANUSCRIPT

Variable

a

Mean

Minimum Maximum

Unit

N

Overall Appearance

1-5 scale 1-5 scale

458 458

3.02 3.37

0.40 0.45

1.00 2.00

Consistency Flavour

1-5 scale 1-5 scale

458 458

3.24 3.17

0.37 0.42

Sensory descriptive Pressure firmness

1-9 scale

459

5.51

Shear firmness Odour intensity

1-9 scale 1-9 scale

459 459

5.51 5.20

Acetic odour Elasticity

1-9 scale 1-9 scale

459 459

Cohesion Firmness chewing

1-9 scale 1-9 scale

Pasty Solubility

StDev

P Sample

P Age

4.00 4.25

0.093 0.005

0.078 0.754

2.33 1.00

4.00 4.00

0.521 0.019

0.146 0.006

0.59

3.80

7.50

<0.001

0.41 0.39

4.50 4.17

6.80 6.33

<0.001 <0.001

1.04 6.15

0.12 0.51

1.00 4.80

2.00 7.67

459 459

4.37 3.89

0.67 0.72

2.83 2.17

6.17 5.83

1-9 scale 1-9 scale

459 459

2.92 4.75

1.31 0.69

1.00 3.00

5.80 6.67

Dry Floury

1-9 scale 1-9 scale

459 459

1.30 2.90

0.36 0.48

1.00 1.33

Grainy Flavour intensity

1-9 scale 1-9 scale

459 459

1.75 5.69

0.74 0.50

1.00 4.00

Aromatic Nutty

1-9 scale 1-9 scale

459 459

3.79 1.02

0.53 0.09

Malty Sweetness

1-9 scale 1-9 scale

459 459

1.18 1.28

Sour Salty

1-9 scale 1-9 scale

459 459

5.89 5.10

Pungent Bitter

1-9 scale 1-9 scale

459 459

Sulphurous

1-9 scale

459

a

<0.001

M AN U

<0.001 <0.001

0.408 <0.001

0.062 <0.001

<0.001 <0.001

<0.001 <0.001

<0.001 <0.001

<0.001 <0.001

2.80 4.20

0.019 <0.001

<0.001 <0.001

4.33 7.00

<0.001 <0.001

<0.001 <0.001

5.20 1.60

0.009 <0.001

<0.001 <0.001

TE D

EP

AC C

2.60 1.00

SC

Quality scoring

RI PT

Table 2 Basic statistics for sensory variables.

0.26 0.27

1.00 1.00

3.25 2.71

<0.001 0.001

0.007 0.009

0.43 0.22

4.75 4.60

7.50 6.33

<0.001 <0.001

<0.001 <0.001

2.06 2.97

0.55 0.55

1.00 1.50

4.00 5.40

<0.001 <0.001

<0.001 <0.001

2.08

0.72

1.00

3.80

<0.001

<0.001

P values from ANOVA with two factors: Sample (n=153) and Age (n=3).

ACCEPTED MANUSCRIPT

Table 3 Cross correlation coefficient (Pearson) for chemical analytes and amino acids in cheese.

Asp

Glu

Asn

Ser

Gln

Thr

Cit

1.00 0.89 0.88 0.94 0.93 0.93 0.96 0.73 0.86 0.97 0.95 0.96 0.79 0.96 0.92 0.96

1.00 0.96 0.93 0.94 0.88 0.92 0.66 0.96 0.94 0.94 0.93 0.80 0.92 0.89 0.90

1.00 0.94 0.93 0.86 0.93 0.72 0.94 0.93 0.94 0.93 0.76 0.91 0.87 0.88

1.00 0.95 0.92 0.97 0.73 0.89 0.97 0.98 0.98 0.85 0.97 0.92 0.93

1.00 0.92 0.97 0.74 0.92 0.97 0.97 0.97 0.81 0.95 0.94 0.95

1.00 0.95 0.69 0.84 0.96 0.93 0.96 0.77 0.97 0.93 0.96

1.00 0.79 0.88 0.99 0.98 0.98 0.78 0.97 0.96 0.96

1.00 0.64 0.74 0.77 0.74 0.46 0.72 0.74 0.71

Tyr

Val

Met

Phe

Orn

Leu

Lys

Pro

1.00 0.94 0.96

1.00 0.95

1.00

M AN U

SC

RI PT

AN

1.00 0.91 0.91 0.91 0.75 0.90 0.85 0.87

1.00 0.99 1.00 0.81 0.99 0.95 0.98

1.00 0.99 0.83 0.97 0.95 0.96

1.00 0.81 0.99 0.95 0.97

EP

53 samples 8, 24, and 40 weeks age. Abbreviations are: SN, soluble nitrogen; AN, amino nitrogen;

AC C

a

SN 1.00 0.34 0.83 0.87 0.92 0.89 0.78 0.89 0.73 0.85 0.88 0.92 0.90 0.73 0.89 0.84 0.83

TE D

SN AN Asp Glu Asn Ser Gln Thr Cit Tyr Val Met Phe Orn Leu Lys Pro

a

1.00 0.78 0.78 0.80

ACCEPTED MANUSCRIPT

Table 4 Cross-validation measures for sensory variables.

SE Mean

PLS2 Opt n PC

RMSECV Chem FTIR NIR

Correlation coefficient, R RPD Fluor Spectr Chem FTIR NIR Fluor Spectr Chem FTIR

NIR

RER Fluor Spectr Chem FTIR

RI PT

Parameter

a

NIR

Fluor

Spectr

4

5

15

6

11

Overall quality

0.15

0.36

0.36

0.37

0.37

0.35

0.41

0.44

0.36

0.39

0.45

1.10

1.11

1.07

1.08

1.12

8.30

8.39

8.07

8.14

8.46

Appearance

0.14

0.44

0.44

0.42

0.44

0.43

0.09

0.13

0.36

0.16

0.25

1.00

1.01

1.07

1.02

1.03

5.06

5.11

5.39

5.12

5.20

Consistency

0.17

0.30

0.29

0.32

0.32

0.29

0.58

0.61

0.47

0.52

0.61

1.23

1.26

1.14

1.15

1.26

5.63

5.74

5.18

5.27

5.75

0.18

0.37

0.36

0.38

0.37

0.36

0.48

0.51

0.44

0.45

0.50

1.14

1.16

1.11

1.11

1.15

8.21

8.35

7.99

8.01

8.32

Pressure firmness

0.23

0.53

0.51

0.57

0.58

0.50

0.43

0.49

0.37

0.32

0.52

1.11

1.14

1.03

1.02

1.17

6.97

7.20

6.52

6.40

7.35

SC

Odour/flavour

0.21

0.36

0.39

0.38

0.38

0.35

0.48

0.29

0.38

0.52

0.50

1.14

1.06

1.08

1.06

1.16

6.44

5.96

6.09

6.01

6.53

Odour intensity

0.26

0.36

0.37

0.36

0.32

0.31

0.31

0.27

0.20

0.53

0.60

1.06

1.04

1.09

1.19

1.26

5.96

5.86

6.10

6.68

7.05

Elasticity

0.23

0.44

0.46

0.47

0.47

0.41

0.53

0.49

0.42

0.45

0.61

1.17

1.13

1.09

1.09

1.25

6.51

6.30

6.06

6.10

6.97

Cohesion

0.25

0.45

0.48

0.53

0.57

0.45

0.74

0.69

0.60

0.55

0.73

1.49

1.37

1.25

1.18

1.46

7.45

6.88

6.25

5.89

7.34

Firmness chewing

0.22

0.45

0.50

0.54

0.60

0.47

0.79

0.72

0.67

0.58

0.76

1.61

1.43

1.33

1.20

1.52

8.24

7.28

6.80

6.12

7.76

Pasty

0.27

0.73

0.85

0.98

1.03

0.80

0.83

0.76

0.67

0.62

0.79

1.78

1.55

1.33

1.27

1.63

6.53

5.68

4.89

4.68

5.98

Solubility

0.25

0.43

0.51

0.53

0.58

0.47

0.78

0.68

0.65

0.56

0.74

1.60

1.36

1.31

1.21

1.49

8.45

7.19

6.91

6.37

7.88

Dry

0.22

0.30

0.30

0.32

0.33

0.30

0.57

0.55

0.49

0.40

0.56

1.22

1.19

1.14

1.09

1.21

6.06

5.91

5.67

5.43

6.00

Floury

0.30

0.42

0.44

0.44

0.45

0.42

0.49

0.37

0.42

0.35

0.49

1.15

1.09

1.10

1.07

1.15

6.89

6.52

6.57

6.39

6.86

Grainy

0.25

0.47

0.54

0.58

0.62

0.52

0.78

0.70

0.63

0.56

0.72

1.57

1.39

1.28

1.19

1.42

7.05

6.21

5.72

5.35

6.38

Flavour intensity

0.24

0.35

0.35

0.39

0.36

0.32

0.72

0.71

0.63

0.70

0.78

1.42

1.42

1.28

1.39

1.58

8.49

8.46

7.62

8.29

9.44

Aromatic

0.37

0.45

0.44

0.46

0.47

0.43

0.54

0.55

0.52

0.51

0.60

1.16

1.18

1.15

1.11

1.22

5.76

5.86

5.68

5.49

6.04

Malty

0.10

0.26

0.26

0.26

0.25

0.25

0.21

0.22

0.22

0.25

0.35

1.00

1.00

1.00

1.02

1.05

8.72

8.71

8.70

8.84

9.10

EP

TE D

M AN U

Shear firmness

0.25

0.25

0.25

0.26

0.25

0.25

0.23

0.09

0.12

0.24

0.22

1.07

1.05

1.03

1.07

1.06

6.91

6.79

6.69

6.92

6.86

0.22

0.35

0.36

0.38

0.39

0.35

0.56

0.53

0.48

0.45

0.58

1.21

1.18

1.14

1.10

1.23

7.77

7.59

7.32

7.10

7.91

Salty

0.08

0.21

0.21

0.21

0.21

Pungent

0.35

0.42

0.43

0.46

0.47

Bitter

0.31

0.42

0.42

0.48

0.49

Sulphurous

0.30

0.61

0.64

0.63

0.64

4

6

13

13

0.22

0.45

0.46

0.54

0.47

3

3

14

16

0.41

0.43

0.49

0.36

PLS1 Opt n PC Firmness chewing PLS1 Opt n PC Bitter a

0.31

AC C

Sweetness Acidic

0.21

0.18

0.24

0.28

0.24

0.29

1.02

1.03

1.04

1.02

1.04

8.08

8.18

8.25

8.08

8.28

0.42

0.63

0.61

0.53

0.50

0.64

1.32

1.28

1.21

1.19

1.33

7.12

6.93

6.52

6.42

7.21

0.43

0.62

0.62

0.44

0.40

0.60

1.31

1.30

1.13

1.12

1.28

9.31

9.26

8.09

7.94

9.10

0.57

0.52

0.47

0.50

0.54

0.61

1.17

1.13

1.15

1.13

1.26

4.55

4.41

4.47

4.39

4.90

0.49

0.78

0.77

0.67

0.76

0.74

1.60

1.56

1.15

1.53

1.47

8.15

7.97

5.85

7.80

7.48

0.43

0.64

0.61

0.42

0.61

0.60

1.34

1.27

1.12

1.52

1.27

9.51

9.07

7.96

10.83

9.07

9

X values of calibrations are chemical variables, Fourier transform infra-red (FTIR), near infra-red (NIR), fluorescence, and all spectral data. SE Mean: standard error of the mean (between assessors), RMSECV: root mean square error of prediction - validated. Explained variance is measured in % of total Y-variance.

ACCEPTED MANUSCRIPT

AC C

EP

TE D

M AN U

SC

RI PT

Ratio performance deviation (RPD) = StDev(ref)/RMSECV, and Range error ratio (RER)=(Max value-Min value)/RMSEP. Opt n PLSn: optimal number of factors in PLS1/PLS2 modelling.

ACCEPTED MANUSCRIPT

Table 5 Cross-validation measures for sensory variables at 40 weeks age modelled from chemical and spectroscopic data.

a

RI PT

SC

M AN U

TE D

NIR 10 0.30 0.42 0.21 0.31 0.57 0.33 0.32 0.41 0.43 0.41 0.70 0.32 0.17 0.35 0.24 0.26 0.53 0.30 0.31 0.38 0.16 0.49 0.46 0.49

Correlation coefficient, R RPD Fluor Spectr 8w sens Chem FTIR NIR Fluor Spectr 8w sens Chem FTIR 6 7 7 0.28 0.28 0.28 0.38 0.38 0.35 0.45 0.46 0.43 1.08 1.07 0.38 0.39 0.38 0.34 0.14 0.30 0.47 0.45 0.43 1.06 1.00 0.21 0.21 0.19 0.54 0.48 0.54 0.50 0.52 0.61 1.19 1.13 0.28 0.28 0.31 0.41 0.41 0.40 0.53 0.54 0.43 1.09 1.09 0.53 0.47 0.50 0.56 0.51 0.24 0.37 0.55 0.44 1.20 1.17 0.32 0.32 0.32 0.60 0.30 0.45 0.49 0.49 0.49 1.25 1.05 0.30 0.30 0.30 0.39 0.01 0.16 0.32 0.31 0.35 1.10 0.99 0.41 0.38 0.39 0.45 0.36 0.37 0.35 0.51 0.38 1.06 1.03 0.44 0.39 0.38 0.51 0.54 0.38 0.28 0.53 0.53 1.15 1.18 0.45 0.37 0.35 0.57 0.67 0.52 0.35 0.63 0.66 1.20 1.33 0.67 0.60 0.56 0.51 0.44 0.31 0.37 0.55 0.64 1.16 1.11 0.31 0.30 0.30 0.46 0.35 0.37 0.38 0.44 0.51 1.12 1.05 0.18 0.17 0.18 0.08 -0.05 0.27 0.01 0.18 -0.14 0.98 0.96 0.34 0.31 0.39 0.59 0.47 0.59 0.62 0.70 0.45 1.27 1.17 0.25 0.23 0.22 0.21 0.21 0.26 -0.05 0.36 0.43 1.07 1.08 0.28 0.28 0.28 0.27 0.27 0.34 -0.01 0.16 0.02 0.99 1.00 0.51 0.52 0.51 0.28 0.07 0.13 0.25 0.16 0.16 1.00 0.95 0.29 0.29 0.30 0.16 0.26 0.22 0.26 0.31 0.25 0.96 0.98 0.30 0.28 0.31 0.31 0.14 0.22 0.32 0.45 0.19 1.12 1.06 0.38 0.33 0.35 0.41 0.43 0.26 0.25 0.54 0.42 1.12 1.12 0.17 0.16 0.17 0.43 0.41 0.51 0.42 0.47 0.37 1.10 1.15 0.50 0.50 0.51 0.53 0.44 0.41 0.35 0.35 0.26 1.23 1.15 0.43 0.45 0.46 0.45 0.13 0.36 0.43 0.38 0.33 1.12 1.00 0.48 0.48 0.46 0.14 0.27 0.15 0.13 0.21 0.29 0.97 1.01

EP

Opt n PC PLS2 Overall quality Appearance Consistency Odour/flavour Pressure firmness Shear firmness Odour intensity Elasticity Cohesion Firmness chewing Pasty Solubility Dry Floury Grainy Flavour intensity Aromatic Malty Sweetness Acidic Salty Pungent Bitter Sulphurous

StDev Range RMSECV Chem FTIR 6 4 0.32 2.17 0.29 0.29 0.44 2.00 0.41 0.44 0.24 1.33 0.21 0.22 0.33 2.33 0.31 0.31 0.57 2.75 0.47 0.49 0.37 1.80 0.29 0.35 0.32 1.83 0.29 0.33 0.42 2.25 0.39 0.40 0.46 2.50 0.40 0.39 0.47 2.83 0.39 0.36 0.72 4.00 0.62 0.65 0.33 2.67 0.30 0.32 0.17 1.00 0.18 0.18 0.44 2.34 0.35 0.38 0.26 1.43 0.24 0.24 0.27 1.60 0.27 0.26 0.51 2.20 0.50 0.53 0.29 2.25 0.30 0.30 0.33 1.71 0.30 0.32 0.4 2.50 0.36 0.36 0.18 1.08 0.17 0.16 0.55 3.00 0.45 0.48 0.48 3.15 0.43 0.49 0.47 2.00 0.49 0.47

AC C

Parameter

a

NIR 1.05 1.03 1.18 1.08 1.01 1.11 1.00 1.02 1.07 1.15 1.04 1.05 1.02 1.27 1.08 1.01 0.96 0.97 1.08 1.05 1.15 1.13 1.06 0.97

Fluor Spectr 8w sens 1.12 1.14 1.15 1.18 1.08 1.14 1.07 1.02 1.04 1.05 1.07 1.07 0.98 1.30 1.04 0.94 1.00 0.99 1.12 1.05 1.10 1.11 1.12 0.98

1.13 1.12 1.17 1.19 1.20 1.14 1.06 1.11 1.18 1.27 1.20 1.11 1.00 1.43 1.14 0.96 0.96 1.00 1.20 1.21 1.13 1.10 1.08 0.99

1.12 1.15 1.28 1.09 1.14 1.16 1.08 1.07 1.21 1.36 1.29 1.13 0.95 1.14 1.19 0.97 0.99 0.99 1.08 1.16 1.10 1.08 1.05 1.04

X values in the calibrations are chemical variables, Fourier transform infra-red (FTIR), near infra-red (NIR), fluorescence, and eight-week sensory data. SE Mean: standard error of the mean (between assessors), RMSECV: root mean square error of prediction - validated. Explained variance is measured in % of total Y-variance. Ratio performance deviation (RPD) = StDev(ref)/RMSECV. Opt n PLSn: optimal number of factors in PLS1/PLS2 modelling.

ACCEPTED MANUSCRIPT

Table 6 Basic statistics and cross-validation measures for chemical variables modelled from spectroscopic data.

a

mg g % of Tot N % of Tot N rel unit rel unit rel unit rel unit rel unit rel unit rel unit rel unit mmol kg-1 mmol kg-1 mmol kg-1 g kg-1 g kg-1 g kg-1

459 5.39 - 5.81 459 39.3 - 45.2 459 7.4 - 30.6 459 2.9 - 23.3 441 0.7 - 5.1 441 69.5 - 1289.6 440 0.4 - 3.4 440 0.2 - 7.2 440 0 - 3.3 440 0 - 10.3 440 0.1 - 4.6 440 0 - 32.5 326 0 - 0.5 326 7.2 - 12.8 326 0 - 1.2 312 0 - 0.3 312 0.1 - 1 312 8.9 - 14.6

% of Tot N

459

2.9 - 23.3

2.95

g kg-1

312

8.9 - 14.6

0.99

-1

0.07 1.05 4.61 2.95 0.52 232 0.48 1.19 0.61 1.44 0.70 4.78 0.12 0.78 0.13 0.05 0.17 0.99

FTIR 10 0.05 0.69 1.86 0.68 0.37 206 0.52 0.88 0.27 0.92 0.73 2.52 0.11 0.65 0.09 0.02 0.14 0.61 6 1.11 13 0.47

NIR 9 0.07 1.00 3.12 1.30 0.42 204 0.45 0.96 0.26 0.87 0.74 2.60 0.10 0.68 0.10 0.02 0.14 0.78 16 1.48 10 0.71

Fluor Spectra FTIR NIR Fluor 5 10 0.06 0.05 0.77 0.12 0.48 1.01 0.85 0.72 0.14 0.08 2.87 2.20 0.84 0.42 0.55 1.12 1.11 0.87 0.35 0.57 0.38 0.39 0.53 0.28 0.51 214 183 0.51 0.51 0.42 0.53 0.38 0.52 0.68 0.50 0.87 0.99 0.50 0.33 0.51 0.28 0.44 0.58 0.60 0.50 0.90 1.23 0.29 0.38 0.24 0.72 0.55 0.50 0.45 0.50 2.59 4.11 0.38 0.26 0.24 0.11 0.08 0.61 0.67 0.54 0.67 0.53 0.63 0.59 0.60 0.10 0.08 0.51 0.35 0.47 0.03 0.03 0.38 0.32 0.14 0.13 0.11 0.53 0.48 0.62 0.75 0.58 0.74 0.49 0.55 16 6 1.20 1.10 0.86 0.74 0.84 10 18 0.69 0.46 0.89 0.71 0.72

RER

Spectra FTIR

NIR Fluor Spectra FTIR

NIR

Fluor Spectra

1.02 1.04 1.48 2.28 1.23 1.14 1.07 1.24 2.35 1.66 0.95 1.84 1.19 1.15 1.31 1.92 1.22 1.27

5.90 5.88 7.44 15.75 10.41 5.99 6.71 7.29 12.49 11.84 6.06 12.51 5.09 8.22 11.80 12.29 6.34 7.27

RI PT

StDev

RPD

0.69 0.58 0.85 0.86 0.64 0.59 0.57 0.53 0.68 0.50 0.61 0.51 0.65 0.59 0.68 0.55 0.68 0.70

1.57 1.51 2.48 4.34 1.39 1.13 0.91 1.35 2.28 1.57 0.97 1.89 1.11 1.20 1.42 1.94 1.26 1.63

1.17 1.04 1.60 2.62 1.38 1.08 0.91 1.37 2.17 1.60 0.98 1.85 1.05 1.16 1.40 1.84 1.38 1.33

1.39 1.23 2.10 2.66 1.33 1.27 1.25 1.20 1.38 1.18 1.29 1.16 1.55 1.47 1.61 1.44 1.64 1.71

9.11 8.49 12.49 30.03 11.73 5.93 5.73 7.91 12.15 11.21 6.17 12.89 4.72 8.60 12.81 12.44 6.51 9.34

6.79 5.84 8.08 18.15 11.68 5.70 5.71 8.05 11.57 11.39 6.23 12.56 4.48 8.35 12.59 11.78 7.15 7.61

8.08 6.94 10.56 18.37 11.25 6.68 7.84 7.05 7.37 8.39 8.25 7.91 6.61 10.54 14.45 9.24 8.51 9.79

0.86

2.66 1.99 2.46

2.68

18.38 13.78 17.00

18.55

0.88

2.11 1.40 1.44

2.16

12.13

12.39

SC

Opt n PC PLS2 pH Tot N Soluble N Amino N Acetalehyd Etanol Aceton i-Propanol n-Propanol 2-butanone Diacetyl 2-butanol Acetoin Acetic acid Butyric acid Pyruvate Succinic acid Lactic acid Opt n PC PLS1 Amino N Opt n PC PLS1 Lactic acid

Min-Max

Correlation coeff., R

M AN U

N

RMSECV

TE D

Basic statistics

EP

Unit

AC C

Parameter

a

8.03

8.26

X values in the calibrations are Fourier transform infra-red (FTIR), near infra-red (NIR), and fluorescence data. Spectra=FTIR+NIR+Fluorescence combined spectra. StDev: Total standard deviation (between samples- reference method). RMSECV: root mean square error of prediction - validated. Ratio

performance deviation (RPD) = StDev/RMSECV. Range error ratio (RER) = (Max-Min)/RMSEP. Opt n PLSn: optimal number of factors in PLS1/PLS2 modelling.

ACCEPTED MANUSCRIPT

Table 7. Basic statistics and crossvalidation measures for chemical variables modelled from spectroscopic data.

53 53 53 53 53 53 53 53 53 53 53 53 53 52 53 53 53

1.42 - 8.74 11.3 - 60.5 9.0 - 29.8 2.46 - 17.3 3.04 - 27.5 2.64 - 26.4 1.74 - 12.7 1.24 - 11.1 1.5 - 12.3 4.94 - 35.1 1.3 - 10.3 1.54 - 21.8 5.4 - 23.0 4.86 - 32.5 13.9 - 58.8 0.04 - 37.6 3.46 - 33.1

1.96 12.61 5.19 3.95 6.90 6.34 3.01 2.08 2.63 8.12 2.59 5.28 4.65 6.92 11.98 10.44 8.10

53

4.94 - 35.1

8.12

53

1.24 - 11.1

2.08

FTIR 3 1.23 8.29 2.90 2.42 3.84 4.08 1.75 1.69 1.74 4.38 1.51 3.24 2.50 5.32 6.14 6.46 4.71 3 4.32 3 1.69

NIR 10 1.53 9.55 3.88 3.13 5.63 4.81 2.26 1.71 2.15 6.05 1.87 3.86 3.51 5.40 9.28 8.16 6.34 12 6.30 6 1.71

Fluor 6 0.90 6.48 2.46 1.82 3.58 3.18 1.34 1.46 1.59 3.65 1.14 2.38 2.23 4.46 6.07 5.75 3.93 9 3.26 5 1.37

Spectra 9 1.08 7.27 2.36 1.97 2.92 3.06 1.33 1.40 1.63 3.37 1.09 2.58 2.00 4.22 5.13 5.32 3.39 10 3.33 7 1.23

FTIR NIR

Fluor Spectra

FTIR

NIR

Fluor Spectra FTIR

NIR

Fluor Spectra

0.78 0.75 0.83 0.79 0.83 0.76 0.81 0.59 0.75 0.84 0.81 0.79 0.84 0.63 0.86 0.78 0.81

0.89 0.86 0.88 0.89 0.86 0.87 0.90 0.71 0.81 0.90 0.90 0.89 0.88 0.76 0.87 0.84 0.88

0.84 0.82 0.89 0.87 0.90 0.88 0.90 0.74 0.79 0.91 0.91 0.87 0.90 0.79 0.90 0.86 0.91

1.60 1.52 1.79 1.63 1.79 1.55 1.72 1.23 1.51 1.85 1.72 1.63 1.86 1.30 1.95 1.62 1.72

1.28 1.32 1.34 1.26 1.23 1.32 1.33 1.21 1.22 1.34 1.39 1.37 1.32 1.28 1.29 1.28 1.28

2.17 1.95 2.11 2.17 1.93 2.00 2.24 1.42 1.65 2.23 2.28 2.22 2.08 1.55 1.97 1.81 2.06

1.82 1.74 2.20 2.00 2.36 2.08 2.27 1.49 1.61 2.41 2.37 2.05 2.32 1.64 2.33 1.96 2.39

5.97 5.94 7.16 6.14 6.39 5.83 6.32 5.88 6.25 6.89 6.00 6.26 7.07 5.20 7.32 5.81 6.29

4.80 5.15 5.34 4.76 4.36 4.95 4.89 5.79 5.05 4.99 4.84 5.25 5.04 5.12 4.84 4.60 4.67

8.11 7.59 8.46 8.18 6.85 7.50 8.23 6.77 6.82 8.27 7.93 8.52 7.93 6.20 7.41 6.52 7.53

6.80 6.77 8.81 7.55 8.39 7.79 8.32 7.09 6.67 8.97 8.26 7.86 8.82 6.55 8.75 7.05 8.73

0.84 0.66 0.92

0.91

1.88

1.29

2.49

2.44

6.99

4.79

9.26

9.06

0.58 0.58 0.76

0.80

1.23

1.22

1.52

1.69

5.87

5.80

7.24

8.07

0.65 0.67 0.68 0.64 0.62 0.67 0.68 0.59 0.61 0.69 0.71 0.70 0.68 0.64 0.66 0.64 0.65

RER

RI PT

StDev

SC

Min-Max

RPD

TE D

N

Correlation coefficient, R

M AN U

RMSECV

EP

PLS2 Opt n PC Asp Glu Asn Ser Gln His +Gly Thr Cit Tyr Val Met Trp+Ile Phe Orn Leu Lys Pro PLS1 Opt n PC Val PLS1 Opt n PC Cit

Basic statistics

AC C

Parameter

a

a Units for amino acids are mmoles kg-1. X values in the calibrations are Fourier transform infra-red (FTIR), near infra-red (NIR), and fluorescence data. Spectra=FTIR+NIR+Fluorescence combined spectra. StDev: Total standard deviation (between samples- reference method). RMSECV: root mean square error of

prediction - validated. Ratio performance deviation (RPD) = StDev/RMSECV. Range error ratio (RER) = (Max-Min)/RMSEP. Opt n PLSn: pptimal number of factors in PLS1/PLS2 modelling.

AC C

EP

TE D

M AN U

SC

RI PT

ACCEPTED MANUSCRIPT

AC C

EP

TE D

M AN U

SC

RI PT

ACCEPTED MANUSCRIPT

AC C

EP

TE D

M AN U

SC

RI PT

ACCEPTED MANUSCRIPT

AC C

EP

TE D

M AN U

SC

RI PT

ACCEPTED MANUSCRIPT

ACCEPTED MANUSCRIPT

b

M AN U

SC

RI PT

a

d

AC C

EP

TE D

c