Namur, 16/04/2015

Grelet C 1, Fernandez J.A 1, Dardenne P 1, Baeten V 1, Vanlierde A 1, Soyeurt H 2 , Darimont C 1 & Dehareng F 1

1 Walloon Agricultural Research Center (CRA-W), Gembloux, Belgique

2 University of Liège, Gembloux Agro-Bio Tech, Gembloux, Belgique c.grelet@cra.wallonie.be

Standardisation of milk MIR spectra,

Development of common MIR equations

0 50 100 150 200 250 300-0.04

-0.03

-0.02

-0.01

0

0.01

0.02

0.03

0.04

0 100 200 300 400 500

-0.1

0

0.1

0.2

0.3

0.4

0.5

Spectra after slave standardization

wavelength (cm-1)

A

0 100 200 300 400 500

-0.1

0

0.1

0.2

0.3

0.4

0.5

Spectra after slave standardization

wavelength (cm-1)

A

0 100 200 300 400 500

-0.1

0

0.1

0.2

0.3

0.4

0.5

Spectra after slave standardization

wavelength (cm-1)

A

0 100 200 300 400 500

-0.1

0

0.1

0.2

0.3

0.4

0.5

Spectra after slave standardization

wavelength (cm-1)

A

0 100 200 300 400 500

-0.1

0

0.1

0.2

0.3

0.4

0.5

Spectra after slave standardization

wavelength (cm-1)

A

0 100 200 300 400 500

-0.1

0

0.1

0.2

0.3

0.4

0.5

Spectra after slave standardization

wavelength (cm-1)

A

OptiMir

Develop MIR models predicting cow state:

Variability is

needed to build robust models

67 instruments

-0.015 -0.01 -0.005 0 0.005 0.01 0.015 0.02

-0.02

-0.015

-0.01

-0.005

0

0.005

0.01

0.015

Scores on PC 4 (0.01%)

Score

s o

n P

C 5

(0.0

1%

)Samples/Scores Plot of X212

Variability of instruments PCA on informative wavenumbers , for 5

common samples analysed on 45 instruments from the same brand

1000 1200 1400 1600 1800 2000 2200 2400 2600 2800 3000-0.2

-0.1

0

0.1

0.2

0.3

0.4

Spectra after log

wavelength (cm-1)

A

Machine Y

Machine X

Wavenumber (cm-1)

Ab

sorb

ance

Instruments are different

Common milks analysed on 2 instruments from 2 brands

0

100

200

300

400

500

600

700

800

900

0 100 200 300 400 500 600 700 800

Mes

ure

d g

CH

4 /

day

Prediction by MIR (g CH4 / day)

Common milks analysed on 2 instruments from 2 brands Methane model (A.Vanlierde, 2013)

Creation and use of common models?

N= 5y= 590.7677+0.074968xR square= 0.0181biais= -766.7733RMSE= 778.462

-400 -200 0 200 400 600 800480

500

520

540

560

580

600

620

640

660

680CH4 (g/day) predictions without standardisation, Master vs BEL0201

Slave predictions

Maste

r pre

dic

tions

Common milks analysed on 2 instruments from 2 brands Methane model

Creation and use of common models?

Machine X : 670 g CH4/day

Machine Y : -230 g CH4/day

Predictions Instrument Y (g CH4/j)

Pre

dic

tio

ns

inst

rum

en

t X

(g

CH

4/j

)

Not possible to use a common model on 2

instruments from 2 brands

N= 5y= -5.90+1.84xRMSE= 200.53

150 200 250 300 350 400 450 500 550 600300

350

400

450

500

550

600CH4 Master vs. CH4 Slave

Slave CH4 predictions (g CH4/day)

Maste

r C

H4 p

redic

tions (

g C

H4/d

ay)

Common milks analysed on 2 instruments from the same brand Methane model

Creation and use of common models?

Machine X : 550 g CH4/day

Machine Z : 290 g CH4/day

Predictions Instrument Z (g CH4/j)

Pré

dic

tio

ns

inst

rum

en

t X

(g

CH

4/j

)

Not possible to use a common model on 2 instruments from

the same brand

Instrument X European Master

3.5

3.6

3.7

3.8

3.9

4

4.1

20

11

12

12

20

11

12

15

20

11

12

21

20

11

12

26

20

11

12

30

20

12

01

04

20

12

01

10

20

12

01

13

20

12

01

20

20

12

01

25

20

12

01

31

20

12

02

10

20

12

02

16

20

12

02

21

20

12

02

24

20

12

03

01

20

12

03

05

20

12

03

07

20

12

03

13

20

12

03

16

20

12

03

22

20

12

03

27

20

12

04

02

20

12

04

03

20

12

04

05

20

12

04

10

20

12

04

12

20

12

04

13

20

12

04

19

20

12

04

24

20

12

04

30

20

12

05

04

20

12

05

10

20

12

05

15

20

12

05

22

20

12

05

25

20

12

06

01

20

12

06

06

20

12

06

13

20

12

06

18

20

12

06

20

20

12

06

22

20

12

06

26

20

12

06

28

20

12

07

03

20

12

07

06

20

12

07

12

Fat

pre

dic

tio

n (

g/1

00

ml)

Fat prediction from raw spectra of an instrument analyzing an UHT milk

Lot 1

Lot 2

Lot 3

Lot 4

6 months

Instruments are not stable in time

Instrument X European Master

Instruments also suffer from big perturbations maintenance operation, piece replacement

3.45

3.5

3.55

3.6

3.65

3.7

3.75

3.8

3.85

3.9

3.95

4

20

11

12

19

20

11

12

21

20

11

12

23

20

11

12

29

20

12

01

02

20

12

01

04

20

12

01

09

20

12

01

13

20

12

01

20

20

12

01

26

20

12

01

30

20

12

02

01

20

12

02

06

20

12

02

09

20

12

02

14

20

12

02

17

20

12

02

21

20

12

03

01

20

12

03

05

20

12

03

12

20

12

03

13

20

12

03

14

20

12

03

15

20

12

03

19

20

12

03

28

20

12

04

03

20

12

04

18

20

12

05

02

20

12

05

08

20

12

05

15

20

12

05

21

20

12

05

30

20

12

06

04

20

12

06

08

20

12

06

13

20

12

06

19

20

12

06

21

20

12

06

22

20

12

06

26

20

12

06

29

20

12

07

06

20

12

07

12

Fat

pre

dic

tio

n (

g/1

00

ml)

Fat prediction from raw spectra of an instrument analyzing an UHT milk

Lot 1

Lot 2

Lot 3

6 months

Instruments are not stable in time

Issues:

• Not possible to create common tools

• Not possible to transfer a model on other instruments

• Instruments not stable in time

Slope/bias correction not possible for models predicting methane, fertility, ketosis… Direct Standardisation of the spectra

A model can be used on only 1 instrument and within a limited time

Brand C

Brand B

Brand A

Variability between instruments

How it works

Selected instruments for Master creation

How it works

Brand C

Brand B

Brand A

Selected instruments for Master creation

Master

How it works

Brand C

Brand B

Brand A

Master

18 Selected instruments for Master creation

How it works

Brand C

Brand B

Brand A

Master

18 Selected instruments for Master creation

How it works

Brand C

Brand B

Brand A

Master

18 Selected instruments for Master creation

How it works

Brand C

Brand B

Brand A

Standardized instruments

Master

Standardised instruments for

creation of equations

Standardised instruments for use

of equations

18 Selected instruments for Master creation

How it works

Brand C

Brand B

Brand A

Absorbance in an area r1 (master) correlated to the absorbance within R2 (slaves)

« Master »

« Slave »

r1

R2

PIECE-WISE DIRECT STANDARDIZATION (PDS)

r1j = R2j bj + b0j

How it works

0

1

2

3

4

5

0 2 4 6

Pro

tein

s %

Fat %

Ring test samples composition

27 Labs 67 Apparatus

How it works

PDS

0 100 200 300 400 500

-0.1

0

0.1

0.2

0.3

0.4

0.5

Spectra after log

wavelength (cm-1)

A

0 100 200 300 400 500

-0.1

0

0.1

0.2

0.3

0.4

0.5

Spectra after slave standardization

wavelength (cm-1)

A

Machine Y

Master

Machine Y

Master

Results

-0.015 -0.01 -0.005 0 0.005 0.01 0.015 0.02

-0.02

-0.015

-0.01

-0.005

0

0.005

0.01

0.015

Scores on PC 4 (0.01%)

Score

s o

n P

C 5

(0.0

1%

)Samples/Scores Plot of X212

Variability of instruments PCA on informative wavenumbers , for 5

common samples analysed on 45 instruments from the same brand

Reduce the variability of instruments

-0.01 -0.005 0 0.005 0.01 0.015 0.02 0.025

-0.02

-0.015

-0.01

-0.005

0

0.005

0.01

0.015

Scores on PC 4 (0.01%)

Score

s o

n P

C 5

(0.0

0%

)Samples/Scores Plot of X212

NONSTD

STD

MASTER

MASTER

N= 5y= 3.15+0.99xRMSE= 21.1816

300 350 400 450 500 550 600300

350

400

450

500

550

600CH4 Master vs. CH4 Slave

Slave CH4 predictions (g CH4/day)

Maste

r C

H4 p

redic

tions (

g C

H4/d

ay)

N= 5y= -5.90+1.84xRMSE= 200.53

150 200 250 300 350 400 450 500 550 600300

350

400

450

500

550

600CH4 Master vs. CH4 Slave

Slave CH4 predictions (g CH4/day)

Maste

r C

H4 p

redic

tions (

g C

H4/d

ay)

Predictions slave (g/d/cow)

Pré

dic

tio

ns

mas

ter

(g/d

/co

w)

Common milks analysed on the master and on a slave instrument from the same brand Methane model

Creation and use of common models?

Without standardisation With standardisaton

Predictions slave (g/d/cow)

Pré

dic

tio

ns

mas

ter

(g/d

/co

w)

Prédiction master : 560 g/d/cow

Prédiction slave : 545 g/d/cow Possible to use a common model on 2 instruments from

the same brand

N= 5y= 590.7677+0.074968xR square= 0.0181biais= -766.7733RMSE= 778.462

-400 -200 0 200 400 600 800480

500

520

540

560

580

600

620

640

660

680CH4 (g/day) predictions without standardisation, Master vs BEL0201

Slave predictions

Maste

r pre

dic

tions

N= 5y= -28.4074+1.0493xR square= 0.89163biais= 7.8444e-013RMSE= 23.091

450 500 550 600 650 700450

500

550

600

650

700CH4 (g/day) predictions after standardisation, Master vs BEL0201

Slave predictions

Maste

r pre

dic

tions

Predictions slave, brand B (g/d/cow)

Pré

dic

tio

ns

mas

ter

(g/d

/co

w)

Common milks analysed on the master and on a slave instrument from another brand Methane model

Creation and use of common models?

Without standardisation With standardisaton

Predictions slave, brand B (g/d/cow)

Pré

dic

tio

ns

mas

ter

(g/d

/co

w)

Prédiction master : 570 g/d/cow

Prédiction slave : 550 g/d/cow Possible to use a common model on 2 instruments from

different brands

47 instruments (7 brand A, 1 brand B, 39 brand C) Methane model RMSE between master and slaves predictions, before and after standardisation

Brand A Brand B Brand C Global average

RMSE before PDS 195.54 778.46 132.15 155.34

RMSE after PDS 35.85 23.09 20.48 22.83

0.00

100.00

200.00

300.00

400.00

500.00

600.00

700.00

800.00

900.00

RM

SE o

f C

H4

(g/

d/c

ow

)

RMSE of CH4 predictions between master and slaves

Creation and use of common models?

47 instruments (7 brand A, 1 brand B, 39 brand C) Poly-Unsaturated Fatty acids model RMSE between master and slaves predictions, before and after standardisation

Brand A Brand B Brand C Global average

RMSE before PDS 0.0749 0.1175 0.0206 0.0308

RMSE after PDS 0.0092 0.0053 0.0032 0.0041

0.0000

0.0200

0.0400

0.0600

0.0800

0.1000

0.1200

0.1400

RM

SE P

UFA

(g/

10

0m

l)

RMSE of PUFA predictions between master and slaves

Creation and use of common models?

• 1827 milk samples (standardized since 2012) • Brand A • Brand C

• Large variability of breeds (17), seasons,

feeding systems and geographical areas:

Belgium and Luxembourg: brand C instruments

France : Brand A and C instruments

Germany: Brand A and C instruments

Ireland: Brand C instruments

Finland: Brand C instruments

UK: Brand C instruments

Fatty acids models Description Mean SECV R²cv RPDcv Use

3.912 0.007 1.00 132.99 +++

C4:0 0.104 0.008 0.93 3.67 +

C6:0 0.072 0.006 0.91 3.32 +

C8:0 0.047 0.004 0.91 3.29 +

C10:0 0.111 0.010 0.91 3.37 +

C12:0 0.133 0.012 0.92 3.62 +

C14:0 0.445 0.031 0.93 3.88 +

C14:1 cis 0.038 0.008 0.68 1.78 -

C16:0 1.192 0.093 0.94 4.18 +

C16:1 cis 0.066 0.013 0.73 1.91 -

C17:0 0.027 0.003 0.80 2.24 0

C18:0 0.393 0.057 0.84 2.51 0

Total of C18:1 trans 0.125 0.026 0.79 2.17 0

C18:1 cis9 0.751 0.062 0.95 4.35 +

Total of C18:1 cis 0.808 0.064 0.95 4.58 +

Total of C18:2 0.096 0.015 0.69 1.79 -

C18:2 cis9, cis12 0.061 0.012 0.72 1.91 -

C18:3 cis9, cis12, cis 15 0.020 0.004 0.68 1.77 -

C18:2 cis 9, Trans 11 0.028 0.010 0.74 1.95 -

Saturated FA 2.689 0.072 0.99 10.22 +++

Mono-unsaturated FA 1.077 0.058 0.97 5.83 ++

Poly-unsaturated FA 0.159 0.021 0.77 2.10 0

Unsaturated FA 1.237 0.065 0.97 5.75 ++

Short chain FA 0.347 0.025 0.93 3.88 +

Mid chain FA 1.982 0.105 0.97 5.53 ++

Long chain 1.580 0.110 0.95 4.52 +

Bbranched FA : iso + anteiso 0.090 0.012 0.75 2.00 0

Omega 3 0.026 0.006 0.66 1.73 -

Omega 6 0.103 0.015 0.72 1.89 -

Odd FA 0.155 0.016 0.83 2.41 0

Trans FA 0.160 0.030 0.80 2.26 0

C18:1 0.936 0.061 0.96 5.18 ++

Cross validation with 4 subsets

R²cv RPDcv Use

1.00 132.99 +++

0.93 3.67 +

0.91 3.32 +

0.91 3.29 +

0.91 3.37 +

0.92 3.62 +

0.93 3.88 +

0.68 1.78 -

0.94 4.18 +

0.73 1.91 -

0.80 2.24 0

0.84 2.51 0

0.79 2.17 0

0.95 4.35 +

0.95 4.58 +

0.69 1.79 -

0.72 1.91 -

0.68 1.77 -

0.74 1.95 -

0.99 10.22 +++

0.97 5.83 ++

0.77 2.10 0

0.97 5.75 ++

0.93 3.88 +

0.97 5.53 ++

0.95 4.52 +

0.75 2.00 0

0.66 1.73 -

0.72 1.89 -

0.83 2.41 0

0.80 2.26 0

0.96 5.18 ++

Cross validation with 4 subsets

Reference values (SF6)

Methane emitted by dairy cows (A.Vanlierde, 2013)

• 84 from Brand B • 368 from Brand C

0

100

200

300

400

500

600

700

800

900

0 100 200 300 400 500 600 700 800

Mes

ure

d g

CH

4 /

day

Prediction g CH4 / day

N = 452

R²c = 0,76

SECV=69

RPD=1.83

Methane model

0

2

4

6

8

10

12

14

16

18

20

0 5 10 15 20

Cit

rate

s R

éfé

ren

ce

Citrates Predictions

Citrates (mmol/L)

BHB, acetone and citrates models

• 536 milk samples all standardized • Brand A • Brand C

From France, Germany and Luxembourg

Monthly since january 2012, 67 instruments into 27 labs Merging database

Creation of common models, more robust

Use of the models by all instruments

Conclusion

Common langage for all instruments

Equations become universal

Standardisation of milk mid-infrared spectra from a European dairy network, C. Grelet, J.A. Fernández Pierna, P. Dardenne, V. Baeten, F. Dehareng,

Journal of Dairy Sciences, 2015, 98 :2150–2160

New tools for a more sustainable dairy sector www.optimir.eu

WITH THE SUPPORT OF

Thank you for your attention