Sugar measurements in soybeans using Near Infrared Spectroscopy

1
Sugar measurements in soybeans using Near
Infrared Spectroscopy
Lidia Esteve Agelet Term project for the
coursework AE 569
,

• Results
• No significant differences in calibrations from
  different instruments or treatment of data.
• Coefficients of determination for PLS regressions
  were very low for both sucrose (best model 0.20)
  and stachyoseraffinose (best model 0.32).
• SEP -1.3 for sucrose and 0.7 for
  stachyoseraffinose.
• Coefficients of determination for MLR calibration
  were higher for both compounds (around 0.70).
  Low coefficients of determination in cross
  validation (around 0.01) indicate overfitting and
  collinearity between spectral data.
• SL-SVM improved the results slightly. Calibration
  model for sucrose with first derivative applied
  gave R20.45, and R20.58 for stachyoseraffinose
  (Fig 4 and Fig 5)
• Calibration results for more uniform data did not
  give significant improvement either for PLS
  calibrations or MLR. (Fig 3)
• The standard errors of lab on the five duplicates
  were very high

• Introduction
• Soluble carbohydrates are the third compound of
  soybeans by weight (11), after protein ( 40)
  and oil (21) (dry basis)
• The major soluble carbohydrates are sucrose (6 -
  8), stachyose (1.4 -4.1), and raffinose (0.1
  0.9 ) (Cicek, 2001).
• High sucrose content is desirable for some
  soyfood production.
• Raffinose and stachyose are undesirable due to
  low digestibility.

Objective 1) Develop a NIR calibrations for
measuring sucrose, raffinose and stachyose in
whole soybeans.
Fig 4. LS-SVM for stachyoseraffinose,
first derivative applied
Fig 5. LS-SVM for sucrose,
first derivative applied

• Conclusions
• No model for sucrose measurement is good enough
  to be used (best r20.45 for SL-SVM)
• The best model for stachyoseraffinose could be
  used for rough screening (r20.58 for SL-SVM)
• The standard error of the lab is high, and is a
  limiting factor to get better models of
  calibration, so higher sample size does not
  ensure a better calibration model (Kovalenko,
  2005)

• Materials and methods
• Samples
• 170 soybean samples ( Crop years 2003 and 2004).
  5 samples duplicated. Samples were selected
  according their protein and oil content, to get a
  uniform sugar distribution.
• Instruments
• Foss Infratec 1229 (fig 1) and Foss Infratec
  1241(fig 2), transmittance units
• 850 -1048 nm, 2nm sampling interval, whole grain
  cell.
• Fig 1. Foss infratec 1229
  Fig 2. Foss infratec 1241
• Methods
• Spectral data were analyzed without treatment,
  and with First and Second Svitsky-Golay
  derivatives
• Linear calibration methods with one-out cross
  validation
• Partial Least Squares and Multiple Linear
  Regression

• Acknowledgements
• The author wants to thank Igor Kovalenko for his
  valuable help in the elaboration of this project,
  as well as Charles R. Hurburgh and Glen Rippke
  for their advices.

Fig 3. MLR calibration for sucrose, uniform
distribution without pretreatment
.
References Cicek, M. 2001. Genetic marker
analysis of three major carbohydrates in soybean
seeds. PhD dissertation. Blacksburg, Virginia
Virginia Polytechnic Institute and State
University, Department of Crop and Environmental
Sciences. Kovalenko, I. 2005. Near-Infrared
(NIR) Spectroscopy and Chemometrics Primer.
Manual for the grain quality laboratory, Iowa
State University.