Premium
Measurement of soybean fatty acids by near‐infrared spectroscopy: Linear and nonlinear calibration methods
Author(s) -
Kovalenko Igor V.,
Rippke Glen R.,
Hurburgh Charles R.
Publication year - 2006
Publication title -
journal of the american oil chemists' society
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.512
H-Index - 117
eISSN - 1558-9331
pISSN - 0003-021X
DOI - 10.1007/s11746-006-1221-z
Subject(s) - partial least squares regression , support vector machine , calibration , stearic acid , chemistry , artificial neural network , analytical chemistry (journal) , mathematics , biological system , chromatography , artificial intelligence , computer science , statistics , organic chemistry , biology
Abstract A key element of successful development of new soybean cultivars is availability of inexpensive and rapid methods for measurement of FA in seeds. Published research demonstrated applicability of NIR spectroscopy for FA profiling in oilseeds. The objectives of this study were to investigate the applicability of NIR spectroscopy for measurement of FA in whole soybeans and compare performance of calibration methods. Equations were developed using partial least squares (PLS), artificial neural networks (ANN), and support vector machines (SVM) regression methods. Validation results demonstrated that (i) equations for total saturates had the highest predictive ability ( r 2 =0.91–0.94) and were usable for quality assurance applications, (ii) palmitic acid models ( r 2 =0.80–0.84) were usable for certain research applications, and (iii) equations for stearic ( r 2 =0.49–0.68), oleic ( r 2 =0.76–0.81), linoleic ( r 2 =0.73–0.76), and linolenic ( r 2 =0.67–0.74) acids could be used for sample screening. The SVM models produced significantly more accurate predictions than those developed with PLS. ANN calibrations were not different from the other two methods. Reduction in the number of calibration samples reduced predictive ability of all equations. The rate of performance degradation of SVM models with sample reduction was the lowest.