Comparison of PLS1 and PLS2 for Prediction of Protein, Oil, Carbohydrate, and Water Contents in Soybean using FT-NIR Spectroscopy
As standard measurement methods of these contents in soybean have several disadvantages of the measurements, such as time consuming, labor intensive and destructive procedure which hamper fast and economic quality evaluation for soybean in massive production. Hence, the demand of rapid, robust, and nondestructive measurement methods for main nutrient components, such as protein, oil, carbohydrate, and water contents of soybean has been increased. The near-infrared spectroscopic technique is well known as methods for the non-destructive quality measurement of a variety of food products. Furthermore, this nondestructive spectroscopic technique with chemometrics can provide physical and chemical characteristic of the information on specimens. The purpose of this study were to develop an optimal prediction model for determining protein, oil, carbohydrate and water content of soybeans using a NIR spectroscopic technique and to compare performance of partial least square regression methods (1 and 2). 207 kinds of soybean crops were used to measure FT-NIR spectrum data. The PLS2 regression had better performance for protein, oil and carbohydrate contents in soybean. In addition, PLS2 model was less complex than the PLS1 models based on number of factors. The standard error of prediction (SEP) of protein, oil, carbohydrate, and water contents in soybean by PLS2 model were 0.506 %, 2.124 %, 2.511 %, and 0.582 %, respectively. This study shows that PLS2 model for several dependant variables has good potential to develop prediction models for major components in soybean.