Determination of main constituents in wheat using near infrared hyperspectral imaging.

摘要

Differentiation of wheat classes and rapid measurement of main constituents (e.g., protein, starch, oil content, and moisture content) in wheat are important challenges to the grain industry. In this study, NIR reflectance and absorbance values of hyperspectral images of wheat samples were used for identifying the Canadian wheat classes at same and different moisture levels and for predicting protein and oil content of wheat. Images of wheat were obtained using a NIR hyperspectral imaging system. Seventy five normalized NIR mean reflectance and NIR absorbance features were extracted from the scanned images of wheat. The extracted features were used to develop classification models and prediction models for identifying wheat classes; and predicting protein and oil contents of wheat, respectively.; Classification accuracies were 100% in classifying Canada Prairie Spring Red (CPSR), Canada Western Extra Strong (CWES), Canada Western Hard White Spring (CWHWS), Canada Western Red Spring (CWRS), Canada Western Red Winter (CWRW), and Canada Western Soft White Spring (CWSWS) wheat; and > 98% for the other two wheat classes (Canada Prairie Spring White (CPSW) and Canada Western Amber Durum (CWAD)) using linear discriminant analysis (LDA) with leave-one-out cross validation. Using quadratic discriminant analysis (QDA) with leave-one-out cross validation, the classification accuracies were > 96% for all wheat classes. The classification accuracies of 80--100% and 89--100% were found for artificial neural network (ANN) models with two different training patterns such as 60% training--30% test--10% validation (60-30-10) and 70% training--20% test--10% validation (70-20-10), respectively.; Classification accuracies of 100% were achieved using LDA with leave-one-out cross validation for CWSWS wheat at 14, 16, 18, and 20% moisture levels with 75 features. And, classification accuracies of > 90% were achieved for all wheat classes except CWES wheat at 20% moisture level and CWHWS wheat at 14% moisture level in LDA with leave-one-out cross validation using 75 features. Plots of the first two canonical variables showed that protein and moisture contents of wheat could be predicted using the NIR absorbance values of hyperspectral images. Principal components analysis (PCA) and STEPDISC procedure were used to find the top wavelengths in wheat class identification.; A 75 feature PLSR model for predicting protein in wheat produced the best standard error of prediction (SEP = 0.68) and a good correlation (r = 0.94) with the measured protein in wheat. Also, the 75 feature PLSR model for predicting oil content in wheat produced the best SEP of 0.10 and a r value of 0.83 with the measured oil content in wheat. Results of this study showed that NIR hyperspectral imaging could be used as an effective method for predicting protein and oil contents in wheat and identifying wheat classes at different moisture levels.