Comparing two statistical discriminant models with a back-propagation neural network model for pairwise classification of location and crop year specific wheat classes at three selected moisture contents using nir hyperspectral images.
Mahesh, S., Jayas, D.S., Paliwal, J., and White, N.D.G. (2014). "Comparing two statistical discriminant models with a back-propagation neural network model for pairwise classification of location and crop year specific wheat classes at three selected moisture contents using nir hyperspectral images.", Transactions of the ASABE, 57(1), pp. 63-74. doi : 10.13031/trans.57.10125 Access to full text
Knowledge of wheat classes and seed moisture contents not only determines the end use of wheat flour but also helps in developing effective storage systems for wheat. Samples of four classes of wheat, including Canada Western Red Spring (CWRS), Canada Western Hard White Spring (CWHWS), Canada Western Soft White Spring (CWSWS), and Canada Prairie Spring Red (CPSR), were obtained from at least five different locations for each class in Manitoba, Saskatchewan, and Alberta for the 2007, 2008, and 2009 crop years and conditioned to moisture contents of 13%, 16%, and 19%. Near-infrared (NIR) hyperspectral images were acquired from bulk samples in the 960-1700 nm wavelength region at 10 nm intervals. The first and second principal component score images were compared for the segmented images of all wheat classes. Pairwise wheat class identification was done using a non-parametric statistical model and a four-layer back-propagation neural network (BPNN) model. The NIR wavelengths of 1260 to 1380 nm had the highest factor loadings for the first principal component using principal component analysis (PCA). The four-layer BPNN model was used for two-class identification of wheat classes. Overall average pairwise classification accuracies of 83.7% were obtained for discriminating wheat samples based on their moisture contents. Average classification accuracies of 83.2%, 75.4%, and 73.1%, were obtained for identifying wheat classes for samples with 13%, 16%, and 19% moisture content (m.c.), respectively. In this study, discriminant models yielded better classification accuracies than BPNN models. Overall average classification accuracies of wheat classes using statistical models were 80.6% for the linear discriminant analysis (LDA) and 76.3% for the quadratic discriminant analysis (QDA). This work showed that NIR hyperspectral imaging can be used as a potential nondestructive tool for classifying moisture-specific wheat classes.
- Date modified: