Classification of the waxy condition of durum wheat by near infrared reflectance spectroscopy using wavelets and a genetic algorithm

Near infrared (NIR) reflectance spectroscopy has been applied to the problem of differentiating four genotypes of durum wheat: 'waxy', Wx A1 null null, wx-B1 null and wild type. The test data consisted of 95 NIR reflectance spectra of wheat samples obtained from a USDA-ARS wheat breeding program. A two-step procedure for pattern recognition analysis of NIR spectral data was employed. First, the wavelet packet transform [14], [15] was applied to the NIR reflectance data using wavelet filters at different scales to extract and separate low-frequency signal components from high frequency noise components. By applying these filters, each reflectance spectrum was decomposed into wavelet coefficients that represented the sample's constituent frequencies. Second, wavelet coefficients characteristic of the waxy condition of the wheat samples were identified using a genetic algorithm for pattern recognition. The pattern recognition GA employed both supervised and unsupervised learning to identify wavelet coefficients that optimized clustering of the spectra by genotype in a plot of the two largest principal components of the data. By sampling key feature subsets, scoring their PC plots, and tracking those genotypes and samples that were difficult to classify, the pattern recognition GA was able to identify a set of wavelet coefficients whose PC plot showed clustering of the wheat samples on the basis of their 'waxy' condition. Object validation was also performed to assess the predictive ability of the proposed NIR method to identify the 'waxy' condition of the wheat. An overall classification success rate of 78% was achieved for the spectral data.