Automatic and rapid discrimination of cotton genotypes by near infrared spectroscopy and chemometrics

This paper reports the application of near infrared (NIR) spectroscopy and pattern recognition methods to rapid and automatic discrimination of the genotypes (parent, transgenic, and parent-transgenic hybrid) of cotton plants. Diffuse reflectance NIR spectra of representative cotton seeds (n = 120) and leaves (n = 123) were measured in the range of 4000-12000 cm(-1). A practical problem when developing classification models is the degradation and even breakdown of models caused by outliers. Considering the high-dimensional nature and uncertainty of potential spectral outliers, robust principal component analysis (rPCA) was applied to each separate sample group to detect and exclude outliers. The influence of different data preprocessing methods on model prediction performance was also investigated. The results demonstrate that rPCA can effectively detect outliers and maintain the efficiency of discriminant analysis. Moreover, the classification accuracy can be significantly improved by second-order derivative and standard normal variate (SNV). The best partial least squares discriminant analysis (PLSDA) models obtained total classification accuracy of 100% and 97.6% for seeds and leaves, respectively.