EFFECTS OF SPECTRAL PREPROCESSING AND WAVELENGTH SELECTION ON DISCRIMINATION OF MAIZE VARIETIES BY NIR SPECTROSCOPY
The discrimination of crop seed varieties is now one of the important subjects in the agricultural product testing. The near-infrared (NIR) spectroscopy as a powerful tool to make the fast, non-destructive testing has recently been applied in the discrimination of seed varieties, e.g. maize, wheat and rice. However, the analysis of NIR spectral data has to depend on chemometrics, in which the data preprocessing and wavelength selection are the foundation of building a model. Therefore, it is needed to investigate the effects of pre-processing and wavelength selection on the qualitative discrimination of seed varieties.In this work we study the effects of data preprocessing (no preprocessing, first derivative and second derivative transformation, standard normal variate transformation (SNV), vector normalization, smoothing,) and wavelength selection on the discrimination of maize seed varieties. Totally 140 samples were obtained from seven varieties of maize seeds (20 samples per variety, 10 for training the model, another 10 for testing the model). NIR spectra were recorded from 4000 cm-1~12000cm-1 at 3.9 cm-1 interval using a FT-NIR spectrometer (resolution 8 cm-1). The raw spectral data are pretreated by the above-referenced methods. The performance of the six preprocessing methods is evaluated on basis of the two data sets: all of the spectral data and the data from the characteristic regions selected by a standard deviation-based feature selection method, respectively. Furthermore, based on principal component analysis (PCA) scores of the processed data, the discrimination models for every variety are built using Biomimetic Pattern Recognition (BPR) method, which has been successfully used to build the discrimination models for maize and wheat seeds. The correct acceptance rate (CAR) (to measure the ability of recognizing the samples of the same variety) and correct rejection rate (CRR) (to measure ability of distinguishing other varieties) for the testing samples were finally calculated.The average CAR and CRR for seven varieties of each kind of model using one combination (one preprocessing method with or without wavelength selection) attain above 80％. The best discrimination model uses first derivative and all spectral data, by which both CAR and CRR for five varieties reach 100％, and the average CAR and CRR attains 98.6％ and 98％, respectively. Except smoothing, other preprocessing methods can get the better CAR and CRR than non-preprocessing. The wavelength selection can only improve CAR of SNV and vector normalization models, which rise CAR to 100％. Because the seeds of different cereal crops share similarity in compositions and traits, our results may be helpful for building the qualitative discrimination models of seed varieties of other cereal crops by NIR spectroscopy.
Near-infrared spectroscopy maize variety discrimination spectral preprocessing wavelength selection
Tingting Guo Wenjin Wu Qian Su Shoujue Wang Dong An
Institute of Semiconductors,Chinese Academy of Sciences,Beijing 100083,China Graduate University of College of Information and Electrical Engineering,China Agricultural University,Beijing 100083,China Institute of Semiconductors,Chinese Academy of Sciences,Beijing 100083,China
国际会议
北京
英文
1-7
2009-10-14（万方平台首次上网日期，不代表论文的发表时间）