Distinguishing varieties of paddy seeds based on Vis/NIRS and chemometrics.

摘要

The potential of visible and near infrared reflectance spectroscopy (Vis/NIRS) was investigated for its ability to nondestructively distinguish the varieties of paddy seeds. A total of 150 samples were prepared for spectra collecting from spectroradiometer (325-1075 nm). Then principal component analysis (PCA) was performed on the spectra of all the samples. PCA compressed hundreds of spectral data into several new variables, which can explain the most variance of original spectra. The 2-dimension plot was drawn with the scores of the first 2 PCs, it provided the clustering information of the varieties of paddy seeds. Principal component analysis showed that the cumulative variance of first 4 principal components (PCs) were 99.6%. So, the first 4PCs were used to replace the original spectra. The first 4 PCs were used as inputs of a back propagation artificial neural network (ANN). One hundred and twenty five samples were selected from five varieties randomly (25 for each variety), then they were used as train samples to develop ANN model. The optimal topology structure was 4-7-1 for three-layer neural network. This model was used to predict the varieties of 25 unknown samples, and the recognition rate was 100%. This model was reliable and practicable, and Vis/NIRS has substantial potential for distinguishing varieties of paddy seeds.