Maize seed classification using hyperspectral image coupled with multi-linear discriminant analysis

Abstract

Seed purity is an important parameter for evaluating seed quality and can be effectively studied by seed classification. Hyperspectral images between 400 and 1000 nm were acquired for 1632 maize seeds (17 varieties) for classifying seed varieties. Fourteen features including a spectral feature and 13 imaging features (i.e., 5 first-order and 8 s-order textural features) were extracted from the hyperspectral image data. A multi-linear discriminant analysis (MLDA) algorithm was developed to select the optimal wavelength and transform/reduce the classification features to improve the acquisition and processing speed of the hyperspectral images. Least square support vector machine was used to develop classification models based on MLDA with spectral features, imaging features, and combination of spectral and imaging features. The effects of MLDA, uninformative variable elimination (UVE) coupled with linear discriminant analysis (LDA), and successive projection algorithm (SPA) coupled with LDA were adopted. Experimental results indicate that the combination feature based on the wavelength selection algorithm of MLDA yielded high classification accuracy under the same number of wavelengths (varying between 5 and 15). Meanwhile, the classification model based on MLDA feature transformation/reduction method achieved superior classification accuracy of 99.13% over SPA coupled with LDA (90.31%) and UVE coupled with LDA (94.17%) and improved by 2.74% relative to that of the mean spectrum of the full wavelength model. The proposed method can be used effectively for seed identification and classification.