Identification of wheat kernel varieties based on hyperspectral imaging technology and grouped convolutional neural network with feature intervals

Abstract

Variety identification of seeds is essential for evaluating seed purity and ensuring crop yields. Hyperspectral imaging is a mainstream nondestructive testing method for seed purity. The hyperspectral data contains spectral and spatial information, which is large and redundant. According to the characteristics of hyperspectral data, this study used spectral feature intervals coupled with a grouped convolutional network to construct a seed variety identification model. Firstly, three spectral wavelength interval selection methods are used to constrain the spectrum, which are interval continuum removal (iCR), backward interval partial least squares (BiPLS), and interval random frog (iRF). Since spectral feature intervals in the hyperspectral image represent different physical and chemical characteristics, it is more effective to extract features of each interval independently by group convolution (GC). Then, the features are integrated into the network structure. The results showed that the combination of feature wavelength and grouping convolution enabled classification models effectively to obtain better classification accuracy and improve the inference speed of the model. The iCR-GC CNN model had the highest classification accuracy with a 4.4% increase and the fastest inference speed with a 44.21% reduction by comparing with the reference CNN classification model.