Hybrid convolutional network based on hyperspectral imaging for wheat seed varieties classification

Abstract

Seed quality and safety are related to national food security, and seed variety purity is an important indicator in seed quality detection. In agricultural production, seeds of higher purity are more accessible with stable genetic traits. Thus, it is necessary to rapidly and non-destructively detect the seed purity. In this study, a novel hybrid convolutional network coupled with hyperspectral imaging was used for the sorting of wheat seeds. Hyperspectral imaging acquires spectral and spatial information simultaneously, which provides a collection of internal and external features of samples. To take full advantage of the spatial and spectral information, the hyperspectral images of wheat seeds were analyzed by hybrid convolution. Firstly, the average spectra were extracted from the region of interests (ROIs) and processed with one-dimensional (1D) convolution to extract spectral features. Meanwhile, the two-dimensional (2D) convolution was used to directly analyze hyperspectral images focusing more on extracting spatial features. Then, the features extracted by hybrid convolutional were integrated into the full connection layer of the network structure. The performance of the proposed models was validated with 8000 samples from the eight varieties of wheat seed. Experiments demonstrate that the hybrid convolutional network outperforms the reference model and gets excellent classification accuracy of 95.65% for wheat seeds.