Identification of wheat seed endosperm texture using hyperspectral imaging combined with an ensemble learning model

Abstract

Differences in wheat endosperm structure contribute to differences in wheat flour texture and directly affect aspects such as flour quality, processing, and use. Therefore, the accurate classification of wheat based on endosperm texture is of immense practical interest. In this study, hyperspectral imaging technology (400–1000 nm) was combined with ensemble learning to classify wheat with different endosperm textures using spectral and shape features. Two feature extraction algorithms, competitive adaptive reweighted sampling and successive projection algorithm, were used to extract feature wavelengths. Furthermore, unknown characteristic data (new varieties of wheat) were fed into the model for classification. The results showed that feature fusion can markedly improve classification accuracy. The full-wavelength, subspace-based ensemble learning model based on the fusion of spectral and shape features had the best performance, and its classification accuracy reached 92.10%. In addition, the accuracy of all models for predicting new varieties decreased. However, the subspace-based ensemble learning model showed the best performance for identifying new wheat varieties with 88.03% accuracy. Thus, ensemble learning effectively classified both multiple known and new varieties of wheat with different endosperm textures. These results and this technology can help farmers and food manufacturers optimize their crop selection and processing strategies.