A lightweight dual-channel feature fusion model for wheat variety identification in hyperspectral images

High-purity wheat varieties can ensure the quality of subsequent production. Hyperspectral images finely characterize the spectral properties of different substances through multi-band reflectance data and are widely used for non-destructive detection of wheat varieties. However, its high-dimensional nature results in substantial processing and computational resource consumption. Pixel-by-pixel and band-by-band calculations are needed for classification with deep learning models, which significantly increase computational complexity. This study proposes a lightweight dual-channel feature fusion model. By designing lightweight feature extraction modules, features are extracted separately from spectral and image information. Subsequently, the features extracted by the two modules are weighted and fused to achieve rapid and non-destructive classification of wheat varieties. The model achieved a classification accuracy of 97.50% for four wheat varieties. Compared with other deep learning models, this model significantly reduces the number of parameters while maintaining performance and improving detection speed. It provides a novel, lightweight architecture for hyperspectral images that enables the detection of different wheat varieties. The lightweight model can be deployed for on-site detection on mobile and embedded devices.