ClusterRiceNet: A novel rice seed variety classification network based on hyperspectral imaging and spectral band clustering

Rice seed variety classification is crucial for agricultural production and seed quality control. Hyperspectral imaging technology enables the detection of subtle spectral variations between varieties but suffers from high dimensionality and information redundancy. Traditional dimensionality reduction methods, such as band selection and feature extraction, often result in the loss of critical details or compromise physical significance of spectral bands. To address these issues, we propose ClusterRiceNet, a novel rice seed classification network based on hyperspectral imaging and spectral band clustering. Within ClusterRiceNet, an Input Space Redefinition algorithm (ISRK) is designed to reduce spectral redundancy while preserving the physical significance of all bands and enabling flexible data augmentation. Furthermore, two feature extraction modules at different granularities, i.e. SFEIDT (Isometric Domain Transformation-based Spectral Feature Extraction) and GDFE (Swin-Transformer-based Global Dependency Feature Extraction), are introduced to capture discriminative spectral details and nonlocal dependencies, respectively. Finally, a fusion module integrates these features into a unified representation for efficient, non-destructive classification. Experiments on two constructed rice seed hyperspectral datasets show that ClusterRiceNet outperforms eight state-of-the-art image classification models and four advanced rice classification networks in terms of accuracy, F1-score, and κ coefficient. The proposed framework provides a new perspective for efficient and physically meaningful hyperspectral image analysis across diverse domains.