Explainable Hybrid CNN-Vision Transformer with Morphometric and Cross-Modal Attention for Simultaneous Classification of Rice Seed Ecology and Variety

Accurate classification of rice seed varieties and ecological classes is essential for seed quality assessment and precision agriculture. This paper proposes an explainable hybrid CNN–Vision Transformer (ViT) framework that combines data augmentation, handcrafted morphometric features, and cross-modal attention for the simultaneous classification of rice seed variety and rice ecology. A dataset comprising 16,000 high-resolution rice seed images collected from multiple agro-ecological zones and representing five rice varieties and three ecological classes, was used alongside biologically meaningful morphometric descriptors capturing seed size, shape, and structural characteristics. Experimental results demonstrate that data augmentation significantly enhances model generalization (2–4% accuracy improvement), while morphometric features provide consistent additional gains (1–3%). The inclusion of attention mechanisms further refines performance by focusing on discriminative seed regions. Among the evaluated models, the proposed smart attention-based hybrid CNN–ViT architecture, integrated with morphometric features and data augmentation, achieved superior performance for rice seed classification. The model obtained 96 ± 0.02% accuracy, 96% precision, 95% recall, and 95% F1-score for single-task variety classification, and 93 ± 0.02% accuracy with 93% precision, recall, and F1-score for single-task ecology classification. In the joint dual-task setting, the model achieved 90 ± 0.03% accuracy, with precision, recall, and F1-score all at 90%, indicating balanced performance across metrics. Furthermore, t-statistic and p-value significance tests were performed to confirm the statistical reliability of the performance improvements. To improve model transparency, Grad-CAM and attention heatmaps are employed to visualize the contribution of both visual and morphometric features, highlighting biologically relevant seed regions that influence predictions. The Explainability analysis confirms the interpretability and reliability of the proposed framework, supporting its applicability in real-world agricultural decision-making.