Husk Species Classification Using a ViT–DenseNet Hybrid Model with Explainable AI

Agricultural husks are valuable agro-industrial byproducts used in energy production, biodegradable packaging, livestock feed, and circular-economy applications. Accurate identification of husk species is essential for improving processing efficiency and supporting sustainable resource utilization, yet manual classification is often unreliable due to high similarity in texture, color, and structure across species. This study presents a hybrid deep learning model that integrates Vision Transformer and DenseNet121 architectures for automated classification of eight major husk types using the BDHusk dataset. A comprehensive preprocessing pipeline that includes color inversion, outlier handling with Modified Z-score, and extensive data augmentation enhances model generalization. The proposed framework achieves a test accuracy of 97.23% with strong reliability scores including a Matthews Correlation Coefficient of 0.9864 and near-perfect AUC values for all classes. Explainable AI methods such as LIME, Grad-CAM, and Grad-CAM++ are employed to ensure that predictions are transparent and grounded in biologically meaningful visual patterns. An ablation study confirms the importance of feature fusion, mixed normalization, GELU activation, and transfer learning in achieving optimal performance. The results indicate that the proposed model is highly effective for fine-grained agricultural image classification providing a strong algorithmic foundation for future hardware-in-the-loop testing in automated sorting feed preparation, and sustainable agro-industrial practices and can support automated sorting, feed preparation, and sustainable agro-industrial practices. This research contributes a scalable and interpretable framework that advances digital agriculture and promotes improved utilization of agricultural byproducts.