An enhanced vision transformer network for efficient and accurate crop disease detection

Crop diseases cause phenomenal damage to overall foodgrain production and affect the country’s food security chain. Prompt, precise and accurate detection of diseases in crops is very crucial for crop disease management. This study proposes an improved vision transformer-based (ViT) network for effective and precise detection of diseases occurring in agricultural crops. The proposed network introduces an improved transformer encoder with triplet multi-head attention (t-MHA) function, a cascading arrangement of attention units with residual connections, enabling the proposed network to progressively refine attention scores across multiple dimensions, learning more fine-grain feature representation of the images. Here, we have considered the two most important crops, viz ., Rice and Apple (aka RicApp dataset), which contribute a lot to the overall GDP of this country. Images of disease-infected rice and apple crops were collected from different agricultural farms under the supervision of agricultural experts. The proposed network obtained a classification accuracy of 97.99% on unseen test data of the RicApp dataset, outperforming the standard Vision Transformer (ViT) by 2.2%. To validate the effectiveness and robustness of the proposed network, thorough comparative analysis was conducted against the popular state-of-the-art CNN and ViT-based architectures. Additionally, feature visualization techniques were employed to demonstrate the explainability of the proposed network’s learning capability on the images. Therefore, the proposed network offers a precise and efficient solution for automated crop disease detection, with promising applications in real-time crop monitoring and precision agriculture.