Innovative Hybrid Approach for Efficient Lightweight Vision Transformer Networks in Plant Disease Detection

The world is currently facing numerous food security challenges.Plant diseases are among the primary issues, recognized as major contributors to crop losses.These losses significantly undermine food production and the global economy.Traditional plant disease detection methods remain reliant on manual assessments by agricultural experts, which can result in prolonged delays and high monitoring costs.Recently, novel solutions have emerged, leveraging the synergy between computer vision and deep learning techniques for early plant disease detection.Recent studies have explored Vision Transformer architecture for disease diagnosis.While these architectures demonstrate significant potential, they also present limitations, such as high computational complexity and a low inductive bias for locality.To address these challenges, we suggest a hybrid model named EcoGreenViT, which combines the power of Vision Transformers with the performance of convolutional neural networks for plant disease detection.Our proposed approach was evaluated on a large public crop disease dataset containing 54305 images of different diseases to assess its performance.Our experimental results indicate that EcoGreenViT achieved the best combination of accuracy and computational cost on the PlantVillage dataset.Our approach clearly outperforms five well-known CNN architectures in terms of accuracy.Additionally, EcoGreenViT surpasses the original vision transformers by achieving the same accuracy with just half the training time by reducing the number of parameters up to 41.86%.These advantages of a reduced number of parameters render our EcoGreenViT model especially ideal for applications in smart agriculture, by facilitating deployment in resource limited environment.