Real-Time Crop Disease Detection using Lightweight Vision Transformers (ViTs) on Drones
Accurate and efficient plant disease detection is crucial for precision agriculture to help minimize yield losses. Nevertheless, existing deep learning models are computationally resource-intensive, making them unsuitable for real-time monitoring with drones. In this work, we propose a lightweight vision transformer with a token pruning and knowledge distillation approach to detect plant diseases. We follow a teacher-student architecture where a vision transformer with complete feature learning ability serves as a teacher model and guides a student vision transformer. Attention-based token pruning helps prune out unimportant patch tokens from the student network during transformer layers and decreases computational redundancy without removing disease-relevant tokens. The proposed approach is evaluated on the PlantVillage color dataset with 54,305 images belonging to 38 classes (disease and healthy). Our experimental study shows that the proposed Student Vision Transformer network achieves a remarkable accuracy score of 96.47% and a macro F1-score of 95.30%, using only 1.28 million parameters. Moreover, our approach decreases the number of parameters by 61%. Our robustness tests under various transformations demonstrate the applicability of our vision transformer with pruning and knowledge distillation for real-time drone-assisted crop disease monitoring.