Rice leaf disease identification by residual-distilled transformer.

As the worldwide planting crop, rice feeds nearly half of the world's population. However, the continuous spread of diseases is threatening rice production. It is of great practical value to identify rice diseases precisely. Recent studies suggest that the computational approaches provide an opportunity for rice leaf disease prediction and achieve a series of achievements. However, the existing works for rice leaf disease identification are still unsatisfactory either in identification accuracy or model interpretability. To address these limitations, a residual-distilled transformer architecture is proposed in this study. Inspired by the early success of transformers in computer vision, the distillation strategy is introduced to distill weights and parameters from the pre-trained vision transformer models. The residual concatenation between vision transformer and the distilled transformer are as residual blocks for features extraction, and then fed them into multi-layer perceptron (MLP) for prediction. Experimental results demonstrate that the presented method achieves 0.89 F1-score and 0.92 top-1 accuracy, outperforms the existing state-of-the-art models on the rice leaf disease dataset which collected in paddy fields. In addition, the proposed architecture provides model interpretability to grasp the key features that are significant for positive prediction results.