Automatic Diagnosis of Plant Diseases via Triple Attention Embedded Vision Transformer Model

Plant disease infestation causes severe crop damage and adversely affects crop yield. This damage can be reduced if diseases are identified in the early stages. Initially, farmers and agricultural scientists used to diagnose plant diseases with their naked eyes. With the dawn of different advanced computer vision techniques, various researchers have utilized these techniques for automatic disease detection in plants using their leaf images. In this research work, a novel triple attention embedded vision transformer is proposed for automatically diagnosing diseases in plants. In the proposed model, channel and spatial attention are embedded in addition to the multi-headed attention of the original vision transformer model. The reason for embedding the channel attention and spatial attention in vision transformer is that the existing multi-headed attention of vision transformer only considers the global relationship between the features and ignores the spatial and channel relationship. Moreover, in order to increase the confidence of farmers and agricultural scientists in predictions of the proposed model, human-understandable visual explanations are also provided with the predictions. These explanations are generated using the local interpretable model-agnostic explanations (LIME) framework. The experimentation of this research work is carried out on one publicly available dataset (PlantVillage dataset) and one real-in-field dataset (Maize dataset) having complex background images. For each dataset, it is experimentally found that the proposed model outperformed other research works found in the literature. Moreover, the visual explanations for the predictions of the proposed model highlight the infected area of leaves for diseased class predictions.