Deep-Learning-Based Diagnosis of Cassava Leaf Diseases Using Vision Transformer

Abstract

Viral diseases are major causes leading to the poor yields of cassava, which is the second-largest source of food carbohydrates in Africa. As symptoms of these diseases can usually be identified by inspecting cassava leafs, visual diagnosis of cassava leaf diseases is of significant importance in food security and agriculture development. Considering the shortage of qualified agricultural experts, automatic approaches for the image-based detection of cassava leaf diseases are in great demand. In this paper, on the basis of Vision Transformer, we propose a deep learning method to identify the type of viral disease in a cassava leaf image. The image dataset of cassava leaves is provided by the Makerere Artificial Intelligence Lab in a Kaggle competition, consisting of 4 subtypes of diseases and healthy cassava leaves. Our results show that Vision-Transformer-based model can effectively achieve an excellent performance regarding the classification of cassava leaf diseases. After applying the K-Fold cross validation technique, our model reaches a categorization accuracy 0.9002 on the private test set. This score ranks top 3% in the leaderboard, and can get a silver medal prize in the Kaggle competition. Our method can be applied for the identification of diseased plants, and potentially prevent the irreparable damage of crops.