A robust vision transformer-based approach for classification of labeled rices in the wild

Since rice is a widely consumed food and can be exposed to different diseases depending on the climatic conditions of the region where it grows, it is very important to monitor it before harvest. Unlike the literature, instead of using an image that contains a single leaf taken with a controlled background, anew database consisting of images taken in uncontrolled field conditions close to its natural state was studied. Another challenge about the images taken from the field is that the five classes in the database are not evenly distributed. In this study, a method based on vision transformers which is robust to the class imbalance problem is proposed for rice leaf images taken in the wild with complex backgrounds. Additionally, the vision transformer-based model was compared with several transfer learning methods with and without finetuning. The results obtained revealed that the proposed approach significantly increased the performance of the sensitivity metric while reaching the highest accuracy rate. Since the images studied are close to images taken with a drone and no preprocessing is applied against the field conditions, the proposed method can be used in the development of a drone-based plant disease early warning system in the future.