Paddy Disease Detection through Vision Transformers

Abstract

Listen

Translate article

Abstract Rice has an indispensable role as part of global food production. Despite being a staple food, Brunei Darussalam struggles with attaining self-sufficiency in rice, relying heavily on imports to meet consumption demands. With an inverse relationship between available resources and global population, there needs to be a focus on ‘doing more with less’ to increase crop output. Paddy diseases inflict volatility in both crop quality and yield. Timely diagnosis of paddy is imperative for reducing losses. Traditional diagnosis methods are both time-consuming and unreliable, urging the need for a faster, more efficient approach. The main objective of the study is to determine the viability and performance of transformer-based models trained on publicly available datasets in paddy disease classification. A comparative study is performed on applying fine-tuning to various Convolutional Neural Networks and Transformer models. The models are evaluated on a dataset of Bruneian diseased paddy. To meet the principles of ethical AI, attention maps are explored to improve the trustworthiness and robustness of the model. Results indicate that the transformer-based models outperformed the CNN models, indicating that transformers have potential in improving existing works in paddy disease detection. Preliminary feedback revealed attention maps increased overall trust in the system’s predictions.