Early detection of tomato leaf diseases using transformers and transfer learning

Tomato, one of the world’s most valuable cash crops and a staple in global cuisine, is susceptible to various diseases that, if not detected early, can lead to significant yield declines and the potential loss of entire hectares. Transformer models have shown substantial performance improvements in image recognition, including early plant leaf disease detection. However, their ability to generalise across different datasets and real-world settings remains uncertain, as they are often trained within similar distributions. Transfer learning, however, enables a model to learn features from a different distribution, enhancing its ability to generalise to new, real-world data. This study proposed three transfer learning approaches (ViT-ImageNet, ViT-Base, and ViT-Small) to predict tomato leaf diseases from images. The proposed methods were evaluated on the widely used PlantVillage dataset and a newly collected dataset, TomatoEbola, which includes subsets from Dikumari, Kukareta, and Kasaisa farms to reflect various environmental conditions. Experimental results demonstrated that the ViT-Base model achieved the highest accuracy of 99.17 % on the PlantVillage dataset and 77.27 % on the Dikumari subset, whereas the ViT-Small model achieved the highest accuracy of 92.73 % on the Kukareta subset and 91.54 % on the Kasaisa subset of the TomatoEbola dataset. These results outperform state-of-the-art methods such as VGG19, EfficientNetB2, InceptionV3, and DMCNN, which typically achieved accuracies below 90 %. Furthermore, the proposed approaches significantly enhanced robustness to environmental variability, reducing error rates by up to 47 % compared to state-of-the-art deep learning methods. These findings highlight the effectiveness and generalizability of the proposed approach, making it a valuable and sustainable tool for early diagnosis and management of tomato leaf diseases.