Detection of Tea Leaf Diseases Using Deep Transfer Learning

Abstract

One of the most serious plant diseases to the production of tea is the tea leaf disease. In the early stages, it often shows no visible symptoms, making early detection difficult until the infection spreads across the leaf. As a result, tea leaf disease is considered a leading cause of crop loss in tea plantations. This paper suggests that deep learning models, specifically the transfer learning and the ensemble approach, could be deployed in the automation of tea leaf disease detection. First of all, the dataset was pre-treated, and several data augmentation methods were implemented, such as resizing, rescaling, flipping, rotation, zooming, and contrast adjustment.We then conducted Error Level Analysis (ELA) to identify any patterns that may have been overlooked in the images. The research explored deep learning models capable of accurately distinguishing between different classes of the disease. Pre-trained models such as EfficientNetB0 and EfficientNetV2B1 were employed for transfer learning. Furthermore, ensemble models combining CNN with EfficientNetB0 and CNN with EfficientNetV2B1 were evaluated. All models were tested using two approaches: ordinal classification and regular classification. Among the transfer learning models, EfficientNetB0 achieved the highest accuracy of 91.91% with ordinal classification. Within the ensemble models, CNN+ EfficientNetB0 reached a peak accuracy of 89.04% under ordinal classification. These models can assist agronomists and tea farmers, particularly those with limited resources, by enabling effective identification and classification of tea leaf disease