Abstract: Tea plants face from multiple disease threats during their growth cyele, such as red leaf spot,algae spot, and anthracnose. Manual inspection methods struggle to achieve accurate identification, lead-ing to the loss of optimal intervention timing. To address this challenge, this study proposes a Transformerneural network model integrating self-supervised learning and an information entropy weighting mecha-nism. First, each tea leaf image in the training dataset was subjected to 75% random masking. Using theself-supervised leaming paradigm, the model extracted features from unmasked regionsand accurately pre-dicts information in masked regions. Based on this pre-trained model, a tea disease detection model wasconstructed via transfer learning. By incorporating an information entropy weighting strategy, the detec-tion model effectively enhances its attention to the most informativeregions in images, thereby significantlyimproving the accuracy of tea disease recognition. Experimental results show that even when trained on alimited dataset, the proposed model achieves a detection accuracy of 93. 7% , significantly outperformingmainstream deep learning algorithms such as ResNet18, VGG16, and VGG19. Furthermore, the modeldemonstrates excellent adaptability and transferability, indicating its applicability to early waring systemsfor diseases in other crops.

Key words: tea diseases, self-supervised learning, information entropy, transformer, deep learning