<![CDATA[Chili plants are a commodity susceptible to plant pest organism attacks that can significantly reduce productivity. Visual identification of chili diseases by farmers is often inaccurate due to symptom similarity across disease categories, necessitating a technology-based approach capable of performing classification automatically and accurately. This study proposes a hybrid model combining EfficientNetB4 and Vision Transformer for chili leaf disease classification into four categories healthy, yellowish, curl leaf, and spot leaf. EfficientNetB4 extracts local features through compound scaling and MBConv blocks, while ViT models global relationships among image regions through self-attention, enabling a semantically meaningful integration of local and global feature representations that addresses the individual limitations of CNN and transformer-based architectures. The dataset integrates 4,000 secondary images from GitHub and 800 primary images collected directly from chili cultivation fields in Central Java, with splitting performed separately per source to ensure proportional distribution across subsets. To evaluate generalization capability, the model was assessed across three scenarios: training and testing on secondary data only 98.25%, testing on primary field data without prior field exposure 87.50%, and training and testing on integrated data 99.17%, with a perfect accuracy of 100% on the primary-only test set. These results demonstrate that incorporating field-collected data into training directly bridges the generalization gap caused by domain shift between laboratory and real-world conditions, outperforming both single-architecture and previous hybrid approaches reported in prior studies. The findings provide a methodological foundation for developing robust automated disease detection systems applicable across diverse agricultural crops and real-world farming environments.]]>
