Journal of Scientific Research and Reports

This study presents a hybrid image-classification framework for distinguishing healthy tea leaves from blister blight using complementary handcrafted and deep visual features. A balanced dataset of 600 RGB tea-leaf images, comprising 300 healthy and 300 blister-blight images, was evaluated using a stratified 70:15:15 train–validation–test split. Handcrafted features captured colour information from RGB, HSV and CIELAB spaces, together with texture patterns derived from the grey-level co-occurrence matrix (GLCM) and local binary patterns (LBP). In parallel, an ImageNet-pretrained EfficientNetV2-S was used as a fixed deep-feature extractor. The 161 handcrafted features and 1,280 deep features were fused to form a 1,441-dimensional representation. A training-only relevance–redundancy feature-selection strategy, combining mutual-information relevance with correlation-based redundancy reduction, selected the top 100 features. An RBF-SVM was then tuned on the validation set and evaluated on the independent test set. The final model showed strong discrimination on the held-out test partition, while a five-fold stability analysis on the training set yielded an accuracy of 98.81% ± 1.06% and ROC-AUC of 99.95% ± 0.07%. These findings show that combining handcrafted and deep visual features can provide a compact and discriminative representation for binary tea-leaf disease classification. The findings should, however, be interpreted within the scope of the curated dataset and binary task. Independent validation using field images from additional gardens, cultivars, seasons, imaging devices and disease severities is required to assess generalisability and practical deployment.