<p>This study proposes a practical and efficient framework for lemon leaf disease classification by combining deep feature extraction with optimization-based feature selection and machine learning classifiers. Instead of relying solely on end-to-end deep learning, the work focuses on improving feature quality and reducing redundancy. DenseNet121 is used to extract 1,024 deep features from a nine-class lemon leaf disease dataset, capturing relevant visual patterns from the images. These features are then refined using Gray Wolf Optimization (GWO), a wrapper-based feature selection method that reduces the feature space to 405 features, achieving a dimensionality reduction of 60.45%. To evaluate the effectiveness of the selected features, both Random Forest and Support Vector Machine (SVM) classifiers are applied to the original and optimized feature sets. The results show that Random Forest achieves 96.97% accuracy with original features and improves to 97.64% after feature selection. Similarly, SVM achieves the highest accuracy of 98.30% with original features and maintains competitive performance after optimization. The findings demonstrate that GWO effectively removes redundant information while preserving discriminative features, leading to improved efficiency without compromising performance. Compared to baseline approaches on the same dataset, the proposed hybrid framework achieves superior classification accuracy and reduced computational cost. This makes the method suitable for real-world agricultural applications where both accuracy and efficiency are important.</p>