A hybrid deep feature-ensemble learning framework for robust tomato leaf disease classification using DenseNet121-EfficientNetB0 integration and SMOTE-optimize multi-classifier

Early and accurate detection of tomato leaf diseases is essential for reducing crop losses and improving agricultural productivity. However, traditional classification approaches often struggle with subtle disease patterns, intra-class variability, and imbalanced datasets, leading to reduced diagnostic performance. This study proposes a hybrid deep feature–ensemble learning framework to improve the reliability and accuracy of automated detection of tomato leaf diseases. The proposed framework integrates two convolutional neural networks, DenseNet121 and EfficientNetB0, to extract complementary hierarchical features from tomato leaf images. To mitigate class imbalance and enhance generalisation, the Synthetic Minority Oversampling Technique (SMOTE) was applied during training. The extracted features were subsequently used to train three machine learning classifiers: Support Vector Machine (SVM), Random Forest (RF), and Extreme Gradient Boosting (XGBoost). The predictions from these classifiers were aggregated using a soft-voting ensemble strategy to improve classification robustness and reduce model bias. Across ten independent experimental runs, the proposed framework achieved a mean accuracy of 98.72% ± 0.18, precision of 98.55% ± 0.21, recall of 98.63% ± 0.19, F1-score of 98.58% ± 0.17, and AUC of 98.21% ± 0.11. The results demonstrate that combining deep feature extraction with ensemble machine-learning classifiers significantly improves performance in tomato leaf disease classification. The framework exhibits strong generalisation across benchmark datasets, indicating its potential for deployment in intelligent crop monitoring systems and computer-aided agricultural diagnostics.