Fig 1.
Overview of the integrated methodology for tea leaf disease classification, from data acquisition and preprocessing to model training, evaluation, and final prediction.
Table 1.
Detailed summary of the curated tea leaf dataset attributes.
Fig 2.
Visual progression of the data pipeline.
Representative samples of Blight, Red Rust, Helopeltis, and Healthy classes showing raw field captures (left) alongside subsequent iterations of the dynamic augmentation pipeline (middle and right), which utilize stochastic rotations, horizontal flips, and brightness/contrast adjustments to enhance model generalization.
Fig 3.
Flowchart of the Proposed Hybrid Feature Fusion Architecture and Implementation.
The diagram illustrates the dual-branch feature extraction strategy running in parallel.
Table 2.
Stratified dataset split (80-10-10) preserving class equilibrium.
Fig 4.
Dataset split distribution showing the proportion of images used for training (80%), validation (10%), and testing (10%).
Fig 5.
Training and validation performance of the customized CNN.
The plots illustrate the loss history (left) and accuracy history (right) over 30 epochs, demonstrating a steady convergence in training while revealing characteristic fluctuations in validation metrics typical of learning on high-variance natural field imagery.
Fig 6.
Performance analysis of the Custom CNN baseline architecture.
a) Confusion matrix for the Custom CNN Model on the stratified test set. b) Bar Chart for the Custom CNN Model.
Fig 7.
Training and validation performance of the refined ResNet50.
The plots show the loss history (left) and accuracy history (right) for the ResNet50 model with its custom MLP head.
Fig 8.
Performance analysis of the ResNet50 baseline architecture.
a) Confusion matrix for the ResNet50 Model on the stratified test set. b) Bar Chart for the ResNet50 Model.
Fig 9.
Training and validation performance of the DenseNet121 model.
The plots show the loss history (left) and accuracy history (right) over 30 epochs, demonstrating high discriminative power with a final validation accuracy.
Fig 10.
Performance analysis of the DenseNet121 baseline architecture.
a) Confusion matrix for the DenseNet121 Model on the stratified test set. b) Bar Chart for the DenseNet121 Model.
Fig 11.
Training and validation performance of the MobileNetV3 model.
The plots present the loss history (left) and accuracy history (right) for the MobileNetV3-Small backbone, showcasing rapid convergence and stable performance through the integration of a custom classification head and AdamW optimization.
Fig 12.
Performance analysis of the MobileNetV3 baseline architecture.
a) Confusion matrix for the MobileNetV3 Model on the stratified test set. b) Bar Chart for the MobileNetV3 Model.
Fig 13.
Training and validation performance of the EfficientNetV2-B3 model.
The plots exhibit the loss history (left) and accuracy history (right) for the EfficientNetV2-B3 architecture, highlighting high initial accuracy and consistent training stability across the 16-epoch convergence period.
Fig 14.
Performance analysis of the EfficientNetV2-B3 baseline architecture.
a) Confusion matrix for the EfficientNetV2-B3 Model on the stratified test set. b) Bar Chart for the EfficientNetV2-B3 Model.
Fig 15.
Training and validation performance of the Vision Transformer (ViT-B16).
The plots illustrate the loss history (left) and accuracy history (right) for the ViT-B16 model.
Fig 16.
Performance analysis of the Vision Transformer (ViT-B16) baseline architecture.
a) Confusion matrix for the Vision Transformer (ViT-B16) Model on the stratified test set. b) Bar Chart for the Vision Transformer (ViT-B16) Model.
Fig 17.
Training and validation performance of the Hybrid Feature Fusion Model.
The plots present the loss history (left) and accuracy history (right) for the proposed dual-branch architecture, showcasing superior convergence stability and achieving the highest overall accuracy.
Fig 18.
Performance analysis of the Hybrid Feature Fusion Model baseline architecture.
a) Confusion matrix for the Hybrid Feature Fusion Model on the stratified test set. b) Bar Chart for the Hybrid Feature Fusion Model.
Table 3.
Performance comparison of different models on tea leaf disease classification.
Table 4.
Comparison of final training and validation loss across different models.
Table 5.
Comparison of Peak Validation AUC across evaluated models.
Table 6.
Computational Complexity of Evaluated Models.
Fig 19.
Performance curves for the Ablation Study baseline.
The plots illustrate the loss history (left) and accuracy history (right) for the unoptimized Hybrid Feature Fusion baseline, revealing significant performance volatility and extreme validation spikes when custom architectural optimizations and regularization are removed.
Fig 20.
Performance analysis of the Hybrid Feature Fusion Model during the Ablation Study (Unaugmented Baseline) baseline architecture.
a) Confusion matrix for the Hybrid Feature Fusion Model during the Ablation Study (Unaugmented Baseline) on the stratified test set. b) Bar Chart for the Hybrid Feature Fusion Model during the Ablation Study (Unaugmented Baseline).
Fig 21.
Cross-dataset validation on the teaLeafBD dataset.
The plots depict the loss history (left) and accuracy history (right) for the Hybrid Feature Fusion Model on the external teaLeafBD benchmark, demonstrating robust generalizability and stable cross-dataset performance.
Fig 22.
Performance analysis of the Hybrid Feature Fusion Model on the teaLeafBD dataset baseline architecture.
a) Confusion matrix for the Hybrid Feature Fusion Model on the teaLeafBD dataset on the stratified test set. b) Bar Chart for the Hybrid Feature Fusion Model on the teaLeafBD dataset.