Explainable Hybrid Deep Learning with EfficientNetB3–DenseNet121 Feature Fusion for Banana Leaf Spot Disease Classification

Banana leaf spot diseases reduce crop quality and yield, making early and accurate diagnosis essential for precision agriculture. This study proposes an accurate and explainable hybrid deep learning framework for multiclass banana leaf spot disease classification. The model combines EfficientNetB3 and DenseNet121 in dual-branch feature-fusion architecture to capture complementary visual features. A stratified 70/15/15 split was used on the BananaLSD dataset (Cordana, Healthy, Pestalotiopsis, and Sigatoka), with class weighting, data augmentation, random erasing, label smoothing, and staged transfer learning to improve generalization. Performance was evaluated using accuracy, precision, recall, F1-score, confusion matrix, and multiclass ROC analysis, while Grad-CAM was used for explainability. The proposed hybrid model outperformed single-backbone baselines (EfficientNetB3 and DenseNet121), achieving 98.58% accuracy with the lowest loss (0.2578) and strong class-wise performance. Grad-CAM visualizations further confirmed that the model focused on disease-relevant lesion regions. These results support the proposed framework as a reliable tool for automated banana disease diagnosis.