Cross-Architectural Mixture-of-Experts with Adaptive Soft Routing for Plant Leaf Disease Classification

This study proposes an adaptive soft Mixture-of-Experts (MoE) framework that integrates EfficientNet-B0, DenseNet-121, and Swin-Tiny to address challenges in plant leaf disease classification under complex backgrounds and class imbalance.

• The framework uses cross-architectural routing to exploit complementary multi-scale, local, and global features from the three models.
• A soft gating mechanism dynamically assigns input-dependent expert weights.
• A two-stage refinement training strategy is employed to improve optimization stability and generalization.
• Experiments on a highly imbalanced potato leaf disease dataset achieved 91.68% recall and 92.62% F1-score, surpassing the strongest individual expert by 5.91% and 5.03% respectively.
• Additional evaluations on durian and sesame leaf datasets yielded F1-scores of 94.03% and 97.04%, demonstrating robust cross-dataset generalization.

The proposed framework demonstrates potential for reliable real-world crop health monitoring by effectively capturing diverse features and maintaining performance across different datasets.