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

Researchers propose an adaptive Mixture-of-Experts framework combining EfficientNet-B0, DenseNet-121, and Swin-Tiny for plant leaf disease classification, achieving 91.68% recall on imbalanced potato leaf datasets. The soft routing mechanism dynamically assigns expert weights to capture multi-scale features, demonstrating superior performance over single-architecture models and strong cross-dataset generalization on durian and sesame leaf diseases.

This research addresses a practical agricultural problem through sophisticated machine learning architecture design rather than incremental model improvements. Plant disease classification directly impacts food security and farmer profitability, yet real-world deployment faces significant obstacles: severe class imbalance (healthy vs. diseased leaves), variable lighting conditions, and complex natural backgrounds that confuse traditional deep learning approaches.

The proposed framework tackles these challenges by combining three complementary architectures—each excelling at different representation levels. EfficientNet-B0 provides efficient multi-scale feature extraction, DenseNet-121 captures dense local patterns optimal for texture-based disease indicators, and Swin-Tiny leverages transformer-based global context modeling. Rather than simple ensemble voting, the soft gating mechanism learns input-dependent routing that dynamically weights expert contributions, allowing the system to select optimal feature pathways for each leaf image.

The empirical results demonstrate substantial practical value. Achieving 92.62% F1-score on potato leaves represents approximately 5% absolute improvement over the strongest individual expert, translating to meaningful reduction in misclassified samples during field deployment. Cross-dataset validation on durian (94.03%) and sesame (97.04%) shows the approach generalizes beyond training distribution, critical for real-world agricultural applications where new crop varieties and regional disease variants constantly emerge.

This work has direct implications for agricultural technology adoption, particularly in developing regions where crop diseases cause significant yield losses. Implementation as a mobile or edge-deployed system could enable farmers to perform rapid disease screening without laboratory infrastructure, though the paper doesn't address computational requirements or deployment considerations necessary for practical field use.

• →Soft Mixture-of-Experts with dynamic routing outperforms individual architectures by 5.91% recall and 5.03% F1-score on imbalanced plant disease datasets.
• →Cross-architectural approach combines EfficientNet, DenseNet, and Swin-Tiny to capture complementary multi-scale, local, and global features.
• →Framework demonstrates strong generalization across potato, durian, and sesame leaf disease datasets with F1-scores exceeding 91%.
• →Two-stage refinement training strategy improves optimization stability when handling severe class imbalance in agricultural datasets.
• →Practical application enables rapid disease screening for precision agriculture and crop protection without specialized laboratory infrastructure.