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

Plant leaf disease classification is often hindered by complex backgrounds, illumination variations, and severe class imbalance. While individual deep learning architectures excel at specific tasks—such as CNNs for local texture modeling and Vision Transformers for global context—they often struggle to capture both simultaneously. This paper investigates whether a dynamic, input-dependent ensemble of heterogeneous architectures can outperform static fusion methods in agricultural disease diagnosis.

The authors introduce an Adaptive Soft Mixture-of-Experts (MoE) framework. Unlike traditional ensembles that assign fixed weights to models, this approach uses a routing network to dynamically assign weights to three experts: EfficientNet-B0, DenseNet-121, and Swin-Tiny. The process involves a Feature Projection Module (FPM) to align the heterogeneous feature maps into a unified space, followed by a gating mechanism that generates continuous weights for each expert based on the input image. The model is trained using a two-stage refinement strategy, where the entire framework is first fine-tuned and then further optimized with a reduced learning rate to improve stability.

The proposed framework was evaluated on a highly imbalanced potato leaf disease dataset, achieving a 91.68% recall and 92.62% F1-score. This performance surpassed the strongest individual expert model by 5.91% and 5.03%, respectively. Further testing on durian and sesame leaf disease datasets yielded F1-scores of 94.03% and 97.04%, suggesting that the model generalizes well across different crop types. Interpretability analysis using Grad-CAM confirmed that the adaptive routing mechanism effectively focuses on relevant lesion areas by combining the complementary strengths of the underlying architectures.