🤖 AI Summary
This work addresses key limitations of conventional soft-routing Mixture-of-Experts (MoE) approaches in multi-crop disease recognition—namely expert collapse, absence of crop-specific semantic alignment, high retraining costs, and unstable out-of-distribution rejection. To overcome these issues, the authors propose a hard-routing MoE architecture featuring a RouterHead that performs crop classification and rejection based on maximum Softmax probability, complemented by an Energy+KNN dual-gating mechanism to detect distribution shifts. Each crop is assigned a dedicated adapter module, enabling disease classification and temperature-scaled calibration while keeping the EfficientNet-B0 backbone frozen. This design eliminates expert collapse, supports crop-wise incremental updates, and reduces training costs to approximately 9% of full-network fine-tuning. On PlantVillage, the method achieves accuracy statistically on par with the best baseline (Macro-F1 gap ≤ 0.24%) while balancing rejection reliability and deployment efficiency.
📝 Abstract
Timely crop-disease identification is critical to food security. Multi-crop recognition suits Mixture-of-Experts (MoE), but conventional soft-routing MoE learns crop assignment freely end-to-end, letting a few experts dominate (expert collapse) with no semantic correspondence to crops, and facing high retraining costs, unstable rejection of non-target inputs, and a saturated accuracy ceiling. We shift the objective from accuracy toward a trade-off among deployment cost, scaling flexibility, and rejection stability, using deterministic hard routing. We propose AdapterMoE: a RouterHead classifies the crop and rejects non-target crops via a Maximum Softmax Probability threshold, with a dual-gate Energy+KNN out-of-distribution module catching distribution-shifted inputs; five per-crop Adapters atop a frozen EfficientNet-B0 backbone discriminate diseases, each calibrated via Temperature Scaling. Because experts are hard-isolated at the data level, the design avoids expert collapse and exposes an add_crop interface for local, per-crop updates instead of full retraining. On PlantVillage (5 crops, 26 classes), across a fair five-system comparison, AdapterMoE attains accuracy statistically indistinguishable from the best baselines (Macro-F1 within a 0.24-point band) while cutting training cost to about 9% of full-network baselines, expanding to a new crop in