Adaptive Confidence and Agreement Guided Hybrid CNN Framework for Plant Disease Classification

Abstract

Automated plant disease identification is central to precision agriculture, given its role in timely intervention and stable crop yields. This paper introduces an Adaptive Confidence and Agreement-Guided Hybrid CNN Framework, coordinating three CNNs of varying complexity ShuffleNetV2, MobileNetV2, and ResNet18 within a difficulty-adaptive inference pipeline. Rather than applying all models uniformly, a rout-ing mechanism assesses prediction confidence and inter-model agree-ment to determine processing depth: high-confidence agreements are classified directly, disagreements are resolved via confidence-weighted fu-sion, and only the most ambiguous samples are escalated to ResNet18. On PlantVillage, the framework achieves 99.70% accuracy, 99.45% F1-score, and 99% AUC; on the more challenging PlantDoc benchmark, it achieves 65.68% accuracy, 65.88% F1-score, and 96.31% AUC. These re-sults demonstrate that adaptive, sample-driven model orchestration can balance accuracy and computational efficiency across both controlled and real-world conditions.