An Adaptive Confidence-Driven Hierarchical Ensemble Framework for Robust Multi-Crop Disease Classification

Abstract

Early and accurate detection of plant diseases is critical for improving agricultural productivity and reducing crop losses. Deep learning techniques have shown promising results in automated plant disease classification; however, performance often degrades when models are applied across multiple crops with visually similar disease symptoms. This paper presents a comprehensive study of three classification strategies: single-model learning, baseline ensemble learning, and an enhanced ensemble learning framework. The proposed enhanced ensemble integrates weighted soft probability fusion, confidence-based filtering, and crop-adaptive decision rules to improve robustness and generalization. Experiments conducted on multi-crop diseases datasets including paddy, potato and banana demonstrate that the enhanced ensemble outperforms individual models and baseline ensembles, particularly in complex and imbalanced disease classes. The crops considered in this study—potato, banana, and paddy—were selected due to their agricultural significance as major vegetable, fruit, and staple crops in the Anand district of Gujarat.