Confidence-Adaptive Fuzzy Ensemble Deep Learning for Imbalanced Potato Plant Disease Classification

Abstract

The accurate detection of diseases of potato plants is very important for reducing their losses and helping achieve precision agriculture. Even though there has been a great improvement in detecting the diseases of plants based on ensembled learning through images, efficiency of ensembled with deep learning in dealing with imbalanced training data is adversely affected, resulting in low sensitivity to minority disease categories.This paper proposes a Confidence-Adaptive Fuzzy Ensemble Deep Learning (FEDL) model that employs heterogeneous deep learning algorithms along with uncertainty-aware decision fusion in imbalanced potato disease classification problems. The model employs EfficientNet-B3, ConvNeXt-Tiny, and Vision Transformer (ViT) to capture complementary feature representations of local and global nature. The order to address with issue of minority-class inequality, a hybrid approach including SMOTE, data augmentation, balanced sampling, and Focal loss is used in the model training phase. While conventional ensemble methods use fixed model weights for each classifier, the proposed system uses a Mamdani Fuzzy Inference System (FIS) to determine the weight of the model based on the confidence, entropy, probability margin, and reliability of the backbones.The performance evaluation of the proposed framework was conducted by applying it to the class-imbalanced PlantVillage potato disease dataset with MCC, AUC, F1 score, Accuracy, Precision, Recall, and Balanced Accuracy. Results from comparative experiments, ablation studies, and Grad-CAM visualization show that the confidence adaptive fuzzy fusion leads to improved classification results, better minority class detection, and interpretability of predictions compared to deep learning classifiers and traditional ensembles. It was shown that the proposed framework can be considered an effective and generalized  problem solution for reliable plant disease classification in class-imbalanced scenarios.