An Interpretable Feature-Enhancement DeepEnsemble Classifier for Brinjal Disease Detection

Abstract

Artificial intelligence has become an effective tool for improving agricultural productivity through automated crop disease diagnosis. Brinjal (eggplant) cultivation suffers substantial yield losses from diseases such as Shoot and Fruit Borer, Wet Rot, Fruit Cracking, and Phomopsis Blight, yet reliable field-based diagnostic systems remain limited. To address this challenge, we introduce \textit{BrinjalFruitX}, a real-world dataset comprising 1,823 annotated images collected under natural farming conditions in Bangladesh across four disease classes and one healthy class. We propose an interpretable hybrid feature-enhancement deep ensemble framework that integrates image preprocessing, transfer learning, traditional machine learning, class imbalance mitigation, and explainable artificial intelligence. Three preprocessing techniques, Gaussian, Laplacian, and Unsharp Masking, are systematically evaluated, while deep features extracted using pre-trained VGG and ResNet models are classified by Random Forest, K-Nearest Neighbors, and a classifier-level ensemble. The optimal Unsharp--ResNet--Random Forest configuration achieved 80.0\% accuracy with an F1-score of 87.0\%. ADASYN applied in the deep feature space improved minority-class sensitivity, while Grad-CAM and Grad-CAM++ enhanced model interpretability. The proposed framework provides an effective and transparent solution for practical field-level brinjal disease detection and precision agriculture.