Automated Eggplant Disease Detection Using Feature Fusion Based CNN with Explainable AI for Precision Agriculture

Automated and accurate detection of eggplant leaf diseases is crucial for improving yield and reducing economic loss, especially for smallholder farmers who lack timely access to expert diagnosis. However, existing vision-based systems often rely on a single deep model and operate as black boxes, which may struggle with complex visual symptoms, class imbalance, subtle inter-class variability, and a lack of interpretability for agronomists and farmers. To address these limitations, this study first conducts a comprehensive benchmark of 12 popular transfer learning architectures (VGG16, VGG19, ResNet50, ResNet152, DenseNet121, DenseNet201, Xception, InceptionV3, NasNetMobile, MobileNetV2, EfficientNetB7, ConvNeXtBase) on a seven-class eggplant leaf disease dataset using standardized preprocessing, data augmentation, and identical training protocols. After identifying ResNet152 and ConvNeXtBase as the top-performing models, their deep feature representations are fused at the feature level to construct a hybrid CNN-based feature fusion model, followed by a classifier head evaluated using accuracy, precision, recall, and F1-score, yielding more than a 7% improvement in overall accuracy compared with the best individual transfer learning backbone while also providing consistent gains across minority disease classes. To make the model transparent and trustworthy, explainable AI techniques such as Grad-CAM-based saliency maps are integrated to highlight the most influential leaf regions for each prediction, enabling visual validation of disease-specific patterns and reducing reliance on opaque decisions. The proposed XAI-enabled feature fusion model is suitable for deployment in real-world mobile or web-based decision-support tools, allowing farmers and agronomists not only to obtain reliable disease predictions from leaf images but also to understand why a particular diagnosis is made, thereby supporting timely, data-driven, and interpretable management actions in precision agriculture scenarios.