A deep ensemble learning and explainable AI framework for accurate bottle gourd disease diagnosis and deployment

Accurate and timely detection of plant diseases is essential for sustainable agriculture and effective crop management. This study presents a comprehensive deep learning and explainable AI (XAI) framework for bottle gourd ( Lagenaria siceraria ) disease classification based on original image samples. The dataset consists of 7,000 images spanning seven classes: five disease categories and two healthy conditions (fresh fruit and fresh leaf). Ten state-of-the-art architectures—ConvNeXt-Tiny, DenseNet121, EfficientNetB3, InceptionV3, MobileNetV3, ResNet50, Swin-Tiny, VGG19, Vision Transformer (ViT), and YOLO11n—were individually trained and evaluated. To improve overall classification performance, three ensemble strategies—soft voting, weighted soft voting, and stacking—were applied. The stacking ensemble achieved the highest accuracy of 99.52%. To enhance model interpretability, four explainability methods—Grad-CAM, Grad-CAM++, Score-CAM, and Eigen-CAM—were utilized to highlight discriminative image regions contributing to model predictions. The final stacking-based ensemble was deployed via a web application that provides real-time image-based diagnosis along with targeted post-infection treatment recommendations for the five disease classes. This integrated framework exhibits strong potential for practical implementation in precision agriculture, especially in resource-limited settings.