Ensemble deep learning with Grad-CAM visualization for olive leaf disease classification

Early and accurate detection of olive leaf diseases is critical for understanding plant health dynamics and improving sustainable crop production. This study presents a comparative evaluation of multiple ensemble deep learning strategies for the classification of olive leaf diseases from leaf images. The proposed framework combines EfficientNet, ResNet, and MobileNet architectures using transfer learning and compares several ensemble strategies, including weighted soft voting, unweighted soft voting, hard voting, and stacking. A publicly available dataset containing 3,400 olive leaf images, consisting of Healthy, Aculus olearius, and Olive Peacock Spot classes, was used. To reduce environmental noise and improve disease-related feature learning, preprocessing steps including leaf–background segmentation and rotation-based data augmentation were applied in the training set. During cross-validation, controlled data augmentation was applied exclusively to the training folds to increase training diversity and reduce the effect of class imbalance. The proposed framework was rigorously evaluated using stratified five-fold cross-validation and a hold-out test set. Each base model was trained and fine-tuned using the training data, after which ensemble predictions were generated and evaluated on the hold-out test set. Among the evaluated ensemble strategies, the stacking ensemble achieved the best overall performance, with an accuracy of 0.9721, a macro-averaged F1-score of 0.9720, a macro-averaged ROC-AUC of 0.9946, and a Cohen’s Kappa score of 0.9578. The results suggest that disease-related visual characteristics can be effectively captured through image-based analysis, while independent validation under diverse field conditions remains necessary before practical application can be considered.