Pomegranate diseases significantly threaten global fruit production, causing substantial yield losses and economic impacts. Traditional manual disease identification methods are time-consuming, subjective, and require specialized expertise. In this study, we present a novel GCLH (Grouped Convolutional Learning Hierarchy) ensemble model for automated pomegranate disease classification compared to individual baseline models, including VGG16 (92.0%), DenseNet (91.66%), and InceptionV3 (92.05%). The proposed GCLH-based ensemble framework achieved a classification accuracy of 99.31%, outperforming individual backbone models, such as VGG16, DenseNet121, and InceptionV3, which achieved approximately 92% accuracy. This represents an improvement of nearly 7% over standalone CNN models and demonstrates enhanced precision (0.9917), recall (0.9929), and F1-score (0.9922). The significant performance gain confirms the effectiveness of multi-backbone feature fusion and hierarchical attention mechanisms for robust pomegranate disease classification.