Lemon leaf diseases threaten global citrus production, causing significant economic and agricultural losses. While deep learning offers solutions, existing models often lack robustness under real-world conditions like variable lighting and disease severity. This study introduces XIMR-Net, an ensemble deep learning framework that synergizes transfer learning and multi-model fusion to achieve unprecedented accuracy in lemon leaf disease classification. Our methodology begins with rigorous preprocessing: resizing images, normalization, and Contrast Limited Adaptive Histogram Equalization, which improves image quality, as validated by the Peak Signal-to-Noise Ratio. We evaluated ten state-of-the-art CNN architectures known as the transfer learning model. The four main models (InceptionV3, Xception, ResNet101V2, MobileNetV2), each exceeding the accuracy 94%, were integrated using weighted averages in XIMR-Net. Hyperparameter optimization, including focal loss and advanced data augmentation, enhanced precision and recall by over 75% for individual models. XIMR-Net achieved 99.28% accuracy, 98.99% precision, 99.07% recall, and 98.99% F1 score outperforming both standalone models and existing ensemble approaches. Five fold cross-validation confirmed robustness, while confusion matrices revealed near-perfect classification across nine disease categories. By addressing critical gaps in scalability and field adaptability, XIMR-Net provides a deployable tool for precision agriculture, compatible with mobile-based monitoring systems. This work advances AI-driven disease management, offering farmers a reliable, early detection solution to mitigate crop losses and promote sustainable practices.