This study proposes a comprehensive framework that integrates Synthetic Minority Oversampling Technique based transfer learning with Explainable Artificial Intelligence for multiclass eye disease classification from external images. The framework addresses some of the significant challenges in medical image classification, including class imbalance, small dataset problems, and model explainability. A novel, annotated dataset of external eye images was developed first, consisting of five classes: Cataract, Conjunctivitis, Eyelid diseases, Uveitis, and Normal eyes. The dataset was preprocessed through duplicate removal and then validated by a clinical practitioner, followed by SMOTE for class distribution balancing. Secondly, transfer learning, combined with data augmentation, was employed to thwart the constraints of having limited available data, which resulted in significant model performance improvement. Third, XAI techniques, such as Gradientweighted Class Activation Mapping (Grad-CAM++), are employed to explain the predictions of the models. These techniques provide valuable insights into the decision-making process and enhance clinical interpretability by highlighting the most influential regions that distinguish between diseased and healthy cases. Experimental experiments demonstrate that the proposed approach achieves 97.24 % accuracy and superior performance metrics compared to baseline models.