Glaucoma is a leading cause of irreversible blindness, often worsened by delayed diagnosis. Traditional diagnostic methods have limitations in early detection and require significant expertise. Recent advancements in machine learning and deep learning have shown promise in enhancing the diagnostic accuracy and efficiency of glaucoma detection using fundus imagery. This study presents a four-stage method for glaucoma classification. Fundus images are processed using VGG16 and ViT for feature extraction, capturing both local and global features. These features are then fused and reduced in dimensionality using Principal Component Analysis. Finally, the reduced features are classified using a Random Forest classifier. The integrated feature fusion model demonstrates significant improvements in diagnostic performance, achieving higher accuracy, specificity, and sensitivity in distinguishing between non-referable and referable glaucoma classes compared to traditional methods. Specifically, our model achieved an accuracy of 94.2%, an F1 score of 94.2%, a sensitivity of 94.44%, and a specificity of 94.0%. The use of both CNNs and ViTs for feature extraction leverages their strengths, resulting in a more effective diagnostic tool. The combination of CNNs, ViTs, and Random Forest classifiers, along with advanced data augmentation techniques, shows substantial potential for early glaucoma detection and ongoing monitoring. This approach addresses the limitations of current machine learning models and enhances diagnostic accuracy and efficiency, making it a promising tool for clinical settings.