Glaucoma is a leading cause of irreversible blindness worldwide, often progressing without noticeable symptoms until significant vision loss occurs. Early and accurate detection is essential to prevent long-term damage. The authors proposed an effective glaucoma classification framework using a Bidirectional Long Short-Term Memory (Bi-LSTM) hybrid deep learning model supported by advanced feature extraction techniques. Instead of using raw retinal fundus images directly, our method applies a Fibonacci spiral-based sub-block segmentation approach and annular ring analysis to extract 270 distinctive features per image across multiple colour channels. These features capture crucial structural and textural patterns, which are then forwarded to a hybrid architecture combining 1D-CNN, Bi-LSTM, and an attention mechanism for classification. The proposed model achieves outstanding performance, with 97.47% accuracy, 99% precision, 97% recall, and a 98% F1-score surpassing several benchmark models such as ResNet-50 V2 and DenseNet-201. These findings highlight the effectiveness of combining handcrafted feature engineering with temporal deep learning networks, offering a dependable solution for the early detection of glaucoma and laying the groundwork for future clinical deployment.