PalmRachis-BiLSTM-Attn: An Anatomically Guided Explainable Deep Learning Framework for Spatial–Temporal Progression Modeling of Date-Palm Leaf Diseases

Early and interpretable detection of foliar diseases in date palms is essential for sustainable agriculture in arid regions, yet conventional deep-learning models treat symptoms as static visual patterns without considering their biological progression along the rachis. This study introduces PalmRachis-BiLSTM-Attn, a bio-inspired and explainable deep-learning framework that models palm-leaf pathology as a spatial–temporal progression process guided by the anatomical ordering of the rachis. The method begins with a CNN backbone that extracts textural cues from leaf images, which are then reorganized into rachis-aligned sequences and processed through a bidirectional LSTM with temporal attention to capture both forward (disease spread) and backward (regression) vascular dynamics. An integrated explainability module combines Grad-CAM spatial maps with temporal-attention timelines, generating biologically meaningful “where–when” visualizations of disease development. Evaluated on an eight-class date-palm leaf dataset, the framework achieves 98.1% classification accuracy and outperforms static CNNs and non-anatomical sequence baselines across all metrics (precision = 0.988, recall = 0.984, F1 = 0.984). By aligning model behavior with real infection pathways, PalmRachis-BiLSTM-Attn advances progression-aware, transparent, and biologically grounded disease monitoring, offering a deployment-oriented foundation for IoT-enabled crop-health systems and future smart-agriculture.