A Synergistic Approach of Coffee Leaf Pathology Segmentation with Self-Attentive U-Net

This study introduces a novel approach to automate the segmentation of stress regions in coffee leaves, addressing a critical challenge in precision agriculture. Early detection and management of biotic stresses such as diseases, pests, and environmental damage are vital to optimizing crop health. To address this, a deep learning model based on the U-Net architecture is proposed and enhanced with a Self-Attention mechanism. This hybrid approach combines U-Net’s feature extraction strengths with the attention mechanism’s ability to focus on critical regions, enabling precise segmentation in complex scenarios. The Coffee Leaf Biotic Stress Dataset was utilized, with images resized to $512 \times 1024 \times 3$ for computational efficiency and pixel-wise annotated masks for supervised learning. Each pixel was categorized into background, leaf, or stress symptom, supporting comprehensive segmentation tasks. The model achieved a mean Dice coefficient score of 0.90965 and the mean IoU of 0.89080, demonstrating strong segmentation performance across all classes. Self-attention improved the model’s ability to detect subtle stress patterns, differentiating stress symptoms from healthy regions and the background. This work highlights the potential of deep learning to automate plant health monitoring, reducing manual efforts and contributing to sustainable farming. Future enhancements, like class balancing and refined attention mechanisms, could further improve performance.