Automated Detection and Classification of Stem Canker in Dragon Fruit Using Computer Vision Techniques

Early detection of stem canker in dragon fruit is essential for preventing yield losses and improving crop management in commercial plantations. This study develops a three-stage, ground-based computer vision framework that automates the localization, segmentation, and severity classification of stem canker symptoms directly from smartphone imagery captured under real field conditions. A YOLOv8m detector first identifies individual cladodes, after which the Segment Anything Model (SAM2) refines each detection through high-resolution instance segmentation. The segmented cladodes are then classified into four severity levels—Healthy, Mild, Moderate, and Severe—using a hybrid VGG16-Vision Transformer model. Field-acquired images were compiled into a custom dataset and manually annotated by a plant pathologist to accurately reflect the true variability of orchards, including clutter, shadows, and occlusions. The proposed pipeline achieved strong performance across all classes, with F1-scores of 98.41% (Healthy), 90.91% (Mild), 89.83% (Moderate), and 95.65% (Severe), demonstrating robustness to complex visual conditions. While the current work concentrates solely on proximal imaging utilizing a handheld smartphone, the geotagged outputs generated by the system lay the groundwork for future incorporation into large-scale disease mapping frameworks and potential integration with UAV or satellite remote sensing layers. Overall, this study offers a reliable, field-ready approach for objective disease assessment and supports the development of scalable precision agriculture tools for dragon fruit production.