<![CDATA[Nilam Plant (Patchouli plant) located in Minahasa Regency can be affected by several diseases that can significantly reduce their essential oil capacity. The common practice of monitoring crops for diagnosis relies on labor-intensive methods that can be variable in accuracy, depending on previous experience of the tended. The main goal of the study is to develop and to compare model performance using Deep Learning computer vision in monitoring conditions of patchouli plants with respect to different conditions (healthy, bacterial wilt, viral, and budok). This study assess and compare MobileNet (a lightweight classifier) against Deep Learning based object detectors (YOLOv5, v8, v11) using an underlying dataset that was created unimpeachably in a natural patchouli field setting, consisting of 3,000 images which contain 3,820 annotated bounding boxes, across 4 classes (Healthy, Bacterial Wilt, Viral, Budok). Evaluation reveals a clear trade-off between the two models. MobileNet finds (nearly perfect) classification accuracy of 94.7% (F1-score>0.90 for all classes), while the faster YOLOv8l yields an 88.2% mAP50. Both models had the hardest time dealing with the "Viral" class due to visual similarities it shared with the healthy class (F1: 0.90, mAP: 0.81). MobileNet produced better accuracy (94.7%) but had slower inference time (3.0s). YOLOv8l provided real-time detection (1.4s) but lower mAP (88.2%). Our recommendation is a hybrid 2-stage system (YOLO-drone scan; MobileNet-farmer confirmation) as an operational approach for Precision Agriculture in Patchouli farming. Overall, the main takeaway is that: MobileNet is intended for diagnostic application and YOLOv8 superior for real-time video-based field monitoring.]]>
Copyrights © 2026
