Herbal Leaf Identification for Balinese Lontar Usada Knowledge Preservation Using YOLOv8 Object Detection

Abstract

Indonesia is a megabiodiversity country with more than 30,000 documented medicinal plant species. Much of this ethnobotanical knowledge is preserved in the Lontar Usada Bali, a traditional Balinese manuscript that records the medicinal uses of plants. However, preserving this knowledge is challenging due to the declining number of traditional practitioners and the difficulty of identifying medicinal plants in natural habitats. This study proposes a deep learning-based medicinal plant detection system using the YOLOv8 architecture to identify 12 classes of medicinal plant leaves in Taman Usada Bali. A total of 1,344 images containing 3,230 annotated leaf objects were collected under diverse lighting and background conditions. To improve model generalization, horizontal flipping, vertical flipping, rotation, Mosaic, and MixUp augmentations were applied. Five YOLOv8 variants (YOLOv8n, YOLOv8s, YOLOv8m, YOLOv8l, and YOLOv8x) were evaluated using Precision, Recall, F1-score, mAP50, and mAP50–95 metrics. Experimental results showed that all models achieved high detection performance, with YOLOv8m obtaining the highest mAP50–95 score of 0.8676. However, Cost Benefit Analysis (CBA) using the Weighted Sum Model (WSM) identified YOLOv8n as the optimal model. Although YOLOv8m achieved the highest accuracy, YOLOv8n obtained the highest WSM score (2.6400) by balancing detection performance (mAP50–95 of 0.8464         ) and computational efficiency. With a 6 MB model size, 2.7 ms inference time, and 1.559 hours of training, YOLOv8n is suitable for real-time mobile and edge-computing applications. The novelty of this study lies in integrating Lontar Usada Bali taxonomy into a structured dataset, applying WSM for model selection, and enhancing detection robustness through Mosaic and MixUp augmentation.

GS Cited Analysis