<![CDATA[Indonesia harbors extensive biodiversity, including many medicinal plants used in health. Accurate and scalable identification remains challenging because of high intra class variation, strong visual similarity among species, and variable capture conditions in the field. This study evaluates three convolutional neural networks for Indonesian medicinal plant classification: ResNet18, VGG16, and EfficientNet_B0, and an average voting ensemble that aggregates their class probabilities. Using a held-out test set, ResNet18 achieves 87.00% accuracy, 91.16% precision, 87.00% recall, and 87.35% F1 score; VGG16 records 63.50% accuracy, 63.89% precision, 63.50% recall, and 61.10% F1 score; EfficientNet_B0 attains 94.00% accuracy, 94.49% precision, 94.00% recall, and 94.01% F1 score; the ensemble reaches 93.00% accuracy, 94.99% precision, 93.00% recall, and 93.22% F1 score. Confusion matrix analysis indicates that remaining errors concentrate on visually similar pairs, notably Aloe Vera versus Green Tea and Basil Leaf versus Green Tea. The findings show that EfficientNet_B0 is the strongest single backbone, while the ensemble improves precision and stability with accuracy close to the best individual model. These results demonstrate a practical pathway for reliable, field deployable recognition of medicinal plants in Indonesia. Future work will expand the dataset across seasons and regions, increase class balance, and explore stronger augmentation, focal or class weighted losses, and probability calibration. We also plan segmentation assisted pipelines, staged fine tuning at higher resolution, and lightweight deployment through pruning and quantization on mobile or embedded devices. To support reproducibility and adoption, the training pipeline and plotting scripts are made available in an open Kaggle notebook.]]>
Copyrights © 2026
