Identifikasi Tanaman Obat Indonesia dengan Vision Transformer dan Augmentasi Adaptif

Abstract

Manual identification of medicinal plants faces serious challenges due to morphological similarities between species, variations in lighting, and limited availability of botanists in the field. This study proposes an identification system for 100 types of Indonesian medicinal plants using the Vision Transformer (ViT) architecture with a stepwise fine-tuning approach and adaptive image augmentation. The model used is vit_base_patch16_224 initialized with ImageNet-1k pretrained weights, equipped with a classifier head consisting of a series of LayerNorm, Dropout (0.3), and Linear (768→100). The training strategy integrates stepwise freezing (freeze-unfreeze) in the first three epochs, the AdamW optimizer with a weight decay of 0.05, label smoothing (ε=0.1), and cosine-based learning rate scheduling to ensure stable convergence on medium-scale datasets. The dataset used consists of 10,000 images divided using a non-stratified random split with a ratio of 70% training data, 15% validation data, and 15% test data. The evaluation results show that the model achieved an accuracy of 97.3% on the test data, with a macro precision of 0.974, a macro recall of 0.975, and a macro F1-score of 0.973. The macro values ​​were calculated by summing the metric values ​​for each class separately, then dividing by the number of classes without considering the number of samples in each class. The training process lasted for 16 epochs before being terminated by the early stopping mechanism with the best validation accuracy of 97.53% at the 11th epoch. These results demonstrate that the ViT stepwise fine-tuning approach is able to address the challenges of multi-class scale classification on medium-sized datasets effectively and reproducibly.