COMPARATIVE PERFORMANCE AND GENERALIZATION ANALYSIS OF MOBILENETV1 AND MOBILENETV2 FOR RHIZOME SPICE CLASSIFICATION

Abstrak

Indonesia's rich biodiversity includes rhizome spices that are often difficult to distinguish manually due to their similar visual characteristics. This study developed and compared MobileNetV1 and MobileNetV2 for classifying four rhizome spice classes, namely ginger, turmeric, galangal, and aromatic ginger, using a dataset of 1,120 images. MobileNetV1 achieved a training accuracy of 0.9611 at a learning rate of 0.001 in 1,522.65 seconds, whereas MobileNetV2 achieved a higher training accuracy of 0.9823 at a learning rate of 0.0002 in 1,444.20 seconds. While MobileNetV2 demonstrated superior classification performance and faster convergence, MobileNetV1 demonstrated stronger generalization capability, indicated by a smaller train-validation accuracy gap (1.21% vs. 1.80%) and more stable validation performance. Both no-dropout models achieved an accuracy, precision, recall, and F1-score of 0.9642 on the 112-image test set. The two best-performing models were deployed in a Streamlit-based web application. The results demonstrate that MobileNetV2 is preferable when maximizing predictive performance, whereas MobileNetV1 offers greater robustness for relatively small datasets. This study contributes to the development of practical AI-based tools for agricultural and spice-identification applications.