<![CDATA[Indonesia is one of the tropical countries with the greatest biodiversity in the world, including around 50,000 plant species and around 7,500 used by the community as raw materials for traditional medicine. However, the large number of plant species poses a challenge in distinguishing between edible plants and those that contain toxins, because herbal and toxic plants often have similar morphological characteristics of their leaves. Therefore, this study was conducted to design a classification system to distinguish between herbal and toxic plant leaves, which is expected to be utilized by the community as a preventive measure to reduce the risk of poisoning due to incorrect plant identification. The system was built using a Convolutional Neural Network (CNN) architecture implemented through the Flask framework and equipped with a rule-based system to determine plant categories based on the model's prediction results. The dataset consists of 960 images with 8 classes of local Indonesian plant leaves, where four categories include herbal plants (moringa, mint, betel, and basil) and four categories of poisonous plants (saga rambat, bandotan, gympie-gympie, and jelatang). This study conducted experiments on several CNN architectures, including custom and pretrained models (EfficientNetB0, MobileNetV2, and ResNet50V2). The best results were obtained from the EfficientNetB0 model trained using images with an input shape of 224×224 pixels, a batch size of 24, and the Adam optimizer, achieving a training accuracy of 99.51% and validation accuracy of 98.96%. This model demonstrated superior accuracy compared to other models, such as the custom model (93.88%), MobileNetV2 (99.50%), and ResNet50V2 (99.38%). The evaluation results show that the EfficientNetB0 model has excellent performance, with an overall classification accuracy of 99.00%, precision of 99.00%, recall of 99.00%, and an F1-score of 99.00%.]]>
Copyrights © 2025
