Abstract
Accurate identification of medicinal plant leaves supports the correct and safe use of herbal remedies, yet manual identification is often unreliable because many species share closely related leaf shapes, textures, and colors. Transfer learning has emerged as a practical strategy for image classification tasks of this kind, since it reuses the feature representations of pre-trained convolutional neural networks (CNNs) and thereby lowers both training time and computational demand. This study aims to compare the performance of three transfer learning backbones—MobileNetV2, ResNet50V2, and EfficientNetB0—on a five-class medicinal plant leaf recognition task covering guava (Psidium guajava), basil (Ocimum basilicum), aloe vera (Aloe vera), betel (Piper betle), and pandan (Pandanus amaryllifolius), using a working dataset of 1,750 leaf images (350 images per class before deduplication), and to statistically verify whether the performance differences among the three architectures are significant. Data were partitioned 80:10:10 for training, validation, and testing, and each architecture was trained across five independent runs under an identical configuration; Accuracy, Precision, Recall, and F1-score were used to quantify performance, followed by the Shapiro–Wilk and Kruskal–Wallis tests. MobileNetV2 produced the strongest mean Accuracy (98.27% ± 0.52), while ResNet50V2 led on Recall. The Kruskal–Wallis results, however, showed no statistically significant gap between the three architectures on any metric (p > 0.05). In other words, MobileNetV2, ResNet50V2, and EfficientNetB0 can be regarded as functionally equivalent for this classification task, and the practical choice among them should therefore weigh factors such as inference speed, model size, and deployment constraints rather than average predictive performance alone.