Plant Species Classification from Leaf Images Using Transfer Learning-Based EfficientNetB0 and MobileNetV2 Models

A ccurate and rapid identification of plant species is important for biodiversity conservation, agricultural applications, ecological research, and medicinal plant studies. In this study, transfer learning-based EfficientNetB0 and MobileNetV2 models were comparatively evaluated for the automatic classification of plant species from leaf images. The publicly available 48 Plant Leaves Dataset, containing 4,427 RGB images from 48 plant classes, was used. The dataset was initially divided using stratified sampling into training, validation, and test subsets. Both models were trained using ImageNet-pretrained weights and a frozen-backbone transfer learning strategy with architecture-specific input preprocessing. To assess the stability of model performance, the complete experimental procedure was repeated using five different random seeds. An exact-duplicate audit was also performed, and test images having byte-identical counterparts in the training or validation subsets were excluded from the primary evaluation, resulting in a duplicate-filtered test set of 651 images. On this test set, EfficientNetB0 achieved an accuracy of 93.09%, macro Precision of 93.51%, macro Recall of 93.17%, and macro F1-score of 92.90%, whereas MobileNetV2 achieved an accuracy of 88.48%, macro Precision of 89.23%, macro Recall of 88.27%, and macro F1-score of 88.10%. Across five repeated experiments, EfficientNetB0 achieved a mean accuracy of 92.81 ± 0.92%, compared with 88.18 ± 0.92% for MobileNetV2. A paired exact McNemar test also indicated a statistically significant difference between the two models on the duplicate-filtered primary test set (p = 0.00122). Class-level analyses further showed that EfficientNetB0 produced more consistent performance across plant classes, whereas MobileNetV2 exhibited greater variability for several classes. Overall, EfficientNetB0 achieved higher classification performance under the experimental conditions of this study; however, the findings are specific to the present dataset, preprocessing pipeline, and frozen-backbone transfer learning protocol and should not be interpreted as evidence of universal architectural superiority.