Dual-stream deep learning for medicinal plant species identification using leaf shape and venation features

Accurate classification of medicinal and aromatic plant species is crucial for biodiversity conservation, pharmacological research, and the sustainable use of botanical resources. This paper proposes a dual-stream deep learning framework that integrates leaf-shape and venation-pattern features to enhance plant species classification. Pretrained MobileNetV2 architectures are employed for both shape and vein feature extraction, and are compared against custom convolutional neural network (CNN) models for performance evaluation. Experimental results on a dataset comprising 39 plant species show that shape-based features extracted with MobileNetV2 achieve a validation accuracy of 94.94%. On the other hand, vein features extracted using CNN and MobileNetV2 achieve lower accuracies of 57.63% and 67.09%, respectively. Feature-level fusion of the shape (MobileNetV2) and vein (CNN) models yields a combined accuracy of 82.28%. Prominently, the dual MobileNetV2 architecture, which combines both modalities, achieves the highest validation accuracy of 98.28% and an F1-score of 98.58%. The experimental results show that the proposed framework leverages the complementary characteristics of leaf shape and venation patterns: shape features capture the global morphological structure of leaves, while venation patterns encode fine-grained internal anatomical structures, such as vein connectivity and branching. By combining these two modalities through a dual-stream architecture, the model learns richer, more discriminative feature representations, thereby enhancing the classification performance. Our model validates the potential of the proposed multimodal deep learning methods for building effective classification systems for medicinal plant species.