Sequential Dual-Model Verification for Medicinal Plant and Leaf Identification

Computerized recognition of medicinal plants from images has significant applications in ethnobotany, biodiversity surveillance, and community medicine. In this paper, we introduce a sequential dual-model pipeline that initially classifies an image of a whole plant and then confirms the prediction with a paired leaf image. The research centers on a curated dual-modality dataset of medicinal plants, screened to exclude classes with poor representation for stable training. Four pre-trained convolutional neural networks were independently trained on each of the plant and leaf modalities with experiments under various training regimes. Among these backbones, DenseNet-121 was found to be the most consistent and robust for plant and leaf modalities, achieving an accuracy and F1-score of 0.94 in a sequential pipeline, while sustaining an accuracy of 0.90 on an external web-scraped test set. Also, ResNet-50 has achieved the highest sequential accuracy of 0.95 out of all these models, although it showed weaker performance in a single modality, therefore underlining the efficiency of cross-modal sequential verification. The models were tested with an internal controlled split and an external web-scraped test set to gauge generalizability. Experiments showed that the sequential pipeline used in this work enhances decision reliability by integrating complementary cues from plant and leaf images. Among the backbones that were tested, certain architectures showed highly robust consistency and validation performance, demonstrating the merit of dual-stage classification for application in the real world. All training and testing were performed on a cloud-based GPU environment. The results highlight the merit of sequential dual-input architectures for stable medicinal plant identification and lay out guidelines for future optimization in field applications.