Comparative Analysis of Machine Learning Classifiers for Medicinal Plant Identification

Abstract

Medicinal plants have played a crucial role in human healthcare for centuries, with more than 80% of the developing world’s population relying on traditional medicine for primary healthcare needs . However, the rapid loss of plant species—estimated at 100 to 1000 times greater than natural extinction rates—threatens both biodiversity and potential drug discovery, with the Earth losing at least one potential major drug every two years . The accurate identification and classification of medicinal plant species by botanist experts remains a complex, time-consuming, and error-prone activity, necessitating automated solutions. This manuscript presents a comprehensive analysis of machine learning classifiers for medicinal plant identification, evaluating eight classical machine learning algorithms and six deep learning architectures on multiple medicinal plant datasets. The proposed system integrates image preprocessing techniques including noise reduction, normalization, and data augmentation, followed by feature extraction using Histogram of Oriented Gradients (HOG), Local Binary Patterns (LBP), Scale-Invariant Feature Transform (SIFT), and deep feature extraction using pre-trained convolutional neural networks. Experimental evaluation on the Central India Medicinal Plant Dataset (CIMPD), comprising 9,130 leaf images across 23 medicinal plant species, demonstrates that DenseNet201 with optimization and histogram equalization achieves the highest accuracy of 99.0%, followed by VGG16 with 98.7% and DenseNet201 with histogram equalization at 98.58% . Among traditional machine learning classifiers, Support Vector Machines with deep features achieve 96.76% accuracy. Transfer learning with pre-trained models was employed in 83.8% of recent studies, with Convolutional Neural Networks (CNNs) used by 64.5% of researchers as the primary deep learning classifier . The findings indicate that deep learning approaches, particularly pre-trained CNN architectures with appropriate preprocessing and data augmentation, significantly outperform traditional machine learning methods, achieving testing accuracies exceeding 90% on plant organs such as leaves and flowers . This research contributes to the development of accurate, scalable, and deployable medicinal plant identification systems for biodiversity conservation, pharmacological research, and traditional medicine preservation.