Transfer Learning Approach for Medicinal Plant Classification using ConvNeXt and SVM
Abstract
Medicinal plant classification is a vital area of study in botany, pharmacology, and traditional medicine. Accurate classification of medicinal plants is essential for drug development, conservation, and effective utilization of natural resources. Traditional classification is based on direct observation and expertise, which may lead to human error. With the introduction of deep learning and computer vision, CNN-based classification of medicinal plants is gaining popularity. Recently, a novel CNN based classification approach, known as ConvNeXt, has been shown to outperform traditional CNN architectures. In this paper, a hybrid framework combining ConvNeXt Base, Squeeze and-Excitation (SE) channel attention, and a Support Vector Ma chine (SVM) classifier head is proposed for robust medicinal plant classification. Evaluation is conducted using a two-phase transfer learning strategy across multiple datasets: a primary 30-class dataset (3,670 images) and a secondary 40-class dataset (11,890 images). The proposed model achieves an outstanding accuracy of 99.22% (95% CI: [99.14%, 99.31%]) on the primary dataset and 98.76% on the secondary dataset. Rigorous 5-fold cross-validation and paired statistical hypothesis testing (p < 0.001) confirm significant superiority over traditional baselines. Furthermore, a comprehensive computational complexity analysis demonstrates an inference throughput of 0.42 ms/image , confirming both effectiveness and deployment feasibility.