Leveraging knowledge distillation for lightweight and interpretable deep learning in Ethiopian medicinal plant classification

• Developed a lightweight and interpretable deep learning model for Ethiopian medicinal plant classification using knowledge distillation. • Employed a teacher-student framework to transfer knowledge from complex models to a compact, computationally efficient model. • Achieved high performance: 99.46% accuracy, 0.9949 precision, 0.9946 recall, and 0.9945 F1-score. • Used a proprietary dataset of 12,470 leaf images representing 44 species, enhanced with image processing and augmentation techniques. • Applied LIME-based interpretability analysis to reveal key leaf features driving classification decisions. • Demonstrated practical implications for integrating traditional medicine into modern pharmaceutical research. Deep learning has emerged as a valuable solution for image-based classification tasks such as identifying plant species. However, applying it to medicinal plant identification remains problematic due to insufficient data, high computing needs, and the difficulty in understanding how the models make conclusions. The prime objective of this study is to develop a lightweight and interpretable deep learning model using knowledge distillation for accurate identification and classification of Ethiopian medicinal plants. To address this objective, the research proposes a novel framework that combines compression techniques with knowledge distillation using teacher-student models. This integration aims to transfer insights from complex teacher models to a more compact student model thereby improving computational efficiency without compromising performance. The study employs a proprietary dataset comprising 12,470 leaf images representing 44 Ethiopian medicinal plant species, ensuring uniformity through image processing techniques and addressing data scarcity with augmentation methods. Training involves optimizing hyperparameters such as epochs, batch size, and activation functions. Experimental results show that the distilled student model attained high accuracy (99.46%), with a precision of 0.9949, recall of 0.9946, and F1-score of 0.9945, closely matching the teacher models. Interpretability analysis using techniques like Local Interpretable Model-agnostic Explanations sheds light on crucial plant features influencing classification decisions, enhancing model transparency and decision understanding. The novelty of this study lies in the combined use of compression, knowledge distillation, and interpretability for medicinal plant classification. The findings demonstrate the potential of interpretable deep learning and transfer learning in medicinal plant classification, offering computationally efficient and transparent solutions. This work provides practical implications for integrating traditional medicine into modern pharmaceutical research and supports the development of standardized botanical classification systems. Future studies are recommended to explore cross-domain datasets, multimodal data integration, and real-time deployment in resource-constrained environments.