Interpretable Deep Learning Approaches For Identification And Classification Of Ethiopian Indigenous Medicinal Plant Species

Abstract

Ethiopia, known for its rich biodiversity, holds significant therapeutic potential in its diverse array of medicinal plants. Traditional medicines serve as cost-effective and culturally accepted healthcare solutions, used by the population in regions with limited healthcare infrastructure. However, identifying and classifying these Ethiopian indigenous medicinal plants species is a complex and time-intensive task requiring specialized scientific expertise. The main objectives of this research work is to identify andclassify Ethiopian indigenous medicinal plants using deep learning and interpretability. The study started using a systematic literature review aimed at investigating deep learning approaches to identifying and classifying medicinal plants. Subsequently, various deep learning approaches were employed to develop an efficient model through transfer learning and ensemble learning for identifying and classifying medicinal plants species. To tackle the interpretability issues of deep learning, interpretable deep learning models were designed using a multiple teacher-student approach with knowledge distillation concepts. It was done to present an integrated framework for identifying and classifying indigenous medicinal plants species using interpretable deep learning approaches. In the experimental phase, a dataset containing 12,438 labeled leaf images was prepared. Employing efficient pretrained models such as VGG19, VGG16, Xception, and InceptionNetV3, was adopted to enhance the model’s performance. In addition an ensemble EfficientNetB0, EfficientNetB2, and EfficientNetB4 are applied for the identification of parts and uses of Ethiopian indigenous medicinal plants. In the interpretable deep learning approach, a novel, distilled student model was designed using a collaborative teacher-student framework. The systematic review revealed disparities in global research due to resource and dataset variations, with most researchers uses private datasets and employing leaf shapes, transfer learning, and pre-trained models. The study effectively addressed the stated challenges and achieving a commendable accuracy of 95% through fine-tuning. The distilled student model attained an exceptional accuracy of 99.83%, facilitated by knowledge transfer metrics like cosine similarity and MSE. Integrating interpretability techniques such as LIME enhances model transparency and reliability, bridging traditional and modern medicine realms. Addressing the lack of globally accessible datasets for medicinal plants is essential to mitigating disparities in the field.