Misidentifications in ayurvedic medicinal plants: Convolutional neural network (CNN) to overcome identification confusions

Abstract

Plants are a vital ingredient of traditional medicine in Sri Lanka, and the quantity of medicinal plants used as a number differs in literature. Field identification of medicinal plants is carried out based on various plant characteristics. Conventional identification keys are available for plant identification, but it is a complex and time-consuming process that relies on manual observation, introducing inherent errors, particularly among individuals lacking extensive botanical expertise. This may cause uncertainty in the identification of plants due to lack of professional training, morphological similarity and nomenclatural confusion of plants. Such uncertainty may result in misidentifying another plant(s) as intended medicinal plants, which may lead to unsafe consequences. The objectives of the study were accomplished with the following three steps; listing the flowering plants used for medicinal purposes in Sri Lanka using multiple detailed botanical literatures, identifying medicinal plants that are confused with other medicinal or non-medicinal plants using literature and a questionnaire survey, and developing Convolutional Neural Networks (CNN) based technology to distinguish confusing plants. The study prepared a list of 1358 flowering plants cultivated and used in Sri Lanka as medicinal plants. Fifty-three medicinal plants that are confused with 63 medicinal and non-medicinal plant species were identified by two surveys. The CNN solution experimented with five species of the Bauhinia genus with close morphologically similar leaves with a high misidentification possibility. Using CNN EfficientNet-B0 Architecture for classification, four models were tested, and Model 4, which was trained using the augmented dataset with white-coloured background, resulted in a validation accuracy of 96.16%. The current study demonstrated the critical role of CNNs and EfficientNets in addressing misidentification issues in morphologically similar medicinal plants.