Sequential Minimal Optimization of Herbal Plants and Bias Treatment of Multi-Feature Attributes for Quantitative Ethnobotany

The unprecedented endemic nature of viral infections and diseases posing global medical challenges in recent time calls for home-grown and inclusive interventions. Nonetheless current clinical efforts by medical and research virologists through orthodox medicine, traditional medicine is an area largely untapped for a broad-based proactive measure and interventions. Whereas acceptance rate of traditional medicine scales gradually, the fundamental problem of identification of herbal plants by end users suffices. Notwithstanding the natural outcrop of Medicinal herbal plants across ethnicities and communities, they are mostly unidentified by and alien to digital natives hence militating against potency awareness and possible usage for basic curative procedures. With advancements in data mining and artificial intelligence however, innovative intelligent systems have been deployed for several interventionist efforts across various professions including the health subsector. These novelties can be further enhanced and tailor-made to address the fast-paced trend in human health challenges. Existing literatures in the area of plant classification for identification often deploy laboratory dataset of plant images for predictive analytics with little or no consideration for field plants in the natural habitat of end users. In this work, machine learning is deployed for a novel leaf-based biometric system for ethno botanical herbal plants using Synthetic Oversampling Minority Technique (SMOTE) and Multi-attribute approach. The featured framework encapsulates herbal plant-image capturing by end users, preprocessing for multi-attribute extraction, bias treatment in training set by SMOTE, and comparative classification across machine learning algorithms each chosen from Functions, Decision trees, Lazy and Bayes classifying categories. We demonstrate with iteration of SMOTE processes that a less performing classification algorithm could surpass threshold by reducing class imbalance. Our multi-attribute approach significantly improved classification precision and accuracy as a clear departure from existing trend in texts. A precision accuracy of 0.993 is accomplished through Sequential Minimal Optimization as the most successful identifier with the next best result achieved by Liblinear with 98.3%. Deploying this proposed model of herbal plant biometrics for software design, which outperforms existing literature, would greatly promote herbal medicine towards finding home-grown solutions to global health challenges.