The Open University
A Robust Framework for Automatic Classification of Medicinal Plants in the Natural Environment
Abstract
dc:description.abstractMedicinal plants provide essential therapeutic compounds with fewer side effects than pharmaceutical drugs. However, misidentifying species can have serious consequences, even fatalities. High intra-class and low inter-class variations among medicinal plants make accurate classification challenging for non-specialists and botanists alike. To automatically classify medicinal plants, researchers are using Machine Learning (ML) and computer vision techniques to process large image datasets. Convolutional Neural Networks (CNNs) excel in automatic image feature extraction and accuracy compared to traditional ML models. Nonetheless, their advancements are often limited by non-representative training images, causing misclassification due to variability within species. Additionally, the opaque nature of CNN models complicates understanding their decision-making processes. To address these challenges and improve classifications, this thesis presents a novel framework that automatically guides the selection of representative image samples for the training, validation, and testing subsets in the dataset partitioning phase, advises on CNN model selection and provides insight into the decisions made by the trained CNN using explainable artificial intelligence techniques. The guided dataset partitioning strategy employs a series of cognate algorithms for selecting representative training dataset samples, using Mutual Information (MI) as a similarity measure in combination with Gaussian mixture models, expectation-maximisation, and principal component analysis. To validate the new framework and gain valuable insights into how the CNN model achieves high accuracy, various state-of-the-art explainable artificial intelligence techniques were critically evaluated. Based on quantitative analysis, Local Interpretable Model-Agnostic Explanations (LIME) was applied, as it enables a more discrete and controlled analysis. The framework consistently demonstrated superior quantitative performance compared to a state-of-the-art approach. The research findings convincingly demonstrate the value of the proposed framework in improving the accuracy and reliability of CNN-based medicinal plant classification on complex, real-world image datasets. The integration of LIME-based quantitative analysis supports model-level explainability by offering insight into the internal decision logic of the network. This enables a more informed evaluation of the model’s behaviour rather than merely justifying its outputs. The framework equips practitioners with a flexible set of tools for robust and explainable automatic identification of medicinal plant species in natural environments. Additionally, the framework is applicable to domains such as wildlife monitoring, facial recognition, and agricultural diagnostics, where high intra-class and low inter-class variations make classification particularly difficult.
Degree
thesis:*- Name dc:type.qualificationname
- phd
- Level dc:type.qualificationlevel
- doctoral
- Grantor dc:publisher.institution
- The Open University
- Year dc:date.issued
- 2025
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Shahmiri, Lida
Rights
- Language dc:language
- en