Convolutional Neural Networks for Robust Fynbos Leaf Classification: Enabling Trustworthy Machine Learning in Botanical Science
| dc.contributor.advisor | Winberg, Simon | |
| dc.contributor.author | Govender, Jarushen | |
| dc.date.accessioned | 2025-08-18T11:26:00Z | |
| dc.date.available | 2025-08-18T11:26:00Z | |
| dc.date.issued | 2025 | |
| dc.date.updated | 2025-08-18T11:23:21Z | |
| dc.description.abstract | The Fynbos Leaf Optical Recognition Application (or FLORA) is a novel machine-learning tool created for the purposes of aiding conservation efforts of the Cape Floral Region, and the species of plants known as Fynbos in particular. Known for their distinctive evolutionary features, the species maintains a revered position in the ecological heritage of South Africa. This version of FLORA intends to make use of a Convolutional Neural Network trained on a dataset of collected leaf images, to correctly classify species of Fynbos using natural images as training data. The thesis intends to combat many of the pitfalls of using CNN technology such as working with small datasets and provides a novel approach for dealing with image quality issues and over-fitting that arise from working with limited data. This project also intends to be scalable, and to be able to grow and become more generalised as more training data are added. The collected data involved manual sample collection using photography equipment and consists of 1,196 field images spread across 35 different species of plants. A part of thesis involved the creation of a novel Image Quality Assessment tool to remove low quality images that negatively influenced the predictive capability of the model. The model evaluation process makes use of SHapley Additive exPlanations (SHAP), a tool for visualising model predictions, to contribute to the explian-ability of the model and to develop trust and confidence in machine-learning algorithms, with the ultimate aim of providing a tool to merge the fields of ecology, botany and electrical engineering. Multiple models were trained and evaluated and the selected model for the project obtained a classification accuracy of 76% on the validation data, and an F1-score of 74%. This was an extremely positive result as the training data consisted of exclusively natural images and no feature engineering was performed. The model was then tuned to specific hyper-parameter values which yielded a small performance boost, and then tested on its ability to generalise. Five new classes were added to the training set and the model performance remained consistent, demonstrating robust generalisation. This project contributes knowledge to the growing field of image recognition, and provides a clear framework for model explain-ability which should benefit future research endeavors. | |
| dc.identifier.apacitation | Govender, J. (2025). <i>Convolutional Neural Networks for Robust Fynbos Leaf Classification: Enabling Trustworthy Machine Learning in Botanical Science</i>. (). University of Cape Town ,Faculty of Engineering and the Built Environment ,Department of Electrical Engineering. Retrieved from http://hdl.handle.net/11427/41608 | en_ZA |
| dc.identifier.chicagocitation | Govender, Jarushen. <i>"Convolutional Neural Networks for Robust Fynbos Leaf Classification: Enabling Trustworthy Machine Learning in Botanical Science."</i> ., University of Cape Town ,Faculty of Engineering and the Built Environment ,Department of Electrical Engineering, 2025. http://hdl.handle.net/11427/41608 | en_ZA |
| dc.identifier.citation | Govender, J. 2025. Convolutional Neural Networks for Robust Fynbos Leaf Classification: Enabling Trustworthy Machine Learning in Botanical Science. . University of Cape Town ,Faculty of Engineering and the Built Environment ,Department of Electrical Engineering. http://hdl.handle.net/11427/41608 | en_ZA |
| dc.identifier.ris | TY - Thesis / Dissertation AU - Govender, Jarushen AB - The Fynbos Leaf Optical Recognition Application (or FLORA) is a novel machine-learning tool created for the purposes of aiding conservation efforts of the Cape Floral Region, and the species of plants known as Fynbos in particular. Known for their distinctive evolutionary features, the species maintains a revered position in the ecological heritage of South Africa. This version of FLORA intends to make use of a Convolutional Neural Network trained on a dataset of collected leaf images, to correctly classify species of Fynbos using natural images as training data. The thesis intends to combat many of the pitfalls of using CNN technology such as working with small datasets and provides a novel approach for dealing with image quality issues and over-fitting that arise from working with limited data. This project also intends to be scalable, and to be able to grow and become more generalised as more training data are added. The collected data involved manual sample collection using photography equipment and consists of 1,196 field images spread across 35 different species of plants. A part of thesis involved the creation of a novel Image Quality Assessment tool to remove low quality images that negatively influenced the predictive capability of the model. The model evaluation process makes use of SHapley Additive exPlanations (SHAP), a tool for visualising model predictions, to contribute to the explian-ability of the model and to develop trust and confidence in machine-learning algorithms, with the ultimate aim of providing a tool to merge the fields of ecology, botany and electrical engineering. Multiple models were trained and evaluated and the selected model for the project obtained a classification accuracy of 76% on the validation data, and an F1-score of 74%. This was an extremely positive result as the training data consisted of exclusively natural images and no feature engineering was performed. The model was then tuned to specific hyper-parameter values which yielded a small performance boost, and then tested on its ability to generalise. Five new classes were added to the training set and the model performance remained consistent, demonstrating robust generalisation. This project contributes knowledge to the growing field of image recognition, and provides a clear framework for model explain-ability which should benefit future research endeavors. DA - 2025 DB - OpenUCT DP - University of Cape Town KW - CNN technology, fynbos LK - https://open.uct.ac.za PB - University of Cape Town PY - 2025 T1 - Convolutional Neural Networks for Robust Fynbos Leaf Classification: Enabling Trustworthy Machine Learning in Botanical Science TI - Convolutional Neural Networks for Robust Fynbos Leaf Classification: Enabling Trustworthy Machine Learning in Botanical Science UR - http://hdl.handle.net/11427/41608 ER - | en_ZA |
| dc.identifier.uri | http://hdl.handle.net/11427/41608 | |
| dc.identifier.vancouvercitation | Govender J. Convolutional Neural Networks for Robust Fynbos Leaf Classification: Enabling Trustworthy Machine Learning in Botanical Science. []. University of Cape Town ,Faculty of Engineering and the Built Environment ,Department of Electrical Engineering, 2025 [cited yyyy month dd]. Available from: http://hdl.handle.net/11427/41608 | en_ZA |
| dc.language.iso | en | |
| dc.language.rfc3066 | eng | |
| dc.publisher.department | Department of Electrical Engineering | |
| dc.publisher.faculty | Faculty of Engineering and the Built Environment | |
| dc.publisher.institution | University of Cape Town | |
| dc.subject | CNN technology, fynbos | |
| dc.title | Convolutional Neural Networks for Robust Fynbos Leaf Classification: Enabling Trustworthy Machine Learning in Botanical Science | |
| dc.type | Thesis / Dissertation | |
| dc.type.qualificationlevel | Masters | |
| dc.type.qualificationlevel | MSc |