Improving Deep Learning Classifiers for Plant Phenotyping using XAI Techniques
dc.contributor.advisor | Mondal, Debajyoti | |
dc.contributor.committeeMember | Stavness, Ian | |
dc.contributor.committeeMember | Sun, Shangpeng | |
dc.contributor.committeeMember | Elshorbagy, Amin | |
dc.contributor.committeeMember | McQuillan, Ian | |
dc.contributor.committeeMember | Vassileva, Julita | |
dc.creator | Mostafa, Sakib | |
dc.creator.orcid | 0000-0002-4777-7832 | |
dc.date.accessioned | 2024-06-18T04:15:51Z | |
dc.date.available | 2024-06-18T04:15:51Z | |
dc.date.copyright | 2024 | |
dc.date.created | 2024-05 | |
dc.date.issued | 2024-06-17 | |
dc.date.submitted | May 2024 | |
dc.date.updated | 2024-06-18T04:15:52Z | |
dc.description.abstract | The world's food security today is under threat and we need to improve the food production and nutritional values of the crops, invent climate-resilient foods, and plan better crop management to ensure food security. Precision agriculture is playing an integral role in food security by integrating technology to analyze large agricultural data to understand phenomic and genomic information, predict new breeds of crops, and improve crop production. Deep learning models are an important part of precision agriculture as they allow the analysis of large amounts of data and achieve impressive results. However, the deep learning models are of a black-box nature as they provide very little information about which features of the data contributed to the results and how such results were achieved. Explainable AI (XAI) allows researchers to investigate deep learning models and provide an explanation of the results which allows building trust in the model and helps improve and invent better models. In plant phenotyping, researchers are increasingly using deep learning models for analysis, but XAI techniques are being explored only recently. This thesis investigates the potential of XAI in plant phenotyping by leveraging XAI techniques for the selection and performance improvement of deep learning models with a focus on classification tasks in plant phenotyping. I start by conducting a comprehensive review of the XAI techniques in plant phenotyping. The review provides a detailed overview of the state-of-the-art and traditional XAI techniques that are used in various domains of research. The XAI techniques are designed to improve the explainability of the deep learning models. Therefore, I reviewed the deep learning models that help plant scientists study plant phenomics. The use of deep learning models and XAI techniques in plant phenotyping is still in its early stages. I provide an overview of the application of deep learning and XAI in plant phenotyping and propose an XAI framework that may be used to ensure the use of XAI for plant scientists. There are a few examples that use XAI techniques for the selection of a deep learning model in plant phenotyping. It is essential for a model to have the right complexity or depth relative to the dataset to achieve optimal performance. Therefore, I investigate ways to utilize XAI for selecting the model depth of deep learning classifiers performing plant phenotyping. I use a popular XAI technique known as Guided Backpropagation to visualize the learning of the intermediate layers of a deep learning model performing plant phenotyping tasks. I study the visualization of the feature maps to understand the relation between the capacity and the features learned by a model. I show that the shallow layers of a deep learning model learn more diverse features than the layers at higher depths. I designed a technique that leverages the Guided Backpropagation based visualization of the layers to provide insights into the model depth that should be used to achieve optimal performance for the dataset being used for model training. Optimization of the deep learning models is considered to be an important part of building the model. I propose two novel approaches that utilize an XAI tool known as the activation pattern of the neurons to improve the accuracy of deep learning classifiers. In the first approach, I embed the neurons' activation probability in the training process as a loss function that increases the classification accuracy of state-of-the-art deep learning models on popular datasets by as high as 4.5%. In the second approach, I developed an auxiliary model based training process that achieves a classification accuracy improvement of as high as 5.2%. Unlike the loss function based approach, the auxiliary model based training is faster and can be used to boost the performance of a pre-trained model. I conduct extensive experiments with diverse deep learning classifiers on traditional and plant phenotyping datasets. The results show that the described techniques can potentially improve the accuracy of the classifiers in all cases. | |
dc.format.mimetype | application/pdf | |
dc.identifier.uri | https://hdl.handle.net/10388/15767 | |
dc.language.iso | en | |
dc.subject | deep learning | |
dc.subject | plant phenotyping | |
dc.subject | explainability | |
dc.subject | xai | |
dc.subject | ai | |
dc.subject | explainable AI | |
dc.subject | data bias | |
dc.subject | agriculture | |
dc.subject | guided backpropagation | |
dc.subject | neural network visualization | |
dc.title | Improving Deep Learning Classifiers for Plant Phenotyping using XAI Techniques | |
dc.type | Thesis | |
dc.type.material | text | |
thesis.degree.department | Computer Science | |
thesis.degree.discipline | Computer Science | |
thesis.degree.grantor | University of Saskatchewan | |
thesis.degree.level | Doctoral | |
thesis.degree.name | Doctor of Philosophy (Ph.D.) |