Plot Detection in Sequential Crop Field Images using Deep Learning
| dc.contributor.committeeMember | Stavness, Ian | |
| dc.contributor.committeeMember | Eramian, Mark | |
| dc.contributor.committeeMember | Gutwin, Carl | |
| dc.creator | Norbu, Phuntsho | |
| dc.date.accessioned | 2024-05-24T22:45:53Z | |
| dc.date.available | 2024-05-24T22:45:53Z | |
| dc.date.copyright | 2024 | |
| dc.date.created | 2024-05 | |
| dc.date.issued | 2024-05-24 | |
| dc.date.submitted | May 2024 | |
| dc.date.updated | 2024-05-24T22:45:53Z | |
| dc.description.abstract | Plant phenotyping, essential for understanding plant growth and development, is hindered by its resource- intensive nature. Recent advancements in non-destructive imaging techniques, particularly remote sensing and proximal sensing, have revolutionized plant phenotyping by enabling high-throughput data collection. However, challenges persist such as the laborious preprocessing demands. This thesis addresses the pre- processing challenge, particularly of selecting centered plot images from overlapping sequential plot images. Leveraging two deep-learning models, an image classification model (ICM) and a video summarization model (VSM), we show the potential of such models to automate this selection process. ICM approaches the task as a classification problem, while VSM adopts a video summary perspective to extract key frames representing centered plots. Both ICM and VSM, trained on a dataset comprising 78,420 canola plot images spanning two different years achieve a notable F1 score of 0.81. Our findings also reveal the robustness of these models to common data collection anomalies, including repeated images and a shuffled dataset where the sequential nature of the images is disrupted. This resilience underscores the practical utility of the models in addressing challenges inherent in real-world plant phenotyping scenarios. Additionally, to streamline the image selection process, we introduce PlotReel, a web application that allows quick sequential crop-field image selection. By predicting and seamlessly navigating to subsequent centered plot/row images based on regular spacing and constant camera speed, PlotReel simplifies the image selection process, saving users from manually sorting through overlapping and redundant images. | |
| dc.format.mimetype | application/pdf | |
| dc.identifier.uri | https://hdl.handle.net/10388/15711 | |
| dc.language.iso | en | |
| dc.subject | Deep learning, Plant Phenotyping, CNN, Video summary | |
| dc.title | Plot Detection in Sequential Crop Field Images using Deep Learning | |
| 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 | Masters | |
| thesis.degree.name | Master of Science (M.Sc.) |