University of Saskatchewan
Using procedural models to improve image-based plant phenotyping
Abstract
dc:description.abstractDigital plant phenotyping plays a crucial role in advancing agricultural productivity. Traditional manual phenotyping methods are time-consuming, labor-intensive, and prone to human error. To develop crop varieties that are more resilient to climate change and adaptable to diverse environmental conditions, plant breeders require accurate insights into phenotype, and its relationship to genotype — and to environment. Image-based high throughput plant phenotyping offers a scalable and efficient solution to support breeding programs aimed at creating climate-resilient and higher-yielding crops. This thesis utilizes various areas of computer science, with a primary focus on procedural plant modelling together with computer vision to improve image-based plant phenotyping. First, we survey the relationship between procedural plant modelling and computer vision, aiming to contribute to advancements in image-based plant phenotyping. Second, a novel technique is developed for leaf counting in maize. This method can predict leaves, sometimes even if they are occluded in the images by using information regarding plant development on top of image processing techniques, and by using a graph-theoretic matching algorithm. Third, we studied a number of important traits of canola, from indoor images of 50 different genotypes of canola. This includes a prediction to identify stress from images using classical machine learning algorithms, and 81% prediction accuracy is obtained. Fourth, we investigate the application of synthetic images generated from procedural plant models for training deep learning networks to do phenotyping tasks. The tasks included leaf counting in maize and inflorescence branch counting in canola. We determine that realism in synthetic images has a substantial effect on their ability to make accurate predictions on real images. Finally, we determine a new method using a genetic algorithm that can find parameters of a procedural plant model that produces a simulation that closely matches a sequence of maize images geometrically. This algorithm uses a novel distance function that uses a matching algorithm from graph theory. This method can be used to find a simulation that closely matches a temporal sequence of real images of a plant. Altogether, these works demonstrate that procedural models can help with phenotyping, especially in combination with other machine learning approaches.
Degree
thesis:*- Name thesis:degree_name
- Doctor of Philosophy (Ph.D.)
- Level thesis:degree_level
- Doctoral
- Discipline thesis:degree_discipline
- Computer Science
- Department thesis:degree_department
- Computer Science
- Grantor
- University of Saskatchewan
- Year dc:date.issued
- 2025
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Khan, Nazifa Azam
- Advisor dc:contributor.advisor
-
- McQuillan, Ian
- Committee members dc:contributor.committeemember
-
- Gutwin, Curl
- Parkin, Isobel
- Eramian, Mark
- Jin, Lingling
- Henry, Christopher
- Mondal, Debajyoti
Subjects
dc:subject × 1Rights
- Language dc:language.iso
- en
Identifiers
dc:identifier.*- Handle dc:identifier.uri
- https://hdl.handle.net/10388/17514