Generalizable Canola Seedling Detection in Aerial Field Imagery

As a major oilseed crop, canola has an important role to play in improving global food security. By making timely management decisions based on accurate assessments of crop health, producers can improve the yield returns of their canola fields. Emergence plant counts are prized as one of the earliest and most useful indicators of crop health. Traditionally, plant population numbers have been estimated by manually counting the number of emerged seedlings at various field locations. Due to the high amount of labour involved, only small areas of a field can be counted with this approach. In recent years, a less labour-intensive, higher-throughput approach to plant counting has been proposed. By applying powerful deep learning- based object detection models to drone-captured field imagery, plant seedlings can be counted automatically. Unfortunately, a number of practical challenges have thus far impeded the widespread adoption of this new approach to plant counting. In particular, long turnaround times and poor model performance in novel image conditions have prevented the realization of a trustworthy, rapid, and autonomous canola detection system. In this thesis, we seek to address these issues. First, we introduce a software tool that enables users to train and apply plant detection models in very few steps. By providing direct access to object detection models through an easy-to-use interface, our tool enables canola researchers and producers to obtain emergence count estimates in time to make critical management decisions. Second, we present a large and diverse dataset of aerial canola seedling images. We show that models trained on this dataset can generalize well to new aerial canola seedling image sets that have been acquired with the same capture setup. Finally, we conduct a series of experiments that investigate several key issues surrounding the training and deployment of plant detection models. We hope that the tool, data, and experiments we present will help spur the development of better autonomous plant detection systems.

Name thesis:degree_name

Master of Science (M.Sc.)

Level thesis:degree_level

Masters

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Saskatchewan

Year dc:date.issued

2024

Author dc:creator

• Andvaag, Erik Arvid

Advisor dc:contributor.advisor

• Stavness, Ian

Committee members dc:contributor.committeemember

• Eager, Derek
• Shirtliffe, Steve
• Makaroff, Dwight

• deep learning
• plant phenotyping
• computer vision
• object detection

Language dc:language.iso

en

Handle dc:identifier.uri

https://hdl.handle.net/10388/15531

OAI identifier oai:identifier

oai:harvest.usask.ca:10388/15531