Quantitative Phenotyping using Deep Learning

Abstract

Phenotyping, the process of measuring characteristics (often referred to as traits), is a cornerstone of modern crop breeding. The automation of phenotyping (known as phenomics) is crucial for accelerating research in the face of climate change and increased pest biodiversity. In image-based phenomics, classical computer vision (CV) and deep learning (DL) have been used successfully for image-level and enumerative traits. However, the automated collection of accurate quantitative trait data at the organ level has not yet been fully explored. Typically, classical CV algorithms often lack the robustness and consistency needed to capture complex traits in variable domains. Deep learning, however, can be limited by the cost of the training data necessary to build a robust model. Concerns also remain about DL accuracy when applied to genetic research. This thesis reports the training of novel DL models, the design of phenotyping pipelines to extract quantitative trait measurements via both traditional and generated training data; relating variation in the traits to variation in the genome, and the creation of a novel weakly supervised training system that pretrains networks to recognise plant morphology. Firstly, synthetic training data was used as part of an active learning system to successfully train a phenotyping model for isolated Brassica pods, starting from a small initial dataset. Secondly, a DL model and pipeline were developed to extract organ-level traits from images of entire Arabidopsis fruiting stems from a multi-parent advanced generation inter-cross (MAGIC) population and successfully relate the variability of these metrics to variability in the genome using QTL analysis. Finally, a novel pre-training system was designed that utilises weakly supervised training to recognise plant structure morphology, guided by classical computer vision pipeline. This system was demonstrated on a new collection of images of both crop and non-crop relatives to pre-train a base model for general plant organ detection, followed by fine-tuning with minimal training data for downstream measurements. This thesis contributes to the development and verification of AI pipelines, paving the way for more efficient (pre-)training systems in image-based plant phenotyping. These advancements offer the potential to scale up phenotyping, accelerating genetic research and plant breeding programmes.

Date of Award	2025
Original language	English
Awarding Institution	Aberystwyth University
Supervisor	John Doonan (Supervisor) & Chuan Lu (Supervisor)