A Latent Variable Model for Plant Stress Phenotyping Using Deep Learning

Abstract

With a growing population and a changing climate, increasing crop yields in a diversity of environmental conditions is becoming increasingly important. Studying genome-by-environment (GxE) effects is a critical path for such improvements, and high-throughput plant phenotyping is necessary for carrying out such experiments at scale. Image-based phenotyping techniques offer a scalable, non-destructive way of quantifying plants' responses to their environment - however, these techniques can be cumbersome and subjective. Each image dataset is unique, and requires either a hand-crafted image processing pipeline or a large annotated training set, which can be expensive and time-consuming. Additionally, researchers must select what feature is to be used to quantify changes due to the treatment, such as biomass, colour, the number of organs, or some other visual indication of the individual's response to its environment.

This dissertation explores image-based plant phenotyping, beginning with a discussion of image processing tools. Deep learning is introduced, with a survey of popular deep learning tasks applicable to plant phenotyping. Deep Plant Phenomics, a novel software platform for deep learning research in plant phenotyping, is introduced. A model of the Arabidopsis thaliana rosette is introduced and it is demonstrated that the use of synthetic data has the potential mediate some of the issues common to plant image datasets in deep learning. Finally, Latent Space Phenotyping (LSP) is introduced. LSP is a novel paradigm for quantifying response to treatment in plants which requires no hand-engineered pipelines or annotation of training data. The ability of LSP to detect arbitrary visual responses to treatment is demonstrated through five case studies involving both real as well as synthetic data. These case studies show that the method replicates two previously identified candidate loci for drought tolerance in an interspecific cross of Setaria, as well as demonstrating the flowering-time dependent drought response of Brassica napus L. The flexibility of the previously described synthetic A. thaliana model facilitates follow-up discussion where the behaviour of LSP is studied in additional experiments.

The techniques described in this dissertation lay the groundwork for future developments in image-based plant phenotyping, particularly in the use of deep learning, simulation, and latent variable models.