Enhancing genetic gains with high-throughput phenotyping and genomic selection: approaches in wheat and maize breeding

Abstract

This work is composed by four chapters: Chapter 1 consists in a pioneer work where we attempted to provide the scientific community with relevant insights on multi- spectral high-throughput phenotyping in tropical wheat breeding aiming at designing indirect selection strategies. The main goals were to determine the best stages to acquire aerial images to extract vegetation indices, investigate key genetic parameter (such as heritability and accuracy) in different stages, understand the relationship between traits and evaluate the efficiency of indirect selection using aerial-based high-throughput phenotyping. Chapter 2 is an extension of Chapter 1. While in Chapter 1 we focused on single-time-point analysis, in Chapter 2 we leveraged linear mixed models to analyze high-throughput phenotyping data considering its longitudinal aspect. We successfully predicted single genotypic values from time- series-derived data and found relevant insights on the optimal number of flights required to acquire high-quality multi-spectral aerial information. We also decomposed the coefficients of correlations between spectral and agronomic data into direct and indirect effects, and used that information to design an indirect selection strategy. In Chapter 3, instead of using the multi-spectral variables to design selection strategies, we focused on implementing a machine learning pipeline to predict untested wheat genotypes. In short, we used spectral bands and vegetation indices as inputs to predict agronomic variables using machine learning models considering a stage-wise approach. Finally, Chapter 4 shows the application of digital phenotyping and genomic selection to predict temperate inbred lines. We integrated environmental information, novel and traditional phenotyping methods and whole-genome resequencing information to predict the performance of a subset of maize diversity panel for stalk lodging resistance. We demonstrate the usefulness of novel phenotyping platforms to access lodging resistance and show how to leverage information between correlated environments to boost predictive-ability. This thesis is composed by several works, where we attempted to integrate the state of art of plant breeding and fundamental concepts of quantitative genetics and statistics to solve real problems often faced by plant breeders. We hope this work serve as a guide for the implementation of future pipelines aiming to improve genetic gains in plant breeding programs of economically important crops. Keywords: Artificial intelligence; Digital phenotyping; Genomic prediction; Genotype- by-environment interaction; Linear mixed models; Triticum aestivum L.; Zea mays.