Implementing sensor technology to evaluate genetic and spatial variability within the Kansas State University Wheat Breeding program

Abstract

Globally wheat is one of the three most important cereal crops globally providing 20% of protein and total calories consumed. In the world as well as the state of Kansas, wheat is planted on more acres than any other crop. Additionally, wheat sales generated $1.27 billion in revenue in 2021 making wheat an economic driver for the entire state. However, the annual genetic gain in wheat is 0.8-1.2% and is not sufficient to support the increasing global population. Therefore, the adoption of new technology and computational methods are critical to increase genetic gain and increase wheat adaptability both globally and in the Central Plains. Proper temporal resolution is critical for quality HTP sensor data collection, as collection at key physiological growing points can increase yield prediction and assist with phenotypic selection. However, growth stages are dependent on weather and fluctuate both across locations and years. This makes day of year or day after sowing a poor phenology metric, particularly with winter wheat where the vernalization requirement compounds phenology prediction challenges and significantly shifts developmental stages relative to calendar days. This study was designed to assess the performance of various phenology models to predict heading time of both historically adapted and experimental genotypes of wheat genotypes in Kansas. The results suggest that full season models with multi-phase coefficients can increase phenology prediction over traditional thermal indices. However, using cumulative thermal times after the vernalization requirements also provided phenology predictions that were statistically similar to the full season phase change models. Genotype by environment interactions is a prominent issue for breeding programs, particularly when performance testing elite lines across multiple locations and years. In addition to macroenvironments, variations in soil properties have shown to develop microenvironments within location years. These soil microenvironments can potentially be quantified through both traditional and precision agriculture tools. Whereas, traditional soil sampling density is limited by cost and time, precision agriculture on-the-go soil sensors have the potential to gather large quantities of data. However, these measurements are often giving only relative measurements. Through this experiment two sensor platforms were evaluated as potential tools to quantify spatial variability within breeding programs. This study showed that soil spatial variability does impact genotype yield performance and that indirect measurements from both sensor platforms can quantify this impact. The continued development of high quality, cost effective multi-spectral imaging devices has led to numerous studies to evaluate this technologies ability to predict traits and grain yield. Despite these advancements the widespread implementation of these tools for selection has been slow and most breeders still rely on harvested grain yield and visual selection for cultivar advancement. The intention of this experiment was to evaluate high spatial resolution data from, multi-spectral sensors at multi-temporal collection points to make yield group rank order selections. Additionally, a random forest algorithm was used to evaluate the potential of incorporating machine learning with HTP data as a selection tool. Although the rank order correlations were higher than the correlation to grain yield, the selection accuracies of random forest were not statistically better than the no-information rate. However, this study does lay the groundwork for future similar studies using alternative sensor aided metrics and machine learning algorithms. Overall, the combined results of these studies show that these precision agriculture tools have to potential to increase genetic gain in plant breeding. However, these studies also show that both sensor and computational limitations still exist. Moving forward it is pivotal that future studies focus on technology combinations that have the potential to easily be implemented within a breeding program.