Recipient Organization
OHIO STATE UNIVERSITY
1680 MADISON AVENUE
WOOSTER,OH 44691
Performing Department
(N/A)
Non Technical Summary
Coffee is one of the most important globally-traded crop commodities with a retail value of $200 billion globally. Hawaii is the main coffee producing state in the US, and it is of high agricultural importance within the state. In 2022 a 17% increase in coffee production is forecasted with an estimated value of $60 million and a downstream economic impact valued at $211 million. Breeding in perennial fruit crops such as coffee involves hurdles that add to the difficulties of most annual crops, including long breeding cycles, larger areas required for field evaluation, smaller populations, and challenges in multiplying individual genotypes to replicate in single or multiple environments. Coffee has the added complication of being grown in different agronomic systems and biophysical environments. There have been relatively few efforts to integrate high-throughput phenomics into coffee selection and breeding, and this project seeks to fill that gap. This project will leverage a new coffee breeding program being developed by World Coffee Research in collaboration with other national programs and research organizations including the USDA/ARS in Hilo Hawaii. We will develop phenomic methods to be applied in this network (and any other coffee breeding programs), to improve selection and increase genetic gain in coffee. Specifically, this project will develop a suite of machine learning algorithms to relate rapid hyperspectral reflectance measurements to key gas exchange and foliar nutrient composition traits and yield, as well as resilience to coffee rust, developed from data spanning multiple sites, growth environments and seasons.
Animal Health Component
30%
Research Effort Categories
Basic
20%
Applied
30%
Developmental
50%
Goals / Objectives
Phenomics is the quantification of plant growth, performance and composition to efficiently connect genomics to plant function and agricultural outputs. Modern phenomics applied to plant breeding includes sensing and evaluation platforms that can be affordably deployed to dramatically improve throughput. Phenomic prediction can allow for more accurate direct selection. It provides for objective evaluation of traits and crop characteristics, at higher frequency and for larger populations across a broader range of environments than traditional breeding selection methods and genomic prediction. Hyperspectral reflectance observations have gained significant momentum as a key tool for phenomic assessment due to dense and diverse information content in highly resolved spectra (order of 1-5 nm per resolved spectral band), allowing for the identification of a variety of traits related to pigment and nutrient contents, leaf structure and moisture content, disease incidence or susceptibility, among others. The overarching goal of this project is to bring the power of hyperspectral reflectance phenotyping to coffee assessment and breeding through the development of coffee-specific algorithms related to gas exchange parameters, foliar nutrient profiles, yield and coffee rust incidence and resilience.There have been relatively few efforts to integrate high-throughput phenomics into coffee selection and breeding. Most applications have been focused on seed composition and origin classification. There have not been any efforts to deploy phenomic technologies to increase coffee seed yield directly, which is the main factor affecting farmer profitability. Assessing plant photosynthetic capacity offers one path to enhancing plant productivity and seed yield. Developing a high-throughput algorithm to quantify coffee photosynthetic capacity, across growth habit and environmental variation, from rapid hyper-spectral reflectance measurements is one of the primary goals of this project, motivated by the fact that model calibration for this approach has not been found transferable between species.Another primary goal of this project is the development of a set of algorithms relating costly and time-consuming leaf nutrient composition observations with rapid hyperspectral reflectance measurements. Leaf nutrient composition is not only important for overall surveys of plant and soil health, but of importance to study the possible confounding interactions of foliar nutrient contents with photosynthetic capacity and overall yield. For example, in coffee high leaf phosphorus content has been associated with high photosynthetic capacity under drought.Another important goal of this project is the development of hyperspectral reflectance algorithms related to the susceptibility and incidence of coffee leaf rust. Coffee leaf rust is the most important pathogen in coffee, reducing both yield and quality. Developing reflectance methods to evaluate and predict the susceptibility of coffee plants to rust damage would be useful for both surveillance of coffee plantations as well as the early evaluation of germplasm with resistance to the pathogen. In other crops and other species of leaf rust, reflectance-based methods have been demonstrated to predict disease severity prior to expert evaluation. In coffee, leaf rust detection through spectroscopy has been focused on a few vegetation indices, not taking advantage of the information in the full visible through shortwave infrared spectrum available through many of our modern field spectrometers. Our approach will utilize the full information in high spectral resolution reflectance spectroscopy measurements for leaf rust incidence detection.
Project Methods
This project will integrate hyper-spectral reflectance data collection with multiple machine learning approaches for optimally extracting data from such high-dimensional datasets. The focus here is on coffee traits spanning foliar gas exchange and nutritional profiles, plant yield and rust resistance. Visible through shortwave infrared (VSWIR) reflectance spectroscopy has been shown to be an immensely powerful sensing technique for characterizing terrestrial vegetation. VSWIR reflectance measurements using a field spectrometer will be a core measurement methodology spanning all goals of this project. These measurements provide a high spectral resolution (order of one or a few nm) view of the interactions of the leaves with light across the 350-2500 nm portion of the electromagnetic spectrum, where the majority of the solar energy is concentrated. These hyper-spectral reflectance observations will be used to develop trait algorithms using Partial Least Squares Regression (PLSR). PLSR has been widely demonstrated to be an effective approach for utilizing all non-redundant information in high dimensional (hyper-spectral) reflectance spectra in the plant sciences. PLSR is a multivariate technique that has been shown to be extremely powerful for predicting a set of dependent variables from a large set of independent variables (i.e. hyperspectral VSWIR reflectance). PLSR is particularly suited to problems in which the matrix of predictors has significantly more variables than observations, where multi-collinearity among the predictor variables can be problematic, introducing redundancy into a model. PLSR models the relationship between the dependent variables and the predictors by projecting both sets of variables onto a set of linearly independent latent vectors. This projection onto a new, linearly independent basis set bears resemblance to principal component regression (PCR) but differs by constructing this new basis set such that it not only optimally reduces the variance in the independent variables, but also explains as much as possible of the covariance between the independent and dependent variables. Rigorous cross-validation and model evaluation will be a central component of the model development aspects of this project in order to guarantee the robustness and broad applicability of the models developed here. Leaf gas exchange (i.e. Licor 6800) measurements will be used to quantify the photosynthetic parameters. Lab analytical analysis will be used to evaluate foliar nutrient concentrations. Leaf rust incidence and severity will be rated by experts in the field for model development and validation purposes.To achieve the goal of determining the genetic bases and accuracy of selection for yield and associated traits using the phenotypic data (yield, spectra, photosynthesis, leaf nutrients, leaf rust severity) of the segregating populations in Hawaii and El Salvador, along with the genotyping data of the mid-density panel, we will test various single trait and multi-trait methods: Bayesian GBLUP, Ridge Regression, partial least squares regression, sparse kernel, multivariate linear mixed models, among others. In this component of the project we will include hyperspectral reflectance information from multiple regions of the spectrum, providing additional reflectance "traits" to explore.Another goal is to identify genomic regions that control phenotypic variation, allowing for the selection of associated markers for Marker-Assisted-Selection and the development of populations to further map the traits of importance for yield. In this project, the genome wide association mapping within each species will be carried out using the GAPIT package using the iterating Fixed and Random Model Circulating Probability Unification method, analyzing each trait separately. Population structure and kinship will be used to account for the confounding effects between marker testing and relatedness estimation. In all cases, the SNP markers will be filtered for a minor allele frequency higher than 5% and less than 10% of missing data. We will build on our previous results to evaluate in coffee the genomic regions that control directly measured and modeled trait values across all coffee hyperspectral reflectance samples (i.e. the spectral library) collected in this project.