Machine Learning Approach for Prescriptive Plant Breeding

We explored the capability of fusing high dimensional phenotypic trait (phenomic) data with a machine learning (ML) approach to provide plant breeders the tools to do both in-season seed yield (SY) prediction and prescriptive cultivar development for targeted agro-management practices (e.g., row spacing and seeding density). We phenotyped 32 SoyNAM parent genotypes in two independent studies each with contrasting agro-management treatments (two row spacing, three seeding densities). Phenotypic trait data (canopy temperature, chlorophyll content, hyperspectral reflectance, leaf area index, and light interception) were generated using an array of sensors at three growth stages during the growing season and seed yield (SY) determined by machine harvest. Random forest (RF) was used to train models for SY prediction using phenotypic traits (predictor variables) to identify the optimal temporal combination of variables to maximize accuracy and resource allocation. RF models were trained using data from both experiments and individually for each agro-management treatment. We report the most important traits agnostic of agro-management practices. Several predictor variables showed conditional importance dependent on the agro-management system. We assembled predictive models to enable in-season SY prediction, enabling the development of a framework to integrate phenomics information with powerful ML for prediction enabled prescriptive plant breeding.