Climate and genetic data enhancement using deep learning analytics to improve maize yield predictability

Abstract

Despite efforts to collect genomics and phenomics ('omics') and environmental data, spatiotemporal availability and access to digital resources still limit our ability to predict plants' response to changes in climate. Our goal is to quantify the improvement in the predictability of maize yields by enhancing climate data. Large-scale experiments such as the Genomes to Fields (G2F) are an opportunity to provide access to 'omics' and climate data. Here, the objectives are to: (i) improve the G2F 'omics' and environmental database by reducing the gaps of climate data using deep neural networks; (ii) estimate the contribution of climate and genetic database enhancement to the predictability of maize yields via environmental covariance structures in genotype by environment (G×E) modeling; and (iii) quantify the predictability of yields resulting from the enhancement of climate data, the implementation of the G×E model, and the application of three trial selection schemes (i.e. randomization, ranking, and precipitation gradient). The results show a 12.1% increase in predictability due to climate and 'omics' database enhancement. The consequent enhancement of covariance structures evidenced in all train-test schemes indicated an increase in maize yield predictability. The largest improvement is observed in the 'random-based' approach, which adds environmental variability to the model.

Keywords: Climate data science; Genomes to Fields (G2F); deep neural network (DNN); genotype by environment (G×E) model; maize yield predictability; train–test schemes.