Analytics for climate-uncertainty estimation and propagation in maize-phenotype predictions

Increasing food demands triggered by population growth and climate change require improved diagnostics and prognostics of crop yields from local to regional to global scales. However, our abilities to predict crop's phenotypes are as constrained as our abilities to forecast sub-seasonal to seasonal hydrometeorological and climate events and identify the role crops' genetics plays. The goal here is to develop a conceptual framework to identify the sources and propagation of uncertainty of environmental variables (temperature, dew point, relative humidity, solar radiation, rainfall, and wind speed and direction) in the environmental covariance matrix as the main driver, along with genetics and their interactions, for the predictability of maize phenotypes. We hypothesize that a select number of environmental variables can drive the predictability of phenotypes (for this research, yield) based on the generation of environmental covariance structures of more than 90 environment records from Genomes to Field (G2F) Initiative in the US. The G2F project has integrated information and recorded datasets (genetics, phenotypes, and weather/climate) for researchers to develop methodologies for improving multiple maize hybrid genomes' traits predictability under different environmental conditions. The objectives of this study are: (1) to improve G2F weather time series and fill the gaps using machine learning techniques, (2) to couple the covariance matrix concept to a Global Sensitivity Analysis (GSA) method called PAWN, (3) to quantify the sensitivity of Environmental Covariates (ECs) to multiple input weather variables, and finally (4) to identify the primary uncertainty sources for environmental similarity based on ECs. By incorporating Environmental Covariates (ECs) with the Genetics by Environments (GxE) statistical model for phenotypic prediction, the predictability of maize phenotypes can be enhanced. The GSA-PWAN sensitivity index report the difference between conditional and unconditional cumulative distribution functions (CDF) of output for each environmental variable based on Kolmogorov-Smirnov statistic (K-S). Here, PAWN is used to predict ECs sensitivity in response to multiple environmental variables. We defined conditional and unconditional of each input (xi), and the environmental covariance is generated under these two situations, then the difference between two CDFs of output (ECs) is considered as a sensitivity response to each of the inputs. Based on PAWN-ECs coupling, the results indicate that temperature (K-S = 0.091), dew point (K-S = 0.089), and solar radiation (K-S = 0.082) are the leading causes of uncertainty in environmental similarities for phenotypic predictability, respectively.