Principal variable selection to explain grain yield variation in winter wheat from UAV-derived phenotypic traits

Automated phenotyping technologies are constantly advancing. However, collecting diverse phenotypic traits throughout the growing season and processing massive amounts of data still take lots of efforts and time nowadays. Selecting minimum number of phenotypic traits that have the maximum predictive power has the potential to largely reduce the phenotyping efforts. The objective of this study was to select principal UAV-derived parameters on winter wheat along the growing season that contribute most in explaining grain yield. Five times of multispectral imagery and seven times of RGB imagery were collected by an UAV system during the spring growing season in 2018, from which phenotypic trait (vegetation index and plant height) maps were generated. From these maps, a total of 172 parameters was calculated for each plot including statistical descriptions of the pixel values and the dynamic growth rate. These variables were import into two variable selection algorithms: LASSO regression (the least angle and shrinkage selection operator) and random forest. Ten variables with highest averaged importance scores were selected by each algorithm. Results showed that most of the selected variables were derived from plant height map, especially related with the plant height measured in the last two data collections in the growing season (grain filling and maturity stages). The methods provided in this study can be applied to larger data set collected from multiple years and locations to narrow down the important phenotypic traits and growth stages to be focused on in the data collection and processing to streamline the breeding process.