Optimization of temporal UAS-based imagery analysis to estimate plant maturity date for soybean breeding

Abstract

Estimating the date of maturity of soybean breeding field plots is necessary for breeding line characterization and for informing yield comparisons among varieties. The main drawback of visually dating soybean maturity is the sheer scale of note recording entailed and the frequency at which these notes need to be taken. The overall aim of this study was to build upon prior work in using low-cost UAS-based RGB cameras to estimate soybean maturity date by examining the effect of vegetation index, summary statistic of the pixel values from each region of interest (plot), statistical model, and flight frequency. Maturity dates collected from five environments with 53 experimental trials (4,415 plots) were both visually dated and imaged using a RGB camera carried by a UAS. Using the mean greenness leaf index on each plot combined with LOESS regression, we achieved high correlations between ground and UAS-based estimates (r = 0.84–0.97). Precision, quantified by broad-sense heritability estimates, was greater for UAS-based dates in 29 of 53 field trials, and nearly equivalent in 11 more field trials. We found that 54% of the significant deviations between ground and UAS-based estimates were caused by inaccurate UAS-based estimates, while errors in the ground-based estimates accounted for 46% of the deviations. Reasons for these inaccurate estimates were attributed to lodging, presence of weeds, low germination, and within-line genetic heterogeneity in the plots. A detailed description of the analysis pipeline, a user-friendly R script, and all of the images and ground data have been made publicly available to help other researchers and breeders test and adopt these methods.