An accurate phenotyping of morphological traits of crops is important, but challenging task in large breeding programs. Field based phenotyping of complex traits is time-consuming, laborious, and highly subjective. Establishment of effective phenotyping methods is increasingly emerging reach areas. In this study, we validated 3D reconstruction through three-dimensional laser multispectral imaging system (CropScanner PS-100) to phenotype important morphological traits such as plant height, leaf length, leaf width, total leaf area, convex hull volume and biomass from booting to late grain filling. For this, a population of recombination inbred lines (RIL) of 269 lines derived from a cross of Zhongmai 578/Jimai 22 was used in glasshouse conditions. We compared the precision of 3D reconstruction-based point clouds which were extracted using the Multi-View Stereo and Structure from Motion (MVS-SFM) algorithm from a RGB-D camera and the CropScanner PS-100 with 3D laser lidar scanning with biomass. The results showed that the CropScanner PS-100 had high consistency compared with the phenotypes extracted from 3D reconstruction based on multi-view imagines. The line who has the higher chlorophyll content, chlorophyll fluorescence and convex hull volume in maturity also has the higher biomass. We suggest that our method can provide a basis for selection of wheat cultivars with high biomass.

Xm-30236

Can we teach the machine to select like the breeder?

Sebastian Michel1, Franziska Loeschenberger2, Christian Ametz2, Hebert Bistrich2, Hermann Buerstmayr1

1Department of Agrobiotechnology, IFA-Tulln, University of Natural Resources and Life Sciences Vienna, Tulln, Austria

2Saatzucht Donau GesmbH & CoKG, Austria

Plant breeding is considered to be the science and art of genetically improving plants according to human needs. Scientific breeding is thereby driven by big data from various omics disciplines, while the art of breeding is determined by breeder´s intuition and experience that leads to selection decisions by the so-called breeder´s eye. For the purpose of this study, we investigated the possibility of predicting both data-driven and knowledge-driven selection decisions with genome-wide distributed markers and various methods of conventional statistical learning, machine learning, and deep learning. The target trait was thereby the retrospective binary classification of selected versus non-selected crosses and progeny, which were part of six different cohorts in an applied wheat breeding program. Genomic predictions of this classification were conducted within each cohort as well as across cohorts by employing, amongst others, random forests, support vector machines, probabilistic neural networks, and convolutional neural networks, whereas the heritability of these selection decisions was estimated by Bayesian logistic models. A low to medium heritability was estimated for selection decisions made in preliminary yield trials (h² = 0.25-0.50), while the heritability was slightly higher for variety selection in multi-locations trials (h² = 0.40-0.50) as well as for choosing crosses in order to build new cohorts (h² = 0.30-0.55). The prediction accuracy of selection decisions was substantially higher when models were trained with lines from the same cohort (r = 0.30) in comparison to a forward prediction across cohorts (r = 0.15). Nevertheless, it should be noticed that the binary classification into selected versus non-selected crosses and progeny can be interpreted as a comprehensive breeder´s index, which comprises the entirety of data-driven and knowledge-driven inputs that are available for decision-making in a breeding program. Our results show that machine and deep learning methods cannot replace actual breeder´s decisions, but have large potential to aid breeders in the decision-making process during the selection of crosses and in the variety development process.