Research Associate University of Wisconsin, Madison Madison, Wisconsin
Body of Abstract: Preharvest sprouting (PHS) in quinoa panicles adversely affects yield. The sprouting of mature seeds on the panicle due to high humidity pre-harvest results in the appearance of thin radicles on the panicle that vary in color from pale white to deep red. PHS screening methods have been developed in major cereal crops, like barley and wheat, but differences in the plant architecture limit its use in quinoa. Our collaborators systematically acquired > 4000 standard RGB images of quinoa panicles ( > 200 genotypes) and hand scored them from 0-9 for PHS. Here, we describe the construction of a high-throughput automated scoring pipeline that utilizes this image collection to predict the scores representative of PHS severity, by statistically characterizing the color and texture features present in the images. Unsupervised clustering of those feature parameters followed by Dirichlet distribution parameter estimation predicts the PHS score. We use z-score normalized local curvature information for each color channel in the image as features, extracted for each pixel in the image. These features are then clustered into k=25 classes using k-means clustering. The class distributions are extracted for n x m image patches and principal component reduction (n=5) is used to remove covariance in the patch distributions. The most likely parameters found for the Dirichlet distribution describe the patch distributions using Nelder-Mead method that minimizes the negative log-of-likelihood. We used robust linear regression to predict the scores from these parameters and provide evidence that our method can successfully account for >80% of the variation present in the panicles. This high-throughput image analysis pipeline is faster than manual scoring and the high-throughput deployment expedites the analysis of new data and reanalysis of existing data. The algorithms developed for the pipeline are generally applicable, thus, increasing the usability in similar phenotypic measurements for other cereal crops.