Figure 1.

Germinability and usability prediction scheme and results. (A) The experimental process consisted of five steps: (1.) Seed-by-seed capturing of lot samples by GeNee Detect, (2.) orderly sowing, (3.) germinability and usability phenotyping, (4.) classifier training with 90% of the image and phenotypic data and classifier testing with the 10% left-out data, and (5) determining classification performance via the percentile bootstrap method and extraction of the average precision, recall at 90% precision and precision at 80% recall measures. The scatter plots at 4. and 5. visualize lot 2603 germinability prediction outputs. A scatter plot is a 2D projection of the multi-dimensional seed variation assigned by the classifier. Each dot represents a single seed and the closer the dots are, the more similar the seeds are. Here, dot colors represent actual phenotypes. At 4., seed dispersal resulted in an apparent, although incomplete, spatial separation between the blue germinating and green non-germinating test set seeds. At 5., setting a 90% precision ratio (upper plot, bold seeds predicted as germinating) resulted in a corresponding 98% recall, with 2% of the germinating seeds mistakenly associated with the seeds predicted as non-germinating (lower plot, bold seeds). (B) Pie charts: all 32 germinability and 47 usability predictions achieved 90% precision, but with various recalls. 56% and 47% of the predictions, respectively (51% in total), were ascribed with high (≥ 80%) recall. Bars: high recall was achieved across a broad basal germinability or usability, as low as 59% and 57%, respectively (here both quantized to 60%). Quantized rates (%) encompass ± 5% values. Scatter plots were generated by Plotly.js version 2.5.1: https://github.com/plotly/plotly.js.