10.1007/s00122-021-03988-8 VCF file containing SNP markers

Roots were digitally phenotyped as described by Brainard et al. (2021). In brief, after being cleaned, roots were QR coded and placed against either a white vinyl or black felt backdrop depending on root pigmentation. Images were acquired using a Nikon 5600 DSLR camera tethered to a computer running macOS 10.14. Python bindings for the OpenCV library were used to create binary masks of the roots by thresholding the hue-saturation-value color space. Custom MATLAB scripts were subsequently used to remove residual curvature in each root, and a random forest classifier was used to remove any unexpanded portion of the taproot. Python scripts for image acquisition and production of binary masks are available at: https://github.com/shbrainard/carrot-phenotyping; MATLAB scripts for straightening binary masks and performing PCA on contours or curvature values are available at: https://github.com/jbustamante35/carrotsweeper.

Following acquisition and pre-processing, phenotypes were extracted from the straightened, de-tipped binary masks. Root length was calculated as the distance from the center of the root crown to the root tip, following both straightening and elimination of the unexpanded, etiolated portion of the root. Maximum width was measured as the distance across the widest portion of the carrot, which is typically located just below the root crown. Total root size was defined as the 2D area of the entire binary mask. Aspect ratio was calculated as the ratio of length and maximum width. In addition, in order to quantify size-independent parameters of contour shape, PCA was performed on the root contour following a normalization procedure whereby each carrot was standardized to have a maximum width of 1, and a length of 1000. The scores along the first principal component quantify the degree of root fill, or how far down the length of the carrot the maximum width of the root is maintained; this trait accounts for over 80% of the variation in size-independent root shape. In addition, curvature values were computed at each point along the root contour in both the shoulder and tip regions as described by Driscoll et al. (2012). PCA of the first and last 50 elements of these curvature profiles was then performed; the first principal component of the former was used as a metric of shoulder broadness, while the first principal component of the latter was used as a measure of tip fill (a schematic workflow of this image acquisition pipeline is shown in Supp. Fig. 1). Together, this suite of root traits has been found to allow for accurate classification of roots, compared to a visual assignment of carrot market class (Brainard et al. 2021). In this study, two carrots were included in each raw image. With this workflow, the image acquisition phase required one minute per root, and one additional minute of computational time was required to perform pre-processing of the binary masks produced during acquisition, and phenotyping of these standardized images using a 3.3 GHz Intel Dual-Core i7 CPU and 16 GB of 2133 MHz LPDDR3 RAM.

Finally, prior to association analyses and construction of genomic prediction models, the diversity panel was restricted to those accessions that exhibited little to no branching of the taproot. Both the nature of the root-straightening algorithm—which depends upon identifying a carrot tip—and many of the phenotypes themselves (length, tip fill), implicitly require that the root be a single unbranched taproot. Small root hairs were removed through smoothing operations, and just as forked or split roots were discarded prior to image acquisition, accessions with highly branched fibrous root systems were also excluded on the basis of being inappropriate to an analysis of market class traits. Together with the failure of some roots to produce new leaf tissue following vernalization, this reduced the total size of the diversity panel used in subsequent analyses to 662 unique cultivated accessions.