An automated, high-throughput image analysis pipeline enables genetic studies of shoot and root morphology in carrot (Daucus carota L.)

Carrot is a globally important crop, yet efficient and accurate methods for quantifying its most important agronomic traits are lacking. To address this problem, we developed an automated analysis platform that extracts components of size and shape for carrot shoots and roots, which are necessary to advance carrot breeding and genetics. This method reliably measured variation in shoot size and shape, leaf number, petiole length, and petiole width as evidenced by high correlations with hundreds of manual measurements. Similarly, root length and biomass were accurately measured from the images. This platform quantified shoot and root shapes in terms of principal components, which do not have traditional, manually-measurable equivalents. We applied the pipeline in a study of a six-parent diallel population and an F2 mapping population consisting of 316 individuals. We found high levels of repeatability within a growing environment, with low to moderate repeatability across environments. We also observed co-localization of quantitative trait loci for shoot and root characteristics on chromosomes 1, 2, and 7, suggesting these traits are controlled by genetic linkage and/or pleiotropy. By increasing the number of individuals and phenotypes that can be reliably quantified, the development of a high-throughput image analysis pipeline to measure carrot shoot and root morphology will expand the scope and scale of breeding and genetic studies.


1
Introduction 28 Carrot is a globally important crop that originated in Central Asia Vavilov, 29 1992) with a secondary center of diversity in Asia Minor (Banga, 1957 Swanton et al., 2010), or the fact that the petioles must be sufficiently strong 37 for the root to be mechanically harvested (Rogers and Stevenson, 2006). Currently, a primary 38 breeding objective is to achieve rapidly growing, sturdy shoots without compromising the size and 39 shape of the storage root. Therefore, methods to measure both shoots and roots more objectively are 40 required (Horgan, 2001). These methods should be quantitative and objective, replacing traditional 41 subjective descriptors such as circular, obovate, obtriangular, and narrow oblong to describe the root 42 profile, or blunt, slightly pointed, and strongly pointed to describe the distal end (or tip) of the storage 43 root. Similarly, methods should characterize shoot architecture more comprehensively than typical 44 measurements of plant height, width, and biomass. 45 Image analysis has proven useful in describing several crop shoot systems while growing in 46 controlled environments, during the field season, and after harvest (Fahlgren et al., 2015;Furbank 47 and Tester, 2011; Lobet et al., 2013). Notably, a similar approach to characterizing carrot shoots must 48 accommodate some special issues. In contrast to many crops, carrots do not produce a shoot structure 49 by erecting a typical stem axis with leaves. Instead, an apical meristem at or beneath the soil 50 produces leaves attached by petioles to internodes that do not elongate during the vegetative phase of 51 the crop cycle. The petiole of each leaf, not the internode, elongates at an angle to lift and spread the 52 leaf blade. Thus, the cluster of petioles attached to the crown of the root is a major architectural 53 feature of the shoot structure that a phenotyping method must capture. 54 In addition to attributes of individual plant parts, allocation of resources between the shoot and root 55 of plants plays a central role in crop fitness and improvement (Lynch, 2007;Poorter et al., 2012). 56 Thus, a phenotyping platform for a root crop such as carrot should measure both shoot and root traits. 57 For instance, what may appear to be a practically helpful change in shoot architecture could 58 negatively impact light interception and therefore photosynthesis (Falster and Westoby, 2003), while 59 altered root structure could influence fibrous root architecture, which plays a critical role in water and 60 nutrient acquisition (Lynch, 1995 suggests potential genetic relationships, but the causal genetic loci, the extent of polygenic control, 75 and the influence of pleiotropy on shoot and root architecture in carrot have not yet been investigated. 76 For the reasons outlined above, carrot breeders are interested to measure carrot root and shoot 77 morphologies, preferably more objectively (Horgan, 2001). More precise and objective data on the 78 traits of interest will increase the ability to leverage genomic data and the potential for genetic gain in 79 breeding projects. Current limitations include the inability to measure some traits of interest and the 80 labor cost to collect hand measurements. These bottlenecks can be addressed using high-throughput 81 image analysis (Fahlgren et al., 2015;Furbank and Tester, 2011). Moreover, increasing precision and 82 sample size through automated image analysis will support practical breeding efforts by decreasing 83 experimental error, thereby improving estimates of heritability, facilitating the detection of causative 84 genetic loci, and expanding our understanding of quantitative inheritance (Kuijken et al., 2015). 85 Here we describe a relatively simple and low cost method to acquire 2D images of whole, excavated 86 carrot plants. This is coupled with a set of custom computer algorithms that quantify shoot 87 architectural features as well as the size and shape of storage roots. Research Station (Hancock, WI, USA). Figure S1 diagrams the sample size and sources of 104 individuals used for imaging and QTL mapping, which are described briefly below. 105 Diallel progenies were grown in a randomized complete block design (RCBD) with two replicates in 106 WI (2015) and CA (2016) (see Turner et al. 2018 for additional details). The F 2 population, L8708 x 107 Z020, was identified from prior field screening as segregating for plant height, shoot biomass, and 108 root storage shape and color. This population was derived from a cross between L8708, an orange 109 inbred line with a medium-long storage root and compact shoots, and Z020, a yellow, cultivated 110 landrace from Uzbekistan with a short, blunt-tipped storage root and broad, prostrate leaves. A single 111 F 1 plant was selected from this cross and selfed to produce the F 2 population used for mapping in this 112 study. F 2 plants were grown at the CA location in 2013 (n = 63) and 2016 (n = 450) and at the WI 113 location in 2016 (n = 77). Additional F 2 plants of the same cross, but derived from a different F 1 114 plant, were also grown at CA in 2016 (n=128) and were used only for validation of image 115 measurements. 116 A total of 1041 carrot plants were measured manually and photographed for the dual purpose of  118  developing an automated phenotyping method and determining the genetic architecture of important  119 traits. Hand measurements were recorded for shoot height (cm), root length (cm), leaf number, shoot 120 biomass (g), and root biomass (g). Unless otherwise specified, the term 'root' will refer to the storage 121 root in this report. Shoot height, measured as the distance from the crown to the tip of the longest 122 leaf, was recorded in the field for three plants per plot of each diallel entry and after harvest for each 123 F 2 individual. Root length was measured as the distance from the crown to the tip of the storage root, 124 defined here as having a diameter greater than 2 mm. Leaf number was recorded as the total number 125 of fully expanded, true leaves. Shoot biomass was sampled by removing all shoot tissue more than 4 126 cm above the crown. For root biomass, fresh weight was recorded for the entire root and for a 127 subsample, which was dried and extrapolated to estimate dry weight for the entire root. Fresh weights 128

Manual Measurements 117
were recorded immediately for both shoot and root tissues. For dry shoot and root weights, samples 129 were dried at 60°C in a forced-draft oven and values were recorded after reaching constant mass. 130 Ground truth data for digital measurements of petiole length and diameter was recorded for a subset 131 of 100 images using ImageJ (Schneider et al., 2012). 132 Otsu threshold method was applied to produce a binary image (MASK) in which pixels belonging to 150 the carrot object were white (1) and background pixels were black (0). Based on the location of the 151 horizontal black line on the baseboard, images were split into shoot and root sections for 152 corresponding morphometric analyses. 153

Computational Workflow 154
As described by Miller et al. (2017), a high-throughput computational workflow was implemented 155 using a community cyberinfrastructure, which is publicly available as a software tool through the 156 CyVerse Discovery Environment web interface (Figure 1). Briefly, image files were uploaded to the integrated rule-oriented data store system (iRODS) (Rajasekar et al., 2010) managed by CyVerse 158 (Merchant et al., 2016) (Figure 1). Each image was processed as a separate computational job using 159 parallel computing enabled by the University of Wisconsin's Center for High-Throughput 160 Computing. Scheduling, resource matching, execution of analyses, and return of results was managed 161 by the HTCondor software (Thain et al., 2005). Results were then returned to the data store holding 162 the original images ( Figure 1A). 163

Image Analysis 164
All images were processed through a two-stage workflow ( Figure 1B) and data was returned as both 165 individual CSV files for each measurement and as an indexable JavaScript Object Notation (JSON) 166 file containing all measurements. For the shoot, root, and whole carrot masks, data output included 167 classic image measurements of a bounding box (used to measure shoot height, root length, and root 168 width), convex hull, eccentricity, equivalent diameter, Euler number, perimeter, and solidity. 169 Measurements of interest included shoot and root biomass profiles, petiole width, petiole number, 170 and petiole length, which are described in detail below. File names, measurements, and data structure 171 are described in Table S1. shoot with overlays of the half elliptical grid and computed biomass profile. The SBP determined in 187 this way formed the basis for subsequent shoot trait extraction methods. 188

Petiole Characteristics 189
To estimate petiole width, a Euclidean distance transformation (EDT) was applied over the entire 190 binary shoot image. The EDT labels each pixel in the plant mask with a value equal to the distance to 191 the nearest contour pixel. Next, the image was skeletonized. The EDT value at each skeleton point 192 was sampled to produce a distribution of values corresponding to each pixel in the mask. This 193 distribution was used as the input for the prediction step using partial least squares (PLS) regression 194 (Wold, 1982;Wold et al., 1984) against the ground truth values from ImageJ. The number of 195 components to retain in the PLS model was assessed using cross-validation with a one-fold holdout. 196 To predict the number of petioles in an image, the digital shoot biomass (i.e. the sum of white pixels 197 in the binary shoot image) was divided by the algorithm-measured petiole width. This was performed for every image of a shoot. The resulting ratio of total mass divided by average petiole width value 199 was the input for PLS regression against the true counts, which were collected by hand at the time the 200 image was acquired. The number of components to retain in the PLS model was assessed using 201 cross-validation with a one-fold holdout. 202 To predict petiole length, the SBP was subjected to principal components analysis. The principal 203 components extracted from the SBP and the ground truth values for petiole length, which were 204 collected from 100 images in ImageJ, were used to train a two-layer feed forward neural network 205 (Bhandarkar et al., 1996). The prediction step was also performed with PLS regression as was done 206 for the petiole number. In this case, the neural network method provided higher correlations than PLS 207 regression. Vectors for petiole counts, width, and length were returned to the data store for 208 subsequent analyses. 209

Root Shape 210
A root biomass profile was generated by recording the number of white pixels along each horizontal 211 sweep, which was returned as a 1000-dimensional vector ( Figure 2B). To focus exclusively on shape 212 differences, the root biomass profile was normalized by both length and width prior to principal 213 components analysis, which was used to examine symmetrical shape variance. The binarized root 214 image with the root outline in green was also returned to the data store for error checking. 215

Correlations and Repeatability 216
All downstream analyses were performed in R 3.3.2 (R Core Team, 2016). Pearson's correlation 217 coefficients (r) and Spearman's rho (ρ) were used to compare manual-and image-measured traits. 218 For manual-measured and digital biomass, correlations were estimated using a linear log-log 219 relationship, following established guidelines for allometric models of biomass partitioning in carrot 220 (Hole et al., 1983) and in seed plants (Enquist and Niklas, 2002). When possible, algorithm-measured 221 values were converted from pixels to centimeters using reference points of known size on the 222 baseboard. 223 Repeatability, which describes the proportion of trait variance attributable to differences among 224 rather than within individuals, was calculated using observations for 336 individual plants 225 representing 42 crosses from a six-parent diallel mating design. Variance components were assessed 226 using the linear mixed-effects model 012 = + 0 + 1 + 2 1 + 01 + 012 , where 012 is the 227 phenotype, 0 is the effect of genotype, 1 is the effect of environment, 2(1) is the effect of 228 replication k within environment j, 01 is the interaction between genotype i and environment j, and 229 012 is the residual error. Repeatability was estimated on an entry-mean basis as . 233

DNA Extraction and Quantification 234
Following image capture, a 1.5 g leaf sample (fresh weight) was collected from each F 2 plant. Total 235 genomic DNA was isolated from ~20 mg of lyophilized leaf tissue using the CTAB method of 236 Murray and Thompson (1980) with modifications by Boiteux et al. (1999). DNA quality was 237 assessed visually using 1% agarose gel electrophoresis and double-stranded DNA was quantified 238 using the Quant-iT™ PicoGreen® dsDNA assay kit (Life Technologies, Grand Island, NY, USA). 239 Concentrations were normalized to 10 ng/µl. 240 genotypes which deviated from expected segregation ratios based on a Chi-square test (P < 0.001) 254

Genotyping-by-Sequencing (GBS)
were excluded. All linkage groups were obtained at a LOD threshold greater than 10. The regression 255 mapping algorithm was used with Kosambi's mapping function to calculate the distance between 256 markers (Kosambi, 1943 Results 281

Image analysis 282
For the 1041 images submitted through the analysis pipeline, 917 (88%) ran successfully and 283 returned data. Of the 124 images that failed, two were also missing hand measurements, eight had 284 root defects such as sprangle (i.e. branching of the root), 60 had poor lighting or shadowing, eight 285 overlapped with the edge of the image or the black line separating the shoot and root, and 46 failed 286 for reasons which were not readily identifiable, with possible explanations including the presence of 287 numerous fibrous roots, interference of labels, and/or diminutive plant size. 288

Correlations between hand and algorithm measurements 289
Overall, traits extracted automatically from images had strong and significant (P<0.001) correlations 290 with their manually measured analogs, ranging from r = 0.77 for leaf number to r = 0.93 for root 291 biomass. Relationships among manual-and image-measured values for shoot height, shoot biomass, 292 root length, and root biomass are detailed in Figure 3. Shoot height and root length each had 293 correlations of r = 0.88 between manual and image measurements, with larger correlations observed 294 for shoot biomass and shoot area (r = 0.91) and between root biomass and root area (r = 0.93). 295 Notably, correlations ranged from low to moderate when comparing shoot to root attributes, such as 296 shoot height and root length (r = 0.18), and the correlation between shoot and root biomass deviated 297 from unity for both manual measurements (r = 0.72) and for algorithm values (r = 0.62). 298 Similarly, Figure 4 presents the strong correlations between manual measurements and algorithm 299 predictions for petiole attributes, with manual measurements of petiole length and width based on 300 ground truth data from images. additional measurements, are provided in Figure S2. 306

Principal components analysis of shoot biomass and root shape 307
For shoot biomass profiles, principal components analysis identified differences in the magnitude and 308 location of biomass ( Figure 5). The first two principal components accounted for 80.3 percent of the 309 variation explained (PVE). Sweeping PC1 detected differences in overall biomass accumulation 310 (43.7 PVE), which is likely a combination of increases in both leaf number and total leaf area. 311 Sweeping PC2 corresponded to decreasing petiole length and overall height (36.6 PVE), capturing 312 variation for shoot compactness. 313 To identify symmetrical differences in root shape, root biomass profiles were rescaled to constant 314 length and width prior to principal components analysis. Principal components detected differences 315 in the contour of the roots, with the first three principal components accounting for 88.6 PVE ( Figure  316 6). Changes in PC1 corresponded to differences in overall shape (conical vs. cylindrical; 66.4 PVE). 317 Variation in PC2 was associated with the shape of the root tip from a tapered shape to a blunt, 318 rounded shape (16.6 PVE). For PC3, changes corresponded to diameter in the longitudinal section 319 (5.6 PVE). 320 Results differed slightly from findings using landmark analysis by (Horgan, 2001) correct for aspect ratio (i.e. the ratio of width to height), which allowed us to explain more variation 325 in shape independent of root length and width. Disparities may also result from differences in 326 measurement technique and in the range of root shapes represented in each study. Interestingly, our 327 results are also similar to findings in Japanese radish (Iwata et al., 1998), which identified principal 328 components for aspect ratio (73.9 PVE), bluntness at the distal end of the root (14.2 PVE), and 329 swelling in the middle of the root (3.9 PVE). 330

Repeatability 331
Estimates of repeatability were moderate for most traits, ranging from low (e.g. root length) to high 332 (e.g. shoot height) and were comparable between manual and image measurements ( Table 1, Table  333 2 Repeatability for root traits ranged from 0.01 for manual measurements of root length to 0.32 for 340 manually measured root biomass, with a value of 0 observed for root PC2 ( Table 2). Observations of 341 low repeatability for root length and shape characteristics may be due to low phenotypic variation 342 among the inbred parents, which were primarily selected for divergent shoot characteristics, and/or 343 genotype by environment interaction (GxE). As observed for shoot traits, estimates of repeatability 344 were generally higher within environments, supporting the importance of GxE for these phenotypes. 345 Compared to manual measurements, image derived values successfully identified the lowest ranking 346 line for shoot height (L6038), shoot biomass (L6038), and root biomass (B7262) ( Table 1 and Table  347 2). Discrepancies between manual and image measurements, for instance between the highest line for 348 shoot height based on manual measurements (Nbh2189A x B7262B) and based on image 349 measurements (Nbh2189A x P6139B), may be due to differences in how the measurements were 350 obtained (e.g. measured at the plot level in the field or for individual plants) and due the prevalence 351 of missing observations in the WI2015 season. 352

Genotyping and genetic linkage map construction 353
A total of 116,030 SNPs were identified for 467 individuals. After filtering for missing data and 354 allele frequency, the final data set contained 15,659 high quality SNPs. The linkage map was 355 constructed using 461 individuals and included a total of 640 high quality SNP markers across nine 356 chromosomes ( Figure S3). The total distance covered was 719 cM with an average marker spacing 357 of 1.1 cM and a maximum marker spacing of 17.7 cM (Table S2). 358

QTL for shoot and root traits 359
Overall, seven significant QTL on chromosomes 1, 2, 3, 4, 5, and 7 were identified for manual 360 measurements of carrot shoot and root traits. Of these, six QTL were also detected for traits extracted 361 computationally from images (Figure 7). Additionally, the use of image based measurements resulted in the identification of two additional QTL for root PC1 and petiole width on chromosomes 363 6 and 8, respectively. Significant QTL, including the most significant marker and corresponding 1.5 364 LOD interval, are described in detail for shoot traits in Table 3 and for root traits in shoot biomass included regions on chromosomes 3 (6 PVE) and 4 (5 PVE), of which only the region 384 on chromosome 3 was found for the image-extracted trait (4 PVE). This same region on 385 chromosome 3 was also identified for petiole length (3 PVE) and for shoot PC2 (5 PVE). For the 386 image measurement of petiole width, two QTL, which were not identified for any hand 387 measurements, were found on chromosomes 4 (5 PVE) and 8 (6 PVE). Despite strong correlation of 388 shoot PC1 with shoot biomass, no QTL were identified for shoot PC1. 389 Root traits: In contrast to the region on chromosome 7 described above, a QTL on the proximal end 390 of chromosome 7 was identified for manually measured root length (4 PVE), but not for the 391 corresponding image measurement. Two other QTL for root length were identified on chromosomes 392 1 and 3 for both manual (9 PVE and 6 PVE, respectively) and image (14 PVE and 7 PVE) 393 measurements. The same QTL on chromosome 3, which was also identified for shoot biomass and 394 petiole length, was detected for root PC2. For image-based measurements of root length and 395 biomass, another QTL was also identified on chromosome 4 (10 PVE and 4 PVE, respectively). To facilitate crop improvement efforts in carrot, we present a pipeline to assess whole-plant 405 morphology, which to date has lacked protocols for standardized, quantitative measurements. This 406 method will enable more in-depth genetic and phenotypic studies in carrot by providing: (1)  and including a marker of known size during imaging to automatically convert pixels to metric units. 426 The high correlation between image-extracted traits and hand-measured analogs (r >0.7) provides 427 evidence that this is a reliable method to capture phenotypic diversity and quantitative trait variation 428 for important breeding targets in carrot. By enabling precise measurements for larger population 429 sizes, the power of subsequent genetic investigations will be improved to enable more precise 430 estimates of heritability and ultimately to better inform breeding strategies to increase genetic gain 431 (Fiorani and Schurr, 2013;Kuijken et al., 2015). Additionally, a distinct advantage of this approach is 432 the ability to measure shape parameters, which do not have an objective or practical hand 433 measurement equivalent. Previous work on carrot shoot morphology includes image analysis of 434 leaflet shape  and an assessment of phenotypic and genotypic diversity for shoot 435 height in commercially available carrot germplasm (Luby et al., 2016). However, this is the first 436 method to implement a high-throughput, quantitative assessment of carrot shoot architecture. The geometric criteria (Koszela et al., 2013). The scope of these approaches was restricted to assessing 445 varietal and quality differences in root shape, independent of haulm characteristics, and was limited 446 to commercially available varieties. We build upon these methods by characterizing root shape 447 without landmarks , expanding the methodology to capture shoot architecture, 448 and demonstrating the detection of subtle but biologically important variation in diverse genetic 449 resource populations. Deviations from previous reports of principal components for carrot root shape 450 can be partly explained by the decision to normalize for root length and width (i.e. aspect ratio), a 451 step which can be omitted if aspect ratio is a trait of interest. It is also worth noting that the scope of 452 our approach could be improved with the inclusion of additional root classes, such as Paris Market 453 and Kuroda types (Simon et al., 2008). 454

Identification of QTL for shoot and root characteristics 455
Vegetative plant organs often evolve as phenotypic modules, and consequently tend to be highly 456 correlated and share evolutionary tracts (Bouchet et al., 2017) We report evidence for the co-localization of QTL for shoot traits (height, leaf number, biomass, 474 petiole width, and petiole length) and root characteristics (length, biomass, and tip fill) on the distal 475 end for the long arm of chromosome 2. This suggests a pleiotropic basis and/or tight genetic linkage 476 for the morphological integration of shoot and root architecture in carrot. This finding is also 477 consistent with the recent identification of a QTL and selective sweep on a nearby region of 478 chromosome 2, which included the identification of a candidate domestication gene in carrot . Interestingly, in this study we also find a member of 486 the AHL gene family within the confidence interval for the QTL identified on chromosome 2 ( Table  487 S3). While our findings support evidence that the region on chromosome 2 is important for carrot 488 growth and development, they differ from the findings of Macko-Podgórni et al. in two important 489 ways: (1) we did not observe overlap between the support intervals of significant QTL on 490 chromosome 2 in this study and the DcAHLc1 gene and (2) we did not find any significant QTL for 491 image-based measurements of root width, although we did observe a significant QTL for root PC2, 492 which captures variation in the amount of tapering (or swelling) at the tip of the root. A likely 493 explanation for not finding the DcAHLc1 gene to contribute to root shape in our study, which used a 494 cross between domesticated breeding stocks, is that Macko-Podgórni et al. (2017) used a wild x 495 domesticated cross (D. carota subsp. commutatus x 2874B), in which the DcAHLc1 gene is 496 segregating. Together, these findings suggest the possibility of additional candidate gene (s) on  497  chromosome 2 and tight linkage among genes influencing carrot shoot and root development, which  498 are inherited together as a suite of traits. 499 By providing a foundation for future genetic mapping and genome-wide association studies, the 500 significant QTL detected in this study will contribute to the development of marker-assisted selection 501 and fine mapping efforts for carrot shoot and root morphology. Further research will be necessary to 502 validate the prevalence and importance these regions in different genetic backgrounds, over the 503 course of developmental stages, and across environments. 504

Conclusions and future directions 505
The development of a high-throughput image analysis pipeline for carrot shoot and root morphology 506 provides new opportunities for crop improvement and to elucidate the underlying genetics for 507 quantitative traits. The design for image collection is simple, low-cost, and could be easily adapted 508 for use in other crops with similar morphology. Ideally, this methodology could be expanded to other 509 important crops, e.g. cassava, beet, radish, and other members of the Apiaceae family, such as celery, 510 parsnip, parsley, and cilantro, which have widespread culinary uses but lack substantial research 511 investment. Images are also an ideal medium to facilitate collaborations, as they transfer 512 multidimensional information for which analysis is standardized and automated (Lobet et al., 2013). 513 As such, the ability to analyze and share carrot images through public repositories is an opportunity 514 to increase the scope, archival, and reproducibility of carrot research. 515 Data from this method can be used in numerous applications for carrot breeding and research. 516 Morphological variation can be rapidly assessed and catalogued for diverse genetic backgrounds, 517 providing a resource to better inform experimental design and population selection for more in-depth 518 analysis. This pipeline can be used in tandem with physiological studies, for instance to evaluate the 519 effects of gibberellic acid and cytokinin, which are known to influence carrot shoot and root 520 morphology (Wang et al., 2015b(Wang et al., , 2015a. Phenotypic data can also be integrated into predictive 521 models for carrot growth and development by imaging plants at various developmental stages, 522 permitting further investigation of allometric relationships between the shoot and root. In future 523 studies, it will also be important to consider the relationship between fibrous root architecture, which 524 provides a source of photosynthates, water, and soil-borne nutrients, and the storage root, which 525 serves as a sink for these metabolites that are essential for vegetative and reproductive growth. 526 This approach is specifically tailored for a carrot breeding program, but could also complement 527 existing image analysis software and methods for detailed analyses. For example, research on the 528 genetic basis of lateral branching in carrot roots is underway using RootNav (Pound et al., 2013) and 529 SmartRoot (Lobet et al., 2011), which are well established methodologies to quantify root system 530 architecture. Potential improvements and expansions of our method include incorporation of uniform 531 lighting and a marker of known size, as well as extension of carrot phenotyping to field-scale 532 measurements over the course of the growing season. 533 The method presented in this study provides an initial step in automated phenotyping for carrot. By 534 enabling rapid, precise measurements of important agronomic characteristics in carrot, this platform 535 will allow carrot breeders to measure greater population sizes, increasing throughput and supporting 536 downstream analyses. 537 to Charlene Grahn and Julie Dawson for helpful advice on the project, and to Rob Kane (deceased) 564 and Tom Horejsi for technical support and field management.   analogous to manual measurements (middle), and traits that were only measured from images 860 (bottom). Arrows designate QTL that were identified by image measurements but not by manual 861 measurements. Horizontal lines indicate the significant LOD thresholds for P<0.05 (solid) and 862 P<0.01 (dashed). 863