See discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/328347226

A Low-Cost Automated System for High-Throughput Phenotyping of Single Oat

Seeds

Article · December 2018

DOI: 10.2135/tppj2018.07.0005

CITATIONS

0READS

596

10 authors, including:

Some of the authors of this publication are also working on these related projects:

Genomic - Environmental error evaluation View project

Breeding Value of Primary Synthetic Wheat Genotypes for Grain Yield View project

James Clohessy

University of Florida

3 PUBLICATIONS   0 CITATIONS   

SEE PROFILE

William Duke Pauli

The University of Arizona

40 PUBLICATIONS   81 CITATIONS   

SEE PROFILE

Edward S Buckler V

Stony Brook University

2 PUBLICATIONS   0 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Michael A. Gore on 08 December 2018.

The user has requested enhancement of the downloaded file.

Page 1 of 13

The Plant Phenome JournalOriginal Research

Core Ideas

• A single-seed analyzer system was modified to provide low-cost seed imaging.

• The system throughput allows rapid, nondestruc-tive phenotyping of single seeds.

• Accuracy of seed shape and size measurements were similar to those obtained manually.

• Repeatability of morphometric seed traits was higher than that of seed color traits.

Copyright © American Society of Agronomy and Crop Sci-ence Society of America. 5585 Guilford Rd., Madison, WI 53711 USA. This is an open access article distributed under the terms of the CC BY-NC-ND license (http://creativecom-mons.org/licenses/by-nc-nd/4.0/)Plant Phenome J. 1:180005 (2018)doi:10.2135/tppj2018.07.0005

Received 18 July 2018. Accepted 9 Oct. 2018. *Corresponding author (jameswclohessy@gmail.com; mag87@cornell.edu)

Abbreviations: 2-D, two-dimensional; BLUE, best linear unbiased estimator; CART, classification and regression tree; CIELAB, Commission Internationale de l’Eclairage L*a*b*; NIR, near-infrared reflectance; RGB, red, green, and blue; SSA, single-seed analyzer.

J.W. Clohessy, D. Pauli, K.M. Kreher, E.S. Buckler V, M.E. Sorrells, and M.A. Gore, Plant Breeding and Genetics Section, School of Integrative Plant Science, Cornell Univ., Ithaca, NY 14853; P.R. Armstrong, USDA–ARS, Center for Grain and Animal Health Research, Man-hattan, KS 66502; T. Wu, Northwest A&F Univ., No. 3 Taicheng Road, Yangling, Shaanxi 712100 China; O.A. Hoekenga, Cayuga Genetics Consulting Group LLC, Ithaca, NY 14850; J.-L. Jannink, USDA–ARS, Robert W. Holley Center for Agriculture and Health, Plant Bree-ding and Genetics Section, School of Integrative Plant Science, Cornell Univ., Ithaca, NY 14853. J.W. Clohessy, current address, North Florida Research and Education Center, Univ. of Florida, 155 Research Road, Quincy, FL 32351; D. Pauli, current address, School of Plant Sciences, 541A Marley Hall, Univ. of Arizona, Tucson, AZ 85721.

1 J.W. Clohessy and D. Pauli contributed equally to this work.

Efforts focused on the genetic improvement of seed morphometric and color traits would greatly benefit from efficient and reliable quantitative phenotypic assessment in a nondestructive manner. Although several seed phenotyping systems exist, none of them combine the cost effectiveness, identity preservation, throughput, and accuracy needed for implementation in plant breeding. We integrated an image analysis compo-nent into a single-seed analyzer (SSA) system that also captures near-infrared reflectance (NIR) and weight data. Through the development and utilization of an open-source computational image analysis pipeline, image data acquired by two cameras mounted on the SSA machine were automatically processed to derive estimates of length, width, height, volume, and color for 96 individual dehulled groats (seeds) of five oat (Avena sativa L.) genotypes replicated across days. With the exception of color, the four traits were found to be strongly correlated with and have repeatability comparable to manual measurements. The seed color values had moderately strong correlation with those measured by a colorimeter, but further improvements to the SSA system are needed to increase measurement accuracy. These results demonstrate that the SSA system has the potential to provide a low-cost solution for the rapid, accurate measurement of mor-phological traits on individual seeds of oat and potentially other crop species, allowing the screening of seeds from numerous genotypes in breeding programs.

Cereal yield, measured as the amount of dry grain over a harvested area (kg/ha), has a complex genetic architecture consisting of numerous small-effect loci that interact with the environment. The grain (or seed) yield of small grain cereals,

including oat, common wheat (Triticum aestivum L.), and barley (Hordeum vulgare L.), can be separated into three primary components that are typically more heritable and genetically tractable than yield itself: individual seed weight, seed number per spike, and spikes or reproductive tillers on a per area basis (Rustgi et al., 2013). Of these three components, the weight of an individual seed is largely determined by both the rate and duration of seed fill that occurs from anthesis to maturity. The accumulation of water and dry matter (e.g., carbohydrates, oil, protein, etc.) throughout this developmental period results in changes to seed length, width, height, and volume, all of which contrib-ute to defining the final dry weight of an individual seed (Xie et al., 2015).

In oat, a single grain consists of a groat (caryopsis) and a hull, with the hull removed by a mechanical dehuller during the milling process for human consumption. The size and shape of the groat is important because it affects the maximum size of oat flakes, milling yield, and test weight (Doehlert et al., 2006; Groh et al., 2001). Additionally, low groat weight has been associated with high oil concentration (Peterson and Wood, 1997). Oat in particular is consumed as a whole grain, and consumers have shown preference for groats based on color (Johnson et al., 1997). Furthermore, constituents such as lignin and suberin may be surface enriched in the oat groat, affecting both color and quality (Miller

A Low-Cost Automated System for High-Throughput Phenotyping of Single Oat Seeds

James W. Clohessy,1* Duke Pauli,1 Kevin M. Kreher, Edward S. Buckler V, Paul R. Armstrong, Tingting Wu, Owen A. Hoekenga, Jean-Luc Jannink, Mark E. Sorrells, and Michael A. Gore*

Published December 6, 2018

Page 2 of 13

and Fulcher, 2016). To improve via direct selection on these traits, oat breeding programs need a phenotyping platform that provides accurate and precise quantitative measures of single dehulled groats (hereafter seeds) in a cost-effective, high-throughput manner.

While it is possible to manually measure the dimensions and weight of a single seed using a digital caliper and analytical laboratory balance, there are restrictions on the number of samples that can be quickly and cheaply processed and the accuracy to which two- and three-dimensional parameters can be estimated. To overcome these limitations, past efforts have focused on the use of digital image analysis for the phenotyping of individual seeds, but these methods are not without weaknesses. Elliptic Fourier analysis implemented in the software SHAPE (Iwata and Ukai, 2002) enables the two-dimensional (2D) closed contour shape variation of seeds to be described by Fourier series expansions (i.e., a series of curves that approximate the curve under consideration) through the generation of elliptic Fourier descriptors. This image-based analysis approach allowed the genetic architecture of rice (Oryza sativa L.) grain shape to be studied, which provided novel insight into the genetics responsible for trait variation (Iwata et al., 2010). However, the sampled seeds had to be carefully oriented for imaging in a manual fashion, and principal components of the coefficients of the elliptic Fourier descriptors rather than physical measurements (e.g., length, width, etc.) were used as morphological phenotypes (Iwata et al., 2010; Williams et al., 2013). More recently, the development of software methods for the analysis of images acquired by consumer-grade flatbed scanners has enabled the measurement of various 2D seed shape parameters and color (Moore et al., 2013; Tanabata et al., 2012; Whan et al., 2014) but has lacked the additional dimension needed to measure seed height.

Instruments constructed to enable nondestructive phenotyping of individual seeds from a range of small- to large-seeded plant species have potential to overcome the drawbacks of early generation digital image analysis systems. As an example, Spielbauer et al. (2009) constructed an instrument that collected near-infrared reflectance (NIR) and weight data on single maize (Zea mays L.) seeds to predict grain quality traits such as protein and starch content. This instrument served as the basis for building a single-seed analyzer (SSA) machine to collect the same data types from individual oat seeds (Montilla-Bascón et al., 2017). However, these and other related high-throughput systems (Melchinger et al., 2018, Rolletschek et al., 2015) for evaluating end-use quality traits do not have the capability to measure seed morphometric characteristics such as length, width, and height. Illustrative of the advantages derived from integrating laboratory automation and digital image analysis, the phenoSeeder system constructed by Jahnke et al. (2016) used a volume-carving method (Roussel et al., 2016) to reconstruct the three-dimensional shape of individual seeds. Although the system had high reproducibility for the measured seed morphological traits such as volume and mass, it was hindered by a less than optimal throughput and complexity of the robotics needed to manipulate individual seeds.

In an effort to address the individual seed phenotyping needs of oat breeding programs, we incorporated an imaging component into the existing low-cost SSA machine of Montilla-Bascón et al. (2017) that already had NIR spectroscopy and weighing capabilities. The objectives of our study were to: (i) develop an open-source computational pipeline for oat seed image analysis, (ii) quantify morphological and color phenotypes of individual seeds from five oat genotypes and compare the results with manual measurements, and (iii) estimate the number of oat seeds that need to be phenotyped to provide high trait repeatability and accurate estimation of trait parameters.

�Materials and MethodsOat Samples for Measurement

In 2015, a set of five different oat genotypes (Ogle, Newdak, Corral, Hidalgo, and SD081107) was evaluated at Cornell University’s Snyder Farm in Ithaca, NY. Conventional oat culti-vation practices for the northeastern United States were used. The plots were mechanically harvested. After harvesting, grain samples were placed in a dryer at 21°C to remove excess moisture and dehu-lled using a mechanical dehuller (Codema LLC). The visual color differences observed among the dehulled seeds from the five oat genotypes spanned a narrow range of hue, saturation, and light-ness values for the color brown. Only unbroken and healthy seeds were evaluated using the SSA. For each genotype, 96 seeds were randomly sampled and individually placed into two 48-well plates for replicated measurements with the SSA.

Single-Seed Analyzer Design, Function, and ControlThe SSA provides NIR spectroscopy paired with weight

measurement and imaging to nondestructively phenotype single seeds rapidly in a batch process that preserves the identity of the seeds by placing them in the same well position of a second receiving plate (Supplemental Fig. S1). This design of the SSA (Supplemental Fig. S2) represents a combination and improvement of two previously constructed seed analysis devices (Spielbauer et al., 2009; Ziegler et al., 2013). The NIR component of the SSA was not a focus of this study but has been previously demonstrated for individual oat seeds by Montilla-Bascón et al. (2017). In this study, individual seeds were processed through the SSA for weight measurement and imaging using gravity-enhanced compressed air and a series of V-shaped trough seed slides.

The controlling software for the SSA used Visual C++ based on Visual Studio 2010 and a SQL Server 2008 database (Microsoft). Multithreading, consisting of five threads, was used for collecting data or controlling each instrument component. These were a primary thread and an X-Y table, NIR spectrometer, analytical balance, and camera thread. The primary thread provided a user interface and interacted with the other four threads that served as subordinate threads. Each subordinate thread undertook separate tasks related to different hardware or sensor requirements as indicated by their names.

Page 3 of 13

Seed Plate Loading and Seed Weight Data CollectionThe SSA was used to analyze 96 or 192 seeds in a single ses-

sion, as seeds placed into two or four 48-well plates were stored on the SSA. A total of three sessions was required to analyze the 10 48-well plates each day. In all, the complete set of 10 plates was analyzed on six different days, for a total of six repeated measure-ments per seed. The run order of the 10 plates was randomized across the three sessions on each day. For each session, two or four 48-well plates were inverted and loaded onto an indexing X-Y table that was used to deliver individual seeds sequentially into the SSA. Single seeds traveled through a glass tube and then dropped onto an analytical balance (GH-200, A&D Company Limited) to obtain a weight measurement. Once a stable seed weight was recorded, a burst of compressed air ejected the seed into a funnel that guided the seed onto a slide and then onto a rotary imaging table.

Seed Image Data CollectionThe SSA used a rotary imaging table (Supplemental Fig. S3)

that rotated the seed to the web camera (Model C615, Logitech) imaging position, enabling the acquisition of top- and side-view seed images and eventual deposition of the seed in the same posi-tion of a second 48-well plate so as to retain seeds in their original order (Supplemental Fig. S2). To block any reflections and non-seed objects from the side and top cameras, top (4.5 by 3 cm) and side (4.9 by 2.7 cm) backgrounds made from matte black construction paper (Mi-teintes 160 GSM no. 425 Stygian Black, Canson) were attached to the rotary imaging table. This allowed a consistent image background at any location where a seed could be imaged on the table, but with a tradeoff of eliminating the potential imaging benefit of a bottom-view camera.

We provided uniform lighting across the seed by construct-ing a lighting system with multiple spectra and variable intensity. Briefly, an aluminum shroud with a diameter of 9 cm and a 1-cm-diameter hole in the center was affixed to each of the web cameras, serving as a mount for strips of light-emitting diode (LED) lights. Each strip consisted of six LED lights. Two sets of three strips were placed around the lens of the top camera, and three strips were placed on the top half of the side camera. The LED strip lights consisted of three light colors: 4000K neutral white (Model 2026NW-40K, Qurosy), 3100K warm white (Model 2026WW-31K, Qurosy), and 395- to 405-nm ultraviolet (UV, Model 5050, Joygo). All of the same color LEDs were on a separate circuit, with each circuit having a potentiometer dimmer (Model T90040, Triangle Bulbs) to control the intensity of light. The UV lights were disabled and intended to detect phenotypes beyond those included in this study.

The warm and neutral white light potentiometers were set to 534 and 557 W of resistance, respectively, for the LEDs to produce a full color gradient on the seeds without overexposure. For the top camera, white balance and exposure values were set manually using the Logitech driver software and images of a standard gray card (Model 24 ColorCard, CameraTrax). Similarly, the focus was manually set for both web cameras using settings in the Logitech driver software and images of 8-point Times New Roman printed

text. The web cameras were instructed to take an image by the SSA control software using the OpenCV Visual C++ library (www.opencv.org). Each webcam provided a 1920 by 1080 single bitmap image that was received by the OpenCV software and saved to a hard-drive directory.

In total, 5760 bitmap images were generated, comprising 2880 sets of top and side images (Supplemental Table S1). To improve portability, all bitmap images were converted to PNG-24 format. The machine technician recorded if any seeds were lost or potentially reordered, resulting in the removal of 90 image sets prior to image processing. The raw image data are available on written request.

Image Color CorrectionThe process to estimate seed length, width, height, volume,

and color from images required a multistep computational pipeline (Supplemental Fig. S4). The web camera software did not allow the application of separate image settings to different web cameras connected to the same computer, thus the color accuracy of the top camera was prioritized over the side camera. Briefly, we initiated color correction of all collected image data using a color reference card (Model 24 ColorCard) imaged by the top camera at the start of each of the 6 d. Each image of the color reference card was analyzed using GNU Image Manipulation Program (GIMP) (www.gimp.org) by selecting all of the pixels contained in the gray color box on the imaged reference card. With the mean red, green, and blue values of the selected area, individual channel offsets were deter-mined by calculating the difference between these mean values and values expected for the reference (red = 162, green = 163, blue = 162). These calculated daily offsets of a few pixels were applied to both top (i.e., color and light fluctuation correction) and side (i.e., only light fluctuation correction) images using a color offset func-tion in IrfanView (www.irfanview.com) (Supplemental Fig. S5).

Top Image Cropping and Seed SegmentationThere was inherent variability in where seeds were positioned

on the matte black background affixed to the rotary imaging table such that not every imaged seed was centered relative to the top web camera. Therefore, image cropping was needed to reduce extrane-ous background area that otherwise significantly increased image processing time. Through implementation of an image prepro-cessing Python script (sapre-processing.PY; https://github.com/GoreLab/SSA_Image_Processor) that utilized OpenCV and Pillow modules, all top camera images were cropped at the extreme top and left sides. Removal of pixels from the right and bottom sides would also be preferred for processing speed, but the variability in seed positions necessitated the inclusion of these sides.

To detect and segment seed pixels from the background area of all top camera images, the EasyPPC software (Guo et al., 2013) used a classification and regression tree (CART) model to differ-entiate seed from background (Supplemental Fig. S6). The CART model was trained using a random selection of 33 color-corrected and six non-color-corrected top camera images. The addition of non-color-corrected images provided a wider color palette,

Page 4 of 13

improving the results for darker and lighter hues caused by shad-ing and glare, respectively.

Side Image Cropping and Seed SegmentationVariability in seed position on the matte black background

was even more extreme in side camera images. Therefore, the image preprocessing Python script was used in combination with the EasyPPC software to crop the side images. The top and bottom rows of each side image were cropped by the Python script, as those areas could not contain a seed. Subsequently, the EasyPCC software used a CART model to detect the matte black background in each side image, resulting in cropped side images that were segmented into non-background and back-ground areas (Supplemental Fig. S7). This CART model was trained using a random selection of 33 color-corrected side images that had their top and bottom rows cropped as described above. Next, all image areas detected as not of interest (i.e., nei-ther the matte black background nor seed) were cropped for a second time by the Python script, producing images of a seed against the matte black background. Finally, the EasyPCC soft-ware used a second CART model trained to detect the seed from the remaining background. We trained this CART model using a random selection of 35 images from the second round of crop-ping (Supplemental Fig. S8).

When training the CART models for side images, the EasyPPC software often detected colors in the regions of interest that were the same or similar to those found in the background area of the images. The occurrence of these false positives was minimized using an iterative process where a CART model was trained, applied to a test set of images, evaluated, and retrained, if needed. The metric for model evaluation was proper detection of the seed boundaries, such that there were fewer regions at the seed–background bound-ary where the seed was improperly labeled as background. If severe encroachment was found, the CART model was retrained with a different background area highlighted. This process was repeated no more than two times. All images used for model training were removed from the result set (Supplemental Table S1).

Image-Based Measurement of Seed Height, Length, and Width

The top and side segmented images were used to measure seed phenotypes. The Python OpenCV module was used in the main-processing Python script (Samain.PY; https://github.com/GoreLab/SSA_Image_Processor) to create a 2D rectangle around the seed in both the top and side images (Supplemental Fig. S9). The edges of the rectangle were bound to the outermost points of the seed. The rectangle was allowed to freely rotate, thus minimiz-ing the size. The length and width of this translated and rotated rectangle in the top image represented the length and width of the seed in pixels. Seed height in pixels was found using the same process of rectangle translation and rotation applied to the side image. Other variables measured from this 2D rectangle included the position in the image and the angle of the midline of the seed.

These two variables were used to convert the measured pixels to millimeters.

Conversion of pixel measurements to seed length in milli-meters was accomplished using scaling variables calculated from calibration images (i.e., images of a ruler) taken before the experi-mental seeds were processed through the SSA. For the top image of the seed, the process of pixel to millimeter conversion was linear because the seed always fell in the same plane. However, for the side image, the position of the seed was subject to change between images due to the physical design of the SSA. To address this, a series of calibration images were taken using the side camera with a ruler placed at several different known distances from the camera. The scaling factors produced from this process allowed a pixel rep-resenting a portion of a seed at a known distance from the side camera to be converted to a millimeter distance. The top camera image contained the position of the seed, and thus the distance between the seed and side camera could be calculated. Using the pixel height of the seed found in the side image, the scaling factors were used to convert the side image height distance from pixels to millimeters. This process of finding the seed location and scaling was repeated for each set of images and, as such, for each seed.

Seed Image Error DetectionThe SSA collected a low number of top and side images that

contained misaligned or missing seeds. Briefly, the main-processing Python script (Samain.PY) detected these error events by analyz-ing several seed phenotypes and their positions in both the top and side images. The position of the seed was saved as a binary 2D array representing the pixels that contain the seed in the image and the non-seed-containing pixels. The perimeter of this array was com-pared with the maximum bounds of the source raw image. If the perimeter intersected with the maximum bounds, then this indi-cated that the seed was only partially in the image. Next, the area of the seed, as defined by the total number of pixels, was found using the binary 2D array. If the total number of pixels comprising the seed was <400 pixels (i.e., an order of magnitude less than an average sized seed), then this indicated that only dust or debris (i.e., no seed) was detected. If the total number of pixels encompassing the seed was an order of magnitude more than an average sized seed, then this indicated a likely non-seed object in the image. Lastly, seed angle measurements from the top image were used to identify potential errors contained in the side image. If the angle of the seed was >80° with respect to the x axis in the top image, then this indicated that the side image would contain only a small cross-section of the seed. Therefore, seed volume could not be accurately estimated. Top and side images having seed perimeter, area, and/or angle values exceed-ing these defined thresholds were flagged in the output data file with a description of the detected errors. Flagged images were removed and not included in the analyzed results (Supplemental Table S1).

Image-Based Seed Volume MeasurementThe volume of an individual seed was estimated using an ellip-

soidal approximation of the seed cross-section in combination with

Page 5 of 13

a Riemann sum across the length of the seed. Briefly, the data from the top and side images were used to create a 2D seed location array, which was then used to measure the height and width of any given cross-section and total length (Supplemental Fig. S10). Notably, the side image array rarely matched the top image array with respect to length because of the variability in seed angle. To resolve this issue, we scaled the side array along the seed length axis, enabling the pixel length of the side array to equal the pixel length of the top array. The seed height was proportionally maintained throughout this process.

The estimated height, width, and rescaled length values were used to estimate volume as

( ) ( )

0Volume

2 2

l L

l

w l h ll

=

=

= p Då [1]

where summation occurs along the length of the 2D array from l = 0 to l = L, where L is the total array length in pixels. The area of an ellipse and, as such, the cross-sectional area at any point in the seed l is given by p[w(l)/2][h(l)/2], where w(l) and h(l) are the width and height, respectively, at point l along the length of the 2D seed location arrays. The values of w(l) and h(l) were calculated first in pixels and then scaled to millimeters using the scale factors described above. The cross-section as defined above was multiplied by the length of each pixel step Dl in millimeters and summed across the length of the seed, generating an estimated value for the volume.

Image-Based Seed Color MeasurementSeed color was measured using the color-calibrated top camera

image and 2D array of pixels that represented the seed location. As an example of this procedure, Supplemental Fig. S11 shows a red line boundary around the pixels used for color detection. The main-processing Python script (Samain.PY) recorded every colored pixel and its color value, followed by the creation of a comma-delimited text file. The comma-separated values (CSV) file contained a row for each seed with three columns containing the average red, green, or blue (RGB) color value across the entire seed. These color value data were only available for the oat genotypes Ogle and Hidalgo on 5 of the 6 d but were obtained for the other three oat genotypes (Newdak, Corral, and SD081107) on all 6 d.

Final Data Output from Image AnalysisThe final phenotypic data file generated by the main-pro-

cessing Python script (Samain.PY) was a CSV file containing individual seeds as rows and the seven seed phenotypes (length, width, height, average red, average green, average blue, and volume) as columns. An additional column included script error messages or warnings that indicated missing image files or data points where the phenotype values were outside expected ranges.

Manual Seed MeasurementsManual measurements of length, height, and width on individ-

ual seeds were performed with a digital caliper (Model 500-171-30, Mitutoyo) having a measurement accuracy of ±0.01 mm. The digital caliper directly sent measurements to a laptop via a USB connection.

All seed positions were maintained within the tray, and seed mea-surements were taken by two different technicians.

Length was defined as the radius parallel to the seed crease, whereas height was defined as the axis that intersects the seed crease at the top and the embryo at the bottom. Width was defined as the remaining axis that is perpendicular to the seed crease. These measurements were used to calculate the volume of an ellipsoid as

4Volume

3abc= p [2]

where a is the radius and b and c are the axes of the ellipsoid.Seed weights obtained by the SSA were independently vali-

dated by two technicians that manually weighed each seed using the same scale (GH-200, A&D Company Limited).

A colorimeter (Model CR-400, Konica Minolta) was used to measure seed color for each genotype by bulking the two trays of 96 seeds for each genotype. The colorimeter was first calibrated using the included white standard panel and set to use a 2° view angle and a D65 illuminant. Each set of seeds was placed within a glass dish, ensuring that there were at least two layers of seeds throughout and no visible bottom. The colorimeter was then placed over the center of the petri dish and moved around the sample to collect a total of 100 color measurements. The colorimeter returned measurements in the L*a*b* color space format. The measurements were made by two different technicians on different days.

Statistical AnalysesTo assess whether the raw SSA phenotypic data contained

significant outliers, we initially fitted a mixed linear model for each of the eight seed traits in ASReml-R Version 3.0 (Gilmour et al., 2009):

( )

( )

genotype day genotype day

tray genotype

ijk i j ij

ijkkl

Y =m+ + + ´

+e [3]

where Yijk is the individual phenotypic observation; m is the grand mean; genotypei is the effect of the ith genotype; dayj is the effect of the jth day of samples being run; (genotype × day)ij is the interac-tion effect between the ith genotype and jth day; tray(genotype)ki is the effect of the kth tray nested within the ith genotype; and eijk is the residual error effect assumed to be independently and iden-tically distributed according to a normal distribution with mean zero and variance se

2, that is ~ iid N(0,se2). The model terms

genotypei, dayj, and (genotype × day)ij were fitted as fixed effects, while tray(genotype)ki was modeled as a random effect according to ~ iid N(0,s2). To detect significant outliers, Studentized deleted residuals (Neter et al., 1996) obtained from these mixed linear models were examined. Data transformations and unequal vari-ance models were not needed because the assumptions of normality and homoscedasticity were met.

Once significant outliers were removed, an iterative mixed linear model (Eq. [3]) fitting procedure was conducted in

Page 6 of 13

ASReml-R Version 3.0. Likelihood ratio tests were conducted to remove terms fitted as random effects that were not significant at a = 0.05 (Littell et al., 2006). The final model for each trait was used to estimate a best linear unbiased estimator (BLUE) for each genotype on individual days and across all 6 d. Sequential tests of fixed effects were conducted, with degrees of freedom being cal-culated using the Kenward–Rogers approximation (Kenward and Roger, 1997). The Kenward–Rogers approximation, which uses a scaled Wald statistic that involves an approximate covariance matrix to account for variability introduced due to estimation of variance components, adjusts the degrees of freedom used in tests of significance together with an F approximation of the sampling distribution for the scaled Wald statistic (Kenward and Roger, 1997). The bias-adjusted degrees of freedom in the Kenward–Rogers approximation results in more conservative estimates of significance values associated with model terms.

The mixed linear model analysis of phenotypic data obtained by a handheld digital caliper and analytical balance was identical to that used for the SSA data, with the exception of having fitted the following model in ASReml-R Version 3.0:

( )

( )

genotype tech genotype tech

tray genotype

ijk i j ij

ijkki

Y =m+ + + ´

+e

[4]

where Yijk is the individual phenotypic observation collected with the digital caliper or analytical balance; m is the grand mean; genotypei is the effect of the ith genotype; techj is the effect of the jth technician who has taken measurements; (genotype × tech)ij is the interaction effect of the ith genotype with the jth technician; tray(genotype)

ki is the effect of the kth tray nested within the ith genotype; and eijk is the residual error effect assumed to be ~ iid N(0,se

2). The model term genotypei was the only fixed effect, while all other terms were modeled as random effects according to ~ iid N(0,s2). Data transformations and unequal variance models were not needed in this analysis. For each trait, a BLUE for each genotype was estimated with the final model fitted to an outlier-screened data set.

The mixed linear model analysis of phenotypic data collected by a handheld colorimeter was also identical to that used for the SSA except for fitting the following model in ASReml-R Version 3.0:

( )genotype tech genotype techij i j ijijY =m+ + + ´ +e [5]

where Yij is an individual phenotypic observation with the col-orimeter; m is the grand mean; genotypei is the effect of the ith genotype; techj is the effect of the jth technician that has taken measurements; (genotype × tech)ij is the interaction effect between the ith genotype and jth technician, and eij is the residual error effect assumed to be ~ iid N(0,se

2). The model term genotypei was the only fixed effect, with all other terms modeled as random effects according to ~ iid N(0,s2). For each trait, a BLUE for each genotype was estimated using the final model fitted to an outlier-screened data set, without a need for data transformations and unequal variance models.

We assessed the validity of phenotypic data collected by the SSA through comparisons with measurements manually obtained by a digital caliper, analytical balance, and colorimeter. The strength of a relationship between the same traits scored by two different methods was measured with Pearson’s correlation coefficient (rp) in R (R Core Team, 2017). Phenotypic data sets with outliers removed (i.e., prior to estimation of BLUEs) were used to calculate the correlation of daily single-seed measurements obtained by the SSA to those collected using the digital caliper and analytical balance. Given that single-seed measurements based on the digital caliper and analytical balance were taken independently by two different technicians, an average single-seed measurement value across technicians was used in pairwise comparisons with SSA measurements across genotypes for each of the 6 d.

To enable the comparison of seed color measurements between the SSA and colorimeter, the BLUEs for the three color traits (RGB) scored by the SSA were used to generate Commission Internationale de l’Eclairage L*a*b* (CIELAB) color space values for each genotype across the 6 d. The conversion from RGB to CIELAB color space values was performed through a web inter-face (http://colormine.org/convert/rgb-to-lab) that implemented the open source ColorMine library. Next, the resultant SSA-based CIELAB color values were tested for correlation (rp) with those obtained on bulked seed samples by the colorimeter.

For each trait, repeatability (r) was estimated to provide a measure of technical performance of the SSA relative to manual measurements for individual seeds. In the context of this study, repeatability is the proportion of the variance in measurements taken by a single instrument on the same seed under the same conditions that is due to permanent, or non-localized, differences between oat genotypes (Falconer and Mackay, 1996; Piepho and Möhring, 2007). To estimate the appropriate variance compo-nents, the final fitted model (Eq. [3] or [4]) for each trait was reformulated so that all terms were modeled as random effects. For the traits measured by the SSA, variance components were then used to estimate r (Holland et al., 2003):

2g

2 2 2 2g gd gt

re

s=s +s +s + s

[6]

where 2gs is the estimated genetic variance, 2

gds is the estimated variance associated with genotype × day interaction, 2

gts is the estimated variance associated with the effect of sampling tray nested within genotype, and 2

es is the residual error variance. With respect to the repeatability estimates for the traits measured by the digital caliper and analytical balance, the equation used was

2g

2 2 2 2g gtech gt

re

s=s + s + s + s

[7]

where 2gs is the estimated genetic variance, 2

gtechs is the estimated variance associated with the effect of genotype nested within research technician, 2

gts is the estimated variance associated with

Page 7 of 13

the effect of sampling tray nested within genotype, and 2es is the

residual error variance. Standard errors for the repeatability esti-mates were approximated using the delta method (Holland et al., 2003; Lynch and Walsh, 1998).

To assess how measurement repeatability impacted the number of phenotypic observations needed to be collected per seed, we calculated the gain in accuracy from a reduction in phe-notypic variation measured as a percentage as more measurements were obtained for an individual seed. With the inclusion of repeat-ability estimates calculated in Eq. [5], the following formula was used to derive the change in phenotypic variance as the number of measurements, n, was increased from 1 to 10 for each SSA trait (Falconer and Mackay, 1996):

( ) ( )p

p

1 1 nV r n

V n+ -

= [8]

where Vp(n) is the phenotypic variance for a number of measure-ments, n, for a trait, Vp is the overall phenotypic variance based on a single measurement, and r is the trait repeatability.

The number of different seeds that need to be measured per genotype by the SSA to reliably estimate seed trait parameters for an oat genotype, as determined by confidence intervals, was also calculated. With 10% of the trait mean as the defined amount of error, the following formula was used to determine seed sample size (Ott and Longnecker, 2010):

( )2 2/2

2

zn

Ea s

= [9]

where Za/2 is the Z value associated with a confidence interval of either 90 or 95% (a = 0.10 or 0.05, respectively), s2 is the variance for the trait, and E is the amount of acceptable error in the trait measurement, which was calculated as 10% of the estimated mean for each trait. To derive means and variances for a trait, an iterative sampling and modeling approach was used. First, for each trait, random sampling without replacement from the outlier-screened data set was conducted to generate sample sizes ranging from 2 to 20 measurements (n) per genotype. Within each subsampled data set, the following model was fitted to derive genotypic BLUEs for each trait within the n = 2 to 20 sample sizes:

genotypei i iY =m+ +e [10]

where Yi is an individual phenotypic observation; m is the grand mean; genotypei is the fixed effect of the ith genotype; and ei is the residual error effect assumed to be ~ iid N(0,se

2). The derived trait BLUEs for each genotype within the individual n sample sizes were used to estimate trait means and variances specific to that n-sized sample. Once all the trait means and variances for the n = 2 to 20 sample sizes were calculated, an overall, across-n sample size average was calculated for both the trait mean and variance. These representative values were used in Eq. [10] for determining the seed sample size for each trait.

�ResultsThe SSA and manual (digital caliper and analytical balance)

measurements of the individual oat seeds revealed highly significant (P < 0.0001) differences among the five genotypes for seed weight, length, width, height, and volume (Table 1). With respect to the per-formance of the SSA across the 6 d on which seeds were measured, there was no effect of day on seed weight measurements, but there were significant differences among the days for length, width, height, and volume (P < 0.05–0.0001; Table 1; Fig. 1). The interaction effect of genotype × day was only significant for volume (P < 0.01, exclud-ing the seed color traits), indicating that the day-to-day variability did not influence the rank order of the genotypes (Supplemental Table S2). For the RGB color values extracted from the SSA-collected images, no significant differences among the five oat genotypes were detected, but the effects of day and genotype × day were both highly significant (P < 0.0001) for all three color value traits.

When comparing the trait BLUEs of seed weight, length, width, height, and volume for the five oat genotypes between the SSA and manual measurements (Table 2), the estimates were not significantly different (two-sided t-test, P > 0.05). Additionally, the correlation (rp) of daily SSA measurements to those obtained with the digital caliper and analytical balance ranged from 0.618 to 0.994, with an average rp of 0.886 across all days and traits (Table 3). We observed an across-day, average correlation between seed weight and volume of 0.888 for the SSA and 0.986 for manual measure-ments (Fig. 2). With the exception of weight and width, the weakest correlations (rp = 0.618–0.900) between the SSA and manual mea-surements were on the fifth day that seed samples were evaluated. Irrespective of this one anomalous day, estimated correlation coef-ficients were found to be consistent across days for these five traits.

Table 1. F values for fixed effects from an analysis of variance (ANOVA) for oat seed traits measured using the single-seed ana-lyzer (SSA), digital caliper, analytical balance, and colorimeter.

Trait Method Genotype Day Genotype × dayWeight (g) SSA 170.800† 1.811 NS 0.211 NS

balance 61.480† – –Length (cm) SSA 140.900† 2.462* 0.650 NS

caliper 57.450† – –Width (cm) SSA 33.860† 9.670† 0.799 NS

caliper 94.970† – –Height (cm) SSA 89.450† 76.720† 1.536 NS

caliper 31.150† – –Volume (cm3) SSA 32.470† 125.400† 1.835**

caliper 59.640† – –Red value SSA 2.258NS 343.100† 9.029†Green value SSA 2.764 NS 293.900† 7.052†Blue value SSA 2.353 NS 193.200† 8.922†L* colorimeter 5.668* – –a* colorimeter 3.591 NS – –b* colorimeter 1.715 NS – –* Significant at the <0.05 level. NS, not significant at the <0.05 level.** Significant at the <0.01 level.† Significant at the < 0.0001 level.

Page 8 of 13

In contrast, the three CIELAB color values determined via the SSA and colorimeter were significantly different (two-sided t-test, P < 0.01) between the two methods (Table 4). However, strong positive correlations (rp = 0.742–0.915) were observed between color values based on the two measurement methods (Table 4).

The estimates of repeatability (r) on an individual seed basis for weight, height, length, width, and volume measured by the SSA ranged from 0.151 to 0.299, whereas for manual measurements

repeatability estimates for an individual seed ranged from 0.147 to 0.377 for the same five traits (Table 2). The repeatability estimates for these five traits were not significantly different (paired t-test, P > 0.05) between the SSA and manual methods. Neither the SSA nor the manual methods consistently had the highest repeat-ability. The SSA measurements exhibited higher repeatability for weight and height; however, manual measurements for seed length, volume, and width had higher repeatability estimates. Conversely,

Fig. 1. Boxplots of five oat seed traits measured by the single-seed analyzer (SSA) across 6 d.

Table 2. Best linear unbiased estimators (BLUEs) for oat seed traits measured using the single-seed analyzer (SSA), digital caliper, and ana-lytical balance, as well as estimates of repeatability (r) on an individual seed basis with their standard error (SE).

Seed trait

Method

Genotype RepeatabilityCorral Hidalgo Newdak Ogle SD081107 Estimate SE

Weight (g) SSA 0.022 0.020 0.027 0.026 0.030 0.281 0.144balance 0.023 0.022 0.028 0.028 0.031 0.272 0.144

Length (cm) SSA 0.651 0.694 0.720 0.736 0.771 0.227 0.126caliper 0.609 0.638 0.672 0.688 0.726 0.252 0.138

Width (cm) SSA 0.271 0.248 0.272 0.284 0.288 0.299 0.153caliper 0.247 0.223 0.247 0.258 0.262 0.377 0.169

Height (cm) SSA 0.207 0.210 0.224 0.208 0.220 0.151 0.094caliper 0.198 0.197 0.213 0.205 0.210 0.147 0.097

Volume (cm3) SSA 0.017 0.016 0.020 0.019 0.022 0.193 0.115caliper 0.127 0.120 0.150 0.154 0.169 0.268 0.142

Red value SSA 137.254 133.227 140.379 141.315 138.827 0.028 0.028Green value SSA 110.472 107.808 116.022 114.956 113.475 0.054 0.054Blue value SSA 69.046 69.417 74.816 73.061 73.359 0.006 0.094

Page 9 of 13

the RGB color values had the lowest repeatability estimates (mean r of 0.029) of any measured trait (Table 2). On average, the repeat-ability for the three color value traits was nearly six- to sevenfold lower than those obtained for the other five non-color traits by the SSA (mean r of 0.230) and manual (mean r of 0.263) individual seed phenotyping methods.

The repeatability estimates for the SSA-measured traits (Table 2) were used to calculate the gain in accuracy expected from a reduction in the phenotypic variance by multiple measurements on an individual seed. The gain in accuracy expected from multiple measurements (n = 1–10) of each seed was highest for the traits with the lowest repeatability and lowest for traits with the highest repeatability (Fig. 3), congruent with expectations. Of the eight SSA-measured traits, the RGB color value traits would benefit the most from multiple measurements. Additionally, there was an observed concomitant decrease in accuracy gains as the number of measurements increased. Generally, the results suggested that two to three measurements per seed by the SSA provided the best balance between gain in accuracy and resource expenditure (Fig. 3), as there was only a minimal reduction in the phenotypic variance with four or more measurements for an individual seed.

In a sampling analysis focused on the accuracy in estimation of morphological parameters, we calculated the number of different seeds per genotype that needed to be measured by the SSA to esti-mate a mean value with an assumed error rate of 10% for the more repeatable non-seed-color traits. The results of this analysis showed that an estimated sample size of only one seed was required for seed height, width, and length, while weight and volume had estimated sample sizes of seven and eight seeds, respectively, to allow for estima-tion of trait values such that the 90% confidence interval was equal to 10% of the trait mean (Table 5). When considering a confidence interval of 95%, the number of estimated seeds that needed to be measured for these five traits increased by one seed for length and three seeds each for weight and volume, whereas no additional seeds were required for the measurements of width and height.

�DiscussionThe quantification of phenotypic trait variation on large

numbers of genetically distinct individuals continues to require a considerable investment of both time and financial resources for plant breeding programs (Araus and Cairns, 2014; Furbank and Tester, 2011; Pauli et al., 2016b; White et al., 2012). To address this challenge, there have been significant efforts focused on the development, testing, and validation of plant phenotyping systems for both greenhouse and field environments (Andrade-Sanchez et al., 2014; Deery et al., 2014; Jiang et al., 2018; Minervini et al., 2017). These phenotyping advancements have shifted the emphasis toward the use of validated systems for comprehensively assessing plant morphological and physiological traits across varying spatio-temporal scales (Fahlgren et al., 2015; Honsdorf et al., 2014; Pauli et al., 2016a). Often of less focus relative to agronomically and economically important phenotypes evaluated under field condi-tions, such as grain and biomass yield, are measurement of traits on a single-seed basis (Jahnke et al., 2016). These traits, which have consistently been shown to have a direct relationship with end-product quality (Melchinger et al., 2018; Montilla-Bascón et al., 2017; Rolletschek et al., 2015; Spielbauer et al., 2009), would significantly benefit from having a low-cost, easy-to-operate seed

Table 3. Pearson’s correlation coefficients (rp) generated from evaluating the linear relationship of single-seed analyzer (SSA) measured oat seed traits with the same traits measured by a digi-tal caliper and analytical balance across 6 d. All correlations were significant at a = 0.05.Day Weight Length Width Height Volume

1 0.941 0.965 0.893 0.707 0.9102 0.983 0.974 0.927 0.719 0.9523 0.979 0.964 0.918 0.723 0.9364 0.975 0.964 0.898 0.687 0.9295 0.993 0.900 0.895 0.618 0.6696 0.994 0.977 0.912 0.741 0.929

Fig. 2. The relationship between seed weight and volume estimated by measurements from (A) the single-seed analyzer (SSA) or (B) digital caliper.

Page 10 of 13

phenotyping system amenable to laboratory settings. To that end, we developed and tested a novel computational image analysis pipeline combined with a SSA machine that enables the accurate measurement of individual oat seed characteristics in a cost-effec-tive and high-throughput manner.

The SSA machine and cameras, which cost an estimated US$20,000, in combination with the implementation of image

analysis methods based on open source software, allowed the detec-tion of significant differences in individual seed weight, length, width, height, and volume parameters among the five oat genotypes evaluated. The estimates of these seed properties were highly concor-dant with those measured manually using a handheld digital caliper. Conversely, the more subtle variation in seed color across the five oat genotypes was not significantly (P < 0.05) differentiated by any of

Fig. 3. Gain in accuracy from multiple phenotypic measurements of an individual seed with the single-seed analyzer (SSA).

Table 4. Best linear unbiased estimators (BLUEs) for oat L*a*b* color space seed traits measured using the single-seed analyzer (SSA) and colorimeter, as well as Pearson’s correlation coefficients (rp) from evaluating the linear relationship of these same traits between the two measurement methods. All correlations were significant at a = 0.10.Color parameter Method Corral Hidalgo Newdak Ogle SD081107 CorrelationL* SSA 48.290 47.138 50.227 50.006 49.346 0.742

colorimeter 56.869 58.062 59.294 59.756 59.156a* SSA 4.935 4.726 3.732 4.569 4.355 0.915

colorimeter 6.062 6.182 5.666 6.089 5.910b* SSA 26.993 25.230 26.276 26.999 25.937 0.857

colorimeter 23.371 22.266 23.538 23.348 22.785

Page 11 of 13

the RGB color values extracted from images collected by the SSA. This is partly explained by the strong effects of day and its interac-tion with genotype, which had considerably less of an influence on the other four image-based phenotypes (length, width, height, and volume). With the exception of seed weight, the day and genotype × day effects are probably the product of uncontrolled fluctuations in ambient overhead lighting conditions that occurred throughout the times of phenotyping. This environmental variation would have directly impacted the collected imagery through either unequal shadow effects for seed morphometric properties or inconsistent color intensity due to differences in absorbed and reflected electro-magnetic radiation. An enclosure could be added to the imaging station to eliminate all outside ambient light. Regardless of the exact mechanism responsible, the resulting variation resulted in weaker statistical power to discriminate seed dimensional properties and color differences among the genotypes investigated.

The correlations between SSA and manual measurements were very strong (rp = 0.618–0.994) for seed weight, length, width, and volume, demonstrating the technical accuracy and precision of the machine for these traits. Slightly weaker correlations for height reflects the greater difficulty in obtaining reliable focus and color contrast for side profile images due to the tradeoff of prioritiz-ing the quality of top images that provide two seed dimensions. If the camera software had allowed separate imaging parameters for both top and side cameras, then this would probably have enabled the realization of stronger correlations for seed height through more accurate data. With respect to the weaker correlations for seed length, height, and volume observed on Day 5, this could be plausibly explained by the SSA technician unknowingly disturb-ing the calibrated position of the cameras while attempting to manipulate a stray seed on the imaging platform. The CIELAB color values based on the SSA strongly correlated with those of bulked seed measured by a colorimeter; however, the color values were significantly different between the two methods. This differ-ence is attributable in part to the narrow dynamic color range of the camera, which limited the capture of color data for extremely bright or dark areas of the seed.

The estimates of repeatability on an individual seed basis for the SSA were not statistically different from those of the manual measurements for seed weight, length, width, height, and volume. Such a finding indicates that the reliability of the manual and SSA phenotyping methods are statistically equivalent, suggest-ing that the SSA is suitable for rapid, accurate phenotyping of physical attributes that influence milling yield and test weight

(Doehlert et al., 2006; Groh et al., 2001). In stark contrast, the six- to sevenfold lower repeatability estimates of the RGB color values strongly imply that a restrictively high number of multiple measurements on a single seed are needed to generate reproduc-ible measurements of seed color using the SSA. Breeding efforts based on groat color have been previously proposed (Johnson et al., 1997) because of both consumer preference and the availability of variation. However, the SSA would need to obtain more highly reproducible color measurements to implement such an approach.

Although our developed seed phenotyping platform is not novel with respect to the task of phenotyping seed per se (Jahnke et al., 2016; Melchinger et al., 2018), it does offer a lower cost, less complex solution. Jahnke et al. (2016) used a sophisticated robotic arm that, while able to collect very accurate data, was limited in its ability to handle hundreds of samples typically found in a breeding program. In contrast, the SSA is a low-cost solution that collects data nearly as accurate as revealed by the correlation between seed volume and weight, with a data collection rate almost threefold faster than our manual measurement methods. Jahnke et al. (2016) obtained correlation values of 0.99 between volume and mass for rapeseed (Brassica napus L.) using only three seeds of a single cul-tivar. In contrast, we obtained an across-day average correlation of 0.888 (Fig. 2) using five cultivars with significantly different seed dimensions, thus demonstrating the utility of the SSA system. Additionally, the correlation value improved to 0.920 when Day 5 was removed, providing a more representative estimation of the instrument’s performance and approaching the 0.986 correlation value obtained with the handheld caliper.

There are several needed modifications to the SSA system for further improving the image-based phenotyping of individual seeds, with higher quality cameras being the most important. Image data quality could be improved through the use of cameras with both individual settings for color as well as exposure and higher resolu-tion to increase the number of pixels per image that are associated with seeds. An increase in camera resolution would help to resolve difficulties in defining the edge of a seed, thus reducing measurement errors attributed to the miss-assignment of background pixels to a seed. Cameras with better autofocusing capabilities would enhance seed edge detection by allowing reliable autofocus on the exact seed midline position, thereby replacing a less than ideal practice of having a fixed focus setting throughout an entire SSA run. The most impor-tant attribute to be improved on is using a larger and higher bit depth camera sensor that will obtain better color accuracy in the imagery. Color accuracy will not only improve the direct color measurements but will also improve the ability of the CART model to segment seed from background using these higher quality color values. The cameras and lighting equipment used in our study consisted of <1% of the total estimated cost of the SSA. We estimate these suggested improve-ments to the cameras to increase the cost of the SSA by 5% minimally, and up to 35% using professional-grade cameras.

In combination with an improved camera, an additional low-cost improvement that could be implemented easily would be to replace the existing matte black paper background with a different

Table 5. Number of seeds needed for measurement by the sin-gle-seed analyzer (SSA) to estimate a mean trait value with an error rate of 10% at a confidence interval (CI) of 90 or 95%.

Trait Mean Error ± 10% of mean n for 90% CI n for 95% CIWeight, g 0.025 0.003 6.738 9.624Length, cm 0.714 0.071 1.333 1.903Width, cm 0.273 0.027 0.919 1.313Height, cm 0.214 0.021 0.549 0.784Volume, cm3 0.019 0.002 8.002 11.430

Page 12 of 13

color that does not occur on the seeds. This would mitigate the problems associated with the darkest colors of the seed color gradi-ent matching values on the matte black background, an issue which can increase the difficulty for training the CART model, especially around the edges of the seed. An issue that occurred for about 8% of all the image sets (Supplemental Table S1) was that the compressed air ejector moved seeds to the far edge of or beyond the image frame, which was found to be more problematic for the two oat genotypes with lighter seed weight, Hidalgo and Corral (Table 2). This issue could be addressed through automated adjustment based on a digi-tally controlled regulator that provides a calculated air pressure burst according to the measured weight of each seed. These technical mod-ifications in addition to those listed above would result in higher quality image data, which could be confirmed with comparisons to the manual measurements through correlation analyses, allowing the SSA to achieve even higher precision and accuracy.

Critical to the use of an instrument in a breeding program is evaluating the amount and quality of data generated vs. the resources spent to obtain those data. To assess this balance for the SSA, we estimated both the number of times a single seed should be measured to achieve a target level of repeatability for each of the seed traits and the number of different seeds that should be sampled per genotype to estimate phenotypic values with a specified level of confidence. With each run of the SSA taking approximately 1.25 h for a set of four 48-well seed plates, the SSA can deliver an average reduction of about 38 and 51% in measurement error by measuring the seeds two and three additional times, respectively, for seed weight, length, width, height, and volume (Fig. 3). This represents a significant gain in data quality, on which critical breeding decisions will be based, for minimal resource expenditure. Given that a maximum of seven to eight seeds sufficiently captures the variance for these five traits, dividing a 48-well plate across six different genotypes would be the most efficient experimental design. Thus, our recommendation is to randomly sample eight seeds from each seed packet (or plot) and run each 48-well plate twice for measurement on the SSA. This is made possible because the SSA places each measured seed in the corresponding well of the plate after measurement. Because the SSA maintains the identity of each seed measured, the breeder can select individual seeds that have the desired quality, size, and color for fur-ther use in the breeding program. This can be especially valuable for selecting seeds in segregating populations.

�ConclusionsThe low-cost hardware of the SSA machine combined with

open-source image analysis software enabled the simultaneous, rapid measurement of multiple traits on individual seeds. The results obtained in our study demonstrate that the SSA system reliably determines the weight, length, width, height, and volume of single oat seeds with equal accuracy as manual measurements. When com-bined with NIR technology for b-glucans, protein, and oil content (Montilla-Bascón et al., 2017), the validated imaging capabilities of the SSA system provide an affordable solution for comprehensively

measuring the morphology of seeds from thousands of oat geno-types. Such capacity for high-throughput quantitative measurements would help accelerate breeding efforts to improve the yield and quality of oat. The overall adaptability of the SSA to other small grain and larger seed crops is achievable because the NIR light tube assembly (Supplemental Fig. S2) works well for soybean [Glycine max (L.) Merr.] and maize (Armstrong, 2006), although different image analysis algorithms would need to be implemented or developed to extract the desired phenotypic information from each seed type.

Author ContributionsJ.W.C., D.P., K.M.K., and M.A.G. co-wrote the manuscript. D.P. led sta-tistical analyses. P.R.A. and T.W. developed and integrated hardware and software for the single-seed analyzer. J.W.C., K.M.K., and E.S.B.V devel-oped and integrated the hardware and software for imaging, constructed the computational image analysis pipeline, and analyzed the image data. M.E.S. conducted and harvested the field trial for the oat seed. O.A.H, J.-L.J., M.E.S., and M.A.G. conceived the experiment. M.A.G conducted overall project management, design, and coordination and oversaw data analysis. All authors contributed to the critical review of the manuscript.

FundingThis research was supported by PepsiCo Global R&D, the USDA–ARS, USDA-NIFA-AFRI 2011-68002-30029 (Triticeae-CAP), USDA-NIFA-AFRI 2017-67007-25939 (Wheat-CAP), USDA-NIFA-AFRI 2017-67007-26502, Cotton Incorporated, and Hatch Project 149-447.

DisclaimerThe views expressed in this article are those of the authors and do not necessarily reflect the position or policy of PepsiCo, Inc. Mention of a trademark or proprietary product does not constitute a guarantee or war-ranty of the product by the USDA and does not imply its approval to the exclusion of other products that may also be suitable. The USDA is an equal opportunity provider and employer.

Conflict of Interest DisclosureThe authors declare that there is no conflict of interest.

AcknowledgmentsWe especially thank Matthew Ming, Nicholas Kaczmar, Fang Liu, Han Rongkui, Matt Baseggio, and Yujie Meng for providing excellent techni-cal expertise in the collection of phenotypic data through the SSA and manual measurements.

Supplemental MaterialThe supplemental material consists of two supplemental tables that have image metrics and phenotypic data, as well as 11 supplemental figures that diagram the single-seed analyzer machine and image analysis pipeline.

ReferencesAndrade-Sanchez, P., M.A. Gore, J.T. Heun, K.R. Thorp, A.E. Carmo-Silva,

A.N. French, et al. 2014. Development and evaluation of a field-based high-throughput phenotyping platform. Funct. Plant Biol. 41:68–79. doi:10.1071/FP13126

Araus, J.L., and J.E. Cairns. 2014. Field high-throughput phenotyping: The new crop breeding frontier. Trends Plant Sci. 19:52–61. doi:10.1016/j.tplants.2013.09.008

Armstrong, P.R. 2006. Rapid single-kernel NIR measurement of grain and oil-seed attributes. Appl. Eng. Agric. 22:767–772. doi:10.13031/2013.21991

Deery, D., J. Jimenez-Berni, H. Jones, X. Sirault, and R. Furbank. 2014. Proximal remote sensing buggies and potential applications for field-based

Page 13 of 13

phenotyping. Agronomy 4:349–379. doi:10.3390/agronomy4030349

Doehlert, D.C., M.S. McMullen, and J.-L. Jannink. 2006. Oat grain/groat size ratios: A physical basis for test weight. Cereal Chem. 83:114–118. doi:10.1094/CC-83-0114

Fahlgren, N., M. Feldman, M.A. Gehan, M.S. Wilson, C. Shyu, D.W. Bryant, et al. 2015. A versatile phenotyping system and analytics platform reveals diverse temporal responses to water availability in Setaria. Mol. Plant 8:1520–1535. doi:10.1016/j.molp.2015.06.005

Falconer, D.S., and T.F.C. Mackay. 1996. Introduction to quantitative genetics. Longman, New York.

Furbank, R.T., and M. Tester. 2011. Phenomics: Technologies to relieve the phenotyping bottleneck. Trends Plant Sci. 16:635–644. doi:10.1016/j.tplants.2011.09.005

Gilmour, A.R., B.J. Gogel, B.R. Cullis, and R. Thompson. 2009. ASReml user guide, Release 3.0. VSN Int., Hemel Hempstead, UK.

Groh, S., S.F. Kianian, R.L. Phillips, H.W. Rines, D.D. Stuthman, D.M. Wesenberg, and R.G. Fulcher. 2001. Analysis of factors influencing milling yield and their association to other traits by QTL analysis in two hexaploid oat populations. Theor. Appl. Genet. 103:9–18. doi:10.1007/s001220100579

Guo, W., U.K. Rage, and S. Ninomiya. 2013. Illumination invariant segmentation of vegetation for time series wheat images based on decision tree model. Comput. Electron. Agric. 96:58–66. doi:10.1016/j.compag.2013.04.010

Holland, J.B., W.E. Nyquist, and C. Cervantes-Martinez. 2003. Estimating and interpreting heritability for plant breeding: An update. Plant Breed. Rev. 22:9–112. doi:10.1002/9780470650202.ch2

Honsdorf, N., T.J. March, B. Berger, M. Tester, and K. Pillen. 2014. High-throughput phenotyping to detect drought tolerance QTL in wild barley introgression lines. PLoS One 9:e97047. doi:10.1371/journal.pone.0097047

Iwata, H., K. Ebana, Y. Uga, T. Hayashi, and J.-L. Jannink. 2010. Genome-wide association study of grain shape variation among Oryza sativa L. germplasms based on elliptic Fourier analysis. Mol. Breed. 25:203–215. doi:10.1007/s11032-009-9319-2

Iwata, H., and Y. Ukai. 2002. SHAPE: A computer program package for quantitative evaluation of biological shapes based on elliptic Fourier descriptors. J. Hered. 93:384–385. doi:10.1093/jhered/93.5.384

Jahnke, S., J. Roussel, T. Hombach, J. Kochs, A. Fischbach, G. Huber, and H. Scharr. 2016. phenoSeeder: A robot system for automated handling and phenotyping of individual seeds. Plant Physiol. 172:1358–1370. doi:10.1104/pp.16.01122

Jiang, Y., C. Li, J.S. Robertson, S. Sun, R. Xu, and A.H. Paterson. 2018. GPhenoVision: A ground mobile system with multi-modal imaging for field-based high throughput phenotyping of cotton. Sci. Rep. 8:1213. doi:10.1038/s41598-018-19142-2

Johnson, N.J., D.J. Moot, and T.N. Lindley. 1997. Oat groat colour: A quality factor. Proc. Agron. Soc. N. Z. 27:79–82.

Kenward, M.G., and J.H. Roger. 1997. Small sample inference for fixed effects from restricted maximum likelihood. Biometrics 53:983–997. doi:10.2307/2533558

Littell, R.C., G.A. Milliken, W.W. Stroup, R.D. Wolfinger, and O. Schabenberger. 2006. SAS for mixed models. SAS Inst., Cary, NC.

Lynch, M., and B. Walsh. 1998. Genetics and analysis of quantitative traits. Sinauer Assoc., Sunderland, MA.

Melchinger, A.E., J. Böhm, H.F. Utz, J. Müller, S. Munder, and F.J. Mauch. 2018. High-throughput precision phenotyping of the oil content of single seeds of various oilseed crops. Crop Sci. 58:670–678. doi:10.2135/cropsci2017.07.0429

Miller, S.S., and R.G. Fulcher. 2016. Microstructure and chemistry of the oat kernel. In: F. Webster and P.J. Wood, editors, Oats: Chemistry and technology. Academic Press, San Diego. p. 77–94.

Minervini, M., M.V. Giuffrida, P. Perata, and S.A Tsaftaris. 2017. Phenotiki: An open software and hardware platform for affordable and easy image‐based phenotyping of rosette‐shaped plants. Plant J. 90:204–216. doi:10.1111/tpj.13472

Montilla-Bascón, G., P.R. Armstrong, H. Rongkui, and M. Sorrells. 2017. Quantification of betaglucans, lipid and protein contents in whole oat groats (Avena sativa L.) using near infrared reflectance spectroscopy. J. Near Infrared Spectrosc. 25:172–179. doi:10.1177/0967033517709615

Moore, C.R., D.S. Gronwall, N.D. Miller, and E.P. Spalding. 2013. Mapping quantitative trait loci affecting Arabidopsis thaliana seed morphology features extracted computationally from images. G3: Genes, Genomes, Genet. 3:109–118. doi:10.1534/g3.112.003806

Neter, J., M.H. Kutner, C.J. Nachtsheim, and W. Wasserman. 1996. Applied 7lels. McGraw-Hill, Boston.

Ott, R.L., and M.T. Longnecker. 2010. An introduction to statistical methods and data analysis. 6th ed. Cengage Learning, Mason, OH.

Pauli, D., P. Andrade-Sanchez, A.E. Carmo-Silva, E. Gazave, A.N. French, J. Heun, et al. 2016a. Field-based high-throughput plant phenotyping reveals the temporal patterns of quantitative trait loci associated with stress-responsive traits in cotton. G3: Genes, Genomes, Genet. 6:865–879. doi:10.1534/g3.115.023515

Pauli, D., S.C. Chapman, R. Bart, C.N. Topp, C.J. Lawrence-Dill, J. Poland, and M.A. Gore. 2016b. The quest for understanding phenotypic variation via integrated approaches in the field environment. Plant Physiol. 172:622–634. 10.1104/pp.16.00592

Peterson, D.M., and D.F. Wood. 1997. Composition and structure of high oil oat. J. Cereal Sci. 26:121–128. doi:10.1006/jcrs.1996.0111

Piepho, H.-P., and J. Möhring. 2007. Computing heritability and selection response from unbalanced plant breeding trials. Genetics 177:1881–1888. doi:10.1534/genetics.107.074229

R Core Team. 2017. R: A language and environment for statistical computing. R Found. Stat. Comput., Vienna.

Rolletschek, H., J. Fuchs, S. Friedel, A. Börner, H. Todt, P.M. Jakob, and L. Borisjuk. 2015. A novel noninvasive procedure for high‐throughput screening of major seed traits. Plant Biotechnol. J. 13:188–199. doi:10.1111/pbi.12245

Roussel, J., F. Geiger, A. Fischbach, S. Jahnke, and H. Scharr. 2016. 3D surface reconstruction of plant seeds by volume carving: Performance and accuracies. Front. Plant Sci. 7. doi:10.3389/fpls.2016.00745

Rustgi, S., M.N. Shafqat, N. Kumar, P.S. Baenziger, M.L. Ali, I. Dweikat, et al. 2013. Genetic dissection of yield and its component traits using high-density composite map of wheat chromosome 3A: Bridging gaps between QTLs and underlying genes. PLoS One 8:e70526. doi:10.1371/journal.pone.0070526

Spielbauer, G., P. Armstrong, J.W. Baier, W.B. Allen, K. Richardson, B. Shen, and A.M. Settles. 2009. High‐throughput near‐infrared reflectance spectroscopy for predicting quantitative and qualitative composition phenotypes of individual maize kernels. Cereal Chem. 86:556–564. doi:10.1094/CCHEM-86-5-0556

Tanabata, T., T. Shibaya, K. Hori, K. Ebana, and M. Yano. 2012. SmartGrain: High-throughput phenotyping software for measuring seed shape through image analysis. Plant Physiol. 160:1871–1880. doi:10.1104/pp.112.205120

Whan, A.P., A.B. Smith, C.R. Cavanagh, J.-P.F. Ral, L.M. Shaw, C.A. Howitt, and L. Bischof. 2014. GrainScan: A low cost, fast method for grain size and colour measurements. Plant Methods 10:23. doi:10.1186/1746-4811-10-23

White, J.W., P. Andrade-Sanchez, M.A. Gore, K.F. Bronson, T.A. Coffelt, M.M. Conley, et al. 2012. Field-based phenomics for plant genetics research. Field Crops Res. 133: 101-112. doi:10.1016/j.fcr.2012.04.003

Williams, K., J. Munkvold, and M. Sorrells. 2013. Comparison of digital image analysis using elliptic Fourier descriptors and major dimensions to phenotype seed shape in hexaploid wheat (Triticum aestivum L.). Euphytica 190:99–116. doi:10.1007/s10681-012-0783-0

Xie, Q., S. Mayes, and D.L. Sparkes. 2015. Carpel size, grain filling, and morphology determine individual grain weight in wheat. J. Exp. Bot. 66:6715–6730. doi:10.1093/jxb/erv378

Ziegler, G., A. Terauchi, A. Becker, P. Armstrong, K. Hudson, and I. Baxter. 2013. Ionomic screening of field-grown soybean identifies mutants with altered seed elemental composition. Plant Genome 6. doi:10.3835/plantgenome2012.07.0012

View publication statsView publication stats