Research

Comprehensive 3D phenotyping reveals continuous

morphological variation across genetically diverse sorghum
inflorescences
Mao Li1* , Mon-Ray Shao1* , Dan Zeng2 , Tao Ju2 , Elizabeth A. Kellogg1 and Christopher N. Topp1
1
Donald Danforth Plant Science Center, St Louis, MO 63132, USA; 2Department of Computer Science and Engineering, Washington University, St Louis, MO 63130, USA

Summary
Author for correspondence:  Inflorescence architecture in plants is often complex and challenging to quantify, particularly
Christopher N. Topp for inflorescences of cereal grasses. Methods for capturing inflorescence architecture and for
Tel: +1 314 587 1609
analyzing the resulting data are limited to a few easily captured parameters that may miss the
Email: ctopp@danforthcenter.org
rich underlying diversity.
Received: 9 December 2019  Here, we apply X-ray computed tomography combined with detailed morphometrics, offer-
Accepted: 23 February 2020
ing new imaging and computational tools to analyze three-dimensional inflorescence architec-
ture. To show the power of this approach, we focus on the panicles of Sorghum bicolor,
New Phytologist (2020) 226: 1873–1885 which vary extensively in numbers, lengths, and angles of primary branches, as well as the
doi: 10.1111/nph.16533 three-dimensional shape, size, and distribution of the seed.
 We imaged and comprehensively evaluated the panicle morphology of 55 sorghum acces-
Key words: inflorescences, panicle, sions that represent the five botanical races in the most common classification system of the
phenomics, phenotyping, sorghum, X-ray. species, defined by genetic data. We used our data to determine the reliability of the morpho-
logical characters for assigning specimens to race and found that seed features were particu-
larly informative.
 However, the extensive overlap between botanical races in multivariate trait space indicates
that the phenotypic range of each group extends well beyond its overall genetic background,
indicating unexpectedly weak correlation between morphology, genetic identity, and domes-
tication history.

2018), resulting in phenomics becoming an increasingly impor-

Introduction
Plant organs show tremendous diversity in size and shape as a
result of differential growth and patterning throughout develop-
ment. Measurements of plant structures such as inflorescences or
leaves for botany, crop breeding, or ecology studies most often
use manual observation or the analysis of two-dimensional (2D)
images, the latter ranging from electron microscopy (e.g. Buzgo
et al., 2004; Vollbrecht et al., 2005; Prenner & Rudall, 2007;
Kellogg et al., 2013) to high-throughput camera imaging (Chen
et al., 2014; Gehan & Kellogg, 2017). More recently, imaging
techniques have been developed that enable the reconstruction
and analysis of plant tissue or structures in three dimensions. At
the microscopic scale, three-dimensional (3D) reconstructions of
cells can be achieved using several techniques, such as confocal or
scanning electron microscopy (Bougourd et al., 2000; Denk &
Horstmann, 2004). On the other hand, at field scales, plant
height and other traits can be measured using 3D laser scanners
or unmanned aerial systems (Friedli et al., 2016; Malambo et al.,
*These authors contributed equally to this work.
tant tool in field research.
Between these two scales, specific structures of and within
plants can be imaged in three dimensions using high-resolution
X-ray computed tomography (XRT; Stuppy et al., 2003; Dhondt
et al., 2010; Perez-Torres et al., 2015; Rogers et al., 2016; Jiang
et al., 2019). Although the processing and analysis of 3D images
for complex biological structures remains nontrivial, automated
computational tools providing precise analysis for 3D imaging
are becoming increasingly powerful, permitting segmentation,
reconstruction, registration, visualization, feature extraction, and
modeling. For example, 3D wheat grains have been segmented
and measured from images of the wheat panicle (Hughes et al.,
2017), and 3D branching topology has been characterized com-
prehensively in grapevine inflorescences (Li et al., 2017, 2019).
Furthermore, software such as CHIMERA (Pettersen et al., 2004)
has been used widely for the visualization and analysis of molecu-
lar structures, density maps, 3D microscopy, and associated data
(Goddard et al., 2017), and topological data analysis tools such
as the medial axis (Blum, 1967) have been used extensively for
analyzing shape structure and for forming shape descriptors. The
medial axis, in particular, is a topological curve skeleton that is a

Ó 2020 The Authors New Phytologist (2020) 226: 1873–1885 1873

New Phytologist Ó 2020 New Phytologist Trust www.newphytologist.com
This is an open access article under the terms of the Creative Commons Attribution License, which permits use,
distribution and reproduction in any medium, provided the original work is properly cited.

1874 Research
complete shape descriptor, meaning that it can be used to recon-
struct inflorescence shape; the geometry and thickness along the
medial axis can then be used to quickly identify the main stalk
and primary branches. Because the internal structures of grass
inflorescences are often occluded in flat 2D scans, these shape
analysis methods can only be fully utilized with 3D images.
Here, we apply these tools to a long-standing problem of mor-
phological diversity and variation within Sorghum bicolor. The
inflorescences of sorghum are referred to as panicles, with a cen-
tral monopodial axis and lateral branches that then further
branch (Bell, 2008). Spikelets on the branches occur in pairs,
with one hermaphrodite and one sterile, and in the
hermaphrodite spikelet is one fertile and one sterile floret. The
terminal spikelet on each branch is morphologically similar to
the sterile spikelet from the immediately proximal pair, giving
the appearance of a triplet of spikelets. From a mature ovary, the
grain that develops is a type of indehiscent fruit known as a cary-
opsis, where the pericarp of the fruit and the seed coat are fused
(Esau, 1977). Although the thickness of the pericarp varies, in
sorghum it shows relatively little deformation; hence, here, we
use the terms ‘fruit shape’ and ‘seed shape’ interchangeably when
referring to phenotype at the macroscopic level.
Domesticated sorghum exhibits kaleidoscopic variation in
inflorescence forms and seed shapes generated by millennia of
human selection and interchange of germplasm (Harlan & De
Wet, 1971). Attempts at formal Linnaean classification led to a
bewildering taxonomy of species, subspecies, varieties, and forms
(Snowden, 1936, 1955) that obscured the interfertility and cyto-
genetic similarity of the various taxa (De Wet, 1978). To gener-
ate a more practical classification, Harlan & De Wet (1972)
included nearly all of Snowden’s taxa in a single, highly polymor-
phic species, S. bicolor, which was then divided into five morpho-
logical groupings referred to as races – Bicolor, Caudatum,
Durra, Guinea, and Kaffir – plus an additional 10 intermediates.
To distinguish among the races, they then developed a scoring
system based heavily on seed shape and inflorescence morphology
(Supporting Information Fig. S1a,b). The Harlan and De Wet
classification receives some support from DNA sequence data
(Morris et al., 2013; Zhang et al., 2015), but it is also clear that
the races overall are not discrete entities, with Bicolor in particu-
lar being especially heterogeneous. A few of the races correspond
to geographic regions; for example, Guinea is most commonly
found in western Africa, and Durra-like plants occur in eastern
Africa and western India (Deu et al., 1994; Morris et al., 2013).
Although several studies have investigated inflorescence architec-
ture in sorghum (Brown et al., 2006; Zhou et al., 2019), these
have not attempted to evaluate the characteristics suggested by
Harlan & De Wet (1972), in part perhaps due to the difficulty in
measuring such features (Fig. S1). Furthermore, the morphologi-
cal diversity among the races has not been investigated rigorously
using contemporary quantitative methods, leaving this an open
area for investigation. In particular, we considered whether high-
dimensional phenotyping and multivariate analysis would mimic
the relationships seen based on genetic background, or if greater
morphological diversity or alternative patterns would be observed
instead.
New Phytologist (2020) 226: 1873–1885
www.newphytologist.com
14698137, 2020, 6, Downloaded from https://nph.onlinelibrary.wiley.com/doi/10.1111/nph.16533, Wiley Online Library on [02/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
New
Phytologist
Materials and Methods
Plant materials and growing conditions
Germplasm representing the Sorghum Association Panel (Casa
et al., 2008) was planted in single-row plots in June 2017 at the
University of Missouri Columbia Bradford Research Center (lati-
tude 38.89°N, longitude 92.20°W). Within each plot from a
single field replicate, a randomly selected panicle was harvested in
October 2017 to be used in analysis. For this study, only acces-
sions whose botanical race assignment was clearly supported by
genetic evidence were considered (Casa et al., 2008). Principal
component analysis (PCA) of the genotypic data presented by
Morris et al. (2013) was used to further confirm accessions that
corresponded best to one botanical race. We chose 55 representa-
tive accessions, comprising 11 Bicolor accessions, 15 Caudatum,
10 Durra, 9 Guinea and 10 Kafir (Table S1). We harvested an
additional panicle from a second field replicate for 34 accessions.
Harvested panicles were placed upright in large boxes to prevent
flattening or overall distortion of panicle shape.
Imaging by X-ray computed tomography
A helical scan using the North Star Imaging X5000 system and the
included eFX-DR software (MathWorks, Natick, MA, USA) was
performed on individual panicles (clamped vertically upright),
with 1000 radiographs generated per sample rotation, for a total
of 6000–8000 radiographs depending on the length of the panicle,
and an average helical pitch of 78.94 mm. The X-ray source was
set to a voltage of 70 kV, current of 1000 µA, and focal spot length
of 70 lm. The radiographs were then combined using the North
Star Imaging eFX-CT software to generate a 3D reconstruction of
each sample, with a final voxel size of 106–109 µm. Three-dimen-
sional reconstructions were then converted to slices for import into
MATLAB and downstream processing (Figs 1a, S2a).
Image processing and feature extraction
To reduce computational time, the margin space and parts of the
stem approximately 1 cm below the first (lowermost) node were
cropped out. The grayscale intensity was then adjusted and stan-
dardized automatically with a marker (a single stem sample used
across all scans) or semi-automatically without the marker
(Fig. S2a). After adjusting and standardizing the grayscale inten-
sity, a fixed and learned threshold (a grayscale value that is gener-
ally larger than air values and smaller than panicle values) was
applied to segment out the sorghum panicle for feature extrac-
tion. The 3D object consisted of voxels each having three coordi-
nates. From this, the eight features PanicleVolume,
PanicleDepth, PanicleMaxWidth, PanicleWidthDepthRatio,
PanicleConvexHullVolume, PanicleSolidity, PanicleElongation,
and PanicleFlatness were calculated to measure the global shape
of the 3D panicle. From the 3D reconstruction, a composite
side-view 2D projected image was generated. From this, nine 2D
panicle features were calculated: Panicle2DArea, Pani-
cle2DMajorAxisLength, Panicle2DMinorAxisLength,
Ó 2020 The Authors
New Phytologist Ó 2020 New Phytologist Trust
 14698137, 2020, 6, Downloaded from https://nph.onlinelibrary.wiley.com/doi/10.1111/nph.16533, Wiley Online Library on [02/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
New
Phytologist Research 1875

(a) X-ray imaging 3D reconstruction 2D projection Seeds Seedless panicle Feature extraction

3D panicle features

2D panicle features

+ + Seed morphology

Seed spatial features

Branch features

(b) Extract seed morphology and seed spatial features

Identify individual seed Quantify individual seed morphology Seed spatial distribution

Fit ellipsoid Radius histogram

+ +

(c) Extract branch features

Seedless panicle Curved skeleton Identify individual primary branch Identify main stalk

Fig. 1 Pipeline to computationally dissect and quantify Sorghum bicolor panicles from three-dimensional (3D) X-ray imaging. (a) From X-ray imaging, a 3D
reconstruction and a composite side-view two-dimensional (2D) projection of the panicle were generated. Each individual seed is then identified and
digitally removed to form a seedless panicle. Five categories of features can be extracted: 3D panicle features, 2D panicle features, seed morphology, seed
spatial features, and branch features. (b) Pipeline for extracting seed profile: based on the intensity of the 3D X-ray imaging, each individual seed can be
captured and quantified by fitting an ellipsoid and computing the histogram of the radius along uniformly and densely distributed directions. Seed spatial
features, such as seed number distributed along the panicle, can be measured by a curve. (c) Skeleton-based pipeline for obtaining branch features:
thresholding the X-ray imaging generates a shape that is thinned to its curve skeleton. Then, geometric algorithms are used to extract the main stalk and
primary branches, and their corresponding measurements.

Ó 2020 The Authors New Phytologist (2020) 226: 1873–1885

New Phytologist Ó 2020 New Phytologist Trust www.newphytologist.com

1876 Research
Panicle2DAspectRatio, Panicle2DConvexHullArea, Pani-
cle2DSolidity, Panicle2DPerimeter, Panicle2DDepth, and Pani-
cle2DCircularity. This 2D projection contains similar
information to that in conventional 2D photograph imaging,
except color (Table S2).
To extract seed morphological features, a fixed and learned
upper threshold value was used to segment the seeds. Any small
piece less than one-eighth of the mean seed volume was treated as
noise and removed. The shape of each seed was also used to
remove outliers, such as very long pieces. When seeds touched
each other, an erosion and dilation step in which seeds were
shrunk and re-expanded (also known as morphological opening;
Fig. S2b) was applied to identify each individual seed. Seed mor-
phological features included SeedNumber, SeedNumber/Depth,
SeedTotalVolume, and SeedAvgVolume. Other seed morphology
features were computed from 100 sampled seeds (Table S2). For
this, each panicle was divided into 10 bins along the rachis, and
within each bin 10 seeds were randomly picked from those whose
volume was larger than the 25th percentile (to avoid measuring
underdeveloped seeds). SeedElongation and SeedFlatness were
measured in two different ways: (1) computing the ratio of the
voxel variation, and (2) fitting an ellipsoid. We also treated the
distance between the ellipsoid and seed (SeedEllipsoidError) as a
seed morphology feature (Table S2). Since seeds are nearly ball-
like shapes, we computed the histogram (SeedshapeRhist1-10) of
the radius along uniformly and densely distributed directions
(Figs 1b, S3). We identified and recorded the centroid position
of each seed, allowing us to calculate the 3D spatial distribution.
The distributions of seed number, total seed volume, and average
seed volume along the panicle are discretized as SeedNum-
bervhist1-10, SeedBiomassvhist1-10, and SeedSizevhist1-10
(Figs 1b, S3). The distances between seeds and the main stalk
permitted computation of three derived traits: MinDis-
tanceSeedMainStalk, AvgDistanceSeedMainStalk and MaxDis-
tanceSeedMainStalk.
Rachis, primary branch, and node features were extracted
through a computational pipeline involving the medial axis. Each
seedless panicle volume was downsampled by 2, and then an inten-
sity threshold was applied in order to eliminate noise and generate
a 3D shape (Fig. S4). A hole-filling algorithm was applied to the
shape in order to consolidate the stem into one single skeleton
curve when the medial axis is computed. The trait
MainStalkDiameter is computed as a by-product during this step.
This shape was thinned to its 2D medial axis using the Voxelcore
method (Yan et al., 2018) and further down to a one-dimensional
(1D), noise-pruned, curve skeleton using the erosion thickness
method (Yan et al., 2016). The resulting skeleton is also equipped
with a radius measure. The stem is identified by applying hysteresis
thresholding of the radius measure to the skeleton to obtain a thick
region, and then finding a high radius path through this region
(Fig. S5). The junctions along the stem form nodes, from which a
search algorithm is applied to identify candidate primary branches
(Fig. S6). From the stem, node, and branch identification steps are
derived seven traits: 1°BranchNumber, 1°BranchNumber/Depth,
First1°BranchLength, 1°BranchAvgLength, LongestIntern-
odeLength, 2ndLongestInternodeLength, and 1°BranchTipAngle
New Phytologist (2020) 226: 1873–1885
www.newphytologist.com
14698137, 2020, 6, Downloaded from https://nph.onlinelibrary.wiley.com/doi/10.1111/nph.16533, Wiley Online Library on [02/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
New
Phytologist
(Table S2). Internodes were defined as the distance between two
adjacent primary branch nodes, regardless of whether these intern-
ode sections are elongated (e.g. panicles with more evenly dis-
tributed nodes, or panicles with clustered nodes resembling
whorls).
Additional information and detailed feature calculations are
described in Methods S1. Raw extracted measurements from
every sample are available in Table S3.
Statistical analysis
For statistical analysis, phenotypic values of replicates were first
averaged for each accession (unless otherwise indicated). For dif-
ferences in means in univariate traits between botanical races, a
Kruskal–Wallis test was first performed; traits with significant
differences between any of the botanical races were then tested
between every pairwise combination of races using a Mann–
Whitney U test. Similarly, for homogeneity of variances between
botanical races in univariate traits, a Bartlett test was first per-
formed; traits with significant differences in variance between any
of the botanical races were then tested between every pairwise
combination of races with the Brown–Forsythe test using the
leveneTest function with center = ‘median’ in the R package CAR
(Fox & Weisberg, 2019). For each statistical test, P-values across
all phenotypes were adjusted together using the Benjamini–
Hochberg method. For distribution (binned) traits, each distri-
bution was treated as a 10-dimensional vector. To determine sta-
tistical significance between the distributions of any two races,
1000 random permutations were run. At each permutation, the
mean vector was calculated for each race, and the distance (Xi,
where i is a single permutation) between mean vectors of
two races was calculated by using the L2 metric. The P-value was
calculated as the proportion of sampled permutations with
Xi > X0, where X0 is the original difference in means between the
two races.
PCA and hierarchical cluster analysis were performed in MAT-
LAB using functions pca and clustergram. The R function cor.mtest
and package CORRPLOT (Wei & Simko, 2017) were used for
Spearman correlation significance tests and correlation matrix
visualization. The function lda in R package MASS (Venables &
Ripley, 2002) was used for linear discriminant analysis (LDA)
with a jackknifed ‘leave-one-out’ cross-validation (LOOCV)
method on a per-accession basis to calculate classification accu-
racy. The function mantel in R package VEGAN (Oksanen et al.,
2018) was used to perform Mantel tests, which compares the cor-
relation between two matrices. This test was used to detect possi-
ble genotype–phenotype associations, with genotype being
represented by the kinship matrix from Shakoor et al. (2016).
For the other classification methods, parameter tuning and
LOOCV were performed using the R package CARET (Kuhn
et al., 2019). For a support vector machine (SVM), the classifica-
tion function within the R package KERNLAB (Karatzoglou et al.,
2004) was used with method = ‘svmRadial’ (which outperformed
‘svmLinear’ or ‘svmPoly’) and a parameter search including
sigma = seq(0.001, 0.1, by = 0.001) and C = 1:30. For random
forest, the classification function within the R package
Ó 2020 The Authors
New Phytologist Ó 2020 New Phytologist Trust

New
Phytologist
RANDOMFOREST (Liaw & Wiener, 2002) was used with a parame-
ter search including mtry = 1:20 and ntree = 1000. For naive
Bayes, the classification function within the R package NAIVEBAYES
(Majka, 2019) was used with a parameter search including
laplace = c(0,1) and usekernel = c(T,F). For k-nearest neighbors,
the classification function within the R package KKNN (Schliep
et al., 2016) was used with a parameter search including
kmax = 1:30, distance = 1:10, and kernel = ‘optimal’.
The k-means clustering for two, three, four, or five clusters was
performed using the base R kmeans function, with the parameter
nstart = 25. Average cluster silhouette values and visualization of
k-means clustering were generated using the fviz_nbclust and
fviz_cluster functions, respectively, from the R package FACTOEX-
TRA (Kassambara & Mundt, 2017).
Workstation, computational time, and code availability
We ran our pipeline on two different workstations. Depending
on the dimension of the image volumes, it took between
c. 20 min and several hours to calculate all the digitally measured
traits, excluding branch traits, on an x64-based PC with an Intel
Xeon E5-2630 CPU at 2.20 GHz. For primary branch and rachis
detection and trait extraction, the running time ranged between 2
and 5 min on an x64-based PC with an Intel Xeon W3690 CPU
at 3.47 GHz.
The full 3D imaging dataset for this work can be downloaded
from: https://www.danforthcenter.org/scientists-research/princ
ipal-investigators/chris-topp/resources. All code used for image
processing, digital feature extraction, and statistical analysis from
this study can be found at the following GitHub repository:
https://github.com/Topp-Roots-Lab/3D-Sorghum-Inflorescence
. the skeleton, were used as starting points for a search algorithm
Results
Three-dimensional imaging and digital phenotyping of
sorghum inflorescences using X-ray tomography
To quantify inflorescence traits and their variation among the five
main botanical races of sorghum, we imaged individual panicles
using XRT (Fig. S1). The average scan time was approximately
10 min per sample, making this approach amenable to moderate
throughput. From the Sorghum Association Panel (Casa et al.,
2008), 34 accessions were selected for imaging with two field
replicates each. Subsequently, an additional 21 accessions were
imaged with a single replicate each, for a total of 55 sorghum
accessions included in this study, consisting of nine or more
accessions from each botanical race. Accessions were chosen solely
based on high genetic support for identity within their respective
botanical races (Table S1), not by any phenotypic criteria,
thereby selecting genotypes that should be unambiguously
among the best genetic representations of each botanical race.
Three-dimensional panicle features, including global shape and
size, were then computed in MATLAB using the image volumes
generated from an automated reconstruction algorithm after
adjusting the grayscale intensity (Fig. S2a). For comparison, a
Ó 2020 The Authors
New Phytologist Ó 2020 New Phytologist Trust
14698137, 2020, 6, Downloaded from https://nph.onlinelibrary.wiley.com/doi/10.1111/nph.16533, Wiley Online Library on [02/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Research 1877
composite side-view 2D projected image was generated from the
3D reconstruction to measure 2D shape and size, features that
more closely resemble traits typically measured in conventional
photographic imaging of sorghum panicles (Fig. 1a).
A major advantage of using XRT for inflorescence phenotyp-
ing is the simultaneous capture of seeds and their 3D shape pro-
file. To identify individual seeds, a threshold (a grayscale value
that is generally larger than branch values and smaller than seed
values) was first used to segment the seeds relative to branch seg-
mentation. When seeds touched each other, seeds were computa-
tionally shrunk and re-expanded to identify a morphological gap
separating individual seeds. Seed morphological features were cal-
culated by various approaches, including fitting an ellipsoid to
the 3D image (see details in the Materials and Methods section).
Since seeds are roughly spherical, we computed the histogram of
the radius along uniformly and densely distributed directions to
extract finer details in seed shape than 2D images can capture
(Figs 1b, S3a). Because the positions of all the seeds were
recorded, we were also able to compute seed spatial features, such
as the distribution of seed number, total seed volume, average
seed volume along the panicle (Figs 1b, S3b), and the distances
between seeds and the main stalk (rachis).
Next, we developed a skeleton-based method based upon the
medial axis (Blum, 1967) to measure features related to the
rachis, primary branches, and nodes after digitally removing the
seeds (seedless panicle; Figs 1c, S7). The thin, homotopy-preserv-
ing properties of the skeleton allow our method to efficiently
obtain morphological branch traits while preserving the topology
of the digital panicle shape; the thickness measure along the
skeleton was then used to identify the main stalk, along which
nodes were identified. The nodes, which form junctions within
to identify primary branches. Branch features were then derived
from the identified stem, nodes, and primary branches.
In total, our methods retrieved 77 inflorescence and seed fea-
tures (37 stand-alone and 40 distributional), more than any other
quantitative study on sorghum inflorescence phenotype to date; a
full list of features, detailed descriptions, and sample measure-
ments are shown in Methods S1 (Tables S2, S3).
Phenotypic validation and relationships
To compare and validate the traits that could be estimated manu-
ally, we measured 10 traits by hand for 20 individual panicle
samples (c. 1.5 h per sample, Methods S1) and calculated the
coefficient of determination R2 with their digital counterparts
(Fig. 2). Three traits commonly and easily measured in breeding
and genetic studies – seed number, panicle depth, and panicle
width – were highly consistent between manual and digital meth-
ods, with R2 values of 0.98, 0.94, and 0.92, respectively. Main
stalk (rachis) diameter had a weaker correspondence between
manual and digital methods (R2 = 0.81), likely due to the rachis
cross-section often being more ellipsoidal or irregular rather than
a perfect circle, creating uncertainty in the manual measurements.
The remaining manually measured features related to primary
branch or internode traits and had R2 values between 0.64 and
New Phytologist (2020) 226: 1873–1885
www.newphytologist.com
 14698137, 2020, 6, Downloaded from https://nph.onlinelibrary.wiley.com/doi/10.1111/nph.16533, Wiley Online Library on [02/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
New
1878 Research Phytologist

Seed number Panicle depth (mm) Panicle width (mm) Main stalk diameter (mm)
R² = 0.98 R² = 0.94 R² = 0.92 14 R² = 0.81
300
3000 150 12
250
10
2000 200 100
8
150
1000 6 RMSEn = 0.05
RMSEn = 0.06 RMSEn = 0.01 50 RMSEn = 0.05
1000 2000 3000 150 200 250 300 50 100 150 6 8 10 12 14

1° Branch number First 1° Branch length (mm) 1° Branch avg length (mm) 1°BranchTipAngle (◦)
Manual traits

100 R² = 0.81 R² = 0.76 150 R² = 0.64 R² = 0.92

150
100
80
100
100
60 60

40 50
50 20
RMSEn = 0.01 RMSEn = 0.11 RMSEn = 0.01 RMSEn = 0.04
20 0
20 40 60 80 100 0 50 100 150 50 100 150 20 60 100

Longest internode length (mm) 2nd Longest internode length (mm)

50 R² = 0.85 40 R² = 0.84
Bicolor
40
Caudatum linear regression line
30
20 Durra
20 Guinea y = x line
10 Kafir
RMSEn = 0.03 0 RMSEn = 0.07
10 20 30 40 50 0 20 40
Digital traits
Fig. 2 Comparison between digitally and manually measured traits in Sorghum bicolor panicles. Linear regression is performed for 10 features, with points
representing individual panicles and colors representing the different botanical races. The white solid line is the regression line, with R2 value and
normalized root-mean-square error (RMSEn) shown. The white dashed line indicates theoretical perfect correspondence.

0.92. Here again, traits that could more easily be measured by
hand (e.g. primary branch tip angle) showed stronger correspon-
dence between manual and digital methods (R2 = 0.92), whereas
more tedious or difficult measurements such as longest internode
length or second longest internode length (where exhaustive
manual searches over the entire rachis are infeasible) showed
lower correspondence, with R2 values of 0.85, and 0.84 respec-
tively. Indeed, most of the digital measurements are likely more
accurate than manual measurements, except for primary branch
number and first primary branch length, where the digital mea-
surement has relatively higher uncertainty due to challenges in
identifying and segmenting all primary branches. Similarly, aver-
age primary branch length showed a lower correspondence
(R2 = 0.64) possibly due to difficulties in both digital and manual
measurements, and manual measurements took the average of
nine random branches as opposed to the larger proportion of
branches measured by the digital method. Overall, however, the
consistency between digital traits and manually measured traits is
high enough to give confidence in the measurements, with most
instances of lower correspondence being explainable by the afore-
mentioned challenges, primarily among manual methods.
Next, we computed the relationship between each pair of the
77 digital traits using Spearman’s rank correlation coefficient q,
with each accession being an observation. From this, we con-
firmed several intuitive or expected correlations, such as between
3D panicle width and 3D panicle convex hull volume (q = 0.95,
P = 1.18 9 10 28), and between panicle flatness and panicle
width (q = 0.49, P = 1.32 9 10 4), which implies flat, wide
panicles are more common than round, broad ones (Fig. S8).
Total seed number was positively correlated with total seed vol-
ume (q = 0.75, P = 3.16 9 10 11), but slightly negatively corre-
lated to average seed volume (q = 0.31, P = 0.02), indicating a
generally inverse relationship between seed number and seed size
in sorghum. On average, seeds also tended to be farther from the
rachis when the panicle is longer (q = 0.71, P = 1.02 9 10 9).
We also found relationships between traits that, though not sur-
prising, are difficult to measure manually, such as between the
longest internode length and average seed-to-rachis distance

New Phytologist (2020) 226: 1873–1885 Ó 2020 The Authors

www.newphytologist.com New Phytologist Ó 2020 New Phytologist Trust
 14698137, 2020, 6, Downloaded from https://nph.onlinelibrary.wiley.com/doi/10.1111/nph.16533, Wiley Online Library on [02/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
New
Phytologist Research 1879

(q = 0.47, P = 3.35 9 10 4), and the inverse correlation between Archetypal panicle morphology for major sorghum
3D panicle solidity and 3D panicle convex hull volume botanical races
(q = 0.83, P = 9.26 9 10 15). Finally, our measurements also
Each of the five major botanical races is commonly associated
indicate that seed shape is generally independent of other panicle
with a stereotypical inflorescence phenotype, sometimes related
features.
to its most prevalent growing conditions (Harlan & De Wet,
Giving confidence to our digital measurements, correlations
1972). For example, Guinea panicles are frequently ‘open’ in
between 3D features and the closest analogous 2D features were
overall architecture and thus presumably more resistant to mold
high, such as in the case of 3D convex hull volume and 2D con-
or pests in wet conditions, whereas Caudatum and Durra pani-
vex area (q = 0.95, P = 5.45 9 10 29), or 3D panicle solidity and
cles are frequently compact due to drier conditions and therefore
2D panicle solidity (q = 0.95, P = 1.78 9 10 29, Fig. S8). A rep-
potentially more amenable to selection purely for high yield. To
resentative variety of topological structures is shown in Fig. S7.
provide a comprehensive empirical evaluation of these associa-
The primary branches and rachis, shown in different colors, are
tions, we aggregated the genotypes by genetically defined botani-
computed using our skeleton-based method. The negative corre-
cal race and compared trait averages and variances (Figs 3, S9) for
lation between primary branch density and panicle depth
the 37 stand-alone traits (i.e. not a distribution across the panicle
(q = 0.62, P = 3.56 9 10 7) is verified by the common observa-
or seed sections). For 14 of these, at least one of the five botanical
tion of panicles that were either short and dense or long and
races was significantly different from the others (Kruskal–Wallis
sparse. Denser primary branches correlated with shorter intern-
test and Mann–Whitney U test, P < 0.05; Table S4). For exam-
ode lengths (q = 0.70, P = 2 9 10 9, between primary branch
ple, within 3D panicle traits, convex hull volume differed signifi-
density and longest internode length), whereas wider panicles
cantly between Guinea and Caudatum or Durra; solidity differed
tend to have greater tip angles and lower primary branch densities
significantly between Durra and Bicolor, Guinea, or Kafir, and
along the stem. This suggests that open branches would result in
between Caudatum and Bicolor or Guinea (Fig. 3). Branch traits
wider and sparser panicles, and that panicles with greater tip
also differed between botanical races; for example, Durra was sig-
angles and branch density also have lower internode lengths.
nificantly different from the other four races in primary branch
Note that our primary branch detection algorithm performs bet-
density, first primary branch length, and longest internode
ter for wider panicles that have fewer intersections between
length. In seed traits, notable differences were identified, includ-
branches.
ing differences in seed number density (Caudatum vs Durra or

Panicle convex hull volume Panicle solidity 1° Branch number 1° Branch avg length Seed number distribution
10 6
ab a a b ab a bc c a ab apex Bicolor abc
100 Caudatum ac
3 0.2 120 Position along panicle as a percentage (%) Durra bc
20 Guinea a
80 100 Kafir b
(mm)
(mm³)

2
0.1 60 80
1 40 40
60

0 0 20

Seed number Seed avg volume Seed flatness Longest internode length
60
a b ab ab a a a b a a
25 50
3000 0.7
20 40
(mm³)

(mm)

0.6 30 80
2000 15
20
10 0.5
1000 10
base
0 200 400 600
D m
G rra

Ka a
da or

fir

D m
G ra

Ka a
da or

fir

D m
G rra

Ka a
da or

fir
D m
G ra

Ka a
da or

fir
ne

ne

ne
ne
tu

tu

tu
au col

ur
au l

tu

au col
ur
au col
C ico
u

u
ui

ui

ui
ui
Bi

Bi
Bi
B
C

C
C

Fig. 3 Distribution and variation for selected traits among the five botanical races in Sorghum bicolor. For boxplots, each dot indicates an outlier beyond
the first/third quartiles  1.5 times the interquartile range. The central horizontal line indicates the median, the bottom and top edges of the box represent
the 25th and 75th percentiles, respectively, and the whiskers extend to the most extreme nonoutlier data point. Each curve in the rightmost column
represents the average seed number distributed along the panicle from apex to the base for each race. The shadow shows the region within one SD. Colors
correspond to races; different letters indicate groups with statistically significant differences at P < 0.05.

Ó 2020 The Authors New Phytologist (2020) 226: 1873–1885

New Phytologist Ó 2020 New Phytologist Trust www.newphytologist.com

1880 Research
Guinea, Durra vs Bicolor or Guinea, and Guinea vs Kafir), seed
flatness (Caudatum vs Bicolor, Durra, or Kafir), and seed ellip-
soid error, which is the distance between the seed and its fitted
ellipsoid (Caudatum vs Bicolor, Guinea or Kafir, and Durra vs
Bicolor, Guinea, or Kafir). Overall, Durra vs Guinea differed in
the greatest number of stand-alone traits (12), followed by Durra
vs Bicolor (10), then Durra vs Kafir or Caudatum vs Guinea
(both eight). Seven traits also differed in variance between at least
one pair among the botanical races (Bartlett test, P < 0.05), and
within these traits all but one had significant differences in vari-
ance between two or more pairs of botanical races (Brown–
Forsythe test, P < 0.05; Table S4). Together, this indicates that
our 3D phenotyping pipeline is sensitive enough to detect fine-
scale differences between the various types of sorghum.
Next, we treated each of the four sets of distributional traits (10
features each) as a vector and tested for significant differences
between botanical races using 1000 random permutations (Figs 3,
S9). For seed number distribution, Kafir is significantly different
from Caudatum or Guinea, and Caudatum and Guinea are differ-
ent from each other (Table S5). For seed biomass distribution,
Bicolor is significantly different from all the others but Guinea;
and Caudatum is different from Kafir. For seed size distribution,
Guinea is significantly different from all the others but Caudatum;
Caudatum is different from Bicolor or Durra. The most striking
result is that all pairs of races, except Durra vs Guinea, differ statis-
tically for the seed shape radius histogram, implying that seed
shape is one of the main morphological features to distinguish
between races. Furthermore, it demonstrates the importance of
employing new 3D imaging techniques and advanced morpho-
metrics for phenotyping, since 3D seed shape is challenging and
time-consuming to quantify by any manual method.
To compare the overall differences in inflorescence traits, we
calculated the mean value of every trait according to botanical
race and performed two-way hierarchical clustering (Fig. 4a).
Cluster analysis identified two major morphological groups made
up of Caudatum plus Durra, and Kafir plus Bicolor and Guinea.
Within the latter group, Kafir and Bicolor were more similar to
each other than either was to Guinea. Four general clusters of
traits could be identified. Two were comprised entirely of seed
traits: one cluster (topmost 15 traits in Fig. 4a) was predomi-
nantly seed spatial features related to size, whereas the other clus-
ter (12 traits, third cluster down in Fig. 4a) was more equally
composed of seed spatial traits and seed morphology traits,
including flatness, shape, and number. However, when the
botanical races are compared, the relative values of traits of the
two seed-dominated clusters were nearly opposite of each other:
Caudatum and Guinea had higher values for the first cluster,
including seed size distribution and seed biomass distribution in
the lower half of the panicle, whereas Durra and Kafir had higher
values for the second cluster with seed number, shape distribu-
tion, and seed biomass distribution in the half top panicle. The
other two clusters of traits were composed of a mix of panicle fea-
tures, branch features, and seed traits, although seed traits still
made up more than half the total for one of these two. This sug-
gests that seed traits can differentiate between genetically defined
botanical races.
New Phytologist (2020) 226: 1873–1885
www.newphytologist.com
14698137, 2020, 6, Downloaded from https://nph.onlinelibrary.wiley.com/doi/10.1111/nph.16533, Wiley Online Library on [02/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
New
Phytologist
Panicle variation within botanical races prevents robust
pedigree-based classification
To assess variation both within and between botanical races
across all phenotypes, we performed PCA using all 77 traits
(Fig. 4b). The first principal component was roughly split across
all types of traits (2D/3D panicle features, seed morphology and
spatial features, and branch features), indicating that all trait
types contribute to the axis of greatest variation (Fig. S10). The
specific traits include panicle 3D depth, volume, and solidity; 2D
area, depth, and solidity; seed number, total volume, distance
between seed to the rachis, and number/biomass distribution;
and first and second longest internode length and primary branch
number/depth. The second principal component, by contrast,
was primarily based on seed morphology and spatial features, and
particularly for individual distribution traits, such as seed shape
radius histogram in bin 2 to bin 5 (Fig. S3). Notably, however,
PCA did not generate obviously distinct clusters; although the
centroids of clusters manually annotated by botanical race did
show relative positions consistent with that of the hierarchical
clustering, all clusters were partially overlapping with each other.
We also did not observe any clusters using nonlinear dimension-
ality reduction techniques such as Isomap (Fig. S11). This lack of
distinct clusters using an unsupervised approach suggests that the
botanical races have no obvious phenotypic boundaries between
one another based on overall morphological trait variation.
We next used supervised approaches: PCA–LDA, SVM, ran-
dom forest, naive Bayes, and k-nearest neighbor. PCA–LDA,
using the values of the first eight principal components from all
77 traits (capturing 84.3% of the total variance) as inputs into an
LDA, performed best (Table S6). Leave-one-out cross-validation
resulted in a 70.9% overall classification accuracy, which, though
higher than expected by random chance, was still relatively mod-
est given our strict genetically based selection process (Table 1).
Unlike with PCA, the LDA loadings were predominantly influ-
enced by seed morphology traits (Fig. S10), especially seed flat-
ness and elongation, and clusters labeled by botanical race were
more distinct, although still clearly intermixed. PCA–LDA was
then performed based on the traits in each feature category (3D/
2D panicle features, seed morphology/spatial features, branch
features), with seed morphology traits performing best (Table 1),
providing another line of evidence that seed morphology domi-
nates the race classification. However, most panicles between the
various botanical races were overlapping in feature space and not
distinct from one another, as PCA–LDA was unable to fully sepa-
rate the races. Additionally, k-means clustering using up to five
clusters showed no evidence of distinct clusters; even the overlap-
ping groups failed to correspond to botanical race (Fig. S12).
We considered whether phenotype reflected genotype inde-
pendently of botanical race by comparing a kinship matrix of all
accessions against distance matrices either based on all 77 traits
or just seed morphology traits (Fig. 5). A Mantel test comparing
the kinship matrix and the total morphological matrix found a
slight but statistically significant correlation (r = 0.1364,
P = 0.002), whereas comparing the kinship matrix with just the
seed morphology matrix had a somewhat stronger, but still
Ó 2020 The Authors
New Phytologist Ó 2020 New Phytologist Trust
 14698137, 2020, 6, Downloaded from https://nph.onlinelibrary.wiley.com/doi/10.1111/nph.16533, Wiley Online Library on [02/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
New
Phytologist Research 1881

(a) 3D panicle feature (3D) (b)

2D panicle feature (2D)

Caudatum
seed morphology (SM) 8

Guinea
Bicolor
seed spatial feature (SS)

Durra

Kafir
branch feature (B) 6
σ
SeedSizevhist1 4
SeedSizevhist2
SeedSizevhist3
0 SeedSizevhist4 SM 2
SeedAvgVolume

PC2 (17.95%)
13%
SeedSizevhist5
SeedSizevhist6 0
SeedSizevhist7
-σ SeedSizevhist8 87%
SeedSizevhist9 –2
SeedSizevhist10
SeedshapeRhist4 SS
SeedBiomassvhist8 –4
SeedBiomassvhist10
SeedBiomassvhist9
SeedNumbervhist10 –6
SeedNumbervhist9
PanicleFlatness
SeedNumbervhist8
SeedshapeRhist9 –8
SeedshapeRhist1
SeedshapeRhist2
Panicle2DAspectRatio 3D –10
SeedshapeRhist7 2D 13% B
SeedEllipsoidError 10% 10%
SeedshapeRhist3 –10 –5 0 5 10
SeedshapeRhist8
PanicleElongation 35% 32%
MainStalkDiameter PC1 (21.52%)
1°BranchNumber SS
1°BranchNumber/Depth SM Bicolor Caudatum Durra Guinea Kafir
SeedNumbervhist5
SeedNumber/Depth
SeedNumbervhist6 (c) predicted race
SeedNumbervhist7 0.8
Panicle2DCircularity 70.9% classification

actual race
PanicleSolidity 0.6
Panicle2DSolidity accuracy rate
SeedElongation2 0.4
SeedElongation1
SeedBiomassvhist7 0.2
SeedBiomassvhist6
SeedBiomassvhist5 0
PanicleVolume 3
SeedTotalVolume
SeedBiomassvhist4
SeedBiomassvhist3
SeedBiomassvhist2
SeedBiomassvhist1 SM 2
SeedNumbervhist3
SeedNumbervhist2
SeedNumbervhist1 42%
SeedNumbervhist4 58%
SeedNumber 1
SeedshapeRhist10 SS
SeedFlatness1
LD2

SeedFlatness2
SeedshapeRhist6 0
SeedshapeRhist5
First1°BranchLength
AvgDistanceSeed-MainStalk
MinDistanceSeed-MainStalk
Panicle2DMajorAxisLength
MaxDistanceSeed-MainStalk –1
Panicle2DDepth
2ndLongestInternodeLength
LongestInternodeLength 3D
PanicleDepth 2D 21% –2
Panicle2DConvexHullArea B
PanicleMaxWidth
PanicleConvexHullVolume 32% 26%
Panicle2DPerimeter
1°BranchAverageLength 5%
16% –3
Panicle2DMinorAxisLength SM SS
1°BranchTipAngle
Panicle2DArea –4 –3 –2 –1 0 1 2 3 4
PanicleWidthDepthRatio
LD1
Fig. 4 Two-way hierarchical cluster analysis, principal component analysis (PCA), and linear discriminant analysis (LDA). (a) Two-way hierarchical cluster
analysis based on the mean value of morphological features (rows) among the five botanical races of Sorghum bicolor (columns). The heat map indicates
an increase (red) or a decrease (blue) with respect to the overall mean value of five races. The features are clustered into four main groups. The pie charts
show the composition of feature categories for each group. (b) PCA plot of all morphological features with 85% confidence ellipses. The percentage
variance for each principal component (PC) is shown in parentheses. Colors indicate botanical race. (c) LDA plot on the first eight PCs (84.3% variance)
with 85% confidence ellipses. Colors the same as in (b). The confusion matrix for predicted species is shown in the upper left corner. Jackknifed ‘leave-one-
out’ cross-validation found 70.9% classification accuracy.

relatively low, correlation (r = 0.2333, P = 0.001). To determine
whether this can be affected by feature selection, we performed
similar comparisons using increasingly stringent subsets of traits
for constructing the morphological matrix, based on the coeffi-
cient of determination R2 between accession replicates
(Table S7). Phenotypic differences among accession replicates
may be due to innate biological (i.e. developmental) variation, as
well as environmental variation present within the field sampling.
However, the kinship-to-morphology Mantel test was not
significantly changed either from removing the features that had
lower between-replicate consistency or by repeating the compar-
ison using only replicate 1 or replicate 2 phenotypic data
(Table S8). In conjunction with the results of PCA and LDA, this
result indicates that morphological differences between either the
botanical races or overall genetic background are quantitative
rather than qualitative, and relatively modest. Instead, panicle
morphology is more accurately described as a continuum of trait
variation without obvious categorical distinctions.

Ó 2020 The Authors New Phytologist (2020) 226: 1873–1885

New Phytologist Ó 2020 New Phytologist Trust www.newphytologist.com
 14698137, 2020, 6, Downloaded from https://nph.onlinelibrary.wiley.com/doi/10.1111/nph.16533, Wiley Online Library on [02/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
New
1882 Research Phytologist

Table 1 Results of principal component analysis–linear discriminant analysis (PCA–LDA) classification for the five botanical races in Sorghum bicolor using
different categories of features, and Mantel tests for correlation between genotype and phenotype.

Discriminant analysis Mantel test with kinship

Trait category No. of traits 8 PCs variance (%) Accuracy rate (%) Correlation r P-value

All features 77 84.3 70.91 0.1364 0.002

3D panicle features 8 100 29.09 0.005127 0.503
2D panicle features 9 99.94 34.55 0.02546 0.227
Seed morphology 19 97.16 54.55 0.2333 0.001
Seed spatial features 33 95.14 45.45 0.0966 0.013
Branch features 8 100 36.36 0.03337 0.203

(a) Kinship matrix (b) All morphological distance matrix (c) Seed morphological distance matrix

Bicolor 2.5 20
10
18
2
Caudatum 16
8
14
1.5
12
Durra 6

1 10

8 4
Guinea
0.5 6

4 2
Kafir 0
2

0 0

Mantel test with – kinship matrix, r = 0.1364, P = 0.002

Mantel test with – kinship matrix, r = 0.2333, P = 0.001
Fig. 5 Mantel test comparing genotype and phenotype, with respect to Sorghum bicolor race. (a) The kinship coefficient matrix, created based on marker
data (identity by state). (b) The pairwise distance matrix, based on all morphological features. (c) The pairwise distance matrix, based on seed
morphological features. Mantel tests compared the (a) negative value of the kinship matrix against (b) and (c) distance matrices with r = 0.1364 and
P = 0.002 and with r = 0.2333 and P = 0.001, respectively.

in orders of branching or internode elongation, which we have

Discussion
Inflorescence architecture in grasses directly controls the number
of spikelets, and therefore the number of seeds. In addition to its
economic and ecological importance, inflorescence form has long
been used to distinguish grasses from each other, and is a major
character for distinguishing genera (Barkworth et al., 2003; Wu
et al., 2006, 2007; Barkworth et al., 2007). Unsurprisingly, grass
inflorescence architecture has been the subject of numerous stud-
ies and reviews, addressing both morphological diversity (Vegetti
& Anton, 1995; Perreta et al., 2009; Reinheimer et al., 2009;
Kellogg et al., 2013; Pilatti et al., 2019) and its genetic control
(Bommert et al., 2005; Malcomber et al., 2006; Kellogg, 2007;
Bartlett & Thompson, 2014; Kyozuka, 2014; Whipple, 2017;
Koppolu & Schnurbusch, 2019). Diversity among grass inflores-
cences can be summarized by several parameters: the number of
orders of branching, the number of branches at each order,
whether the inflorescence axis terminates in a spikelet, and the
number of spikelets per branch (Kellogg, 2000; Doust et al.,
2005). In addition to variation in numbers of orders of branch-
ing, grass inflorescences vary in the length of the internodes.
Internode elongation occurs late in development and is thus sepa-
rable from branching pattern. Therefore, relatively small changes
captured with XRT, can lead to inflorescences that look substan-
tially different from each other.
Despite the rich literature describing inflorescence architecture
in rice and maize, sorghum has received much less attention,
either developmentally or genetically, although important recent
papers include those by Brown et al. (2006), Morris et al. (2013),
Hmon et al. (2014), and Zhou et al. (2019). Sorghum has three
orders of branching before the spikelet pairs, and the primary
branches form in a spiral phyllotaxis. However, the number of
branches at order of branching, and the extent of elongation of
internodes, is quite variable and challenging to capture in a quan-
titative way. Instead, the architecture of sorghum panicles is typi-
cally described using intuitive but imprecise and qualitative terms
(De Wet, 1978).
Using recent advances in 3D phenotyping, we analyzed a
diverse range of sorghum panicles, chosen solely based on their
genetic background with no prescreening based on phenotype to
bias the outcome. Though differences in the average panicle mor-
phology between botanical races were detectable, multivariate
analysis shows that panicle architecture is indeed continuous, and
relying on discrete class descriptors should be avoided in situations
where specificity is important. This morphological continuum

New Phytologist (2020) 226: 1873–1885 Ó 2020 The Authors

www.newphytologist.com New Phytologist Ó 2020 New Phytologist Trust

New
Phytologist
could reflect local genetic introgression rather than broader
genomic identity corresponding to race, potentially due to the
complex history of breeding and local adaptation in domesticated
sorghum. Although 2D image analysis may have come to a simi-
lar conclusion for some overall shape features (such as panicle
area), our analysis using XRT 3D imaging additionally demon-
strates a phenotypic gradient even for specific inflorescence fea-
tures (such as branch lengths) that cannot be reliably measured
using 2D imaging due to the occlusion of branches, which pre-
vents accurate representation of the panicle branching hierarchy.
When compared with manual branch phenotyping methods,
our skeleton-based pipeline calculates morphological traits from
branches that more completely and accurately represent each
panicle. Our computational method nondestructively captures
branches evenly distributed throughout the panicle, as opposed
to hand-picked branches that may be biased. By using the
thinned, 1D curve structure of the skeleton, the topology of the
shape and the semi-automatic annotation of the stem and pri-
mary branches are more amenable for characterization, while
retaining information of value. An additional advantage of digital
characterization is improved objectivity, as manual measurement
of traits such as branch angle, selection of the nine primary
branches for average lengths, and node distances can be subjec-
tive. Although digital methods also need individual users’ choice
of parameters, the space of such choices is significantly smaller
than those in manual measurements.
In panicles with large tip angles, the longest-path approach of
our skeleton-based method performs exceptionally well because
of the lack of intersections between different branches within the
shape. However, for compact panicles, the path of a single branch
may appear to intersect with several others in the image. This
may cause the detected branch to wind between several actual
branches. Although branch length, curvature, tortuosity, and tip
angle constrain the predicted branch path to be approximately
correct, small deviations will inevitably cause parts of multiple
branches to be included in the same path. In future work, we plan
to explore methods from combinatorial optimization to better
address such topological issues, potentially by clustering ‘branch-
like’ parts of the skeletons together.
Another advantage of XRT imaging for inflorescence pheno-
typing is that all the seeds within a panicle can be segmented and
isolated with their entire 3D shape, a task that would otherwise
be laborious, if not impossible, by optical imaging without the
aid of robotics. Our analysis found that seed shape and size are
among the more distinctive traits of the sorghum botanical races,
thus illustrating the potential for high-resolution mapping of seed
traits. Additionally, within the genotypes analyzed, we found
some support for a trade-off between seed number and seed size,
a long-standing concept in biology that is often difficult to con-
firm empirically due to confounding factors (Smith & Fretwell,
1974; Paul-Victor & Turnbull, 2009). However, because XRT
imaging relies on object density, information such as the color of
the seeds or glumes is not captured. Although there is little evi-
dence that seed color is associated with botanical race, it may be
of interest in other applications, such as predicting seed tannin
content. On the other hand, glume morphology and the presence
Ó 2020 The Authors
New Phytologist Ó 2020 New Phytologist Trust
14698137, 2020, 6, Downloaded from https://nph.onlinelibrary.wiley.com/doi/10.1111/nph.16533, Wiley Online Library on [02/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Research 1883
or absence of awns may be useful for distinguishing between
sorghum botanical races (Harlan & De Wet, 1972). As they pre-
sent special challenges for segmentation, we did not include them
for analysis here, although future efforts may be able to incorpo-
rate these features.
Aside from these limitations, sorghum inflorescence phenotyp-
ing via XRT represents a more powerful, objective, and quantita-
tive way of measuring panicle features compared with previous
approaches. From the combined methods yielding 77 total inflo-
rescence traits, we find interesting relationships between numer-
ous features, but ultimately no compelling evidence that panicle
architecture reflects botanical race. Indeed, in sorghum, botanical
race is not a formal taxonomic rank and has no standing in a Lin-
naean classification, despite some support for genetic separation.
In any case, our results suggest that the complex history and
interbreeding of sorghum germplasm has resulted in panicle
architecture overall being mostly dissociated from botanical race.
This is potentially advantageous for applications such as genome-
wide association studies, in which population structure would be
predicted to have less of an impact on genetic mapping. Likewise,
we found no support for alternative categorical schemes (such as
the common classification of panicle shapes into ‘open’, ‘closed’,
‘intermediate’, etc.) across our samples, indicating that, despite
their convenience, they do not accurately reflect the quantitative
nature of inflorescence variation present in sorghum. Future
studies can be scaled to diversity panels or other mapping popula-
tions with reasonable expectations for finding potentially novel
quantitative trait loci controlling the myriad of sorghum inflores-
cence features we have described here.
Acknowledgements
We thank Keith Duncan for guidance with XRT and Greg
Ziegler for sharing the kinship matrix data. This project was sup-
ported by NSF grants IOS-1638507 to CNT, DBI-1759796 to
CNT and TJ, RI-1618685 to TJ, and DEB-1457748 and IOS-
1822330 to EAK, and NIH grant CA233303-1 to TJ. DZ was
supported in part by an Imaging Sciences Pathway fellowship
from WUSTL. Plant growth and field support were provided by
Felix Fritschi at University of Missouri Columbia and Todd
Mockler at the DDPSC.
Author contributions
CNT, EAK and TJ designed the experiment. M-RS performed
sample collection and imaging. ML, DZ and M-RS performed
feature extraction and analysis. M-RS, ML and DZ wrote the
manuscript with contributions from CNT, EAK and TJ. ML
and M-RS contributed equally to this work.
ORCID
Tao Ju https://orcid.org/0000-0002-1848-1012
Elizabeth A. Kellogg https://orcid.org/0000-0003-1671-7447
Mao Li https://orcid.org/0000-0002-5964-1764
Mon-Ray Shao https://orcid.org/0000-0003-0773-7024
New Phytologist (2020) 226: 1873–1885
www.newphytologist.com

1884 Research
Christopher N. Topp https://orcid.org/0000-0001-9228-
6752
Dan Zeng https://orcid.org/0000-0002-2636-3729
References
Barkworth ME, Capels KM, Long S, Anderton LK. 2007. Magnoliophyta:
Commelinidae (in part): Poaceae (part 1). In: Flora of North America Editorial
Committee, eds. Flora of North America, vol. 24. New York, NY, USA: Oxford
University Press, 1–908.
Barkworth ME, Capels KM, Long S, Piep M. 2003. Magnoliophyta:
Commelinidae (in part): Poaceae (part 2). In: Flora of North America Editorial
Committee, eds. Flora of North America, vol. 25. New York, NY, USA: Oxford
University Press, 1–781.
Bartlett ME, Thompson B. 2014. Meristem identity and phyllotaxis in
inflorescence development. Frontiers in Plant Science 5: e508.
Bell AD. 2008. Plant form: an illustrated guide to flowering plant morphology.
Portland, OR, USA: Timber Press Inc.
Blum H. 1967. A transformation for extracting new descriptors of form. In:
Wathen-Dunn W, ed. Models for the perception of speech and visual form.
Cambridge, MA, USA: MIT Press, 362–380.
Bommert P, Satoh-Nagasawa N, Jackson D, Hirano HY. 2005. Genetics and
evolution of inflorescence and flower development in grasses. Plant and Cell
Physiology 46: 69–78.
Bougourd S, Marrison J, Haseloff J. 2000. Technical advance: an aniline blue
staining procedure for confocal microscopy and 3D imaging of normal and
perturbed cellular phenotypes in mature Arabidopsis embryos. The Plant
Journal 24: 543–550.
Brown PJ, Klein PE, Bortiri E, Acharya CB, Rooney WL, Kresovich S. 2006.
Inheritance of inflorescence architecture in sorghum. TAG. Theoretical and
Applied Genetics. 113: 931–942.
Buzgo M, Soltis PS, Soltis DE. 2004. Floral developmental morphology of
Amborella trichopoda (Amborellaceae). International Journal of Plant Sciences
165: 925–947.
Casa AM, Pressoir G, Brown PJ, Mitchell SE, Rooney WL, Tuinstra MR,
Franks CD, Kresovich S. 2008. Community resources and strategies for
association mapping in Sorghum. Crop Science 48: 30–40.
Chen D, Neumann K, Friedel S, Kilian B, Chen M, Altmann T, Klukas C.
2014. Dissecting the phenotypic components of crop plant growth and
drought responses based on high-throughput image analysis. The Plant Cell 26:
4636–4655.
De Wet JMJ. 1978. Systematics and evolution of Sorghum sect. Sorghum
(Gramineae). American Journal of Botany 65: 477–484.
Denk W, Horstmann H. 2004. Serial block-face scanning electron microscopy to
reconstruct three-dimensional tissue nanostructure. PLoS Biology 2: e329.
Deu M, Gonzalez-de-Leon D, Glaszmann JC, Degremont I, Chantereau J,
Lanaud C, Hamon P. 1994. RFLP diversity in cultivated sorghum in relation
to racial differentiation. TAG. Theoretical and Applied Genetics. 88: 838–844.
Dhondt S, Vanhaeren H, Van Loo D, Cnudde V, Inze D. 2010. Plant structure
visualization by high-resolution X-ray computed tomography. Trends in Plant
Science 15: 419–422.
Doust AN, Devos KM, Gadberry M, Gale MD, Kellogg EA. 2005. The genetic
basis for inflorescence variation between foxtail and green millet (Poaceae).
Genetics 169: 1659–1672.
Esau K. 1977. Anatomy of seed plants. New York, NY, USA: John Wiley & Sons Inc.
Fox J, Weisberg S. 2019. An R companion to applied regression, 3rd edn. Thousand
Oaks, CA, USA: SAGE Publications.
Friedli M, Kirchgessner N, Grieder C, Liebisch F, Mannale M, Walter A. 2016.
Terrestrial 3D laser scanning to track the increase in canopy height of both
monocot and dicot crop species under field conditions. Plant Methods. 12: e9.
Gehan MA, Kellogg EA. 2017. High-throughput phenotyping. American Journal
of Botany 104: 505–508.
Goddard TD, Huang CC, Meng EC, Pettersen EF, Couch GS, Morris JH,
Ferrin TE. 2017. UCSF CHIMERA: meeting modern challenges in visualization
and analysis. Protein Science. 27: 14–25.
New Phytologist (2020) 226: 1873–1885
www.newphytologist.com
14698137, 2020, 6, Downloaded from https://nph.onlinelibrary.wiley.com/doi/10.1111/nph.16533, Wiley Online Library on [02/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
New
Phytologist
Harlan JR, De Wet JMJ. 1971. Toward a rational classification of cultivated
plants. Taxon 20: 509–517.
Harlan JR, De Wet JMJ. 1972. A simplified classification of cultivated sorghum.
Crop Science 12: 172–176.
Hmon KPW, Shehzad T, Okuno K. 2014. QTLs underlying inflorescence
architecture in sorghum (Sorghum bicolor (L.) Moench) as detected by
association analysis. Genetic Resources and Crop Evolution 61: 1545–1564.
Hughes N, Askew K, Scotson CP, Williams K, Sauze C, Corke F, Doonan JH,
Nibau C. 2017. Non-destructive, high-content analysis of wheat grain traits
using X-ray micro computed tomography. Plant Methods 13: e76.
Jiang N, Floro E, Bray AL, Laws B, Duncan KE, Topp CN. 2019. Three-
dimensional time-lapse analysis reveals multiscale relationships in maize root
systems with contrasting architectures. The Plant Cell 31: 1708–1722.
Karatzoglou A, Smola A, Hornik K, Zeileis A. 2004. KERNLAB – an S4 package
for kernel methods in R. Journal of Statistical Software 11: 1–20.
Kassambara A, Mundt F. 2017. FACTOEXTRA: extract and visualize the results of
multivariate data analyses. R package v.1.0.5. [WWW document] URL https://
CRAN.R-project.org/package=factoextra [accessed 6 April 2020].
Kellogg EA. 2000. The grasses: a case study in macroevolution. Annual Review of
Ecology and Systematics 31: 217–238.
Kellogg EA. 2007. Floral displays: genetic control of grass inflorescences. Current
Opinion in Plant Biology 10: 26–31.
Kellogg EA, Camara PE, Rudall PJ, Ladd P, Malcomber ST, Whipple CJ,
Doust AN. 2013. Early inflorescence development in the grasses (Poaceae).
Frontiers in Plant Science 4: e250.
Koppolu R, Schnurbusch T. 2019. Developmental pathways for shaping spike
inflorescence architecture in barley and wheat. Journal of Integrative Plant
Biology 61: 278–295.
Kuhn M, Wing J, Weston S, Williams A, Keefer C, Engelhardt A, Cooper
T, Mayer Z, Kenkel B, Core Team R, Benesty M, Lescarbeau R, Ziem A,
Scrucca L, Tang Y, Candan C, Hunt T. 2019. CARET: classification and
regression training. R package v.6.0-84. [WWW document] URL https://
CRAN.R-project.org/package=caret [accessed 6 April 2020].
Kyozuka J. 2014. Grass inflorescence: basic structure and diversity. Advances in
Botanical Research 72: 191–219.
Li M, Duncan K, Topp CN, Chitwood DH. 2017. Persistent homology and the
branching topologies of plants. American Journal of Botany 104: 349–353.
Li M, Klein LL, Duncan KE, Jiang N, Chitwood DH, Londo J, Miller AJ, Topp
CN. 2019. Characterizing 3D inflorescence architecture in grapevine using X-
ray imaging and advanced morphometrics: implications for understanding
cluster density. Journal of Experimental Botany 70: 6261–6276.
Liaw A, Wiener M. 2002. Classification and regression by RANDOMFOREST. R
News 2: 18–22.
Majka M. 2019. NAIVEBAYES: high performance implementation of the naive Bayes
algorithm in R. R package v.0.9.6. [WWW document] URL https://CRAN.R-
project.org/package=naivebayes [accessed 6 April 2020].
Malambo L, Popescu SC, Murray SC, Putman E, Pugh NA, Horne DW,
Richardson G, Sheridan R, Rooney WL, Avant R et al. 2018. Multitemporal
field-based plant height estimation using 3D point clouds generated from small
unmanned aerial systems high-resolution imagery. International Journal of
Applied Earth Observation and Geoinformation 64: 31–42.
Malcomber ST, Preston JC, Reinheimer R, Kossuth J, Kellogg EA. 2006.
Developmental gene evolution and the origin of grass inflorescence diversity.
Advances in Botanical Research 44: 425–481.
Morris GP, Ramu P, Deshpande SP, Hash CT, Shah T, Upadhyaya HD, Riera-
Lizarazu O, Brown PJ, Acharya CB, Mitchell SE et al. 2013. Population
genomic and genome-wide association studies of agroclimatic traits in
sorghum. Proceedings of the National Academy of Sciences, USA 110: 453–458.
Oksanen J, Blanchet FG, Friendly M, Kindt R, Legendre P, McGlinn D,
Minchin PR, O’Hara RB, Simpson GL, Solymos Pet al. 2018. VEGAN:
community ecology package. R package v.2.5-3. [WWW document] URL https://
CRAN.R-project.org/package=vegan [accessed 6 April 2020].
Paul-Victor C, Turnbull LA. 2009. The effect of growth conditions on the seed
size/number trade-off. PLoS ONE 4: e6917.
Perez-Torres E, Kirchgessner N, Pfeifer J, Walter A. 2015. Assessing potato
tuber diel growth by means of X-ray computed tomography. Plant, Cell &
Environment 38: 2318–2326.
Ó 2020 The Authors
New Phytologist Ó 2020 New Phytologist Trust
 14698137, 2020, 6, Downloaded from https://nph.onlinelibrary.wiley.com/doi/10.1111/nph.16533, Wiley Online Library on [02/02/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
New
Phytologist Research 1885

Perreta MG, Ramos JC, Vegetti AC. 2009. Development and structure of the Fig. S1 Examples of sorghum panicle morphology.
grass inflorescence. Botanical Review 75: 377–396.
Pettersen EF, Goddard TD, Huang CC, Couch GS, Greenblatt DM, Meng EC,
Ferrin TE. 2004. UCSF CHIMERA – a visualization system for exploratory
Fig. S2 Additional details on X-ray imaging workflow.
research and analysis. Journal of Computational Chemistry 25: 1605–1612.
Pilatti V, Muchut SE, Uberti-Manassero NG, Vegetti AC, Reinheimer R. 2019. Fig. S3 Diagrammatic description of seed distribution features.
Comparative study of the inflorescence, spikelet and flower development in
species of Cynodonteae (Chloridoideae, Poaceae). Botanical Journal of the Fig. S4 Thresholding considerations with primary branches.
Linnean Society 189: 353–377.
Prenner G, Rudall PJ. 2007. Comparative ontogeny of the cyathium in
Euphorbia (Euphorbiaceae) and its allies: exploring the organ–flower– Fig. S5 Stem identification process.
inflorescence boundary. American Journal of Botany 94: 1612–1629.
Reinheimer R, Zuloaga FO, Vegetti A, Pozner R. 2009. Diversification of the Fig. S6 Branch identification process.
inflorescence development in the PCK clade (Poaceae: Panicoideae: Paniceae).
American Journal of Botany 96: 549–564.
Rogers ED, Monaenkova D, Mijar M, Nori A, Goldman DI, Benfey PN. 2016.
Fig. S7 Primary branches and rachises of representative sorghum
X-ray computed tomography reveals the response of root system architecture to panicles.
soil texture. Plant Physiology 171: 2028–2040.
Schliep K, Hechenbichler K, Lizee A. 2016. KKNN: weighted k-nearest neighbors. Fig. S8 Trait correlation plot.
R package v.1.3.1. [WWW document] URL https://CRAN.R-project.org/pac
kage=kknn [accessed 6 April 2020].
Shakoor N, Ziegler G, Dilkes BP, Brenton Z, Boyles R, Connolly EL, Kresovich
Fig. S9 Distribution and variation for all traits among the five
S, Baxter I. 2016. Integration of experiments across diverse environments races.
identifies the genetic determinants of variation in Sorghum bicolor seed element
composition. Plant Physiology 170: 1989–1998. Fig. S10 Principal component analysis and linear discriminant
Smith CC, Fretwell SD. 1974. The optimal balance between size and number of analysis loadings.
offspring. American Naturalist 108: 499–506.
Snowden JD. 1936. The cultivated races of Sorghum. London, UK: Allard and Son.
Snowden JD. 1955. The wild fodder sorghums of the section Eu-sorghum. Fig. S11 Isometric feature mapping with all traits.
Journal of the Linnean Society of London, Botany 55: 191–260.
Stuppy WH, Maisano JA, Colbert MW, Rudall PJ, Rowe TB. 2003. Three- Fig. S12 K-mean clustering.
dimensional analysis of plant structure using high-resolution X-ray computed
tomography. Trends in Plant Science 8: 2–6.
Vegetti AC, Anton A. 1995. Some evolution trends in the inflorescence of
Methods S1 Additional methods on primary branch and rachis
Poaceae. Flora 190: 225–228. trait extraction, skeleton generation, stem identification, primary
Venables WN, Ripley BD. 2002. Modern applied statistics with S. New York, NY, branch identification and measurements, and manual panicle
USA: Springer-Verlag. measurements.
Vollbrecht E, Springer PS, Goh L, Buckler ES, Martienssen R. 2005.
Architecture of floral branch systems in maize and related grasses. Nature 436:
1119–1126.
Table S1 Germplasm and genotypes.
Wei T, Simko V.2017. R package ‘CORRPLOT’: visualization of a correlation matrix
(v.0.84). [WWW document] URL https://github.com/taiyun/corrplot Table S2 Trait descriptions.
[accessed 6 April 2020].
Whipple CJ. 2017. Grass inflorescence architecture and evolution: the origin of Table S3 Complete trait value data.
novel signaling centers. New Phytologist 216: 367–372.
Wu Z, Raven PH, Hong D, eds. 2006. Flora of China. Beijing, China: Science
Press. Table S4 P-values for univariate traits.
Wu Z, Raven PH, Hong D, eds. 2007. Flora of China illustrations. Beijing,
China: Science Press. Table S5 P-values for distribution traits.
Yan Y, Letscher D, Ju T. 2018. Voxel cores: efficient, robust, and provably good
approximation of 3D medial axes. ACM Transactions on Graphics 37: e44.
Yan Y, Sykes K, Chambers E, Letscher D, Ju T. 2016. Erosion thickness on
Table S6 Classification results from supervised methods.
medial axes of 3D shapes. ACM Transactions on Graphics 35: e38.
Zhang D, Li J, Compton RO, Robertson J, Goff VH, Epps E, Kong W, Kim C, Table S7 Coefficient of determination between replicates.
Paterson AH. 2015. Comparative genetics of seed size traits in divergent cereal
lineages represented by sorghum (Panicoidae) and rice (Oryzoidae). G3: Genes, Table S8 Kinship-to-morphology Mantel test.
Genomes, Genetics 5:1117–1128.
Zhou Y, Srinivasan S, Mirnezami SV, Kusmec A, Fu Q, Attigala L, Salas
Fernandez MG, Ganapathysubramanian B, Schnable PS. 2019. Please note: Wiley Blackwell are not responsible for the content
Semiautomated feature extraction from RGB images for sorghum panicle or functionality of any Supporting Information supplied by the
architecture GWAS. Plant Physiology 179: 24–37. authors. Any queries (other than missing material) should be
directed to the New Phytologist Central Office.
Supporting Information
Additional Supporting Information may be found online in the
Supporting Information section at the end of the article.

Ó 2020 The Authors New Phytologist (2020) 226: 1873–1885

New Phytologist Ó 2020 New Phytologist Trust www.newphytologist.com