High-Throughput Measurements of Stem Characteristics to
Estimate Ear Density and Above-Ground Biomass
Xiuliang Jin1,2,⋆, Simon Madec1, Dan Dutartre3,BenoitdeSolan
4, Alexis Comar3,and
1INRA EMMAH, UMR 1114 228 route de l’A´
erodrome, 84914 Avignon, France
2Institute of Crop Sciences, Chinese Academy of Agricultural Sciences/Key Laboratory of Crop Physiology and Ecology, Ministry of
Agriculture, Beijing 100081, China
3HIPHEN, Rue Charrue, 84000 Avignon, France
4ARVALIS-Institut du V´
etal, Station Exp´
erimentale, 91720 Boigneville, France
⋆Correspondence should be addressed to Xiuliang Jin; jinxiuxiuliang@.com
Received March ; Accepted May
Copyright © Xiuliang Jin et al. Exclusive Licensee Nanjing Agricultural University. Distributed under a Creative Commons
Attribution License (CC BY .).
Total above-ground biomass at harvest and ear density are two important traits that characterize wheat genotypes. Two experiments
were carried out in two dierent sites where several genotypes were grown under contrasted irrigation and nitrogen treatments.
A high spatial resolution RGB camera was used to capture the residual stems standing straight aer the cutting by the combine
machine during harvest. It provided a ground spatial resolution better than . mm. A Faster Regional Convolutional Neural
Network (Faster-RCNN) deep-learning model was rst trained to identify the stems cross section. Results showed that the
identication provided precision and recall close to %. Further, the balance between precision and recall allowed getting accurate
estimates of the stem density with a relative RMSEclose to % and robustness across the two experimental sites. e estimated stem
density was also compared with the ear density measured in the eld with traditional methods. A very high correlation was found
with almost no bias, indicating that the stem density could be a good proxy of the ear density. e heritability/repeatability evaluated
over genotypes in one of the two experiments was slightly higher (%) than that of the ear density (%). e diameter of each
stem was computed from the prole of gray values in the extracts of the stem cross section. Results show that the stem diameters
follow a gamma distribution over each microplot with an average diameter close to . mm. Finally, the biovolume computed as the
product of the average stem diameter, the stem density, and plant height is closely related to the above-ground biomass at harvest
with a relative RMSE of %. Possible limitations of the ndings and future applications are nally discussed.
Ear density (the numbers of ears per m2) is generally
well correlated with above-ground biomass and grain yield
at maturity of wheat [,]. However, the correlation may
depend on environmental conditions as well as genotypes.
Most stems observed at harvest bear an ear: stem density
(the number of stems per m2) appears thus as a good proxy
of the ear density . Stem density depends thus both on
plant density and on the number of stems per plant which
is quantied by the tillering coecient. e environmental
conditions experienced by the crop and the genotype control
the tillering coecient . erefore, several studies report
the interest of the ear and stem density as traits to be used
plant height and stem diameter are highly correlated with
the above-ground biomass in wheat [–]. erefore, stem
density, ear density, plant height, and stem diameter are thus
highly desired to score the performances of a genotype in
wheat crop breeding programs.
e number of stems per plant is dicult to evaluate
when plants start to produce tillers since plants are oen
intricated and hardly identiable. Further, the number of
stems per plant may change with time due to possible tiller
regression during tillering and stem elongation stages. Aer
the owering stage, most stems bear an ear and the stem
density therefore provides a good proxy of the ear density.
Ear and stem densities are therefore usually measured at
maturity by manual counting over a given sample area. e
stem diameter is rarely measured since it is very tedious
and time consuming. Similarly, above-ground biomass is
rarely measured extensively for the same reasons. Crop height
at harvest is most frequently measured in the eld using
a ruler. In addition to the limits of these low-throughput
invasive measurements that require large human resources
to be completed, the small sampling area used and errors
associated with the manual measurements may result in
signicant uncertainties on these variables that would limit
experimental observations. It appears therefore necessary
to develop new methods for accurate measurements of the
stem density, crop height, and stem diameter for wheat crops
within large eld phenotyping experiments.
e recent advances in high-resolution imaging systems
and computing capacity as well as image processing algo-
rithms oer great opportunities to develop nondestructive
high-throughput methods. Jin et al. andLiuetal.
 have demonstrated that the plant density could be
estimated at early stages in wheat crops from high-resolution
imagery. Direct estimates of the tillering coecient at the
end of the tillering stage were investigated by several authors
with application to the management of nitrogen fertilization
for stable crops. Vegetation indices computed from the
reectance measurements have been empirically related to
the tiller density [–]. However, reectance measurements
are mainly sensitive to the amount of green foliage, which
is loosely related to the stem density. Alternatively, several
authors have developed algorithms for estimating wheat stem
density at early stages from high-resolution imagery .
Unfortunately, this method, applied to plants in pots grown
under greenhouse conditions, is dicult to transfer to eld
conditions. Further, the number of stems at relatively early
stages may overestimate the actual stem density at harvest
Previous scientists have used algorithms for estimating wheat
ear density in-eld conditions using RGB or thermal imagery
[–]. However, these techniques, operated from the top of
number of ears are laying in the lower layers of the canopy.
Previous studies have also demonstrated that above-ground
biomass (AGB) can be estimated using dierent remote
sensing platforms [–]. However, the correlative nature of
these relationships questioned their robustness when applied
outside the domain where they have been calibrated.
e aim of this study is to develop and evaluate a method
to estimate stem density aer the harvest. Images of the
remaining stems cut by the combine machine during harvest
show a clear circular cross section at their tip that could
be identied by machine vision techniques. Further, the
diameter of the stem could be also measured to tentatively
estimate the AGB by combining the average stem diameter
with the stem density and plant height. High-throughput
estimates of plant height have become now a standard trait
easy to compute from D point clouds derived from LiDAR or
standard cameras aboard drone . e main objectives of
this study are therefore () to develop a method for identifying
stems from postharvest submillimetric RGB images and
F : Visual stem identication. Each stem identied corre-
sponds to a green bounding box. Note: the image is actually cropped
from original image by x pixel.
F : Application of the stem detection using Faster-RCNN
algorithm to an image extract in Gr´
eoux. Each yellow bounding box
corresponds to the identied stem and is associated with its score
corresponding to the probability of containing a stem.
compute the stem density; () to compare the estimated stem
density with the ear density measured with traditional inva-
sive methods; () to estimate the stem diameter and describe
their distribution; and () to investigate the capacity of stem
density, stem diameter, and plant height to provide a proxy
of AGB. e eld experiments and data acquisition are rst
described. e developed methods are then presented, and
their performances to estimate stem density, stem diameter,
and AGB are nally evaluated and discussed.
2. Materials and Methods
2.1. Experimental Sites and Ground Measurements. e
eoux and Clermont sites located in France (Table )were
hosting wheat phenotyping experiments with about one
thousand microplots of rows by m length (Gr´
or rows by . m length (Clermont). For both sites,
Pixel position on the diameter
45 degree direction
135 degree direction
F : e extraction of stem diameter of each sub-window
image using diameter gray level prole.
0100 200 300 400 500 600 700 800 900 1000
Estimated stem density (stems/Ｇ2)
Measured stem density (stems/Ｇ2)
F : Comparison between the stem density estimated using
Faster-RCNN method calibrated over the pooled Cgc dataset and
the stem density evaluated visually over the images. e black line
corresponds to the : line; the red and blue circles correspond,
respectively, to the Gr´
eoux (Vg) and Clermont (Vc) validation
rows were spaced by . cm. A subsample of microplots
(Table ) was selected in both sites for the development
and validation of the method. ey included genotypes with
contrasted tillering capacity and plant architecture as well
as variation in irrigation (Gr´
eoux) and nitrogen (Clermont)
300 400 500 600 700 800 900
Measured ear density (Ears/Ｇ2)
Estimated stem density (stems/Ｇ2)
F : Relationship between the estimated stem density and the
measured ear density at the Gr´
eoux (blue dots) and Clermont (red
dots) datasets. e black line corresponds to the : line.
e ear density (ears/m2) was measured in the eld at
maturity for the (Gr´
eoux) or (Clermont) microplots
considered, by counting the ears over three samples of two
rows by . m length corresponding to a . m2sampled area.
e AGB (g/m2) was measured in Gr´
eoux over microplots
by collecting all the plants within three samples of two rows
by . m length. e samples were then oven-dried at ∘Cfor
three days and nally weighed. e height (cm) of the plants
was measured using two LMS LiDARs (SICK, Germany)
xed on a phenomobile, i.e., a robot rover that automatically
details on height measurements are given by Madec et al. in
2.2. Image Acquisition and Visual Labeling of Stems. ACanon
els equipped with a mm focal lens was xed on a pole and
maintained at a . m distance from the ground at the Gr´
experimental site. e camera was set to speed priority. e
same operating mode was used in Clermont, except that
the camera was Sony ILCE- with by pixels
equipped with a mm focal length lens and maintained
at . m from the ground. e images were recorded in
JPG format on the SD memory card. Measurements were
completed under cloudy illumination conditions with light
wind. ree (Gr´
eoux) or four (Clermont) images were taken
over each microplot. A subsample corresponding to four rows
by . m length for Gr´
eoux and four rows by . m length for
Clermont (Table ) was extracted in the center of each image.
is oered the advantage of minimizing image deformation
observed mostly on the borders of the whole image. e
0.2 0.4 0.6
200 400 600
r = 0.12 r = 0.01
r = -0.01 r = 0.09
r = -0.27∗
r = -0.24∗
0.2 0.4 0.6
0 500 1000
r = 0.24∗r = -0.10∗r = -0.05
r = 0.15∗r = -0.13∗
r = -0.49∗∗
F : Correlation and distribution between stem density, average stem diameter, shape, and scale parameters over Gr´
eoux (a) and
Clermont (b) experimental sites. e correlation coecient, r, is given in the upper triangular matrix with ∗∗ and ∗corresponding,
respectively, to signicant values at . and . probability levels.
500 1000 1500500 1000 1500
MLR with all
0 1000 2000
10 15 20
0.6 0.8 1
200 400 600
200 400 600
MLR with all
r = 0.767∗∗ r = 0.806∗∗ r = 0.638∗∗r = 0.889∗∗ r = 0.476∗∗
r = 0.407∗∗ r = 0.603∗∗
r = 0.724∗∗
r = 0.541∗∗r = 0.642∗∗
r = 0.928∗∗ r = 0.781∗∗
r = 0.798∗∗
r = 0.585∗∗
r = 0.719∗∗ r = 0.598∗∗
r = 0.386∗r = 0.560∗∗
r = 0.719∗∗
r = 0.763∗∗
r = 0.9215∗∗
F : Correlation matrix between the AGB and the six variables investigated. Note: ∗∗ and ∗mean correlation signicant at the . and
. level of probability, respectively.
T : Characteristics of the Gr´
eoux and Clermont experimental sites.
Sites Latitude Longitude Number of plots Sowing date Sowing density (seeds/m2)
T : Characteristics of the images taken over the two experimental sites.
Sites Date of images Distance to ground (m) Ground resolution (mm) Sampled area (m2)
eoux // . . .
Clermont // . . .
eoux images were rst resampled using a bicubic inter-
polation algorithm to provide the same resolution as that of
e good quality of images provided strong condence in
the visual identication of the stems (Figure ). A bounding
box was interactively drawn around each stem identied in
the images. e bounding box used to identify each stem was
designed to include enough elements surrounding the stem
(Figure ). A total of images were visually annotated to be
used for the calibration and validation of the method.
2.3. Object Detection Using Faster-RCNN. Convolutional
Neural Networks (CNNs) are powerful machine learning
methods . ey are widely used to extract imagery infor-
mation features and then classify objects. CNNs were trained
using large collections of diverse images to extract more eec-
tively rich feature representations. ese CNNs features oen
outperform handcraed ones such as histogram of oriented
gradients (HOG), local binary patterns, or speeded up robust
features . e TensorFlow (https://www.tensorow.org/)
implementation of Faster Regional Convolutional Neural
Network (Faster-RCNN) by the object detection application
programming interface (API)  was achieved here. Faster-
RCNN has been widely used to detect objects . e region
proposal network (RPN) branch was inserted between the
conv and conv blocks. e Inception-Resnet-V model was
used as it obtained the best accuracy among several modern
object detectors . An anchor was set at each location
considered by the convolution maps of the RPN layer. Each
anchor was associated with a size and aspect ratio. A set of
anchors with dierent size and aspect ratio were assigned
at each location, following the default setting. e number
of proposed regions per patches was set to , which was
consistent with the expected number of stems per patch.
since the memory requirements were too demanding for
larger images. e original images were thus split into
x patches, keeping % overlap between neighboring
patches to minimize possible problems associated with the
borders. e batch size was xed to and the threshold value
for the non-maxima suppression with an IOU (Intersection
Over Union) was set to .. e model was trained with a
learning rate of . and a momentum of .. e model
(COCO) dataset to provide the starting point. e COCO
dataset  contains . million images with . million
of object instances belonging to object categories. More
details on the Faster-RCNN used could be found in Madec et
al. . e pretrained model was then ne-tuned over the
calibration image extracts. It identied and localized stems
using a bounding box associated with a condence score
varying between . and ..
e trained model was nally applied to all the image
extracts available. When identied stem bounding boxes were
overlapping, a minimum of . overlap fraction was used
to eliminate one of the overlapping bounding boxes. Finally,
bounding boxes with a condence score value smaller than
. were not considered as stems. is score threshold value
was optimized to get the best stem density estimation per-
formances. An example of the Faster-RCNN stem detection
result is presented in Figure . e estimated stem density
(stem/m2) was eventually computed by dividing the number
of stems identied over the image extracts of a microplot by
the size of the extracts (Table ).
2.4. Estimating the Stem Diameter and Biovolume. e
bounding box of the identied stems was rst transformed
into gray images using the value (V) component of the HSV
transform : V=.R+.G+.B, where R, G,
of the RGB images coded in bits. e gray value proles
were then extracted along four compass directions: ∘,
∘(Figure ). e gray level proles show typical
patterns with high values corresponding to the border of the
stem and lower values outside and inside the stem (Figure ).
e two borders of the stem were thus identied using
the two maximum gray values. e distance between the
maximums values was computed and then averaged over
the four compass directions to provide an estimate of the
e stem diameter was used to compute the area of the
section of the stem. e basal area of each microplot was then
computed as the average area of the stem section multiplied
by the stem density. Finally, the biovolume was computed as
from the LiDAR measurements.
2.5. Statistical Analysis. Both Gr´
eoux and Clermont datasets
were randomly split into / for model calibration and /
for validation. A rst global training (called here Cgc) was
investigated by pooling the calibration datasets of Gr´
T : Characteristics of the data sets used for the calibration and validation of the algorithm. Statistics of the stem density are indicated
for each data set, including minimum (Min), mean (Mean), maximum (Max), range (Range), standard deviation (SD), and coecient of
variation (CV) of the stem density.
Number of image extracts Stem density (stem/m2)
Dataset Name Gr´
eoux Clermont Min Mean Max Range SD CV (%)
Calibration Cgc .
Validation Vg c .
Calibration Cg .
Validation Vg .
Calibration Cc .
Validat io n Vc .
T : Accuracy of stem identication using the Faster-RCNN method. Results are presented for three calibration datasets (Cg, Cc, and
Cgc). e evaluation is achieved on the validation datasets (Vg, Vc, Vgc).
Calibration dataset Validation dataset Precision Recall Bias
Vg . . -.
Vc . . -.
Vgc . . -.
Vg . . -.
Vc . . .
Vgc . . .
Vg . . -.
Vc . . .
Vgc . . .
datasets Vgc. e performances of this global calibration
(Cgc) were also evaluated on both the Gr´
eoux (Vg) and
Clermont (Vc) validation datasets. en, a cross-validation
was also investigated to better evaluate the robustness of the
classication: the calibration was completed on the Gr´
(Cg) or Clermont (Cc) calibration datasets and validated
on the Gr´
eoux (Vg) and Clermont (Vc) validation datasets.
Table presents the several cases considered.
A detected stem bounding box (i.e., with a score >.) was
considered correct (true positive, TP) if its IOU with a labeled
stem bounding box was larger than the IOU threshold value.
Otherwise, the detected stem bounding box was considered
as false positive (FP). e proposed bounding boxes with a
score<. (i.e., not considered as stems) with IOU larger than
e IOU threshold value was set to the usual value of ..
e precision (TP/(FP+TP)), recall (TP/(FN+TP)), and bias
(-(precision/recall)) were also calculated.
2.6. Heritability Computation. e broad sense heritability
(H2) evaluates the repeatability of the stem or ear density
estimates. It was computed as the percentage of the genotypic
variance, Vg, to the total variance, Vg+Ve, where Ve is the
variance due to the environment . e heritability of
the stem density and ear density was computed over sixteen
wheat genotypes ( plots) selected from the Clermont
experimental site where each genotype was replicated six to
3. Results and Discussion
3.1. Stems Are Accurately Identied Using the Faster-RCNN
Model. To evaluate the robustness of the RCNN model, it
eoux (Cg), Clermont (Cc), or both
datasets (Cgc). Performances computed over the validation
datasets were very good with . <precision<. and . <
recall <. (Table ). Precision and recall were well balanced
with a small bias: -. <bias<..
e results showed that the classication accuracy of
stem identication was very high based on the precision
and recall values over the same experiments (Table ). e
robustness of the classication was further investigated by
comparing the precision and recall values computed over
the validation datasets coming from the other experiments.
Results show that the classication evaluated over the same
experiment used to calibrate the model was always per-
forming the best (Table ). e classication performances
single experiment was validated on the other experiment.
is may be explained both by the limited sample size of the
calibration dataset and also by the specic features associated
with each experiment, including the spatial resolution (Tables
and ). However, when the calibration was completed over
the pooled experiments (Cgc), the precision and recall values
decreased only slightly when evaluated over each individual
experiment (Vg or Vc) (Table ). e model captured the key
information common to the two experiments to provide a
consistent stem identication. It conrmed the eciency and
robustness of the Faster-RCNN method.
T : Performances of the stem density estimation when using Faster-RCNN method for the postclassication step. e evaluation is
achieved on the three validation data sets (Vg, Vc, and Vgc).
Calibration dataset Validation dataset Sample size slope intercept R2RMSE (stems/m2) RRMSE (%)
Vg . . . . .
Vc . . . . .
Vgc . . . . .
Vg . . . . .
Vc . . . . .
Vgc . . . . .
Vg . . . . .
Vc . . . . .
Vgc . . . . .
T : Statistics of the relationships between the estimated stem density and the measured ear density. e Faster-RCNN was trained over
the Cg+Cc dataset.
Datasets Slope Intercept R2RMSE (stems/m2) RRMSE (%)
eoux . . . . .
Clermont . . . . .
eoux & Clermont . . . . .
T : Biomass regression models derived from stem density, ear
density, stem area, height, and biovolume at the Gr´
mental site. Note: ∗∗ means model signicant at the . level of
probability. e R2, RMSE, and RRMSE are averaged R2, RMSE, and
RRMSE values of leave-one-out cross-validation methods.
Variabl e s R 2RMSE (g/m2) RRMSE (%)
Stem density .∗∗ .
Ear density .∗∗ .
Plant height .∗∗ .
Basal area .∗∗ .
Biovolume .∗∗ .
All .∗∗ .
3.2. Stem Density Is Accurately Estimated. e consequences
of the identication performances of the Faster-RCNN model
discussed previously were evaluated in terms of plant density
at the image extract level. For the sake of consistency, several
calibration and validation datasets were considered to further
evaluate the robustness of the model. Results showed RRMSE
values ranging from .% to .%. Calibrating over the
pooled datasets (Cgc, Table )providedthebestperfor-
mances with RRMSE lower than %. A slight degradation
of the performances was observed when calibrating over a
single dataset. Calibrating over the Gr´
eoux dataset provided
the worst performances when validated over the Clermont
and variation in the cutting height and inclination of the
stems during harvest between Gr´
eoux and Clermont sites. In
the pooled Gr´
eoux and Clermont calibration datasets (Cgc)
that provided more robust performances.
When considering the calibration over the pooled dataset
(Cgc) that provided the overall best performances, very small
biases were observed with points closely distributed around
the : line (Figure ).escatteraroundthe:lineappeared
to be relatively independent of the stem density (Figure ).
3.3. e Stem Density Is a Good Proxy of the Ear Density. e
stem density estimated with the Faster-RCNN model cali-
brated over the Cgc dataset was compared to the ear counted
visually at the ground level. Both quantities were evaluated on
dierent samples, expected however to represent the average
microplot value. Results showed that the estimated stem
density based on the Faster-RCNN model was very consistent
(Figure ) with the measured ear density at the Gr´
(Table ) and Clermont (Table ) experimental sites. e
scatter between ear and stem densities appeared to increase
with the density: this was obvious between the Gr´
(<density<) and Clermont (<density<) sites.
Part of the larger scatter observed over the Clermont site
visual counting (. m2). e scatter between ear and stem
densities seems to increase with the density within the
Clermont site between the low and high densities (Figure ).
Nevertheless, the good agreement found between ear and
stem densities was thus conrming the results of Siddique et
Previous studies demonstrated that the RGB imagery
can be used to estimate ear density using image processing
algorithms [–]. However, ear density estimation perfor-
mances were generally limited to a comparison between the
ears detected by the machine learning algorithm and those
that can be visually identied by and operated on the image.
Some discrepancies could appear compared with the actual
ear density, particularly when some ears are lying in the
lower canopy layers and could not be easily seen from the
top of the canopy. Counting ears from the stem sections
appears therefore preferable under such conditions. Further,
stem sections are relatively simpler objects to identify as
compared to ears that may show a large aspect variability.
Additionally, ears can frequently overlap in the eld, making
their identication more complex as compared to stem
sections that never overlap.
3.4. Stem and Ear Densities Are Highly Heritable. e heri-
tability (H2) values of the stem density and ear density were
compared at the Clermont experimental site where several
replicates of genotypes were available. Results show that
the H2values of stem density (.%) and ear density (.%)
were high and close together. is is consistent with the
strong relationship found between both quantities (Figure ).
ese heritability values agreed well with the values provided
by Madec et al. () . e H2value of the stem density
was slightly higher than that of the ear density, probably
because of the larger sample size used for estimating the stem
density from the RGB images, which makes the values more
repeatable. e high values of heritability found suggested
that the proposed method will be well suited to serve the
3.5. Stem Diameter Follows a Gamma Distribution. e
distribution of stem diameter was investigated at Gr´
and Clermont experimental sites, respectively, on and
microplots. e distribution of the stem diameter may
be a pertinent trait describing the structure of the tiller
population that may be impacted by the growth conditions.
e distribution of the stem diameter of each microplot was
adjusted either to a normal or to a gamma distribution.
e corresponding p values associated with the t of each
distribution was computed. Results show that the p value of
the gamma distribution was larger than that of the normal
distribution for % of the microplots for Gr´
eoux and %
of the microplots for Clermont. e gamma distribution
characterized by a scale and a shape parameter was therefore
selected to describe the stem diameter distribution over each
e average stem diameter of each microplot ranged from
. to . mm, with a median value close to . mm for both
sites (Figure ). e average stem diameter was loosely but
positively correlated to the stem density (Figure ): the stress
experienced by the plants was aecting both the density and
the diameter of the stems, with no apparent compensations
between these two traits. e stem diameter distribution
for each microplot as described by a gamma function was
further investigated: shape parameters were slightly smaller
for the Gr´
eoux site (<shape<) as compared to those of
theClermontsite(<shape<). Conversely, scale parameters
were slightly larger for the Gr´
eoux site (.<scale<.) as
compared to the Clermont site (.<scale<.). e distri-
bution of the diameters was more concentrated around the
average for the Clermont site as compared to the Gr´
site where a larger range of diameters was observed. is
may be related to the stress conditions that were stronger in
eoux, particularly during the stem elongation phase. is
was also reected by the stem density that was more impacted
eoux. e scale and shape parameters were negatively
correlated for both sites, with a stronger correlation for Cler-
mont (Figure ). Since the average of a gamma distribution
is dened by the product of the shape and scale parameters,
the negative correlation between the two parameters was
explained by the constraint to keep the average close to .
mm. erefore, both parameters could be equally used to
describe the “atness” of the stem diameter distribution.
Biomass. A total of microplots from the Gr´
were used to relate the measured AGB with the ear density
and the four structural traits derived from high-throughput
measurements: stem density, stem basal area computed as
the product of the average stem diameter and the stem
of the basal area and plant height. Results show that all
these traits are strongly correlated with AGB (Figure and
Table ). e best relationship is however obtained using
the biovolume that combines the three main original traits:
stem density and average stem diameter that are combined
into the basal area; and plant height. Note that these traits
are relatively independent: stem density and plant height are
loosely correlated (Figure ,r
2=.); plant height and basal
area are also loosely correlated (r2=.).
Because the dataset used was limited, the predictive
performances of the relationships observed between the AGB
validation method (Jin et al., ). e best determination
coecients were observed consistently for the biovolume
(Table ) with a relative error of .%, i.e., within the order
of magnitude of the accuracy with which AGB was measured.
Our results are very consistent with those presented by Aziz
et al. () and Pittman et al. ().
To further improve the predictive model, we used all
the ve traits together within a multiple linear regression
model. Marginal improvement of the model performances
was observed (Figure and Table ). is may be explained
by the strong relationships between the ve traits used, as
well as the decrease in the degree of freedom induced by
the increase of the number of coecients to be adjusted
(six coecients instead of two needed when using only the
biovolume). e biovolume appeared therefore as a very
sound proxy of the AGB. Previous results suggested that
AGB could be estimated using dierent optical techniques
and technologies [–]. Our study further conrmed these
results. e results demonstrated that the estimation accurac y
of AGB could be improved by combining LiDAR data and
RGB imagery. However, the stability of the relationship found
over the limited sample used in this study should be further
evaluated with emphasis on the possible dependency on the
environmental and management conditions, as well as on
dierences between genotypes.
is study demonstrated that the identication of the stems
aer the harvest was possible using deep-learning approaches
applied to RGB images. is requires the spatial resolution
to be sucient, i.e., around . mm since the stem diameters
are around . mm. It ensures that the objects to be identied
within the image are represented with an optimal number of
pixels ranging betweenandpixelsasadvisedbyMadec
et al. . Such a high resolution could be achieved using a
high-resolution RGB camera xed either on a pole, on a cart,
on a phenomobile, or even on a UAV ying at low altitude
as already demonstrated by Jin et al. . Alternatively, a set
of RGB cameras could be mounted on the combine machine
and provide, in near real time, an estimate of the stem density.
e method requires the stems not to be covered by the
straw rejected by the combine machine. Further, too inclined
stems due to the harvest process or some postharvest practice
may result in degraded performances since the sections of
the tip of the stems will not be viewed by the camera or
will be strongly deformed. Further, the proposed method
may be not suitable under stem lodging situations where the
stem sections will show unexpected patterns. Nevertheless,
would indicate that the Faster-RCNN model trained over
the illumination conditions may have little impact of the
stem identication since the objects are mostly identied by
the relative brightness of the pixels, with the color itself
bringing very little information. We demonstrated therefore
that the stem density is accessible with high-throughput,
relatively low cost and with a very good accuracy. Further,
the capacity to sample large area to estimate the stem density
will minimize the impact of the spatial variability within a
Although Madec et al.  among others demonstrated
that similar deep-learning techniques could be applied e-
ciently to estimate the ear density, ear identication is more
complex because of strong dierences of the ear aspect
demonstrated in this study that the stem density was a very
close proxy of the ear density although some discrepancy is
expected under specic environmental conditions. In such
circumstances, the distribution of the diameter of the stems
could potentially provide the necessary information to get a
better estimate of the ear density from the stem density and
Once the stem is identied, we demonstrated that the
diameter could be easily measured. e distribution of the
stem diameters followed a gamma function with an average
diameter close to . mm. e distribution of the stem diame-
ters may be indicative of the structure of the tiller population
that may be governed by the genetics in interaction with
conditions experienced by the plants. Finally, the biovolume
computed as the product of the average stem diameter, the
stem density, and plant height was demonstrated to be a
close proxy of the above-ground biomass. is opens very
attractive potential for the breeders to get high-throughput
estimates of the total plant biomass at harvest and possibly
quantify the radiation use eciency and the harvest index
assuming that the yield will be measured anyway. Neverthe-
less, these promising results should be veried under a much
larger number of situations to verify that the correlations are
not too dependent on the environmental conditions as well
as on the genotype.
Conflicts of Interest
e authors declare no conicts of interest.
is study was supported by “Programme d’Investissement
d’Avenir” PHENOME (ANR--INBS-) with participation
of FranceAgriMer and “Fonds de Soutien `
Clermont who participated in the experiments. e work was
completed within the UMT-CAPTE funded by the French
Ministry of Agriculture.
spacing and seeding rate eects on yield and yield components
of so red winter wheat,” Agronomy Journal,vol.,no.,pp.
and R. K. Scott, “e physiological response of winter wheat to
reductions in plant density,” Annals of Applied Biology,vol.,
no. , pp. –, .
 K. H. M. Siddique, E. J. M. Kirby, and M. W. Perry, “Ear:
Stem ratio in old and modern wheat varieties; relationship with
improvement in number of grains per ear and yield,” Field Crops
 J. Hiltbrunner, B. Streit, and M. Liedgens, “Are seeding densities
an opportunity to increase grain yield of winter wheat in a living
mulch of white clover?” Field Crops Research,vol.,no.,pp.
 C. M. Donald, “e breeding of crop ideotypes,” Euphytica,vol.
relationship between height and yield in wheat,” Heredity,vol.
 C. Law, J. Snape, and A. Worland, “Aneuploidy in wheat and
 A. J. King, L. R. Montes, J. G. Clarke et al., “Identication
of QTL markers contributing to plant growth, oil yield and
fatty acid composition in the oilseed crop Jatropha curcas L,”
Biotechnology for Biofuels,vol.,no.,p.,.
 S. Arnoult, M.-C. Mansard, and M. Brancourt-Hulmel, “Early
prediction of miscanthus biomass production and composition
based on the rst six years of cultivation,” Crop Science,vol.,
no. , pp. –, .
Alvaro, L. F. Garc´
ıa del Moral, S.
Dreisigacker, and C. Royo, “Changes in durum wheat root
and aerial biomass caused by the introduction of the Rht-Bb
dwarng allele and their eects on yield formation,” Plant and
 X. Jin, S. Liu, F. Baret, M. Hemerl´
e, and A. Comar, “Estimates of
plant density of wheatcrops at emergence from very low altitude
UAV im age r y,” Remote Sensing of Environment,vol.,pp.–
 S. Liu, F. Baret, D. Allard et al., “A method to estimate plant
density and plant spacing heterogeneity: Application to wheat
crops,” Plant Methods,vol.,no.,p.,.
 M. Flowers, R. Weisz, and R. Heiniger, “Remote sensing of win-
ter wheat tiller density for early nitrogen application decisions,”
 S. B. Phillips, D. A. Keahey, J. G. Warren, and G. L.
Mullins, “Estimating winter wheat tiller density using spec-
tral reectance sensors for early-spring, variable-rate nitrogen
applications,” Agronomy Journal,vol.,no.,pp.–,
 I. M. Scotford and P. C. H. Miller, “Estimating tiller density and
leaf area index of winter wheat using spectral reectance and
ultrasonic sensing techniques,” Biosystems Engineering,vol.,
no. , pp. –, .
 R. D. Boyle, F. M. K. Corke, and J. H. Doonan, “Automated
estimation of tiller number in wheat by ribbon detection,”
Machine Vision and Applications,vol.,no.,pp.–,
 J. A. Fernandez-Gallego, S. C. Kefauver, N. A. Guti´
errez, M. T.
Nieto-Taladriz, and J. L. Araus, “Wheat ear counting in-eld
conditions: High throughput and low-cost approach using RGB
images,” Plant Methods,vol.,no.,p.,.
 J. Fernandez-Gallego, M. Buchaillot, N. Aparicio Guti´
Nieto-Taladriz, J. Araus, and S. Kefauver, “Automatic wheat ear
counting using thermal imagery,” Remote Sensing,vol.,no.,
p. , .
 T. Liu, C. Sun, L. Wang, X. Zhong, X. Zhu, and W. Guo, “In-
eld wheatear counting based on image processing technology,”
Nongye Jixie Xuebao/Transactions of the Chinese Society for
Agricultural Machiner y, vol. , no. , pp. –, .
 S. Madec, X. Jin, H. Lu et al., “Ear density estimation from
high resolution RGB imagery using deep learning technique,”
Agricultural and Forest Meteorology,vol.,pp.–,.
 R. Ballesteros, J. F. Ortega, D. Hern´
andez, and M. A. Moreno,
“Applications of georeferenced high-resolution images obtained
with unmanned aerial vehicles. Part I: Description of image
acquisition and processing,” Precision Agriculture,vol.,no.,
pp. –, .
 S. Gao, Z. Niu, N. Huang, and X. Hou, “Estimating the leaf
area index, height and biomass of maize using HJ- and
RADARSAT-,” International Journal of Applied Earth Observa-
tion and Geoinformation,vol.,no.,pp.–,.
 X. Jin, G. Yang, X. Xu et al., “Combined multi-temporal optical
and radar parameters for estimating LAI and biomass in winter
wheat using HJ and RADARSAR- data,” Remote Sensing,vol.
, no. , pp. –, .
 X. Jin, Z. Li, G. Yang et al., “Winter wheat yield estimation based
on multi-source high-resolution optical and radar imagery
and AquaCrop model using particle swarm optimization algo-
rithm,” ISPRS Journal of Photogrammetry and Remote Sensing,
vol. , pp. –, .
Hu, “Estimating crop stresses, aboveground dry biomass and
yield of corn using multi-temporal optical data combined with a
radiation use eciency model,” Remote Sensing of Environment,
vol. , no. , pp. –, .
 N. Tilly, D. Homeister, Q. Cao et al., “Multitemporal crop sur-
face models: Accurate plant height measurement and biomass
estimation with terrestrial laser scanning in paddy rice,” Journal
of Applied Remote Sensing,vol.,no.,p.,.
 J. Yue, G. Yang, C. Li et al., “Estimation of winter wheat above-
ground biomass using unmanned aerial vehicle-based snap-
shot hyperspectral sensor and crop height improved models,”
 S. Madec, F. Baret, B. De Solan et al., “High-throughput phe-
notyping of plant height: Comparing unmanned aerial vehicles
and ground lidar estimates,” Frontiers in Plant Science,vol.,p.
 A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classi-
cation with deep convolutional neural networks,” Advances in
neural information processing systems, pp. –, .
 J. Deng, W. Dong, R. Socher et al., “ImageNet: a large-
scale hierarchical image database,” in Proceedings of the 2009
IEEE Conference on Computer Vision and Pattern Recognition
(CVPR), pp. –, Miami, Fla, USA, June .
 J. Huang, V. Rathod, C. Sun et al., “Speed/accuracy trade-
os for modern convolutional object detectors,” in Proceedings
of the 2017 IEEE Conference on Computer Vision and Pattern
Recognition (CVPR ’17), pp. -, Honolulu, HI, USA, July
 S. Ren, K. He, R. Girshick, and J. Sun, “Faster R-CNN: towards
real-time object detection with region proposal networks,” in
Advances in Neural Information Processing Systems,pp.–,
 G. W. Meyer and D. P. Greenberg, “Perceptual color spaces for
computer graphics,” ACM SIGGRAPH Computer Graphic,vol.
 M. R. Dohm, “Repeatability estimates do not always set an
upper limit to heritability,” Functional Ecology,vol.,no.,pp.
. Please check and conrm the author(s) rst and last names
and their order which exist in the last page.
. We made the highlighted change(s) for the sake of clar-
ity/correctness. Please verify and check similar highlighted
. Please check the format of Table 3.
. We changed Table 8to Table 7and changed its citation in
the text. Please check.
. We removed the bold font from Figures 4,5,6,and7.
It is very important to conrm the author(s) last and rst names in order to be displayed correctly
on our website as well as in the indexing databases:
Given Names: Xiuliang
Last Name: Jin
Given Names: Simon
Last Name: Madec
Given Names: Dan
Last Name: Dutartre
Given Names: Benoit
Last Name: de Solan
Given Names: Alexis
Last Name: Comar
Given Names: Frédéric
Last Name: Baret