Content uploaded by Jianfeng Zhou
Author content
All content in this area was uploaded by Jianfeng Zhou on Jul 30, 2019
Content may be subject to copyright.
Transactions of the ASABE
Vol. 62(2): 393-403 2019 American Society of Agricultural and Biological Engineers ISSN 2151-0032 https://doi.org/10.13031/trans.13067 393
C
OTTON
Y
IELD
E
STIMATION FROM
UAV-B
ASED
P
LANT
H
EIGHT
A. Feng, M. Zhang, K. A. Sudduth, E. D. Vories, J. Zhou
A
BSTRACT
. Accurate estimation of crop yield before harvest, especially in early growth stages, is important for farmers
and researchers to optimize field management and evaluate crop performance. However, existing in-field methods for esti-
mating crop yield are not efficient. The goal of this research was to evaluate the performance of a UAV-based remote
sensing system with a low-cost RGB camera to estimate cotton yield based on plant height. The UAV system acquired images
at 50 m above ground level over a cotton field at the first flower growth stage. Waypoints and flight speed were selected to
allow >70% image overlap in both forward and side directions. Images were processed to develop a geo-referenced ortho-
mosaic image and a digital elevation model (DEM) of the field that was used to extract plant height by calculating the
difference in elevation between the crop canopy and bare soil surface. Twelve ground reference points with known height
were deployed in the field to validate the UAV-based height measurement. Geo-referenced yield data were aligned to the
plant height map based on GPS and image features. Correlation analysis between yield and plant height was conducted
row-by-row with and without row registration. Pearson correlation coefficients between yield and plant height with row
registration for all individual rows were in the range of 0.66 to 0.96 and were higher than those without row registration
(0.54 to 0.95). A linear regression model using plant height was able to estimate yield with root mean square error of 550 kg
ha
-1
and mean absolute error of 420 kg ha
-1
. Locations with low yield were analyzed to identify the potential reasons, and
it was found that water stress and coarse soil texture, as indicated by low soil apparent electricity conductivity (EC
a
), might
contribute to the low yield. The findings indicate that the UAV-based remote sensing system equipped with a low-cost digital
camera was potentially able to monitor plant growth status and estimate cotton yield with acceptable errors.
Keywords. Cotton, Geo-registration, Plant height, UAV-based remote sensing, Yield estimation.
ield estimation at early growth stages is an im-
portant objective for precision agriculture as it
can assist in optimizing field management, espe-
cially for management-intensive crops such as
cotton (Gossypium hirsutum L.), and in evaluating crop per-
formance (Komm and Moyer, 2015). For many crops, plant
height is an important indicator of yield because larger and
healthier plants usually have higher yield than smaller plants
in the same field. For example, Chang et al. (2017) found
that plant height was a good indicator of sorghum growth
status and yield. Many other studies also showed that plant
height has potential to estimate crop yield or biomass in dif-
ferent crops, including onion (Ballesteros et al., 2018), corn
(Sharma et al., 2016), barley (Bendig et al., 2014, 2015;
Brocks and Bareth, 2018), and wheat (Walter et al., 2018).
However, excessive height in cotton (i.e., rank growth) can
reduce yield, making the relationship weaker. For example,
Huang et al. (2016) reported that cotton in the middle of the
height range had higher yield than taller or shorter cotton.
Cotton plant height is usually adequately controlled with
plant growth regulators; in those cases, cotton yield has been
positively correlated with plant height (Sui et al., 2012; Chu
et al., 2016).
Plant height has been measured using different ground-
based sensing technologies, such as a mobile platform with
ultrasonic sensors (Scotford and Miller, 2004; Sui et al.,
2012; Sharma et al., 2016) and tractor-mounted light detec-
tion and ranging (LiDAR) sensors (Zhang and Grift, 2012).
However, ground-based sensing methods have limitations in
travel speed and may not be efficient for data collection in
large fields. In recent years, unmanned aerial vehicle (UAV)
technology has been used as a high-throughput data collec-
tion tool in many studies due to its flexibility and ability to
cover large areas quickly (Yang et al., 2017; Zhou et al.,
2018). A number of different sensors, including imaging
sensors (e.g., visible cameras, multispectral and hyperspec-
tral cameras, and thermal cameras) and LiDAR sensors, can
be integrated with a UAV platform to measure target crop
traits, such as canopy color, vegetation indices, and canopy
Submitted for review in August 2018 as manuscript number ITSC
13067; approved for publication by the Information Technology, Sensors,
& Control Systems Community of ASABE in January 2019.
Mention of company or trade names is for description only and does no
t
imply endorsement by the USDA. The USDA is an equal opportunity
p
rovider and employer.
The authors are Aijing Feng, Graduate Research Assistant, Division o
f
Food Systems and Bioengineering, University of Missouri, Columbia,
Missouri; Meina Zhang, Research Assistant, Institute of Agricultural
Facilities and Equipment, Jiangsu Academy of Agricultural Sciences,
N
anjing, China; Kenneth A. Sudduth, Research Agricultural Engineer,
and Earl D. Vories, Research Agricultural Engineer, USDA-ARS
Cropping Systems and Water Quality Research Unit, Columbia, Missouri;
Jianfeng Zhou, Assistant Professor, Division of Food Systems an
d
Bioengineering, University of Missouri, Columbia, Missouri.
Corresponding author: Jianfeng Zhou, 211 Ag Engineering Building,
University of Missouri, Columbia, MO 65211; phone: 509-882-2495;
e-mail: zhoujianf@missouri.edu.
Y
394 TRANSACTIONS OF THE ASABE
structural features, including plant height (Sankaran et al.,
2015). UAV-based plant height measurement using a Li-
DAR sensor may be very expensive (Sankey et al., 2018;
Wang et al., 2017). In recent years, low-cost cameras
mounted on a UAV to measure plant height have been stud-
ied in field conditions (de Castro et al., 2018; Jay et al.,
2015). Sequential images collected with UAV systems have
been used to reconstruct the 3D structure of individual plants
or plant rows based on the structure-from-motion (SfM)
method (Westoby et al., 2012), which is able to develop dig-
ital elevation models (DEMs) for the scouted fields using
overlapped images (Smith et al., 2016; Glendell et al., 2017).
These DEMs can be used to extract the height of objects,
including plants (Chang et al., 2017; Malambo et al., 2018;
Walter et al., 2018).
A UAV-based plant height map includes high-resolution
and site-specific information that can be used for estimation
of crop yield based on calibrated estimation models. How-
ever, development of calibration models requires accurate
alignment or registration of UAV-based height data with
ground-based yield data (Yang et al., 2017). Although geo-
referenced height and yield can be registered based on coor-
dinates from their respective GPS systems, this process may
lead to large errors in registration due to low accuracy of the
GPS receivers (Jiao et al., 2018). Low-accuracy geo-refer-
encing information plays a less important role in plot-based
research, where the plots are easily segmented from images
and registered to labeled ground data (Bendig et al., 2014;
Liu et al., 2016; Du and Noguchi, 2017; Duan et al., 2017;
Brocks and Bareth, 2018). These data are typically analyzed
at plot level, i.e., mean values of imagery features and
ground data in each plot are calculated to represent all plants
in the plot, and then the reduced number of observations
(equal to the number of plots) are used to develop estimation
models (Huang et al., 2013, 2016; Van Der et al., 2017).
Similarly, some studies have manually split a large field into
multiple plots by adding ground reference points that can be
identified in images, and then used single data points
(means) from those segmented plots for further analysis
(Duan et al., 2017; Maimaitijiang et al., 2017; Vega et al.,
2015). However, the spatial resolution of the acquired da-
tasets is relatively low, and the averaging process may re-
duce the variability to where it is not possible to develop re-
liable models.
The goal of this study was to estimate cotton yield using
imagery data from a UAV-based low-cost visible camera us-
ing a large data set. Geo-registration between UAV-esti-
mated plant height and ground-based yield data was con-
ducted. The specific objectives of this study were to: (1) ex-
tract and calibrate cotton plant height from UAV-based im-
agery data, (2) register geo-referenced yield data with UAV-
based plant height data, and (3) evaluate the feasibility of
estimating cotton yield using the low-cost imagery-based
plant height.
MATERIALS AND METHODS
EXPERIMENTAL FIELD AND GROUND
DATA COLLECTION
This study was conducted in a cotton research field at the
Fisher Delta Research Center of the University of Missouri,
located in the upper Mississippi Delta region near Portage-
ville, Missouri (36.411° N, 89.696° W). The spatial variabil-
ity of soil texture in the field is quite large due to both allu-
vial and seismic activities. Because soil texture is strongly
related to apparent soil electrical conductivity (ECa; Sudduth
et al., 2003), soil ECa data were collected in May 2016 using
a Veris 3100 instrument (Veris Technologies, Salina, Kan.)
following procedures described by Sudduth et al. (2003).
Cotton cultivar PHY 333WRF (Dow Agrosciences, Indian-
apolis, Ind.) was planted on May 23, 2017, using a commer-
cial planter (Almaco, Nevada, Iowa) at a planting rate of
12 seed m-1 and a row width of 0.97 m, resulting in 152
north-south rows over the field area of 350 m 170 m
(~6 ha).
The cotton was irrigated using a 150 m center-pivot irri-
gation system (8000 Series, Valley Irrigation, Valley, Neb.)
following a predefined protocol. The portion of the field un-
der the center pivot was divided into 15 plots with a range of
irrigation treatments, including areas with no irrigation.
There were also areas in the northeast and southeast portions
of the field that were not irrigated by the center-pivot system.
Because the objective was to estimate yield with height, in-
dependent of other factors, the irrigation treatments were not
considered in developing the yield model.
The field was harvested on October 6-7, 2017, using a
cotton harvester (1996 Case IH 2155, Racine, Wisc.)
equipped with an Ag Leader Insight yield monitor system
that measures the cotton yield using optical emitter and de-
tector arrays on opposite sides of each picker chute, for a
total of eight pairs of sensor arrays (Ag Leader Technology,
Ames, Iowa) and an Ag Leader GPS 1500 receiver. The sys-
tem recorded geo-referenced yield data points separately for
each row at 1 s intervals at approximately 6.5 km h-1, or ap-
proximately 1.8 m of row per data point. The spatially refer-
enced yield data were managed using Ag Leader SMS Basic
version 18.50 and ArcGIS 10.4 for Desktop (ESRI, Red-
lands, Cal.). The yield monitor was calibrated by weighing
loads of cotton in a boll buggy equipped with scales (Master
Scales, Greenwood, Miss.), and the resulting calibration er-
ror based on nine of the ten loads averaged 2.7% with a max-
imum of 3.6%. Some of the seed cotton from the first load
tended to fill voids in the picker basket, so the subsequent
loads emptied more consistently; therefore, the weight of the
first load was not included in the calibration. After harvest-
ing, yield data artefacts were removed with Yield Editor
software (Sudduth et al., 2012). The average seed cotton
yield on the study field was 2530 1130 kg ha-1 (mean with
standard deviation of all data points), and all yield data re-
ported in this study were seed cotton yield.
UAV DATA ACQUISITION
The UAV-based remote sensing system consisted of a
UAV platform (Matrice 600 Pro, DJI, Shenzhen, China) and
an RGB camera (HERO 5, GoPro, San Mateo, Cal.) with a
62(2): 393-403 395
resolution of 3000 4000 pixels, resulting in a spatial resolu-
tion of 26 mm pixel-1 at a flight altitude of 50 m above ground
level (AGL) from the take-off location. The RGB camera was
set to capture still images at a rate of 2 frames per second (fps),
and images were saved to an on-board SD card. Flight of the
remote sensing system was controlled using a UAV ground
control app (Autopilot, Hangar Technology, Austin, Tex.) that
allowed planning the flight path and setting waypoints, flight
speed, and altitude. Prior to flight, the flight path (waypoints)
and speed (10 km h-1) were planned according to the flight
altitude (50 m AGL) and minimum overlap requirement
(>70% in the forward and sideward directions). After the
flight, images and GPS coordinates of the flight path recorded
by the on-board system were downloaded for further pro-
cessing. Image data were collected near noon central daylight
time (CDT) on August 12, 2017, during the cotton flowering
period. In total, 1352 images were acquired by the UAV sys-
tem over the field. Environmental data on the test day were
obtained from a weather station (http://agebb.mis-
souri.edu/weather/stations/pemiscot/framepvl.htm) located
approximately 400 m from the study site. Table 1 shows the
environmental conditions during UAV data acquisition.
Twelve ground reference points (GRPs) consisting of fence
posts (~1.1 m above ground) topped with white polytechnic
boards (about 30 30 cm) were distributed at the edges and
in the middle of the field to validate the height measurement
from the UAV-based sensing system.
IMAGE PROCESSING AND PLANT HEIGHT EXTRACTION
Sequential images were processed using Agisoft Pho-
toScan Pro (ver. 1.4.2, Agisoft LLC, St. Petersburg, Russia)
to generate dense point cloud data based on the SfM method.
GPS coordinates of the flight path from the UAV’s on-board
GPS receiver were extracted and matched with each image
based on their time stamps. The geo-referenced images were
uploaded to Agisoft PhotoScan to generate an orthomosaic
image and DEM map of the study field, as shown in figure 1.
The DEM of the field included elevation information (rel-
ative to mean sea level) for the crop canopy, the bare soil
surface, and the GRPs. Plant height was calculated as the
difference in elevation between the crop canopy and the
nearby bare soil surface. The soil surface elevation could be
measured prior to crop planting; however, that was not done
in this study. Therefore, soil elevation data were extracted
from the same DEM map using procedures described below.
Although there were errors in the measured elevation due to
the UAV having a low-accuracy GPS, it was assumed that
the measurement error over a small region would be similar
for the soil surface and the plant canopy or other objects.
Therefore, the study field was split into 2,550 small plots by
evenly dividing the field with 150 lines in the north-south
direction and 17 lines in the east-west direction (red lines in
fig. 2), resulting in an area of 2 m 8.5 m for each plot. It
Table 1. Environmental data during data collection (0.5 h period).
Air
Temp.
(°C)
Relative
Humidity
(%)
Dew
Point
(°C)
Wind
Speed
(m s-1)
Wind
Direction
(degrees)
Solar
Irradiance
(W m-2)
26.8 63 19.1 1.3 1 989
Figure 1. Results of processed sequential images from UAV system: (a) orthomosaic image of the entire field, (b) dense point cloud of a section o
f
the field, and (c) plant height map of the same area as in (b).
396 T
RANSACTIONS OF THE
ASABE
was assumed that the soil elevation was uniform within a
single plot. The dimension of each plot in the north-south
direction was determined based on the number of yield data
points in each row (~150 yield data points in each row), and
the dimension in the east-west direction was selected to in-
clude a sufficient number of rows to increase the potential of
bare soil being visible within the plot. The elevation infor-
mation for each pixel in the plots was extracted from the
DEM map, and the mean value of the lowest 1% of the ele-
vation information for each plot was used as the average el-
evation of the soil surface in that plot.
Although each plot included about nine cotton rows, as
shown in figure 2, bare soil was not visible in the plots with
full canopy closure. To identify the closed-canopy plots, the
RGB orthomosaic image was converted to the CIE 1976
L*a*b* color space, where channel a* represents the green-
red color component, with negative values corresponding to
green pixels and positive values to the background (Liu et
al., 2017). Therefore, channel a* was used to discriminate
crop and soil in the orthomosaic image. The canopy cover
ratio was calculated using the ratio of the number of crop
canopy pixels to the total number of pixels in each plot. In
this study, plots with a canopy cover ratio of 0.95 or higher
were considered to have a fully closed canopy, where bare
soil could not be seen. To calculate the plant height in the
plots without a visible soil surface, a linear interpolation of
the soil surface elevation between adjacent plots was used to
calculate the approximate elevation of the soil surface. As
shown in figure 2, some portion of the soil in plots 1, 2, and
6 (canopy cover ratio < 0.95) was visible, while the soil in
plots 3, 4, and 5 (canopy cover ratio > 0.95) was not visible.
Linear interpolation between the soil surface elevations of
plots 2 and 6 was used to estimate the soil surface elevation
of plots 3, 4, and 5. A plant height map was then generated
for each plot by subtracting the plot-mean soil elevation
from the plant canopy elevation of every pixel. A section of
a plant height map is shown in figure 1c.
G
EO
-R
EGISTRATION AND
R
OW
S
EPARATION
Because neither the UAV nor the yield monitor used in
this study employed a RTK-GPS system, their horizontal po-
sitioning accuracies were 1 to 2 m (about one to two cotton
rows) and sub-meter (about half to one cotton row), respec-
tively. Therefore, the geo-referenced data for UAV-based
plant height and for yield were not properly aligned (regis-
tered) in some rows when only GPS coordinate information
was used. In this study, an image feature-based geo-registra-
tion method was developed to improve the registration accu-
racy. The major steps included: (1) obtaining the GPS coor-
dinates of the GRPs using Google Earth (Mohammed et al.,
2013; Benker et al., 2011) and identifying the coordinates of
the plant height map, (2) converting the GPS coordinates of
yield data to the image coordinates of the plant height map,
(3) segmenting individual crop rows, and (4) adjusting the
image coordinates of yield data to the corresponding row po-
sition in the plant height map.
The positions of five GRPs located in the four corners and
middle of the field were identified in the plant height map,
and their GPS coordinates were matched to the image coor-
dinates of the plant height map. To convert the GPS coordi-
nates to image coordinates, a conversion scale was defined
and calculated using equation 1:
1 2 5
gk gm
k
ik im
gk gm
k
ik im
xx
xxx,m,k , , , ,m k
yy
yyy
(1)
where (x
gk
, y
gk
) and (x
ik
, y
ik
) are the GPS coordinates and im-
age coordinates of the kth GRP, respectively; and (x
gm
, y
gm
)
and (x
im
, y
im
) are the GPS coordinates and image coordinates
of the mth GRP (m k). Ten sets of conversion scales were
obtained, and their mean value was used as the conversion
scale to convert all yield GPS coordinate points to image co-
ordinate points with equation 2:
00
00
ii gi g k i
ii gi g k i
xxx xx
yyy yy
(2)
where (x
gi
, y
gi
) and (x
ii
, y
ii
) are the GPS coordinates and the
converted image coordinates of the ith yield datum, respec-
tively; ( ,
kk
x y
) is the average conversion scale between the
two coordinate systems; and (x
i0
, y
i0
) and (x
g0
, y
g0
) are the
image coordinates and GPS coordinates of a selected refer-
ence point.
To improve the registration accuracy, individual crop
rows were segmented from the plant height map and used to
align the corresponding rows of the yield data. The purpose
of row segmentation was to segment individual crop rows in
the images. The segmented rows were then directly regis-
tered to the yield monitor data, instead of relying only on the
GPS coordinates. Crop rows in the plant height map were
segmented based on the variation in elevation of the crop
rows and inter-rows. The elevation of crop rows was consist-
ently higher than that at inter-row locations, even with a
closed canopy (fig. 3). To search for inter-row locations,
strips with 50 pixels in the north-south (vertical) direction
and the same width as the whole plant height map (east-west
direction) were defined (shown as red boxes in fig. 3a). The
Figure 2. Section of the field showing plots used for DEM analysis. The red lines divided the field into 2550 small plots, each 2 m 8.5 m. Soil is
visible in plots 1, 2, and 6 but not in plots 3, 4, and 5. The canopy cover ratio for plots 1 to 6 was 0.67, 0.63, 1.00, 0.99, 0.97, and 0.91, respectively.
62(2): 393-403 397
sum of all plant height pixel values in the north-south direc-
tion in each strip was then calculated and plotted along the
east-west direction, as shown in figure 3b. The local valleys
in each curve in figure 3b, i.e., the lowest plant heights, rep-
resented the inter-row locations, which were used to segment
the crop rows, as shown in figure 3c. The coordinates of each
crop row in the plant height map were recorded for further
processing. Meanwhile, the yield data were also divided into
individual rows based on their GPS coordinates. The crop
rows and yield rows were aligned based on their order from
one edge of the field to the other. This created a one-to-one
correspondence between the image rows and the yield rows,
and the position of every yield GPS point (x
gi
, y
gi
) calculated
from equation 2 was adjusted to match the corresponding
image row position.
The geo-referenced yield data were recorded at 1 s inter-
vals during harvesting, resulting in about 25,000 data points
with each representing the cotton weight in an area of about
2 m
2
. However, data from the plant height map had a spatial
resolution of 2.36 10
-3
m
2
pixel
-1
, which resulted in 7559
3880 data points (pixels). To match the two data sets, a small
region of interest (ROI) with 30 30 pixels (equivalent to
about 2.3 m
2
) centered on the coordinate of each yield data
point in the plant height map was selected. The average crop
height in each ROI was extracted from the plant height map
and matched with the yield data for further analysis. During
harvesting, the yield monitor measured the yield separately
for each of the four rows harvested in one pass (harvest
pass), and the software applied offsets from each row to the
GPS receiver to determine the locations of the individual
rows. However, the yield monitor was calibrated using the
total yield of all four rows, and separate calibration for each
row was not possible. Therefore, the data from the 38 four-
row harvest passes were believed to be more accurate than
individual-row data and were used in the subsequent analy-
sis.
D
ATA
A
NALYSIS
The accuracy of the image-based height measurement
was evaluated by comparing the height extracted from the
DEM with that manually measured at the 12 GRPs using a
paired t-test. The Pearson correlation coefficient was calcu-
lated to evaluate the correlation between image-based plant
height and yield. Both processes were conducted in R soft-
ware (RStudio Desktop 1.1.453, RStudio, Boston, Mass.). A
yield prediction model using plant height was developed us-
ing linear regression. The yield estimation model was writ-
ten as equation 3 (Härdle, 1990):
ii
ˆ
ymX
(3)
where
i
ˆ
y is the estimated yield (kg ha
-1
), X
i
is a vector of the
image-based plant height, is the error (or noise) of the data,
and m(X
i
) is the model to be developed. A bin smoothing
method was used to smooth the raw data
1
,
n
ii
i
X y
, where
n is the number of raw data points (Irizarry, 2001), to reduce
the fluctuations and measurement error from the UAV sen-
sors and yield monitor so that the main relationship was eas-
ier to obtain. Prior to modeling, the plant height data corre-
sponding to yield data
1
,
n
ii
i
X y
were sorted in ascending
order based on the values of X
i
, and a bin with a length of
15 data points was used to smooth the raw data
1
,
n
ii
i
X y
.
The mean
, Xy
of each bin was obtained, and all the
1
,
n_bin
ii
i
Xy
made up a new set of data, i.e., smoothed data,
where n_bin is the number of bins. The relationship m(X
i
)
between yield and plant height was fit based on the smoothed
data using a least squares regression model (Hastie et al.,
2001). Residuals between the measured yield obtained from
the yield monitor and the predicted yield obtained from the
regression model were calculated using equation 4:
Residuals
ii
ˆ
yy (4)
where y
i
is the yield (kg ha
-1
) from the yield monitor, and
i
ˆ
y
is the predicted yield from equation 3.
Yield data less than 150 kg ha
-1
were removed from the
data set because those data only occurred at the edges of the
field or in non-cropped areas, where the measured yield
should be zero. Model performance was assessed using root
mean square error (RMSE) and mean absolute error (MAE),
as shown in equations 5 and 6 (Willmott and Matsuura,
2005):
2
1
RMSE
n
ii
i
ˆ
yy
n
(5)
1
MAE
n
ii
i
ˆ
yy
n
(6)
where y
i
and
i
ˆ
y
are the ith observed yield data and predicted
yield, respectively, and n is the total number of observations
Figure 3. Segmentation of crop rows: (a) section of the plant height map showing the field divided into east-west strips perpendicular to the crop
rows, (b) accumulated value of each north-south pixel column in each strip, and (c) plant height map with segmented crop rows.
398 T
RANSACTIONS OF THE
ASABE
in the data set. Although both error statistics evaluate the dif-
ference between predicted yield and measured yield, RMSE
imposes a higher penalty than MAE for predictions that are
far from the measured value (Willmott and Matsuura, 2005).
A 10-fold cross-validation was used to evaluate the pre-
diction performance of the model. The raw data
1
,
n
ii
i
Xy
were randomly partitioned into ten equal-size subsamples. In
each test, a single subsample was retained for testing the pre-
diction ability of the model, and the remaining nine subsam-
ples were used as training data. This process was repeated
ten times, with each of the ten subsamples used exactly once
as the test data. Bin smoothing was conducted on the training
data to obtain a smoothed data set
1
,
n_bin
ii
i
Xy
, and then
a linear model was built based on the smoothed data. The
test data were used to determine the model’s RMSE and
MAE.
R
ESULTS
A
CCURACY OF
DEM-D
ERIVED
H
EIGHT
The image-based height measurements of the twelve
GRPs and the manual measurements are shown in table 2. A
paired t-test indicated that there was no significant difference
in height due to measurement methods. The measurement
errors were in the range of -0.1 to 0.16 m with an average
value of 0.07 m, equivalent to a 6% error relative to the man-
ual measurement. This indicated that the UAV-based imag-
ing measurement method had the potential to quantify plant
height with less than 10% error.
C
ORRELATION
A
NALYSIS
Figure 4 shows the Pearson correlation coefficients (r) be-
tween cotton yield and plant height with and without row
segmentation for the 38 four-row harvest passes. The red
curve shows the correlation coefficient of each pass calcu-
lated based on GPS alignment, while the values on the black
curve were calculated based on row separation. The overall
correlation coefficient was improved from 0.81 to 0.83 after
row separation was added. Although the change in the over-
all correlation was not large, the correlation coefficients of
the rows with low correlation increased substantially from
about 0.7 to 0.8 (blue circles in fig. 4). The GPS coordinates
in the rows with the lowest correlation coefficients had
larger offsets than other rows, and aligning the yield data
more accurately could improve the correlation between plant
height and yield. The results indicate that the geo-registra-
tion based on row separation made an improvement in cor-
relation in these cases.
Figure 4 also shows that the Pearson correlation coeffi-
cients fluctuated over a large range from pass to pass. To
identify possible reasons for the low correlation in some
passes, a detailed analysis was conducted on two passes
(passes 31 and 38), which had the lowest correlation coeffi-
cients using the row separation method (fig. 4). Figure 5a
shows the yield and plant height in pass 31 (r = 0.66), where
points 135 to 145 and points 157 to 162 were in regions of
bare soil, and the yield and plant height were both zero.
Points 1 to 20 in pass 31 and points 10 to 40 in pass 38 had
low yield but taller plants. These points are in the lowest por-
tion of the field near the drain. It is likely that good moisture
conditions early in the season favored taller plants in this
area, but waterlogging later in the season caused boll shed-
ding, resulting in lower yield. The locations of pass 31 and
points 60 to 80 in the pass were identified on the orthomosaic
Table 2. Height measured from DEM and manually.
Ground
Reference
Point
Height
from DEM
(m)
Manual
Measurement
(m)
Absolute Error
(m) (%)
1 0.94 1.04 0.10 11
2 1.10 1.00 0.10 10
3 1.02 1.02 0.00 0
4 1.20 1.04 0.16 13
5 1.16 1.09 0.07 6
6 0.98 1.04 0.06 6
7 1.10 1.07 0.03 3
8 1.09 1.04 0.05 5
9 0.99 1.02 0.03 3
10 1.14 1.02 0.12 11
11 1.13 1.07 0.06 5
12 1.07 1.07 0.00 0
Average 1.08 1.04 0.07 6
Figure 4. Pearson coefficients of linear correlation between yield and plant height in all four-row harvest passes. The numbers on the dashed lines
show the overall correlation coefficients. The red line with circle markers represents the correlation with only GPS alignment, while the blac
k
line with triangle markers represents the correlation using GPS alignment plus position adjustment based on row separation. The blue circles
mark the passes where the correlation improved substantially after applying row separation.
62(2): 393-403 399
image, as shown in figure 6. The points coincide with a rain-
fed portion of the field, and water stress likely decreased the
cotton yield, reducing the correlation between plant height
and harvested yield. However, other rainfed areas had higher
correlations, so the soil texture probably exacerbated the
drought stress response in this area of pass 31.
Rainfall was adequate for early-season growth and devel-
opment of the cotton crop; however, only 2 mm of rain was
recorded between July 9 and August 5. The first irrigation
was applied to portions of the field on July 19, and up to five
irrigations were applied to parts of the field by the time
25 mm of rain was recorded on August 6. No additional irri-
gations were applied.
Y
IELD
E
STIMATION
There were over 25,000 data points in 152 crop rows of
harvested yield data, which was reduced to about 6500 data
points after combining the rows into 38 four-row harvest
passes and removing outliers. The data in each pass were
then smoothed with a 15-point bin smoothing method, re-
sulting in about 450 data points. A 10-fold cross-validation
was conducted, resulting in average errors of RMSE =
Figure 6. Explanation of low correlation between yield and plant height in pass 31 (rows 120 to 124). The red lines in (a) and (c) mark the four
rows of pass 31, and the blue numbers relate to data points in figure 5a. The graph (b) shows the cumulative precipitation from the date o
f
planting. Image data were collected on August 12, 2017, and the crop was harvested in early October (marked with dashed lines in the graph).
Figure 5. Details of cotton yield and plant height in two harvest passes with low Pearson coefficients (after row separation). Blue circles mark the
data points that have inconsistent trends between yield and plant height: (a) pass 31 (r = 0.66) and (b) pass 38 (r = 0.67).
400 T
RANSACTIONS OF THE
ASABE
550 kg ha
-1
(RMSE/mean = 21%) and MAE = 420 kg ha
-1
(MAE/mean = 16%). Figure 7 shows a scatter plot of cotton
yield versus plant height for all the smoothed and un-
smoothed data, in which the blue line shows a linear model
based on the smoothed data, and p-values < 0.01 were ob-
tained for both the plant height coefficient and intercept of
the linear model. Errors obtained for the full data set were
RMSE = 630 kg ha
-1
and MAE = 450 kg ha
-1
.
While the regression model established the general rela-
tionship of crop height and yield, it was instructive to con-
sider the points below the regression line, where the har-
vested yield was much lower than that estimated from plant
height (the unsmoothed data below the green dashed line in
fig. 7). To explore the causes of low yield and high negative
residuals, figure 8 shows the regions with low yield and the
regions with high negative residuals in the field. Figure 8a is
a residual plot of the unsmoothed data in which the red line
separates the points with negative residuals greater than
1000 kg ha
-1
. The high-residual points are plotted as red dots
in figure 8b, which shows that some of these points are in
rainfed regions and some are located adjacent to areas of yel-
low points that had yield less than 1500 kg ha
-1
. The yellow
points were in areas of lowest EC
a
, corresponding to areas
with a high-profile sand content (Sudduth et al., 2003) and
therefore a low soil water holding capacity. This resulted in
short plants and low yield, although the negative residuals
Figure 7. Scatter plot of seed cotton yield versus plant height. The linear function and R
2
value were obtained based on the smoothed data. The
points below the green dashed line are those with negative residuals greater than 1000 kg ha
-1
.
Figure 8. Unsmoothed data points with high negative residuals are below the red line in (a) and are shown as red dots in (b). The yellow points in
(b) have yield less than 1500 kg ha
-1
. “R” marks rainfed regions with no irrigation. The white circles mark regions of lowest apparent soil electrical
conductivity (EC
a
). The region within the purple circle has relatively low EC
a
compared to other parts of the field.
62(2): 393-403 401
were not greater than 1000 kg ha-1 because the plant height
and yield were both small. The red points located on the
boundaries of the sandy soil regions are where the water
shortage was not as severe or persistent, resulting in reduc-
tions in yield with less reduction in plant height and creating
highly negative residuals. Meanwhile, the red points in the
purple circle were in an area with low soil ECa, where the
coarse soil texture (i.e., high sand content) led to water
stress, even with irrigation, and contributed to low yield and
high negative residuals.
DISCUSSION
The primary goal of this study was to evaluate the perfor-
mance of using plant height acquired with UAV-based im-
aging data for cotton yield estimation. Plant height is an im-
portant index for plant growth status, biomass (Brocks and
Bareth, 2018), and yield (Sharma et al., 2016). The plant
height measurement accuracy obtained in this study
(RMSE = 0.08 m) is comparable to that of other studies us-
ing similar methods. For example, RMSE of 0.08 to 0.09 m
was obtained by Brocks and Bareth (2018), and RMSE of
0.05 to 0.1 m was obtained by Chang et al. (2017). With
RMSE values less than 0.1 m, UAV-based height measure-
ment has potential to replace manual measurement. Bendig
et al. (2015) showed that plant height was more accurate than
spectral indices for biomass estimation, although the combi-
nation of plant height and spectral indices was better than
either alone. Meanwhile, Ballesteros et al. (2018) found that
plant height had high correlation with spectral indices and
canopy cover. Therefore, improved methods of plant height
acquisition could provide another approach for cotton yield
estimation.
A unique aspect of this study was that UAV-based sens-
ing data (height, potentially other sensing data, such as veg-
etation indices, and temperature) were registered with
ground data (yield, irrigation, soil, and fertilizer) row-by-
row, which provided high-resolution site-specific infor-
mation. Other studies have correlated yield and remote sens-
ing image data at a lower resolution (e.g., plot level) than
this study. For example, Huang et al. (2016) analyzed the
correlation between yield and UAV-based vegetation indi-
ces from 20 plots where one yield point was acquired from
each plot. Van Der Meij et al. (2017) divided a field into
42 equal plots, and one yield point was acquired from each
plot. Similarly, Vega et al. (2015) used one data point from
each of 32 plots. In contrast, this study connected image data
and geo-referenced ground data using row separation at a
scale of 30 30 pixels, resulting in about 7000 data points.
Although the geo-registration might be improved using a
more accurate RTK-GPS unit, the row separation method
provides a low-cost alternative. The row-based geo-registra-
tion method provided flexibility in estimating yield for one
row or for the whole field. Meanwhile, this method makes it
easy to interpret and explore the data in each row, as shown
in figures 4 and 5.
A bin smoothing method was used to smooth the plant
height and corresponding yield data, which increased the
correlation between plant height and cotton yield. Figure 7
shows that the plant height and yield of cotton had a linear
relationship with R2 = 0.96 for the smoothed data. This result
was different from the results reported by Huang et al.
(2016), where cotton in the middle of the height range had
higher yield compared to cotton with low and high plant
heights (R2 = 0.43). The main reason for the different results
from the two studies may be that the height of the cotton in
this study was within the preferred range, without excessive
growth. Excessive height in cotton (i.e., rank growth) can be
detrimental but is usually adequately controlled with plant
growth regulators. Possible reasons for low yield were ana-
lyzed, and water stress and coarse soil texture, as indicated
by low soil ECa, might be the major reasons for the areas of
low yield. This result was consistent with the findings re-
ported by Vories et al. (2015).
CONCLUSION
A method was established for plant height extraction and
seed cotton yield estimation based on plant height data from
UAV-based remote sensing. A UAV equipped with a low-
cost RGB camera was used to collect images of cotton dur-
ing the flowering period. A DEM was generated from the
RGB images based on the SfM method. Plant height was ex-
tracted from the DEM by calculating the difference between
the canopy height and soil surface. Linear interpolation was
used to estimate the soil elevation in regions where the soil
was invisible due to canopy closure. A geo-registration
method based on row separation was developed to improve
the accuracy of geo-referencing the ground and UAV-based
sensing data. When the row separation method was included
in the geo-registration process, the alignment result was bet-
ter and the correlation between yield and plant height in-
creased from 81% to 83%, showing the effectiveness of row
separation. A bin smoothing method was used to smooth the
plant height corresponding with yield data. A linear model
showed that plant height and cotton yield had a positive re-
lationship. Yield estimation errors of RMSE = 550 kg ha-1
and MAE = 420 kg ha-1 were obtained using 10-fold cross-
validation. From the above results, we conclude that a UAV
equipped with a low-cost RGB camera has potential to map
plant height for cotton yield estimation. Data points with low
yield and high negative residuals (i.e., tall plants with low
yield) from the regression estimate were examined, and it
was found that water stress and low soil water holding ca-
pacity, as indicated by low soil ECa, might be the main con-
tributing factors.
ACKNOWLEDGEMENTS
The authors would like to thank the University of Mis-
souri Agricultural Experiment Station for partly funding this
project. We also thank colleagues Jing Zhou, Chin Nee
Vong, and Dr. Xiuqing Fu for their help in data collection,
data analysis, and improvement of the manuscript.
REFERENCES
Ballesteros, R., Ortega, J. F., Hernandez, D., & Moreno, M. A.
(2018). Onion biomass monitoring using UAV-based RGB
402 TRANSACTIONS OF THE ASABE
imaging. Prec. Agric., 19(5), 840-857.
https://doi.org/10.1007/s11119-018-9560-y
Bendig, J., Bolten, A., Bennertz, S., Broscheit, J., Eichfuss, S., &
Bareth, G. (2014). Estimating biomass of barley using crop
surface models (CSMs) derived from UAV-based RGB
imaging. Remote Sensing, 6(11), 10395-10412.
https://doi.org/10.3390/rs61110395
Bendig, J., Yu, K., Aasen, H., Bolten, A., Bennertz, S., Broscheit,
J., ... Bareth, G. (2015). Combining UAV-based plant height
from crop surface models, visible, and near-infrared vegetation
indices for biomass monitoring in barley. Intl. J. Appl. Earth
Obs. Geoinfo., 39, 79-87.
https://doi.org/10.1016/j.jag.2015.02.012
Benker, S. C., Langford, R. P., & Pavlis, T. L. (2011). Positional
accuracy of the Google Earth terrain model derived from
stratigraphic unconformities in the Big Bend region, Texas,
USA. Geocarto Intl., 26(4), 291-303.
https://doi.org/10.1080/10106049.2011.568125
Brocks, S., & Bareth, G. (2018). Estimating barley biomass with
crop surface models from oblique RGB imagery. Remote
Sensing, 10(2), 268. https://doi.org/10.3390/rs10020268
Chang, A., Jung, J., Maeda, M. M., & Landivar, J. (2017). Crop
height monitoring with digital imagery from unmanned aerial
system (UAS). Comput. Electron. Agric., 141, 232-237.
https://doi.org/10.1016/j.compag.2017.07.008
Chu, T., Chen, R., Landivar, J., Maeda, M. M., Yang, C., & Starek,
M. J. (2016). Cotton growth modeling and assessment using
unmanned aircraft system visual-band imagery. J. Appl. Remote
Sensing, 10, 036018. https://doi.org/10.1117/1.JRS.10.036018
De Castro, A. I., Torres-Sanchez, J., Pena, J. M., Jimenez-Brenes, F.
M., Csillik, O., & Lopez-Granados, F. (2018). An automatic
random forest-OBIA algorithm for early weed mapping between
and within crop rows using UAV imagery. Remote Sensing,
10(2), 285. https://doi.org/10.3390/rs10020285
Du, M., & Noguchi, N. (2017). Monitoring of wheat growth status
and mapping of wheat yield’s within-field spatial variations
using color images acquired from UAV-camera system. Remote
Sensing, 9(3), 289. https://doi.org/10.3390/rs9030289
Duan, T., Zheng, B., Guo, W., Ninomiya, S., Guo, Y., & Chapman,
S. C. (2017). Comparison of ground cover estimates from
experiment plots in cotton, sorghum, and sugarcane based on
images and ortho-mosaics captured by UAV. Funct. Plant Biol.,
44(1), 169-183. https://doi.org/10.1071/FP16123
Glendell, M., McShane, G., Farrow, L., James, M. R., Quinton, J.,
Anderson, K., ... Brazier, R. E. (2017). Testing the utility of
structure-from-motion photogrammetry reconstructions using
small unmanned aerial vehicles and ground photography to
estimate the extent of upland soil erosion. Earth Surf. Proc.
Landforms, 42(12), 1860-1871. https://doi.org/10.1002/esp.4142
Härdle, W. (1990). Applied nonparametric regression. Econometric
Society Monograph No. 19. Cambridge, UK: Cambridge
University Press. https://doi.org/10.1017/CCOL0521382483
Hastie, T., Tibshirani, R., & Friedman, J. (2001). The elements of
statistical learning: Data mining, inference, and prediction.
New York, NY: Springer Science+Business Media.
Huang, Y., Brand, H. J., Sui, R., Thomson, S. J., Furukawa, T., &
Wayne Ebelhar, M. (2016). Cotton yield estimation using very
high-resolution digital images acquired with a low-cost small
unmanned aerial vehicle. Trans. ASABE, 59(6), 1563-1574.
https://doi.org/10.13031/trans.59.11831
Huang, Y., Sui, R., Thomson, S. J., & Fisher, D. K. (2013).
Estimation of cotton yield with varied irrigation and nitrogen
treatments using aerial multispectral imagery. Intl. J. Agric. Biol.
Eng., 6(2), 37-41.
Irizarry, R. A. (2001). Applied nonparametric and modern statistics.
Baltimore, MD: Johns Hopkins University.
Jay, S., Rabatel, G., Hadoux, X., Moura, D., & Gorretta, N. (2015).
In-field crop row phenotyping from 3D modeling performed
using structure from motion. Comput. Electron. Agric., 110, 70-
77. https://doi.org/10.1016/j.compag.2014.09.021
Jiao, C., Chen, C., McGarvey, R. G., Bohlman, S., Jiao, L., & Zare,
A. (2018). Multiple instance hybrid estimator for hyperspectral
target characterization and sub-pixel target detection. ISPRS J.
Photogram. Remote Sensing, 146, 235-250.
https://doi.org/10.1016/j.isprsjprs.2018.08.012
Komm, B., & Moyer, M. (2015). Vineyard yield estimation.
EM086E. Pullman, WA: Washington State University
Extension.
Liu, Q.-S., Li, X.-Y., Liu, G.-H., Huang, C., & Guo, Y.-S. (2016).
Cotton area and yield estimation at Zhanhua County of China
using HJ-1 EVI time series. Proc. 3rd Annual Conf. Info. Tech.
and Applications (ITA 2016), article 09001.
https://doi.org/10.1051/itmconf/20160709001
Liu, S., Baret, F., Andrieu, B., Burger, P., & Hemmerle, M. (2017).
Estimation of wheat plant density at early stages using high-
resolution imagery. Front. Plant Sci., 8, 739.
https://doi.org/10.3389/fpls.2017.00739
Maimaitijiang, M., Ghulam, A., Sidike, P., Hartling, S.,
Maimaitiyiming, M., Peterson, K., ... Fritschi, F. (2017).
Unmanned aerial system (UAS)-based phenotyping of soybean
using multi-sensor data fusion and extreme learning machine.
ISPRS J. Photogram. Remote Sensing, 134, 43-58.
https://doi.org/10.1016/j.isprsjprs.2017.10.011
Malambo, L., Popescu, S. C., Murray, S. C., Putman, E., Pugh, N.
A., Horne, D. W., ... Bishop, M. (2018). Multitemporal field-
based plant height estimation using 3D point clouds generated
from small unmanned aerial systems high-resolution imagery.
Intl. J. Appl. Earth Obs. Geoinfo., 64, 31-42.
https://doi.org/10.1016/j.jag.2017.08.014
Mohammed, N. Z., Ghazi, A., & Mustafa, H. E. (2013). Positional
accuracy testing of Google Earth. Intl. J. Multidisc. Sci. Eng.,
4(6), 6-9.
Sankaran, S., Khot, L. R., Espinoza, C. Z., Jarolmasjed, S.,
Sathuvalli, V. R., Vandemark, G. J., ... Pavek, M. J. (2015).
Low-altitude, high-resolution aerial imaging systems for row
and field crop phenotyping: A review. European J. Agron., 70,
112-123. https://doi.org/10.1016/j.eja.2015.07.004
Sankey, T. T., McVay, J., Swetnam, T. L., McClaran, M. P.,
Heilman, P., & Nichols, M. (2018). UAV hyperspectral and lidar
data and their fusion for arid and semi-arid land vegetation
monitoring. Remote Sensing Ecol. Cons., 4(1), 20-33.
https://doi.org/10.1002/rse2.44
Scotford, I. M., & Miller, P. C. (2004). Combination of spectral
reflectance and ultrasonic sensing to monitor the growth of
winter wheat. Biosyst. Eng., 87(1), 27-38.
https://doi.org/10.1016/j.biosystemseng.2003.09.009
Sharma, L. K., Bu, H., Franzen, D. W., & Denton, A. (2016). Use
of corn height measured with an acoustic sensor improves yield
estimation with ground-based active optical sensors. Comput.
Electron. Agric., 124, 254-262.
https://doi.org/10.1016/j.compag.2016.04.016
Smith, M. W., Carrivick, J. L., & Quincey, D. J. (2016). Structure
from motion photogrammetry in physical geography. Prog.
Phys. Geogr., 40(2), 247-275.
https://doi.org/10.1177/0309133315615805
Sudduth, K. A., Drummond, S. T., & Myers, D. B. (2012). Yield
Editor 2.0: Software for automated removal of yield map errors.
ASABE Paper No. 121338243. St. Joseph, MI: ASABE.
Sudduth, K. A., Kitchen, N. R., Bollero, G. A., Bullock, D. G., &
Wiebold, W. J. (2003). Comparison of electromagnetic
induction and direct sensing of soil electrical conductivity.
Agron. J., 95(3), 472-482.
62(2): 393-403 403
https://doi.org/10.2134/agronj2003.4720
Sui, R., Thomasson, J. A., & Ge, Y. (2012). Development of sensor
systems for precision agriculture in cotton. Intl. J. Agric. Biol.
Eng., 5(4), 1-14.
Van Der Meij, B., Kooistra, L., Suomalainen, J., Barel, J. M., & De
Deyn, G. B. (2017). Remote sensing of plant trait responses to
field-based plant-soil feedback using UAV-based optical
sensors. Biogeosciences, 14(3), 733-749.
https://doi.org/10.5194/bg-14-733-2017
Vega, F. A., Ramirez, F. C., Saiz, M. P., & Rosua, F. O. (2015).
Multi-temporal imaging using an unmanned aerial vehicle for
monitoring a sunflower crop. Biosyst. Eng., 132, 19-27.
https://doi.org/10.1016/j.biosystemseng.2015.01.008
Vories, E. D., Stevens, W. E., Sudduth, K. A., Drummond, S. T., &
Benson, N. R. (2015). Impact of soil variability on irrigated and
rainfed cotton. J. Cotton Sci., 19(1), 1-14.
Walter, J., Edwards, J., McDonald, G., & Kuchel, H. (2018).
Photogrammetry for the estimation of wheat biomass and
harvest index. Field Crops Res., 216, 165-174.
https://doi.org/10.1016/j.fcr.2017.11.024
Wang, D., Xin, X., Shao, Q., Brolly, M., Zhu, Z., & Chen, J. (2017).
Modeling aboveground biomass in Hulunber grassland
ecosystem by using unmanned aerial vehicle discrete lidar.
Sensors, 17(1), 180. https://doi.org/10.3390/s17010180
Westoby, M. J., Brasington, J., Glasser, N. F., Hambrey, M. J., &
Reynolds, J. M. (2012). ‘Structure-from-motion’
photogrammetry: A low-cost, effective tool for geoscience
applications. Geomorphology, 179, 300-314.
https://doi.org/10.1016/j.geomorph.2012.08.021
Willmott, C. J., & Matsuura, K. (2005). Advantages of the mean
absolute error (MAE) over the root mean square error (RMSE)
in assessing average model performance. Climate Res., 30(1),
79-82. https://doi.org/10.3354/cr030079
Yang, G., Liu, J., Zhao, C., Li, Z., Huang, Y., Yu, H., ... Yang, H.
(2017). Unmanned aerial vehicle remote sensing for field-based
crop phenotyping: Current status and perspectives. Front. Plant
Sci., 8(1111). https://doi.org/10.3389/fpls.2017.01111
Zhang, L., & Grift, T. E. (2012). A LIDAR-based crop height
measurement system for Miscanthus giganteus. Comput.
Electron. Agric., 85, 70-76.
https://doi.org/10.1016/j.compag.2012.04.001
Zhou, J., Khot, L. R., Boydston, R. A., Miklas, P. N., & Porter, L.
(2018). Low-altitude remote sensing technologies for crop stress
monitoring: A case study on spatial and temporal monitoring of
irrigated pinto bean. Prec. Agric., 19(3), 555-569.
https://doi.org/10.1007/s11119-017-9539-0