[8 126-137최종]김학진 use of unmanned aerial vehicle for multi
TRANSCRIPT
Use of Unmanned Aerial Vehicle for Multi-temporal Monitoring of
Soybean Vegetation Fraction
Hee Sup Yun1, Soo Hyun Park
2, Hak-Jin Kim
1,3*, Wonsuk Daniel Lee
4, Kyung Do Lee
5,
Suk Young Hong5, Gun Ho Jung
6
1Deptartment of Biosystems & Biomaterials Science and Engineering, Seoul National University, Seoul 08826, Korea
2KIST Gangneung Institute of Natural Products, Gangneung 25451, Korea
3Research Institute for Agriculture and Life Sciences, Seoul National University, Seoul 08826, Korea
4Department of Agricultural and Biological Engineering, University of Florida, Gainesville, FL 32611, United States5Deptartment of Agricultural Environment, National Academy of Agricultural Science, Jeollabuk-do 55265, Korea
6Upland Crop Research Div., National Institute of Crop Science, Gyeonggi-do 16429, Korea
Received: March 7th
, 2016; Revised: May 12th
, 2016; Accepted: May 20th
, 2016
Purpose: The overall objective of this study was to evaluate the vegetation fraction of soybeans, grown under different
cropping conditions using an unmanned aerial vehicle (UAV) equipped with a red, green, and blue (RGB) camera. Methods:
Test plots were prepared based on different cropping treatments, i.e., soybean single-cropping, with and without herbicide
application and soybean and barley-cover cropping, with and without herbicide application. The UAV flights were manually
controlled using a remote flight controller on the ground, with 2.4 GHz radio frequency communication. For image
pre-processing, the acquired images were pre-treated and georeferenced using a fisheye distortion removal function, and
ground control points were collected using Google Maps. Tarpaulin panels of different colors were used to calibrate the
multi-temporal images by converting the RGB digital number values into the RGB reflectance spectrum, utilizing a linear
regression method. Excess Green (ExG) vegetation indices for each of the test plots were compared with the M-statistic
method in order to quantitatively evaluate the greenness of soybean fields under different cropping systems. Results: The
reflectance calibration methods used in the study showed high coefficients of determination, ranging from 0.8 to 0.9,
indicating the feasibility of a linear regression fitting method for monitoring multi-temporal RGB images of soybean fields.
As expected, the ExG vegetation indices changed according to different soybean growth stages, showing clear differences
among the test plots with different cropping treatments in the early season of < 60 days after sowing (DAS). With the
M-statistic method, the test plots under different treatments could be discriminated in the early seasons of <41 DAS,
showing a value of M > 1. Conclusion: Therefore, multi-temporal images obtained with an UAV and a RGB camera could be
applied for quantifying overall vegetation fractions and crop growth status, and this information could contribute to
determine proper treatments for the vegetation fraction.
Keywords: Barley cover cropping, Excess green, Image processing, M-statistic method, UAV, Vegetation index
Original Article Journal of Biosystems Engineering
J. of Biosystems Eng. 41(2):126-137. (2016. 6)http://dx.doi.org/10.5307/JBE.2016.41.2.126
eISSN : 2234-1862 pISSN : 1738-1266
*Corresponding author: Hak Jin Kim
Tel: +82-2-880-4604; Fax: +82-2-873-2049
E-mail: [email protected]
Copyright ⓒ 2016 by The Korean Society for Agricultural MachineryThis is an Open Access article distributed under the terms of the Creative Commons Attribution Non-Commercial License (http://creativecommons.org/licenses/by-nc/3.0)
which permits unrestricted non-commercial use, distribution, and reproduction in any medium, provided the original work is properly cited.
Introduction
In recent years, UAVs have been commonly used for
low-altitude and high-resolution remote sensing applications
in precision agriculture owing to several advantages such
as their versatile applicability, lightweight, and low
operational costs (Hunt et al., 2010; Garcia-Ruiz et al.,
2013). In particular, UAVs have been successfully used to
assess the vegetation status of crops and to predict yields,
because the operational flexibility of vertical takeoff and
landing platforms with various image sensors makes it
easy to fly over agricultural fields and acquire aerial
images with high spatial and temporal resolutions (Torres-
Yun et al. Use of Unmanned Aerial Vehicle for Multi-temporal Monitoring of Soybean Vegetation FractionJournal of Biosystems Engineering • Vol. 41, No. 2, 2016 • www.jbeng.org
127
Sánchez et al., 2014).
The acquired aerial images can help farmers to evaluate
the crop growth status, while providing useful information
about canopy greenness, leaf area, and water stress
assessment, as well as about various geographical conditions
such as crop areas, digital surface models (DSM), and
depth contour lines (Jimenez and Agudelo, 2015). Recently,
many research results for agricultural applications that
used aerial images obtained from UAVs have been reported,
such as the monitoring of early stage crop growth (Torres-
Sánchez et al., 2014) and the development of a weed
detection method (Peña et al., 2013).
A common use of remote sensing is to evaluate crop
growth status based on canopy greenness by quantifying
the distribution of the vegetation index (VI) in the crop
field. Various vegetation indices, including the Normalized
Difference Vegetation Index (NDVI) and Excess Green
(ExG), are defined as reflectance values of the vegetation
canopy in a given ground area (Woebbecke et al., 1995).
However, since remote sensing images are affected by
changes in illumination, atmospheric conditions, or
viewing geometry, it is necessary to perform various
image- processing steps prior to the vegetation index
analysis, including radiometric calibration and geometric
correction methods. In particular, when investigating
temporal changes in time-based datasets of remotely
sensed images, and because vegetation indices are easily
affected by radiometric disturbances, a careful use of a
radiometric calibration is required for obtaining time-
invariant normalized images. A number of image calibration
and geometric correction methods for small-sized
unmanned vehicles have been developed in order to
account for variations in atmospheric transmission, sun
azimuth, and aircraft attitude (Berni et al., 2009).
Separating images into different classes for extracting
meaningful features, or image thresholding, is a common
task in image-processing applications that is usually
performed based on a certain characteristic of the image
pixels, such as threshold, peak, valley, and clustering.
Various classification algorithms, such as the Otsu and
the valley-emphasis methods, are used to automatically
perform clustering-based image thresholding. The Otsu
method is one of the most commonly used because of its
simplicity to calculate the optimum threshold that separates
the two classes in bimodal histograms. This method
assumes that an image contains two classes of pixels and
calculates an optimum threshold by minimizing the
combined spread (Torres-Sánchez et al., 2014).
The objective of this research was to investigate the
potential of a multirotor-based UAV equipped with a
low-cost digital RGB camera, along with the appropriate
calibration methodologies for quantifying temporal and
spatial variations in the vegetation index of soybeans
grown under different cropping conditions. The treatments
consisted of four cropping systems: soybean single-
cropping with and without herbicide application and
soybean cropping with barley as a cover crop, with and
without herbicide application. The assessment was conducted
based on a comparison of the vegetation index maps
obtained from the different crop canopies.
Materials and Methods
Test plots
UAV remote sensing was conducted in a soybean field
of the National Institute of Crop Science (NICS), Rural
Development Administration (RDA), in 2014. The soybean
field was located in 894-51, Gosaek-dong, Gwonseon-gu,
Suwon-si, Gyeonggi-do, Korea (378530E, 4125344N, 30
m above sea level, UTM zone 52N). The effects of
herbicide and the barley crop cover were tested for the
soybean field experiment, and the effects were compared
focusing on soybean growth status and the overall
vegetation fraction.
A soybean crop (Glycine max L. Merrill) was planted on
May 22. A split plot design was used for the herbicide
treatment and for the barley crop cover arrangement.
The test site was divided into 14 individual plots of 30×14
m, corresponding to a total area of about 0.3 ha (Figure 1).
A conventional amount of herbicide for soil treatment
(mixture of Ethalfluralin and Linuron) was applied to the
main plot, and barley (Fenn beard-less, 15 kg/10a) was
planted as a cover crop in the subplot. The subplots contained
three replications for the barley-covered soybean and
four replications for the conventional soybean field.
The barley cover crop in test plots was expected to
suppress the development of weed in the early season
and maintain a proper vegetation fraction (Kobayashi et
al., 2004). In addition, it was reported that the management
of a proper vegetation fraction is important for upland
fields to prevent soil erosion in the fallow period and
early season (Lu et al., 2000; Poggio, 2005).
Yun et al. Use of Unmanned Aerial Vehicle for Multi-temporal Monitoring of Soybean Vegetation FractionJournal of Biosystems Engineering • Vol. 41, No. 2, 2016 • www.jbeng.org
128
Figure 1. Split plot design for the soybean experiments with two treatments and three replications. The ‘N’ letter indicates No-herbicide sprayed, while ‘H’ letter indicates Herbicide sprayed. The ‘B’ letter indicates Barley cover crop, and ‘A’ letter indicates soybean alone.
Figure 2. Views of the UAV platform equipped with a camera (a), the UAV flying over the experimental crop field and the color reference panels (b).
Table 1. Specification of the UAV platform used in the study
Item Specifications
Airframe DJI F550 Hexa-rotor
Flight controller DJI NAZA M Lite
Diagonal
wheelbase550 mm
Propeller 8 × 4.5 in (8045)
Battery 4S Li-Po 6000 mAh
Motor Stator size: 22 × 12 mm, KV: 920 rpm/V
ESC 30 A OPTO, Signal frequency: 30 Hz ~ 450 Hz
Takeoff weight 2,272 g
Maximum flight
time15 min (approx.)
UAV platform and image sensor
The UAV airframe system used in this study was built
with a hexa-rotor airframe model (F550, DJI, China)
equipped with a low-cost digital RGB camera (HERO 3+,
GoPro, USA). The UAV flight was manually controlled
using an R/C flight controller (NAZA M Lite, DJI, China) on
the ground through a 2.4 GHz radio frequency communica-
tion. Figure 2 shows the UAV (Figure 2a) and the color
reference panels used in the study for calibration (Figure
2b). Detail specifications of the UAV platform are
presented in Table 1.
The camera used in the study to capture RGB images
provided 12 megapixels, with a SONY IMX 117 image
sensor and a 14 mm focal length lens. According to the
datasheet provided by the company, as shown in Figure 3,
different spectral responses to different wavelengths of
red, green, and blue bands were considered in image
processing.
Image acquisition and preprocessing
Prior to the image acquisition with the UAV, real-time
Yun et al. Use of Unmanned Aerial Vehicle for Multi-temporal Monitoring of Soybean Vegetation FractionJournal of Biosystems Engineering • Vol. 41, No. 2, 2016 • www.jbeng.org
129
Figure 3. Relative spectral responses to different wavelengths of the SONY IMX117 RGB sensors.
Figure 4. Footprints for the images acquired during a flight plan conducted by changing the flight altitudes over the soybean test field. Green rectangles correspond to single image frames.
Figure 5. The original image acquired with the GoPro camera (a) and the image corrected from lens distortion (b).
images were monitored to determine a proper flight
altitude and field of view. For this purpose, a Wi-Fi
remote application (GoPro App, GoPro, USA) was used,
connecting the camera image into an Android smartphone.
The images were acquired based on a time-lapse function
that took one image each three seconds, while varying the
altitude and orientation of the UAV (Figure 4). The UAV
aerial images of the test field were taken on a 2-week
interval for four months, beginning in the period from
June to October 2014.
As shown in Figure 5, in the first step of image pre-
processing, the acquired images were pre-treated using a
fisheye-removal function provided by a commercial
photo-handling software (Adobe Lightroom 5.5) in order
to generate images corrected from lens distortion. The
images were then georeferenced with ground control
points (GCPs) collected with Google Maps. Finally, image
mapping was performed using the GCPs based on a
WGS84 UTM 52N coordinate system (Figure 6).
Radiometric calibration
A radiometric calibration of the RGB digital numbers in
each original image, captured under different outdoor
illumination conditions, was conducted using the empirical
line method (Smith and Milton, 1999). Using commercial
color tarpaulins with uniform reflectance characteristics,
seven 600×900 mm calibration targets were placed in a
location within the flight path of the UAV platform. As
shown in Figure 7, the reference spectral reflectance of
Yun et al. Use of Unmanned Aerial Vehicle for Multi-temporal Monitoring of Soybean Vegetation FractionJournal of Biosystems Engineering • Vol. 41, No. 2, 2016 • www.jbeng.org
130
Figure 6. Ground Control Points used for image mapping (a) and result image (b).
Figure 7. Reflectance spectrum of the calibration reference targets.
Table 2. Calculated mean spectral reflectance of the calibration targets
Calibration target Calibration target
Red tarpaulin 0.33 0.14 0.11 Gray tarpaulin 0.28 0.25 0.26
Green tarpaulin 0.30 0.38 0.28 Black tarpaulin 0.09 0.09 0.09
Blue tarpaulin 0.11 0.10 0.12 White tarpaulin 0.89 0.89 0.89
Yellow tarpaulin 0.43 0.26 0.15
the calibration targets was measured in the 400-700 nm
spectral range with a spectrophotometer (ColorMate,
Scinco Co., Korea). To determine the RGB reflectance
spectra to be used in the empirical line calibration
method, the mean reflectance values of the calibration
targets were calculated using Eq. 1 (Table 2).
(1)
where,
: Sepctral reflectance of the calibration target,
calibration targets
: Relative spectral sensitivity of the image sensor,
R, G, B bands
The empirical line calibration method derived the
coefficients needed to fit the uncalibrated RGB imagery
acquired with the GoPro camera to the lab-measured
reflectance spectra of the field targets. It was assumed
that the reflectance values of the calibration targets were
proportional to the RGB band digital numbers (Wang and
Myint, 2015) (Eq. 2).
(2)
where,
and : Coefficients of the linear equation obtained
from regression.
: Reflectance, : digital number for each band.
Figure 8 shows sample calibration curves relating
digital numbers to reflectance for each of the RGB bands.
The UAV images were acquired from the standard panels
placed on the ground, thereby exhibiting coefficients of
determination higher than 0.9.
The results of the linear line fitting obtained with this
empirical method demonstrated that the spectral reflectance
was successfully calibrated, as compared with the standard
spectrophotometer measurements, yielding in most
instances coefficients of determination between 0.85 and
0.99 (Table 3).
Vegetation index calculation and evaluation
A vegetation index of ExG (Excess Green, Eq. 3) was
used in the study for quantifying the vegetation fraction
Yun et al. Use of Unmanned Aerial Vehicle for Multi-temporal Monitoring of Soybean Vegetation FractionJournal of Biosystems Engineering • Vol. 41, No. 2, 2016 • www.jbeng.org
131
Figure 8. Sample calibration curves of the RGB sensors that show linear equation relationship between the digital numbers and reflectance spectra, with coefficients of determination > 0.9 (20141015, 146 days after sowing).
Table 3. Results of the calibration curves of the RGB sensor for images obtained at different dates
Image date Image date
20140619
(28 DAS)
R 0.0250 0.093 0.9120140823
(93 DAS)
R 0.0033 -0.193 0.89
G 0.0020 0.125 0.99 G 0.0021 0.015 0.97
B 0.0021 0.163 0.88 B 0.0016 0.067 0.54
20140702
(41 DAS)
R 0.0026 0.069 0.9820140904
(105 DAS)
R 0.0024 0.049 0.95
G 0.0029 -0.013 0.96 G 0.0026 -0.012 0.95
B 0.0022 0.042 0.96 B 0.0020 0.012 0.91
20140716
(55 DAS)
R 0.0023 0.131 0.9720140922
(123 DAS)
R 0.0018 0.246 0.97
G 0.0022 0.109 0.96 G 0.0019 0.193 0.94
B 0.0017 0.129 0.85 B 0.0015 0.188 0.90
20140730
(69 DAS)
R 0.0033 -0.155 0.9520141007
(138 DAS)
R 0.0018 0.226 0.98
G 0.0032 -0.144 0.91 G 0.0019 0.167 0.97
B 0.0025 -0.065 0.79 B 0.0015 0.150 0.93
20140812
(82 DAS)
R 0.0036 -0.275 0.9120141015
(146 DAS)
R 0.0019 0.237 0.95
G 0.0019 -0.008 0.95 G 0.0021 0.156 0.94
B 0.0016 0.036 0.74 B 0.0017 0.149 0.90
in the test plots, since it was reported that the ExG values
could effectively visualize the vegetation fraction and
vegetation growth based on RGB images (Torres-Sánchez
et al., 2014). In this study, the vegetation index was
calculated from the radiometrically calibrated RGB reflectance
values, instead of from the RGB digital numbers.
(3)
where,
and : reflectance values of an image
Vegetation fraction
In order to perform an image classification, the value of
each grayscale image pixel was compared with a prefixed
threshold; if the pixel value was higher than the threshold,
then it was recognized as a vegetation pixel. Once the
image pixels were classified, the vegetation fraction (VF)
was determined as the percentage of pixels classified as
vegetation per unit area of ground surface, which was the
ratio between pixels classified as vegetation and total
pixels in the delimited area (Torres-Sánchez et al., 2014).
The VF values for each frame were calculated using each
of thresholds values. The Otsu threshold was applied in
the greyscale images of every temporal acquisition. The
Otsu threshold is widely used to perform clustering-based
image classification automatically, so it was chosen to
evaluate the feasibility of an automatic threshold determi-
nation. The obtained threshold values were compared to
the original color image to determine whether the binary
Yun et al. Use of Unmanned Aerial Vehicle for Multi-temporal Monitoring of Soybean Vegetation FractionJournal of Biosystems Engineering • Vol. 41, No. 2, 2016 • www.jbeng.org
132
Figure 9. Sample histograms used in the M-statistic applied to the test plots (20140702, 41 days after sowing).
Figure 10. Flowchart of the general image processing steps including image acquisition, image preparation, vegetation index calculation, and data analysis.
image could explain the vegetation fraction and whether
this methodology was applicable to quantify VF using
UAV imagery.
Treatment separability
A statistical index named M-statistic, utilizing histograms
obtained from different treatments, was used to assess
the vegetation fractions in the test plots under different
treatments, both in terms of temporal and spatial variability
(Eq. 4, Kaufman and Remer, 1994).
In principle, M-Statistic quantifies the degree of
discrimination between two different pixel groups,
assuming that they are normally distributed. An M value
between the pixel groups lower than 1 means that the
two histograms overlap each other and have a poor
separability, whereas an M value higher than 1 means
that the histograms are well discriminated. (Pereira et al.,
1999).
Figure 9 shows sample histograms used in the M-statistics.
(4)
where,
,
: Mean vegetation index values for the treatments
,
,
: Standard deviation of vegetation index values
for the treatments ,
Figure 10 shows the flow chart of the general steps
conducted in the study, including image acquisition,
image preprocessing, calculation of vegetation fraction,
and data analysis.
Results and Discussion
Multi-temporal soybean field images
Figure 11 shows changes in the soybean field images
taken with the multirotor-based UAV on a 2-week interval,
along with appropriate image calibrations and removal
of lens distortion. As expected, the color images clearly
changed over time according to varying soybean and
barley growing stages. This experiment demonstrated
that RGB images onboard an UAV could provide useful
information about how canopies of soybeans change over
time under different cropping conditions, thus enabling
the identification of vegetation fraction changes in soybean
fields.
Changes in ExG vegetation index over time
Figure 12 shows ExG vegetation index maps of the
soybean test plots calculated from the multi-temporal
images taken at different dates. As expected, they show
that the vegetation fractions changed over time, due to
changes in the soybean canopy greenness. Furthermore,
there were clear differences in ExG values among individual
plots cultivated under varying cropping conditions,
especially in the early season of < 55 days after sowing
(DAS). However, the ExG values reached steady-state
responses, showing no difference between the test plots,
particularly in the late season of > 120 DAS.
In order to compare the ExG values obtained from the
different treatments, mean values and standard deviations
were calculated using the UAV images from each of the
test plots. Similar to the graphical observation in Figure
12, there were clear differences in mean ExG values
among the test plots when image maps from the early
season of < 60 DAS were compared (Figure 13). At 28
Yun et al. Use of Unmanned Aerial Vehicle for Multi-temporal Monitoring of Soybean Vegetation FractionJournal of Biosystems Engineering • Vol. 41, No. 2, 2016 • www.jbeng.org
133
Figure 11. Multi-temporal soybean field images taken with a UAV, along with radiometric calibrations and corrections for lens distortion, at different days after sowing.
DAS, the ExG values for the test plots without herbicide
application, NA and NB, were higher than those for the
herbicide-sprayed fields, HA and HB. In addition, at the
early stage of fields with the same herbicide treatment,
higher ExG values were measured for the barley cover
cropped field than for those with a soybean monoculture.
At 41 DAS, the order of the ExG mean values changed,
with decreasing values for no herbicide-treated and
barley cover cropped plots and increasing values for the
test plots without herbicide application and barley cover
crop.
The vegetation fractions, derived from the ExG vegetation
index and the Otsu method, showed significant changes
only in the early season, because after the mid-growth
season they were saturated. As shown in Figure 14, in
images taken 28 days after sowing, the herbicide-sprayed
plots showed a lower vegetation fraction than did the
non-sprayed plots, because the herbicide suppressed
overall vegetation growth, including the soybean crop.
Barley cover cropped plots showed a higher vegetation
fraction, because barley grows faster than soybean and
maintains a high vegetation fraction in the spring season.
As for images taken 41 days after sowing, the vegetation
fraction was saturated in non-sprayed fields. In herbicide
sprayed fields, barley cover cropped plots showed higher
vegetation fraction values than those from soybean mono-
culture plots.
Yun et al. Use of Unmanned Aerial Vehicle for Multi-temporal Monitoring of Soybean Vegetation FractionJournal of Biosystems Engineering • Vol. 41, No. 2, 2016 • www.jbeng.org
134
Figure 12. ExG vegetation index maps of the soybean test plots, calculated from multi-temporal images taken at different days after sowing.
Figure 13. Multi-temporal change of mean ExG values under differenttreatments.
Separability tests
M-statistic values were calculated using the means and
standard deviations of ExG pixel groups for different test
plots (Table 4). In the early season, it appeared that the
treatment effect is well defined by the ExG vegetation
index, since the M-Statistic values were the highest
() at 41 DAS. However, after 82 DAS, the M-Statistic
values were lower (≅) and the treatment separability
became poor. Hence, it was thought that the treatments
could not be discriminated in the late season when using
the M-statistics values.
The differences in the ExG vegetation index as well as
the treatment separability found in the early season were
verified using low flight altitude images (<10 m) (Figure
15), assuming that ExG values are proportional to the
vegetation fraction or vegetation growth (Torres-Sánchez
et al., 2014). Barley was a dominant crop in the test fields,
growing faster than soybean in the early stages, thus
supporting the result that barley cover-cropped plots
have a higher vegetation fraction than soybean monoculture
Yun et al. Use of Unmanned Aerial Vehicle for Multi-temporal Monitoring of Soybean Vegetation FractionJournal of Biosystems Engineering • Vol. 41, No. 2, 2016 • www.jbeng.org
135
Figure 14. Vegetation fraction in monoculture and barley cover cropped fields in the early season, 28 and 41 days after sowing.
Table 4. Comparison of M-Statistics values obtained from four different test plots at different crop growth stages
DAS NA-NB NA-HA NA-HB NB-HA NB-HB HA-HB
28 -0.26 0.73 0.23 1.51 0.73 -0.82
41 1.15 2.25 2.00 1.50 1.07 -0.62
55 1.15 0.04* 0.89 -0.93 -0.27 0.72
69 -1.01 0.26 0.55 1.25 1.20 0.37
82 0.25 -0.41 0.05* -0.63 -0.16 0.36
93 0.03* 0.51 0.88 0.54 0.98 0.29
105 0.34 -0.41 -0.03* -0.70 -0.31 0.34
123 0.19 0.03* 0.15 -0.14 -0.08* 0.09*
138 -0.01* 0.59 0.61 0.67 0.70 -0.05*
146 0.01* 0.31 0.12 0.34 0.13 -0.22
Bold for M > 1 (Good Separability) and * for M ≅(Bad Separability).
Figure 15. Low altitude images as ground truth images: 28 DAS (left) and 41 DAS (right).
plots.
Moreover, it was clearly shown that the field with
herbicide application showed a lower vegetation fraction
than did the field without herbicide, due to the herbicide
effect on grass suppression. The ‘NA’ treatment showed
the highest ExG value at 41 DAS, and such a high index
might be related to the presence of uncontrolled weeds
(Figure 15). As shown in Figure 15, the barley cover
cropped field showed the highest vegetation fraction,
compared to the soybean monoculture field at 28 and 41
DAS. Furthermore, it shows that the herbicide treatment
suppressed the overall vegetation, including soybean,
barley, and weed, thus supporting the vegetation fraction
results in Figure 12.
In addition, the effect of barley crop cover on ExG was
less significant after the early season, since barley
Yun et al. Use of Unmanned Aerial Vehicle for Multi-temporal Monitoring of Soybean Vegetation FractionJournal of Biosystems Engineering • Vol. 41, No. 2, 2016 • www.jbeng.org
136
Figure 16. Comparison of ground truth images, 55 days (left) and 69 days (right) after sowing.
perished after 55 DAS (Figure 16). However, the barley
cover crop had no effect on the soybean crop after 55 DAS,
indicating that barley in its early growth stage could
suppress the growth of other plants, including soybean,
when they are at similar growth stages (Todd et al., 1999;
Corre-Hellou et al., 2011). In summary, our study using
the UAV low-flight altitude RGB images for data analysis
of ExG values demonstrated that ExG values could be
applicable in identifying the influence of herbicide use and
barley cover cropping on plant growth.
Conclusion
We focused on visualizing multi-temporal soybean
growth using an UAV and an RGB camera. The acquired
images were corrected for lens distortion and then
georeferenced. The images were calibrated using an
empirical method based on linear regression to generate
ExG image maps, which could then be used to quantify the
vegetation fraction of soybean. The temporal ExG
values obtained from different soybean test plots were
considerably affected by herbicide and barely cover crop
treatments in the early season of < 60 DAS. Therefore,
multi-temporal images obtained with an UAV and a RGB
camera could be applied for quantifying overall vegetation
fractions and crop growth status, and this information
could contribute to determine proper treatments for the
vegetation fraction.
Conflict of Interest
The authors have no conflicting financial or other
interests.
Acknowledgement
This paper was supported by NIA (The National
Information Society Agency) and IPET (Korea Institute of
Planning and Evaluation for Technology of Food, Agriculture,
Forestry and Fisheries) in 2014-2016.
References
Berni, J. A., P. J. Zarco-Tejada, L. Suárez and E. Fereres,
2009. Thermal and narrowband multispectral remote
sensing for vegetation monitoring from an unmanned
aerial vehicle. IEEE Transactions on Geoscience and
Remote Sensing, 47(3):722-738.
Corre-Hellou, G., A. Dibet, H. Hauggaard-Nielsen, Y. Crozat,
M. Gooding, P. Ambus and E.S. Jensen. 2011. The
competitive ability of pea–barley intercrops against
weeds and the interactions with crop productivity
and soil N availability. Field Crops Research 122(3):
264-272.
Garcia-Ruiz, F., S. Sankaran, J. M. Maja, W. S. Lee, J.
Rasmussen and R. Ehsani. 2013. Comparison of two
aerial imaging platforms for identification of
Huanglongbing-infected citrus trees. Computers and
Electronics in Agriculture 91:106-115.
Hunt, E. R., W. D. Hively, S. J. Fujikawa, D. S. Linden, C.S.
Daughtry and G.W. McCarty. 2010. Acquisition of
NIR-green-blue digital photographs from unmanned
aircraft for crop monitoring. Remote Sensing 2(1):
290-305.
Jimenez, P. L. and D. Agudelo 2015. Validation and
calibration of a high resolution sensor in unmanned
aerial vehicles for producing images in the IR range
utilizable in precision agriculture. In: Proceedings of
Yun et al. Use of Unmanned Aerial Vehicle for Multi-temporal Monitoring of Soybean Vegetation FractionJournal of Biosystems Engineering • Vol. 41, No. 2, 2016 • www.jbeng.org
137
the AIAA SciTech, Paper No. 2015-0988. Kissimmee,
FL: AIAA Infotech @ Aerospace.
Kaufman, Y. J. and L. A. Remer. 1994. Detection of forests
using mid-IR reflectance: an application for aerosol
studies. IEEE Transactions on Geoscience and Remote
Sensing 32(3):672-683.
Kobayashi, H., S. Miura and A. Oyanagi. 2004. Effects of
winter barley as a cover crop on the weed vegetation
in a no‐tillage soybean. Weed Biology and Management
4(4):195-205.
Lu, Y. C., K. B. Watkins, J. R. Teasdale and A. A. Abdul-Bakil.
2000. Cover crops in sustainable food production.
Food Reviews International 16(2):121-157.
Poggio, S. L. 2005. Structure of weed communities
occurring in monoculture and intercropping of field
pea and barley. Agriculture, Ecosystems and Environment
109(1):48-58.
Peña, J. M., J. Torres-Sánchez, A. I. de Castro, M. Kelly, and
F. López-Granados. 2013. Weed mapping in early-
season maize fields using object-based analysis of
unmanned aerial vehicle (UAV) images. PLoS One
8(10):e77151.
Pereira, J. M., A. C. Sá, A. M. Sousa, J. M. Silva, T. N. Santos,
and J. M. Carreiras. 1999. Spectral characterization
and discrimination of burnt areas. In: Remote sensing
of large wildfires, pp. 123-138. Springer-Berlag Berlin
Heidelberg.
Smith, G. M., and E. J. Milton. 1999. The use of the empirical
line method to calibrate remotely sensed data to
reflectance. International Journal of remote sensing,
20(13):2653-2662.
Todd, A. P., C. B. Orvin, and H. O. James. 1999. Increasing
crop competitiveness to weeds through crop breeding.
Journal of Crop Production 2(1):59-76.
Torres-Sánchez, J., J. M. Peña, A. I. de Castro, and F.
López-Granados. 2014. Multi-temporal mapping of
the vegetation fraction in early-season wheat fields
using images from UAV. Computers and Electronics in
Agriculture 103:104-113.
Wang, C. and S. W. Myint, 2015. A simplified empirical line
method of radiometric calibration for small unmanned
aircraft systems-based remote sensing. IEEE Journal
of Selected Topics in Applied Earth Observations and
Remote Sensing, 8(5):1876-1885.
Woebbecke, D. M., G. E. Meyer, K. Von Bargen, and D. A.
Mortensen, 1995. Color indices for weed identification
under various soil, residue, and lighting conditions.
Transactions of the ASAE 38(1):259-269.