Available via license: CC BY
Content may be subject to copyright.
CORRELATION OF UAV-BASED MULTISPECTRAL VEGETATION INDICES
AND LEAF COLOR CHART OBSERVATIONS FOR NITROGEN
CONCENTRATION ASSESSMENT ON RICE CROPS
.
C. M. Bacsa 1, R. M. Martorillas 1, L. P. Balicanta 1, A. M. Tamondong 1
1Department of Geodetic Engineering, University of the Philippines – Diliman, Quezon City, Philippines - (cmbacsa,
rmmartorillas, amtamondong, lpbalicanta)@up.edu.ph
Commission IV
KEY WORDS: Leaf Color Chart (LCC), Unmanned Aerial Vehicle (UAV), Nitrogen Monitoring, Fertilizer Application, Rice
Crops, Vegetation Index (VI)
ABSTRACT:
Fertilizer application is a crucial farming operation for regulating crop health thus crop yield. Optimal fertilizing doubles
agricultural production subsequently raising farmers’ income, food security and economic agriproducts. To optimize the
application of fertilizers, initial monitoring of the current nutrient status of the crops is required. This research will focus on
Nitrogen (N), the most extensive fertilizer nutrient in crop cultivation. Conventional N monitoring involves the use of Leaf
Color Charts (LCC) wherein leaf color intensity is associated with the N content of the crops. Despite its ability to quantify
the optimal amount of needed fertilizers, the LCC method requires extensive on-site labor and lacks accuracy. This study
developed a method that incorporates capabilities of Unmanned Aerial Vehicles (UAVs) equipped with a multispectral sensor
in N monitoring specifically in rice crops, a major agricultural product in the Philippines. In situ N level information collected
through LCC was correlated with remote sensing data, particularly vegetation indices (VIs) extracted from UAV multispectral
imagery of a rice plantation in San Rafael, Bulacan. Several VIs sensitive to crop N content were tested to determine which
has the highest correlation with the LCC data. Through Pearson correlation and regression analysis, NDVI Red Edge was found
to be the most strongly correlated with LCC data suggesting its potential in mapping variability in fertilizer requirements. An
equation modelling LCC observations and NDVI Red Edge values that estimates the N levels of an entire rice plantation was
generated along with the N concentration map of the study area.
1. INTRODUCTION
1.1 Background of the Study
Nutrient management is an integral part of fertilizer
application. Practical methods in balancing nutrients
optimize crop yield, maintain soil health and reduce the
chances of surface and ground water contamination.
Optimizing fertilization involves monitoring the existing
nutrient status of the present vegetation stage. The level and
concentration of crop nutrients are initially assessed to
quantify sufficient amounts of needed fertilizers.
Nitrogen (N) is a primary crop nutrient essential for rapid
growth, grain yield and quality. Among the fertilizers used
by local farmers, 39.4% of which are N-based making it the
most commonly used nutrient in farming (Cruz P. S., 1997).
These type of fertilizers, when applied at optimal amounts
significantly increase agricultural output at low cost.
However both excessive and deficient fertilizer application
result to imbalances in the agricultural setting. Thus, the
research industry continues to seek efficient ways for N
management in farmlands. Through the collaboration of the
International Rice and Research Institute (IRRI) and
PhilRice, the Philippine agricultural sector adopted the use
of Leaf Color Charts (LCCs). These are four-panel tools with
varying shades of green conventionally used to determine
leaf color intensity directly related to crop N level (Cruz &
Obien, 1997). Similar to other conventional ground-based
methods, the use of LCCs require extensive on-site labor and
involves subjective measurements which are highly prone to
inaccuracies. Several studies suggest the integration of these
conventional methods with Unmanned Aerial Vehicle
(UAV) capabilities. Multispectral imagery captured by
UAVs has been proven to be capable of providing
information on crop conditions necessary for nutrient
management. This study proposes a method that incorporates
drone technology with the conventional LCC method to
develop an efficient N monitoring scheme for optimal
fertilizer application on rice plantations.
1.2 Research Objectives
This research aims to develop a method of crop N monitoring
using remote sensing technology particularly UAVs.
Primarily, the goal is the integration of a conventional
ground-based farming practice, the LCC method, and drone
technology. This study intends to provide a more accurate
and modernized version of the LCC method through the
incorporation of UAV imagery analysis. Through the
correlation of LCC observations and parameters extracted
from aerial imagery related to N content, a predictive model
that estimates N levels of rice plantations will be generated.
Specific objectives of the study include the assessment of N-
sensitive Vegetation Indices (VIs) and the determination of
the VI which is the most strongly correlated with LCC
values. An equation modelling LCC readings and the best VI
will be generated to predict the N levels throughout the study
area. The output of this study is an N concentration map of a
rice plantation in San Rafael, Bulacan. Such map aims to
serve as guide for crop N management useful for fertilizer
application. Furthermore, the outputs of this study may serve
as supplementary information to assist agricultural sectors in
implementing policies that will reduce fertilizer expenses
and the adverse effects of N use in farming.
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLII-4/W19, 2019
PhilGEOS x GeoAdvances 2019, 14–15 November 2019, Manila, Philippines
This contribution has been peer-reviewed.
https://doi.org/10.5194/isprs-archives-XLII-4-W19-31-2019 | © Authors 2019. CC BY 4.0 License.
31
1.3 Significance of the Study
Conventional ground-based farming methods for N
monitoring are often laborious and time consuming.
Moreover, ground operations, if not costly, lack accuracy in
measuring crop N levels. Farmers also deserve the
convenience and benefits of technological advancements,
thus, through the incorporation of UAV remote sensing, N
monitoring may be made more convenient and efficient. This
study assessed the capability of UAV imagery for N
management. This integration of agriculture and remote
sensing will allow farmers to manage crop’s N level for
optimal fertilization. Optimal fertilizing doubles agricultural
production subsequently raising farmer’s income, food
security and economic agricultural products. This study
envisions a system of N monitoring which will significantly
raise agricultural output. This study proposes a method of
using UAVs that could be a potential long term investment
for agricultural municipalities to centralize N monitoring
covering plantations within the municipality’s boundaries.
1.4 Related Literature
Crop N monitoring is an essential requisite for optimizing
fertilizer management. Through an efficient assessment of
the current N status, required fertilizer inputs can be
quantified to meet the sufficient nutrient requirement of
crops without compensating the environment (Rochester, et
al., 2009). Conventional methods include the actual on-site
field observations using leaf sampling for laboratory analysis
and chlorophyll level inspection through visual inspection of
leaf greenness index. However, these procedures exhibit
restrictions including time, productivity and cost. In
acknowledgment of this, recent technological advancements
showed an increase in the use of UAV capabilities in
precision agriculture. As compared to satellites and other
spaceborne platforms which requires scheduled overpass, a
low-cost UAV allows time flexible data acquisition at a
higher resolution.
Agricultural sectors have been widely utilizing UAVs for
convenience and efficiency in drought stress detection, yield
prediction and nutrient status assessment (Maes & Steppe,
2018). In particular, several studies has been suggesting the
potential of drone and multispectral sensor integration in
estimating crop N levels. Pölönen (2013) demonstrated
image-based crop biomass and N content estimations
through imagery acquired from a sensor mounted on a light-
weight UAV. VIs extracted from the imagery were
correlated with biomass and N content (Pölönen, 2013). Liu
(2018) conducted a study on a winter oilseed rape farm and
used optimal VIs from in situ hyperspectral data along with
multispectral UAV imagery to estimate the crops’ N
nutrition index (NNI). Results of the study showed that
among the tested VIs, the Normalized Difference VI
(NDVI), Modified Soil-Adjusted VI 2 (MSAVI2) and Red
Edge Chlorophyll Index (CI Red Edge) showed the most
accurate NNI estimations (Liu, et al., 2018). Pagola (2009)
incorporated digital color image analysis on the estimation
of crop N status. Correlation between VIs and Soil and Plant
Analysis Development (SPAD) measurements were
calculated proving the potential of VIs to predict N
deficiencies (Pagola, et al., 2009).
Walsh, et al. (2018) assessed the N concentration level of a
rice crop field through determining the most appropriate
vegetation index (VI) extracted from UAV imagery. In the
same study, the study area was grouped into test and training
data sets. Statistical analysis was performed to determine the
R2 and RMSE to evaluate each model equation. The study
proved that remotely sensed VIs can provide valuable
information on Nitrogen status (Walsh, et al., 2018). The
methodology and parameters used for this research from the
data acquisition up to the final map generation were all based
from related studies regarding N monitoring using UAV
imagery analysis.
2. METHODOLOGY
2.1 Field Visits and Planning
Consultation with research advisers and experts in the field
of agriculture guided the methodological framework.
Different factors that may affect the study were considered
such as the study area and crop to be assessed, instruments
and software availability, and the available budget to be used
throughout the study. Through the Agricultural Municipal
Office of San Rafael, Bulacan request for site access on a rice
plantation was granted. Site inspection of the rice fields
situated in Brgy. Caingin was done under the supervision of
the municipal’s agricultural technologist.
2.2 Study Area
The study area is a specific agricultural site situated in
Barangay Caingin, San Rafael, Bulacan. The chosen region
is specifically a rice crop plantation encompassing an
approximate area of four (4) hectares. The area comprises of
three (3) farmlands with varying fertilizer management
practices particularly in terms of types and quantities of N-
based fertilizers. The rice by the time of data acquisition were
medium early maturing variety crops. The crops are at their
active tillering stage which means that they have been
cultivated for 24-28 days.
Figure 1. Delineation of Study Area in San Rafael, Bulacan
Figure 2. Rice Plantation Documentation
2.3 Flight Planning
Before the actual UAV employment, necessary flight
parameters were set through flight planning and design. The
flight path of the UAV covering the desired area was
established by mission planning using the eMotion software.
Wind condition, climate, obstructions, and other regions of
high risk were taken into account. The final flight altitude
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLII-4/W19, 2019
PhilGEOS x GeoAdvances 2019, 14–15 November 2019, Manila, Philippines
This contribution has been peer-reviewed.
https://doi.org/10.5194/isprs-archives-XLII-4-W19-31-2019 | © Authors 2019. CC BY 4.0 License.
32
used in the operation on the region of interest was 74.3 m
generating an output ground spatial distance resolution of 7.0
cm. The total flight area covered approximately eleven (11)
hectares. Three (3) ground control points were established
situated outside the study area boundary.
Figure 3. Flight path represented in red lines
.
2.4 Instruments
2.4.1 Parrot Sequoia
Parrot Sequoia has been proving the worth of its existence in
the field of agriculture. This commercial sensor allows
absolute reflectance measurements even without the use of
reflectance targets. In this study, this multispectral sensor
was used to monitor crop N status for optimization of
fertilizer strategies to improve overall farm productivity. It is
a 28 x 41 x 59 mm unit weighing 107 grams. It captures four
different bands (Red, Green, Red Edge and Near Infrared)
and has an additional 16 megapixel RGB camera. Parrot
Sequoia comes with a sunshine sensor that records light
conditions in the same spectral bands as the multispectral
sensor. It records the irradiance for each image in each band
and allows the normalization of the varying illumination
during the data acquisition. Moreover, the sunshine sensor
enables GPS data recording. Together, the two sensors obtain
the true value of light and location at any given period. This
unit was acquired through Envisage under the Department of
Geodetic Engineering, University of the Philippines
Diliman.
2.4.2 Sensefly eBee
For the image acquisition, the UAV used in the study was
Sensefly eBee, a fixed wing drone, mounted with a Sequoia
multispectral sensor. This is a fully autonomous mapping
drone suited for various agricultural purposes. This unit
weighing 700 grams can cover a maximum area of 1200 ha
in a single automated mapping flight. The system includes
batteries, an RGB camera, radio modem and an eMotion
software. This drone enables the acquisition of high
resolution imagery which can be further transformed into
orthomosaic and 3d maps.
2.5 Data Acquisition
2.5.1 Ground Control Point (GCP) Establishment
Three (3) GNSS receivers were stationed on the field using
a range pole and bipod setup. The tip of the metal shoe of the
range pole was positioned at the center of the GCP marker.
The bipod’s legs were adjusted until the setup was stable.
The receiver was then mounted on the pole. The receiver and
controller were connected via Bluetooth prior to the actual
fieldwork. Local coordinates were used in the study. Using
the Trimble controller, parameters and configurations were
changed to make use of Post Process Kinematic (PPK) as
survey style. Locations where recorded using the GPS
controller simultaneously with the LCC observations. The
exported file from the receiver were processed using the
Trimble Business Center (TBC).
Figure 4. Setting up the GNSS Receiver (left) and actual
target on GCP (right)
.
2.5.2 UAV Multispectral Imagery Acquisition
The project site was inspected to find an optimal location for
the drone launching. An open space with less obstruction and
low risk was chosen. Relative humidity and wind situation
were taken into account. Upon finish of the GPS
observations on each GCP, multispectral image acquisition
using the fixed wing drone was performed. Imagery of the
rice crop plantation was obtained through the equipped
multispectral sensor. The drone was employed by a licensed
pilot in autonomous mode. A total of three (3) flights with
high overlap was executed by the drone to cover the study
area.
2.6 In Situ N Level Data Collection
In obtaining ground values for N content, the conventional
method of determining the N status of crops using an LCC
was used. A total of twenty-two (22) sampling portions
randomly placed on the field were established. Each
sampling portion encompasses an area of 1 square meter. The
coordinates of each corner of the plot was obtained using
GNSS PPK and was recorded. For every plot, ten (10)
healthy plants were selected. The topmost, fully-expanded
and healthy leaf of each of the ten (10) plants was compared
to the leaf color chart to assess the level of N content. When
observing, the LCC was made sure to not be exposed to
direct sunlight. Only one researcher examined and took LCC
readings from the first to the last plot. The LCC used was
acquired from the municipal agriculture office of San Rafael
Bulacan.
Figure 5. Stakes for marking sampling quadrants (left) and
LCC measurement (right)
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLII-4/W19, 2019
PhilGEOS x GeoAdvances 2019, 14–15 November 2019, Manila, Philippines
This contribution has been peer-reviewed.
https://doi.org/10.5194/isprs-archives-XLII-4-W19-31-2019 | © Authors 2019. CC BY 4.0 License.
33
As part of the validation process, leaf samples from two (2)
quadrants on the field were obtained. A total of two (2)
kilograms of leaf samples was gathered and submitted to the
Plant Tissue Analysis Laboratory of the Philippine Coconut
Authority. The samples were limited to two quadrants to
comply with the allowed kilogram of plants to be obtained in
the farm. Leaf sample collection is destructive for the crops
thus it was greatly reduced.
2.7 Data Processing
2.7.1 GNSS Data Processing
The data acquired from the Static GNSS survey was
processed in TBC. Necessary parameters including antenna
height was fixed to 2.00 meters. Upon processing and
adjustment of the network, the processed data containing the
coordinates of the GCPs was exported. Same method was
done to extract the coordinates of the sampling quadrant
corners from the Kinematic dataset.
2.7.2 Image Post-processing
Multispectral image sets from the first and third flights were
processed in Pix4Dmapper Pro. All 2780 images were
geolocated with output coordinate system auto detected to be
WGS84/ UTM zone 51N. The processing option used was
Ag Multispectral, a template that uses sensors in a camera to
generate radiometrically calibrated reflectance index,
classification, and application maps for precision agriculture.
The compatible cameras for AG Multispectral include Parrot
Sequoia, Micasense Red Edge, and Airinov multiSPEC. For
this study, Parrot Sequoia was used as the multispectral
sensor. The final output are the reflectance tiles for each band
(green, red, red edge and NIR). These reflectance tiles were
used in the generation of the vegetation index maps.
2.7.3 GCP Application for Precise Image Positioning
After initial processing, GCPs were marked in the rayCloud
of images. The adjusted coordinates exported from the Static
Data Processing were used as GCPs for post-processing the
multispectral imagery. Exact locations of GCPs on the image
were marked. At least two (2) marked images were done to
enable automatic marking of the GCPs. The software
automatically selected the location of the GCP on every
image through color correlation of the marked pixel on the
rest of the images. Upon marking all GCPs on the image set,
the processing was reoptimized.
.
Figure 6. GCP marking in rayCloud view Pix4D
.
2.7.4 Generation of Vegetation Index Maps
Post-processed multispectral images were used to generate
the VI maps processed using ArcMap 10.3. Image tiles
having the same band were mosaicked to create a full band
image. Using combinations of these mosaicked images (Red,
Red Edge, Green, and NIR), VI maps were generated using
a raster calculator. This tool builds and executes an
expression using a Python syntax in a calculator-like
interface. Mathematical formulas for each VI sensitive to N
content (CI Green, CI Red Edge, MTCI, NDVI, NDVI Red Edge, and
RTVI Core) were based from Walsh, et.al (2018).
Table 1. Vegetation Indices sensitive to crop Nitrogen
Feature polygons representing the sampling quadrants in the
field were established and was overlaid in each VI map.
These features were created using the Northings and Eastings
of the corners obtained by GNSS. The mean VI for each of
the twenty-two (22) quadrants were extracted using Zonal
Statistics.
2.8 Statistical Analysis
2.8.1 Pearson Correlation
Pearson correlation was used to determine the best VI which
has the strongest correlation with LCC value. Generally, the
strength of correlation between remote sensing data and in
situ ground data was determined by calculating the
correlation coefficient (r). The VI and LCC values were
treated as the independent (x) and dependent (y) variables
respectively. The mathematical formula for computing r is
shown in equation 1. The variable n represents the total
sample population. The VI with the highest r value was
chosen to generate the final model equation.
2.8.2 Regression Analysis
Upon determining the VI with the highest correlation with
LCC values, regression analysis was performed. Regression
was done to find the line of best fit for the two sets of data,
VI and LCC datasets. This is the most commonly used
modeling method. Mathematical trend lines including linear,
exponential, logarithmic, polynomial and power equations
were fitted with the data. The trend line and its respective
equation, which exhibited the best fit and obtained the
highest coefficient of determination (r2) was used as the final
model. This model shall predict LCC values representing
Nitrogen concentration levels using the best vegetation
index.
(1)
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLII-4/W19, 2019
PhilGEOS x GeoAdvances 2019, 14–15 November 2019, Manila, Philippines
This contribution has been peer-reviewed.
https://doi.org/10.5194/isprs-archives-XLII-4-W19-31-2019 | © Authors 2019. CC BY 4.0 License.
34
2.8.3 Leave One-Out Cross-Validation
Leave-one-out cross-validation (LOOCV) was used to assess
the performance of the model equation. This model
validation technique measures how accurate the predictive
model is. In LOOCV, the dataset was divided into training
and test sets. Each iteration excluded one observation to be
the test set, all the remaining belonged to the training set and
was used to generate a model equation. The model tested the
excluded observation. The root mean squared error (RMSE)
for each iteration was computed and averaged to determine
the coefficient of variation (CV). The lower the value of the
CV, the more precise the estimate is. The formula for RMSE
is shown in Equation 2. The variable n is the total sample
population.
2.9 Generation of Nitrogen Concentration Map
The resulting model equation from the best relationship fit
between the LCC observations and UAV-based VI was used
to generate the N concentration map. After processing, mean
values of VIs for each quadrant were recalculated using for
quantitative validation. The resulting map showed the LCC
levels throughout the study area which represents the N
concentration level.
3. RESULTS AND DISCUSSION
.
3.1 LCC Observations and Laboratory Results
A total of twenty-two (22) sampling plots were observed
from the three farmlands comprising the study area (Farms
A, B, and C). Based from the Philippine Rice Research
Institute (PhilRice) guidelines regarding the use of LCC, if
most of the LCC readings are 4 and 5, then the plot where
the LCC observations were obtained have sufficient
Nitrogen level while those plots with mostly 2 and 3 readings
have Nitrogen deficiency. A 30kg N/ha of fertilizer must be
applied during dry season and 23 kg N/ha during the wet
season.
Since the study area is not an experimental field, obtaining 1
kg of leaf samples on each sampling quadrant is impossible
because it may affect the health of the rice crops. Hence, the
2 sets of leaf samples were gathered randomly from Farms A
and B only and submitted to the laboratory. According to
Munson (1998) the nitrogen concentration in plant tissue is
deficient (if less than 2.50%), sufficient (if 2.50% to 4.50%)
and excessive (if greater than 6.00%) (Munson, 1998).
Table 2. Summary of Nitrogen Level in Farms A and B
Table 1 shows the summary of N level in Farms A and B.
Following the PhilRice guidelines on using the LCC, the
sampling quadrants in Farm A have sufficient N level while
Farm B has N deficiency. From the laboratory analysis using
UV-Vis Spectrophotometry, Farm A with 2.761% N total has
sufficient nitrogen level while Farm B with 1.921% N total
is deficient in N level. The results from the two different
methods of N assessment are comparable. The average
values of the LCC observations in both farms are relative to
the laboratory result.
3.2 GNSS Data Processing
This study doesn’t involve global coverage and is not
concerned with the position relative to a datum, thus Local
Positioning System (LPS) was used during the fieldwork.
Problems encountered during the GPS data processing
include error in the network adjustment due to cycle slips.
The discontinuity in the receiver’s phase lock was fixed by
editing the data manually. Table 2 shows the adjusted grid
coordinates used for the GCP marking in the image post-
processing.
Table 3. Adjusted GCP grid coordinates
3.3 Sampling Quadrants
The coordinates from the exported PPK file served as the
corners of the sample quadrants where the LCC observations
were obtained. The northings and eastings data were used to
create polygons in ArcMap 10.3 representing the sampling
quadrants in the map. Figure 7 shows the distribution of
sampling quadrants in Farms A and B plotted in the
mosaicked Red Edge band imagery. Farm A has 11 sample
points while Farm B has 9 sample points. The LCC readings
were mostly obtained near the embankment between the rice
fields since it is the most accessible area for observation.
.
Figure 7. Sampling quadrants in Farm A and B
3.4 Image Post-processing Quality Report
The Ag Multispectral template in Pix4D Mapper was used to
process the flights to generate the reflectance tiles per band.
Upon completion of the post-processing method, a quality
report (shown in Table 4) assessing the output images was
generated.
Images
Median of 6362 keypoints per image
Dataset
2696 out of 2780 images calibrated
Camera
Optimization
0.02% relative difference between
initial and optimized internal camera
parameters
Matching
Median of 1363.72 matches per
calibrated image
Georeferencing
No 3D GCPs
Table 4. Image quality specifications
All 2804 images from flights 1 and 3 are geolocated. Since
the number of keypoints is less than 10000, this suggests that
less visual content could be extracted from the images.
Farm
LCC Observation
Laboratory Result
Farm A
3.91
2.761
Farm B
2.86
1.921
GCP
Easting (m)
Easting
Error
Northing (m)
Northing
Error
1
277636.91
0.001
1655670.25
0.000
2
277613.53
0.001
1655782.74
0.001
3
277804.48
-------
1655610.28
-------
(2)
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLII-4/W19, 2019
PhilGEOS x GeoAdvances 2019, 14–15 November 2019, Manila, Philippines
This contribution has been peer-reviewed.
https://doi.org/10.5194/isprs-archives-XLII-4-W19-31-2019 | © Authors 2019. CC BY 4.0 License.
35
Camera parameters such as shutter speed and exposure time
should be adjusted. This variable however has a small impact
on the overall analysis since the study does not dwell on
visual inspection but on the actual pixel information to be
extracted. 96% of all images are calibrated in a single block.
The difference between the initial and optimized internal
camera parameters is 0.02%. This falls under the 5%
standard for relative difference of the optimized value. A
median of more than 1000 matches in this case 1363.72
denotes that the results in the calibrated regions are likely to
be of best quality
The reflectance tiles (red, green, red edge, NIR) generated
after processing demonstrated good quality specifications
denoting reliability for further analysis.
3.5 Vegetation Index Maps
The Parrot Sequoia multispectral sensor used in the study
captured four (4) discrete spectral bands: Red, Green, NIR,
and Red Edge. Figure 8 shows the mosaicked images of the
four bands. These mosaicked images were used to generate
the vegetation index maps (shown in figure 9).
Figure 8. Mosaicked tile images per band
.
Subsequently, the mosaicked images were used to generate
the vegetation index maps. The N-sensitive VI maps shown
in figure 9 were created as mathematical combinations of the
different mosaicked image bands. By visual inspection, the
output VI maps vary from each other. No VI performed the
same since each is a different combination of various
multispectral bands.
Figure 9. Vegetation Index Maps
3.6 Zonal Statistics
Upon generation of all necessary VI maps, the mean VI (CI
Green, CI Red Edge, MTCI, NDVI, NDVI Red Edge, and RTVI Core)
for each quadrant was extracted using the ArcMap zonal
statistics tool. The range of mean values for each vegetation
index is shown in table 5.
.
Table 5. Mean Vegetation Indices
Each VI exhibits a certain range of values. The NDVI values
ranged from 0.6 to 0.9 which are values close to 1. This
suggests that the regions where the VIs were extracted are
areas of vegetation which holds true since the quadrants were
established within the plantation. The NDVI Red Edge provides
better measurements for latter stage crops since it can
measure further down into the plant canopy. The values
range from 0.27 to 0.49 which falls within the standard range
for green vegetation for NDVI Red Edge. The Chlorophyll
Index (CI) is used to calculate the chlorophyll content in
leaves. The values of these indices are sensitive to even the
slightest variations in chlorophyll content. The mean VIs for
CI Green falls within 2.25 to 7.16 while CI Red Edge ranged from
0.77 to 1.94. The MTCI is a VI highly sensitive to
chlorophyll concentration since reflectance values from NIR,
Red and Red Edge bands were used in the calculation. The
mean index values for each quadrant ranged from 0.61 to
1.69. The RTVI Core is another index that incorporated the
reflectance values of the red edge band. The VI values for
RTVI Core are relatively smaller than the other VIs.
3.7 Statistical Analysis
3.7.1 Pearson Correlation of LCC and VI datasets.
Pearson correlation was used to determine the VI which has
the strongest correlation with in situ data for N concentration.
The correlation coefficient between the VI and the LCC
datasets were calculated to measure the degree of
relationship between the two variables. Figure 10 shows the
correlation coefficients (r) visualized as scatter plots. Among
the VIs, NDVI Red Edge exhibited the strongest correlation
with LCC observations with an r equal to 0.97.
.
Figure 10. Pearson correlation visualized as scatter plot
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLII-4/W19, 2019
PhilGEOS x GeoAdvances 2019, 14–15 November 2019, Manila, Philippines
This contribution has been peer-reviewed.
https://doi.org/10.5194/isprs-archives-XLII-4-W19-31-2019 | © Authors 2019. CC BY 4.0 License.
36
3.7.2 Regression Analysis
Regression analysis was performed to determine the line of
best fit with the NDVI Red Edge values (independent variable)
and the LCC values (dependent variable). Different trend
lines including linear, exponential, logarithmic, polynomial
and power were fitted to obtain the equation that will model
the LCC and NDVI Red Edge values. The whole datasets for
both dependent and independent variables were used to
generate the final model. The coefficient of determination
(R2) was also determined per best fit line. A high value of R2
indicates that one variable is highly predictable from the
other variable. The polynomial equation obtained the highest
R2 value of 0.9731 therefore, this was used as the final model
equation for estimating LCC levels.
.
Table 6. Regression equations
.
3.7.3 Leave One-Out Cross-Validation
Using LOOCV, the performance of the final model equation
was tested. LOOCV involves the calculation of RMSE per
iteration. This was used to compute for the coefficient of
variation (CV). The CV was computed to be 0.03, in
statistics, a CV < 1 indicates low variance thus the generated
equation is suggestive of a good model fit. Estimates that will
be predicted by the model are precise and reliable.
LCC = -35.312 (NDVI Red Edge) 2 + 41.362 (NDVI Red Edge) – 6.9594 (3)
3.8 Nitrogen Concentration Map
Using the final model equation in (3), the N concentration
map for the surveyed rice plantation in San Rafael, Bulacan
dated February 5, 2019 was generated. Non-vegetation
pixels were masked. The output map shows the
corresponding predicted LCC values throughout the area.
Interpreting this, portions exhibiting LCC 2 and LCC 3
indicates Nitrogen deficiency. According to the PhilRice
guidelines, a 30 kg/ha of Nitrogen must be applied to these
areas. Portions with LCC 4 and LCC 5 values denote
sufficient Nitrogen levels.
Figure 11. Nitrogen concentration map and corresponding
fertilizer quantities according to PhilRice
4. CONCLUSION AND RECOMMENDATIONS
.
4.1 Conclusion
Integration of agriculture and remote sensing by UAV
technology is a potential strategy for boosting agricultural
output. Through efficient Nitrogen monitoring, fertilizer
application may be optimized. This study proved an existing
correlation between ground-based Nitrogen data obtained
through Leaf Color Chart (LCC) observations with
parameters extracted from drone imagery. Specifically,
vegetation indices (VIs) acquired from the drone image
processing were assessed to see the VI which is the most
highly correlated with the LCC dataset. Primarily, the results
of the study showed the possibility of using UAV-based
Nitrogen sensitive VIs to quantify actual crop Nitrogen
content in an evidently more efficient and reliable way as
compared to most existing Nitrogen monitoring methods.
UAV integrated with a multispectral sensor having four
discrete spectral bands (Green, Red, NIR, and Red Edge) is
a promising technology specifically for precision agriculture.
It was proved that UAVs are good aerial platforms for
Nitrogen monitoring and a multispectral sensor is highly
suitable for the determination of crop nutrient status. The
methodology used for the assessment of Nitrogen
concentration of rice crops in Farms A and B was ideal to
reduce the detrimental effects of high levels of Nitrogen on
the environment at the same time increase the economic yield
of farmers. Lastly, the study is an advantage for large-scale
Nitrogen monitoring. The method proposed in this research
could be a long-term investment for agricultural
municipalities to centralize Nitrogen monitoring covering all
plantations within its boundary.
Upon the assessment of the six vegetation indices, NDVI Red
Edge had the highest correlation with the LCC observations.
This denotes NDVI Red Edge is a better indicator of Nitrogen
level. This study proved findings on related literature
suggesting that NDVI Red Edge is the best VI to use when
mapping variability in fertilizer requirements. The Nitrogen
concentration map of a rice farm was generated from the best
fit model equation along with the corresponding
interpretation regarding the amount of needed fertilizer per
portion. The output map showed its potential as reference for
efficient Nitrogen management.
4.2 Recommendations
Since NDVI Red Edge had the highest correlation with the LCC
dataset, it is recommended to use a sensor with red edge
band. In addition, LCC observations must be obtained in the
morning, 8:00 – 10:00 AM, to avoid inaccuracies since the
sunlight's glare affects the reflectance on the leaf samples.
Moreover, LCCs must be measured in several sampling
quadrants for sufficient training and test datasets. To have
better validation, leaf samples must be gathered from the
sampling quadrants where the LCC readings were obtained.
However, due to its destructive effect, experimental fields
must be used as study area. Other major agricultural crops
such as corn and sugarcane may be explored and used as
subject crops. Furthermore, the rice plantation used as study
area only involved crops in their active tillering stage.
Further studies may incorporate the growth stage and
varieties of the subject crop as model variables.
NDVI Red Edge
Trendline
Equation
R2
Linear
y = 14.536x – 1.9352
0.9431
Exponential
y = 0.6628e4.4001x
0.9089
Logarithmic
y = 5.4637ln(x) + 8.9132
0.964
Polynomial
y = -35.312x2 + 41.362x – 6.9594
0.9731
Power
y = 18.011x1.6723
0.9499
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLII-4/W19, 2019
PhilGEOS x GeoAdvances 2019, 14–15 November 2019, Manila, Philippines
This contribution has been peer-reviewed.
https://doi.org/10.5194/isprs-archives-XLII-4-W19-31-2019 | © Authors 2019. CC BY 4.0 License.
37
REFERENCES
Cruz, P. S. (1997). Aquaculture Feed and
FertilizerResources Atlas of the Philippines. Food and
Agriculture Organization of the United Nations.
Cruz, R., & Obien, S. (1997). Leaf Color Chart. Rice
Technology Bulletin.
Liu, S., Li, L., Gao, W., Zhang, Y., Liu, Y., Wang, S., & Lu,
J. (2018). Diagnosis of NItrogen Status in Winter Oilseed
Rape (Brassica napus L.) Using in-situ Hyperspectral Data
and Unmanned Aerial Vehicle (UAV) Multispectral Images.
Computers and Electronics in Agriculture, 185-195.
Maes, W., & Steppe, K. (2018). Perspective for Remote
Sensing with Unmanned Aerial Vehicles in Precision
Agriculture. Trends in Plant Science, 1-11.
Munson, A. (1998). Nitrogen and Phosphorus Release from
Humus and Mineral Soil under Black Spruce Forests in
Cnetral Quebec. Soil Biology and Biochemistry, 1491-1500.
Pagola, M., Ortiz, R., Irigoyen, I., Bustince, H., Barrenechea,
E., Aparicio-Tejo, P., . . . Lasa, B. (2009). New Method to
Assess Barley Nitrogen Nutrition Status based on Image
Colour Analysis Comparison with SPAD-502. Computers
and Electronics in Agriculture, 213-218.
Pölönen, I. (2013). Hyperspectral Imaging based Biomass
and Nitrogen Content Estimations from Light-Weight UAV.
Proc. SPIE.
Rochester, I., Ceeney, S., Maas, S., Gordon, R., Hill, J., &
Hanna, L. (2009). Monitoring Nitrogen Use Efficiency in
Cotton Crops.
Walsh, O. S., Shafian, S., Marshall, J., Jackson, C., Blanscet,
S., Swoboda, K., . . . Walsh, W. (2018). Assessment of UAV-
Based Vegetation Indices for Nitrogen Concentration
estimation in Spring Wheat. Advances in Remote Sensing,
71-90.
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Volume XLII-4/W19, 2019
PhilGEOS x GeoAdvances 2019, 14–15 November 2019, Manila, Philippines
This contribution has been peer-reviewed.
https://doi.org/10.5194/isprs-archives-XLII-4-W19-31-2019 | © Authors 2019. CC BY 4.0 License.
38
