Computers and Electronics in Agriculture 196 (2022) 106775

Contents lists available at ScienceDirect

Computers and Electronics in Agriculture

journal homepage: www.elsevier.com/locate/compag

UAV-based chlorophyll content estimation by evaluating vegetation index

responses under different crop coverages
Lang Qiao a, Weijie Tang a, Dehua Gao a, Ruomei Zhao a, Lulu An b, Minzan Li a, b, Hong Sun a, b, *,
Di Song a
a
Key Laboratory of Modern Precision Agriculture System Integration Research, Ministry of Education, China Agricultural University, Beijing 100083, China
b
Key Laboratory of Agricultural Information Acquisition Technology, Ministry of Agriculture and Rural Affairs, China Agricultural University, Beijing 100083, China

A R T I C L E I N F O A B S T R A C T

Keywords: Efficiently estimating chlorophyll content is important in monitoring the photosynthesis capacity and growth
Unmanned aerial vehicle (UAV) status of maize canopy in precision agriculture management. Vegetation index (VI) easily obtained by proximal
Vegetation index remote sensing has been used as a non-destructive and high-throughput way in crop monitoring, especially in
Coverage difference
chlorophyll estimation. However, the estimated results of the field chlorophyll content by VIs always face
Chlorophyll content
challenges from soil background inhibition and estimation stability under the dynamic changes of vegetation
Maize
biomass. Thus, an unmanned aerial vehicle (UAV)-based chlorophyll content estimation was conducted by
evaluating VI responses under different crop coverages. An analysis was conducted on 36 VIs under different crop
coverage conditions to explore their response differences and robustness for chlorophyll estimation. This work
focused on the three kinds of VIs named simple vegetation index, modified vegetation index, and functional
vegetation index. In 2020, at the experimental station of Dryland Farming Institute of Hebei Academy of Agri­
culture and Forestry Sciences, UAV carrying multispectral sensor was used to collect visible and near-infrared
images of the canopy at the jointing stage of maize under six fertilization levels to obtain VIs. After the UAV
fled, ground calibration and sample collection were performed simultaneously, and chlorophyll content was
measured. For data processing, correlation coefficient method (CCM) and maximal information coefficient (MIC)
were first used to analyze the correlation response characteristics of VIs and chlorophyll content under three
different coverage levels. The results showed that when the level of canopy coverage was increased, the linear
correlation between VIs and chlorophyll content was substantially reduced. The MIC response indicating linear
and non-linear combination relationship was more robust. In addition, the VIs obtained by UAV had a significant
linear correlation with maize canopy chlorophyll under low (0.05–0.35) and medium (0.35–0.48) coverage, but
an obvious non-linear correlation under high (0.48–0.75) coverage. Chlorophyll-sensitive parameters were then
screened based on methods of CCM, MIC, and random frog method (RFM), respectively. Partial least squares
regression (PLS) and random forest (RF) algorithms were used to establish the maize canopy chlorophyll content
detection models. The findings showed that when Green minus red vegetation index (GMR), Red light
normalized value (NRI), Normalized difference red edge (NDRE), Modified simple ratio with red edge
(MSRREG), Enhanced Vegetation Index (EVI), Normalized red green difference vegetation index (NDIg),
Normalized red blue difference vegetation index (NDIb), Soil-adjusted vegetation index (SAVI), Optimized soil-
adjusted vegetation index with red edge (OSAVIREG), Soil-atmospherically resistant vegetation index (SARVI)
were selected based on RFM as the optimal spectral variables, the chlorophyll content detection model con­
structed based on PLS had the least numbers of characteristic variables and the best model accuracy. The training
set R2 and RMSE were 0.753 and 2.089 mg/L, respectively, and the verification set R2 and RMSE were 0.682 and
2.361 mg/L, respectively. Field chlorophyll content and detection error distribution maps were also drawn and
combined with the distribution of fertilization management to provide support for the UAV monitoring of crop
growth in the field and variable fertilization management decisions.

* Corresponding author.
E-mail address: [email protected] (H. Sun).

https://doi.org/10.1016/j.compag.2022.106775
Received 3 November 2021; Received in revised form 28 December 2021; Accepted 31 January 2022
Available online 20 March 2022
0168-1699/© 2022 Elsevier B.V. All rights reserved.
L. Qiao et al. Computers and Electronics in Agriculture 196 (2022) 106775

1. Introduction Table 1
Application of vegetation index in crop growth information monitoring.
Chlorophyll content is one of the important indicators for photo­ Crop Parameter Applied vegetation index Reference
graphing photosynthetic capacity and nutritional status. Traditional
Maize Chlorophyll Normalized difference Chaoyang et al.
crop chlorophyll detection relies on chemical analysis, which requires content vegetation index (NDVI), (2008) Chaoyang
destructive capabilities, takes a long time, and cannot be applied for Chlorophyll index with green et al. (2010)
rapid field monitoring (Liu et al., 2019; Zhao et al., 2016). Canopy (CIgreen), Modified soil adjusted
spectral imaging analysis based on crop light absorption detection is vegetation index (MSAVI),
Modified Simple Ratio (MSR)
combined with drone technology to realize crop near-ground perception et al.
analysis, a non-destructive and efficient method that shows potential in Chlorophyll Modified Soil Adjusted He et al. (2015)
the field of agricultural monitoring (Alessandro et al., 2015; Zhang et al., content Vegetation Index (MSAVI),
2016). Normalized difference
vegetation index (NDVI) et al.
High-resolution spectroscopy research showed that crop has strong
Chlorophyll Ratio vegetation index (RVI), Lang et al. (2020)
absorption characteristics in blue light region (400–450 nm) and red content Normalized difference
light region (650–700 nm) due to the influence of pigment groups such vegetation index (NDVI),
as chlorophyll and carotenoids in crop, and has strong reflection char­ Normalized difference red edge
acteristics in near-infrared region (700–1050 nm) due to the internal (NDRE), Soil-adjusted
vegetation index (SAVI) et al.
structure of leaves (Aasen et al., 2018). With its absorption and reflec­ Leaf area Simple ratio (SR), Green Nguy-Robertson
tion characteristics, the vegetation index (VI) calculated by linear or index chlorophyll index (CIgreen), Red et al. (2014)
non-linear combination of reflectance of two or more bands shows great Edge chlorophyll index (CIred-
potential in monitoring the nutrition, physiology, pest stress, and dis­ edge) et al.
Biomass Enhanced vegetation index Y. Zhang et al.
ease information during crop growth (I et al., 2020; He et al., 2015).
(EVI), Visible atmospherically (2021)
However, the stability of its field application is affected by the differ­ resistant index (VARI)
ences of crop varieties and environment. Thus, it is critical to evaluate VI Nitrogen Normalised redness intensity Lu et al. (2021)
responses and select the ones that could overcome challenges from the (NRI), Atmospherically resistant
soil background inhibition and has estimation stability under dynamic vegetation index (ARVI), Green
red ratio vegetation index
changes of vegetation biomass for chlorophyll estimation. This study
(GRRI)
aims to evaluate the response stability of VI under different environ­ Nitrogen Water resistance nitrogen index Wei et al. (2016)
mental differences, optimize canopy spectral parameters, and improve (WRNI)
the stability of field chlorophyll content detection model based on VI. Wheat Chlorophyll Modified chlorophyll absorption Haboudane et al.
content ratio index/Optimized soil- (2008)
Hundreds of VIs have been formed. According to calculation formula
Adjusted vegetation index
and functional requirements, VIs could be divided into three types: (MCARI/OSAVI) et al.
simple vegetation index, modified vegetation index, and functional Nitrogen Excess green index (EXG), Ratio Schirrmann et al.
vegetation index (Jinru et al., 2017; He et al., 2015). The first type is a of the blue and green channel (2016)
simple VI constructed by the combination of the difference or ratios (BG) et al.
Biomass Visible atmospherically resistant Lu et al. (2019)
between bands (Jinru et al., 2017). Jordan (1969) first proposed the
index (VARI), Excess green index
application of the Difference Vegetation Index (DVI) and the Ratio (EXG) et al.
Vegetation Index (RVI) to crop condition detection. Normalized Differ­ Leaf area Modified triangular vegetation Yao et al. (2017)
ence Vegetation Index (NDVI), a typical parameter for the analysis of index index (MTVI2)
water stress Normalized water index-1(NWI- K. et al. (2014)
ground differences and crop development processes, quantifies the
1), Water index (WI)
reflection characteristics of crops by measuring the difference between Rice Chlorophyll Modified chlorophyll absorption Guan et al. (2009)
the near-infrared and red bands. However, NDVI is sensitive to the ef­ content in reflectance index (MCARI)
fects of soil brightness, canopy coverage, and shadow (Jinru et al., Nitrogen Ratio vegetation index (RVI), Wang et al. (2012)
2017), its value gradually approaches saturation when the biological Soil adjusted vegetation index
(SAVI) et al.
coverage conditions change from medium to high ground (Duan et al.,
Leaf area Visible atmospherically resistant Li et al. (2019)
2019). Therefore, parameters for correcting, suppressing various envi­ index index (VARI), Woebbecke index
ronmental interference, and enhancing the robustness of this index have (WI) et al
emerged in an endless stream (Martina et al., 2018). Some modified Yield Normalized difference Wan et al. (2020)
yellowness index (NDYI),
vegetation indices, such as Enhanced Vegetation Index (EVI) and Excess
Normalized difference
Red Index (EXR) (Jibo et al., 2019; John & Dan, 2016), have been vegetation index (NDVI) et al
proposed to intensify the difference in visible light and near-infrared Biomass Normalized difference texture Hengbiao et al.
spectral reflectance for an improved crop detection (Chaoyang et al., index (NDTI), Optimized soil (2019)
2008). Functional vegetation index, such as Soil-adjusted Vegetation adjusted vegetation index
(OSAVI) et al
Index (SAVI), Atmospherically Resistant Vegetation Index (ARVI), have
Potato Chlorophyll Modified difference vegetation Ning et al. (2020)
been proposed to suppress the influence of soil background and atmo­ content index (MDVI)
sphere on the detection of crop parameters (I. et al., 2020). Leaf area Simple ratio (SR), Green Nguy-Robertson
These proposed VIs are widely used in crop monitoring. Table 1 index chlorophyll index (CIgreen), Red et al. (2014)
briefly lists the literature on VI application in chlorophyll content, ni­ Edge chlorophyll index (CIred-
edge) et al.
trogen, leaf area index, biomass, and yield. For chlorophyll content Biomass Excess green index (EXG), Liu et al. (2020a),
estimation, He Yong et al. (He et al., 2015) compared the effects of Visible atmospherically resistant Liu et al. (2020b)
Modified Soil Adjusted Vegetation Index (MSAVI), NDVI, Chlorophyll Index (VARI), Excess red index
Index with Green (CIgreen), and EVI on the detection of maize chloro­ (EXR) et al.
Rape Leaf area Red edge vegetation index (CIred- Peng et al. (2019)
phyll content. The results showed that MSAVI and NDVI were saturated
index edge), Normalized difference
when the chlorophyll content reached 40 mg/m2, and the model based vegetation index (NDVI) et al
on CIgreen had the highest accuracy (R2 = 0.86). Wu et al. (Chaoyang (continued on next page)
et al., 2008) found that comparing with NDVI and Modified Simple Ratio

2
L. Qiao et al. Computers and Electronics in Agriculture 196 (2022) 106775

Table 1 (continued ) establishing a high-precision crop chlorophyll detection model (Wang

Crop Parameter Applied vegetation index Reference et al., 2020).
The ability of various VIs to respond remains to be resolved in the
Yield Normalized difference red edge Peng et al. (2019)
(NDRE), Enhanced vegetation
application of canopy chlorophyll detection. Therefore, the goal is to
Index (EVI) et al explore a solution that aims to clarify the response differences of VIs to
Canopy Normalized difference Index Karla et al. (2008) soil disturbance and canopy biomass dynamic and select VIs to construct
parameters (NDI), Simple ratio (SR) a robust field maize chlorophyll content detection model. Field vege­
tation coverage describes the area proportion of vegetation per unit area
and can be used to quantify the proportion of soil background in the
Table 2 measured area and indicate the difference of crop canopy biomass (Yue
Reflectance calculation formula. et al., 2021). With maize under different fertilization levels as the
Bands Formula research object, this study analyzed the response rule of different types
of VI and maize canopy chlorophyll content under different coverage
B 475 nm y = 0.0032x − 0.1824
G 560 nm y = 0.0035x − 0.1278 conditions and further discussed the influence of spectral parameter
R 668 nm y = 0.0035x − 0.1320 optimization method on chlorophyll content detection. The proposed
REG 717 nm y = 0.0035x − 0.1287 strategies are as follows. First, according to different crop coverage
NIR 840 nm y = 0.0039x − 0.2197
conditions, the response relationship among the three types of VI
(simple vegetation index, modified vegetation index, and functional
(MSR), the good chlorophyll content estimation ability of comprehen­ vegetation index) and chlorophyll content of maize was deeply
sive index considering various interference factors including shadow analyzed. Chlorophyll-sensitive spectral variables were then optimized
and soil background, such as the Transformed chlorophyll absorption by discussing different variable screening methods. Finally, chlorophyll
reflectance index (TCARI), Modified Transformed Chlorophyll Absorp­ content detection models based on different modeling methods were
tion Vegetation Index (MCARI), and Optimized Soil-adjusted Vegetation constructed to provide support for field maize growth evaluation and
Index (OSAVI). Yu et al. (Yu et al., 2014) monitored the chlorophyll management.
content of barley in multiple growth periods based on MCARI/OSAVI,
TCARI/OSAVI, and other comprehensive VIs and found that differences 2. Materials and methods
in canopy structure affect the model monitoring results. In addition, EVI
and EXR show response limitations in the later stages of crop growth 2.1. Experimental program and data processing flow
monitoring (Jibo et al., 2019; John & Dan, 2016). Despite the differ­
ences in the application effects of various VIs, the inhibition of VI on the Field experiments were conducted in 2020 at the Dry Farming and
soil background and its stability under the dynamics of biomass changes Water-saving Agricultural Experimental Station (latitude
during the growth period remain key factors that determine the accu­ 37.9035011244, east longitude 115.7083898640) in Hengshui, Dryland
racy and robustness of the chlorophyll content estimation model. Farming Institute of Hebei Academy of Agriculture and Forestry Sci­
Screening and optimizing modeling around multiple variables can ences, China. The field experiment had 0.4 ha total area and included six
eliminate invalid or low-sensitivity VIs. Statistical regression and ma­ fertilization levels: A1: (nitrogen 0 kg; phosphorus 0 kg)/ha; A2: (ni­
chine learning methods are also important tools for the establishment of trogen 90 kg; phosphorus 60 kg)/ha; A3: (nitrogen 180 kg; phosphorus
high-precision crop chlorophyll content detection models (Zhang et al., 120 kg)/ha; A4: (nitrogen 360 kg; phosphorus 240 kg)/ha; A5: (nitrogen
2021a; 2021b). In addition to using Pearson correlation coefficient to 540 kg; phosphorus 360 kg)/ha; and A6: (nitrogen 720 kg; phosphorus
evaluate the linear correlation degree between characteristic variables 480 kg)/ha (Lang et al., 2022). Each fertilization level was repeated 12
and target objects, Liu et al. (2020a), Liu et al. (2020b) constructed times for a total of 72 plots, and sampling was equally spaced three times
potato chlorophyll detection model by random frog method (RFM) in each plot to obtain 216 samples. For a reduced effect of different
screening chlorophyll-sensitive spectral characteristic variables and fertilization levels on the inter-districts, a 0.8 m buffer zone was set
obtained better results than the chlorophyll detection model based on between the districts, and no sampling was performed in the buffer zone.
Monte Carlo uninformative variables elimination or competitive adap­ The experiment was conducted at the Jointing stage of the maize.
tive reweighted sampling (CARS). The results showed good high- The experimental data acquisition distribution and data processing flow
dimensional spatial non-linear correlation between crop spectral char­ are shown in Fig. 1. a: Data acquisition, including UAV image acquisi­
acteristics and chlorophyll. Zhang et al. (2021) compared the first de­ tion and ground data acquisition; b: Image preprocessing, including
rivative, successive projections algorithm (SPA), CARS, and other image stitching and correction, band fusion, region of interest extrac­
feature screening methods and found that the crop canopy biomass tion, and canopy coverage calculation; c: Radiometric calibration and
detection model established by combining the feature variables screened type spectrum VI extraction; d: Response correlation analysis and
by SPA and CARS has the best accuracy (R2 = 0.89). Yu Lei et al. (2018) feature variable screening; and e: Model construction. Each stage was
analyzed the correlation between hyperspectral and chlorophyll content subsequently introduced in detail.
index (SPAD value) of soybean leaves and reported that the chlorophyll
estimation model based on iteratively retains informative variables 2.2. Data collection
method has higher accuracy than that only using the screening charac­
teristics of strong correlation coefficient method. The research showed The experiment was conducted at the jointing stage of maize on July
that weak correlation coefficient also contributed to the analysis of 15, 2020. The data acquisition process is as follows: (1) the position
chlorophyll index. Chen et al. (2020) compared three feature screening information of ground control points (GCPs) and the central position
methods, namely, RFM, K-means clustering, and mean impact value information of chlorophyll measurement area were obtained by using
(MIV) and found that the spectral variables screened by the random GPS system (S-GNSS, China); (2) UAV multispectral image acquisition;
forest algorithm have advantages for crop biomass evaluation. These (3) ground chlorophyll data acquisition. GCPs were used to correct the
studies showed that using variable selection algorithm to fully excavate geographical location of UAV multispectral image. The central location
potential linear and non-linear correlation between spectral variables of each sampling point was used to match UAV image and chlorophyll
and crop chlorophyll and other growth parameters can analyze the measurement coordinately.
response difference of VI to crop canopy and serve as the basis for

3
L. Qiao et al. Computers and Electronics in Agriculture 196 (2022) 106775

Fig. 1. Flow chart of test and data processing.

2.2.1. UAV image acquisition 2.3. Image preprocessing and coverage calculation
The DJI M600 Pro drone (10 kg UAV weight, 5 kg maximum load
weight) equipped with a multi-spectral camera (RedEdge-MX, Mica­ Original image data were spliced and corrected by using the Pix4D
Sense, USA) was used to obtain crop canopy remote sensing data. The mapper software and the ground control points (GCPs). The orthophoto
quality of the RedEdge-MX multi-spectral camera is 232 g, and the maps of five bands were generated separately and then fused using ENVI
resolution is 1280*960 pixels (the sensor size is 8.7 cm*5.9 cm*4.54 5.3 software. According to the position information of 216 sample
cm). The camera can collect spectral images in five bands: blue (B), points, the fused multi-spectral image was divided. Each sample point
green (G), red (R), red edge (REG), and near infrared (NIR). The center corresponds to a multi-spectral image of 50 × 50 pixels.
wavelength of each band is 475, 560, 668, 717 and 840 nm with Canopy coverage is the ratio of the vertical projection area of the
bandwidth of 20, 20, 10, 10, and 40 nm, respectively. The flight oper­ vegetation to the ground area. Referring to Lang et al. (2020), VI seg­
ation time was from 11:00 to 13:00, the flying height of the UAV was 30 mentation was used to segment the crop canopy multispectral image.
m, the flying speed was 4 m per second, and the image overlap was set to This method realizes the segmentation of crop and soil background by
80%. using the reflection difference characteristics of crop and soil in different
spectral bands and combining the maximum between-class variance
2.2.2. Ground sampling and chlorophyll measurement method. NDVI was selected as the segmentation index. According to the
Three maize plants were equidistantly selected in each experimental calculation formula of NDVI in Table 3, the pixels belonging to the crop
plot, and the location of sampling points of each plant was calibrated canopy were set to 1, and those belonging to the soil background were
using GPS positioning system. Canopy leaves were placed into sealed set to 0 in all 2500 pixels. Formula 4 was then used to calculate the
bags, refrigerated, and brought back to the laboratory to determine the canopy coverage of the sampling area. For simplified calculation, the
chlorophyll content via stoichiometry and the use of a spectrophotom­ coverage value was normalized within the range of 0 to 1.
eter (Lang et al., 2020). First, a piece of 4 cm × 4 cm (without veins) of
Sc
leaf tissue was cut from the middle of each leaf. The chopped leaves were C= (4)
Sa
then submerged in 40 ml of 95% ethanol solution and then soaked in the
dark for 48 h. After chlorophyll extraction, solution absorbance was where Sc and Sa represented the number of crop pixels identified as crop
measured with a UV1800 spectrophotometer at wavelengths of 649 and pixels and the number of all image pixels, respectively, and C repre­
665 nm. The formulas for calculating chlorophyll content are as follows: sented the crop canopy coverage. According to the statistics of maize
Ca = 13.95A665 − 6.88A649 (1) canopy coverage, the three coverage gradients was classified as low
coverage (C1), medium coverage (C2), and high coverage (C3).
Cb = 24.96A649 − 7.32A665 (2)
2.4. Radiation calibration and type spectrum vegetation index extraction
Ct = Ca + Cb (3)

where A649 and A665 are the absorbance of 649 and 665 nm, respec­ Radiometric calibration converts the gray value of the multi-spectral
tively; Ca is chlorophyll a content (mg/L); Cb is chlorophyll b content image into reflectance information. Three standard reflective cloths
(mg/L); and Ct is the total chlorophyll content (mg/L). with reflectivity of 10%, 30%, and 70% were placed on a flat and un­
shaded area around the test field to ensure that the crop and the standard
reflective cloths simultaneously appeared in the multispectral image.
The standard reflective cloth is 1 m * 1 m in size and made of lambert

4
L. Qiao et al. Computers and Electronics in Agriculture 196 (2022) 106775

Table 3
Spectral vegetation index.
Classify Type Vegetation index Formula Reference

Simple Vegetation Difference Vegetation Index Difference vegetation index (DVI) DVI = NIR − R He et al. (2015)
Index Green difference vegetation index (DVIGRE) DVIGRE = NIR − G *
Red edge difference vegetation index DVIRED = NIR − REG *
(DVIRED)
Green minus red vegetation index (GMR) GMR = G-R He et al. (2015)
Excess green index (EXG) EXG = 2G-R-B I. et al. (2020)
Green Difference Vegetation Index (GDVI) GDVI = NIR-G Sripada (2005)
Ratio Vegetation Index Ratio vegetation index (RVI) RVI = NIR/R He et al. (2015)
Green Ratio Vegetation Index (GRVI) GRVI = NIR/G Jinru et al. (2017)
Chlorophyll index with red edge (CIredege) CIredege=(NIR/REG) − 1 I. et al. (2020)
Chlorophyll index with green (CIgreen) CIgreen=(NIR/G) − 1 Dash et al. (2010); I.
et al., (2020)
Green red ratio vegetation index (GR) GR = G/R Jinru et al. (2017)
Green blue ratio vegetation index (GB) GB = G/B Jinru et al. (2017)
Green Chlorophyll Index (GCI) GCI=(NIR/G)-1 Anatoly et al. (2003)
Red light normalized value (NRI) NRI = R/(R + G + B) Silleos et al. (2006)
Green light normalized value (NGI) NGI = G/(R + G + B) Silleos et al. (2006)
Modified Vegetation Normalized difference vegetation index NDVI=(NIR − R)/(NIR + R) Li et al. (2014)
Index (NDVI)
Normalized difference red edge (NDRE) NDRE=(NIR − REG)/(NIR + REG) Li et al. (2014)
Green normalized difference vegetation GNDVI=(NIR − G)/(NIR + GRE) Gitelson et al. (1996)
index (GNDVI)
Red difference vegetation index (RDVI) RDVI=(NIR − R)/(NIR + R)^0.5 Jean-Louis and François-
Marie (1995)
Red difference vegetation index with red RDVIREG=(NIR − REG)/(NIR + REG)^0.5 *
edge (RDVIREG)
Modified simple ratio (MSR) MSR=(NIR/R − 1)/(NIR/R + 1)^0.5 Jing (2014)
Modified simple ratio with red edge MSRREG=(NIR/REG − 1)/(NIR/REG + 1) *
(MSRREG) ^0.5
Meris terrestrial chlorophyll index (MTCI) MTCI=(NIR − REG)/(NIR − R) Zhang & Liu (2014)
Enhanced Vegetation Index (EVI) EVI = 2.5(NIR-R)/(NIR + 6R-7.5B + 1) Martina et al. (2018)
Normalized red green difference vegetation NDIg=(R-G)/(R + G + 0.01) Wu (2016)
index (NDIg)
Normalized red blue difference vegetation NDIb=(R-B)/(R + B + 0.01) Wu (2016)
index (NDIb)
Functional Adjust Soil Vegetation Index Soil-adjusted vegetation index (SAVI) SAVI = 1.5(NIR − R)/(NIR + R + 0.5) Huete (1988)
Vegetation Index Soil-adjusted vegetation index with green SAVIGRE = 1.5(NIR − G)/(NIR + G + 0.5) *
(SAVIGRE)
Optimized soil-adjusted vegetation index OSAVI = 1.16(NIR − R)(NIR + R + 0.16) Wan et al. (2019)
(OSAVI)
Optimized soil-adjusted vegetation index OSAVIGRE=(1 + 0.16)(NIR − G)(NIR + G *
with green (OSAVIGRE) + 0.16)
Optimized soil-adjusted vegetation index OSAVIREG=(1 + 0.16)(NIR − REG)(NIR *
with red edge (OSAVIREG) + REG + 0.16)
Modified Soil Adjusted Vegetation Index MSAVI = 1.5(NIR-R)/(NIR + R) + 0.5 Wan et al. (2019)
(MSAVI)
Anti-atmospheric Interference Atmospherically resistant vegetation index AVRI=(NIR-RRB)/(NIR + RRB); in which Gitelson et al. (2002)
Vegetation Index (ARVI) RRB = R-γ(R-B)
Soil-atmospherically resistant vegetation SARVI = 1.5(NIR-RRB)/ (NIR + RRB + 0.5); Gitelson et al. (2002)
index (SARVI) in which RRB = R-γ(B-R)
Green Atmospherically Resistant Index GARI=(NIR-G + 1.7(B-R))/(NIR + G-1.7 Anatoly et al. (1996)
(GARI) (B-R))
Visible Atmospherically Resistant Index VARI=(G-R)/(G + R-B) A. et al. (2002)
(VARI)

Note:B, G, R, REG, and NIR represent the reflectance of blue, green, red, red edge, and near-infrared bands, respectively; “*” means red edge and green band in­
formation multi-spectral index based on red-band information.

rubber. The conversion formula, which was used to calculate the 2.5. Response association analysis and feature variable screening
reflectance of the spectral image as shown in Table 2, was calculated by
the reflectance of the standard reflective cloth and the gray value of its Correlation Coefficient method (CCM) and Maximal Information
image. Coefficient method (MIC) were selected to analyze the response rela­
where y is the reflectance of the standard reflective cloth; x is the tionship between VI and chlorophyll to clarify the response law of VI and
gray value of standard reflective cloth image. chlorophyll under different maize canopy coverage. CCM measures the
Given the spectral reflectance difference of different wavebands of degree of linear correlation between characteristic variables and
crop, the linear or non-linear combination of reflectance of B, G, R, REG, response variables and its value rang is [-1,1] (Liu et al., 2020). In this
and NIR wavebands were used to calculate different types of spectral VIs study, − 0.7 and 0.7 were set as the thresholds of CCM. MIC extracts the
as shown in Table 3, including Simple Vegetation Index (Difference mutual information coefficients between different variables by calcu­
Vegetation Index and Ratio Vegetation Index), Modified Vegetation lating the approximate probability distribution density of different fea­
Index, and Functional Vegetation Index (Adjust Soil Vegetation Index tures and its value rang is [0, 1]. This method considers the linear
and Anti-atmospheric Interference Vegetation Index). relationship between variables, measures their non-linear relationship,
and evaluates the linear and non-linear combination relationship

5
L. Qiao et al. Computers and Electronics in Agriculture 196 (2022) 106775

between variables (Shao & Liu, 2021). In this study, 0.5 was set as the absorption with reflectivity was 0.18, whereas the B and R bands
threshold of MIC. exhibited strong absorption with reflectivity of 0.12 and 0.14, respec­
Given the potential high-dimensional non-linear relationship be­ tively. The reflectivity of REG and NIR bands increased to 0.26 and 0.28,
tween VIs and chlorophyll, RFM was introduced to screen characteristic respectively. This change trend was in line with typical vegetation
variables. This method is a bionic algorithm, which calculates the spectroscopy, that is, pigments such as chlorophyll a, chlorophyll b and
probability of selection between characteristic variables and response carotenoids absorb blue and red light and reflect green light. The
variables by simulating a Markov chain that obeys a steady-state dis­ reflectance increased at the REG band, and the NIR region was relatively
tribution and is often used to screen chlorophyll content spectral vari­ highly reflective due to the influence of the canopy and leaf structure
ables and its value rang is [0, 1] (Liu et al., 2020). In this study, 0.45 was (Lang et al., 2020). Further analysis on the statistical distribution of
set as the threshold of RFM. reflectance of different samples in each band showed that the distribu­
tion interval of R band reflectance that characterizes chlorophyll ab­
sorption was [0.06, 0.28], which had higher discreteness compared with
2.6. Model establishment and accuracy evaluation
other bands. In addition, the visible light B, G, and R band data had
significant outliers, indicating their susceptibility to interference from
The 216 samples were divided into a training set and a verification
factors such as soil background, canopy structure, and environment (He
set based on the ranking method at the ratio of 3:1, that is, 162 samples
et al., 2015). This feature affects the confidence of using single-band
in training set and 54 samples in verification set. The chlorophyll con­
reflectance to detect the chlorophyll content of maize canopy in the
tent detection model was constructed based on Partial Least Squares
field. Therefore, the VI from the combination of multi-band reflectance
Regression (PLS) and Random Forests (RF) respectively. PLS is a
ratios and differences can effectively offset potential multiplicative or
multivariate statistical data analysis method that combines the advan­
additive interference factors (Chaoyang et al., 2008), and we would
tages of principal component analysis, canonical correlation analysis,
further analyze it.
and linear regression analysis. This technique has the advantage of
The 216 samples were divided by the ranking method. The results are
dealing with multiple correlations among independent variables in a
shown in Table 4. The chlorophyll content of the total samples ranged
small sample. RF is an ensemble learning method based on multiple
from 20.17 to 37.35 mg/L, the average value was 30.6 mg/L and the
decision trees that combines regression sub-models to model a large
standard deviation value was 4.21 mg/L. The training set contained 162
number of interrelated input variables.
samples, and the validation set contained 54 samples. Both sets con­
The model determination coefficient (R2) and the root mean square
tained most of the chlorophyll contented values in the total sample data.
error (RMSE) of the measured and predicted values were used to eval­
The divided sample sets would be used for subsequent training and
uate the detection ability of the model using formulas 5 and 6. R2 rep­
verification.
resents the degree of fit of the model. When its value is close to 1, the
accuracy of model detection is high. RMSE indicates the dispersion of
experimental results. When its value is small, the model effect is good. 3.2. Correlation analysis of vegetation index and chlorophyll content
response under different coverage
∑n
(yi − ̂y i )2
R2 = 1 − ∑i=1 (5)
n
i=1 (yi − yi )
2
The distribution of maize canopy coverage under 6 different fertil­
ization levels is shown in Fig. 3 (b). In general, the interval range of
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
√∑
√n coverage values were between 0.0648 and 0.7082. With the increase of
√ (yi − ̂
√i=1
y i )2 fertilization levels, the average values of coverage gradually increased,
RMSE = (6) reached the maximum point at A5 level and decreased slightly at A6
n
level. According to the statistics of maize canopy coverage, 216 samples
were divided into three coverage gradients: low coverage (C1:
3. Results 0.05–0.35), medium coverage (C2: 0.35–0.48), and high coverage (C3:
0.48–0.75). The chlorophyll distribution map under different coverages
3.1. Statistics of original spectral reflectance and chlorophyll content is shown in Fig. 3 (b). For C1 containing 59 samples, the interval range of
chlorophyll was 20.17–32.55 mg/L, the chlorophyll average value was
The spectral reflectance characteristics of the maize canopy B, G, R, 25.33 mg/L, and the chlorophyll standard deviation was 3.10 mg/L. For
REG, and NIR were first analyzed. As shown in Fig. 2, the median fit line C2 including 82 samples, the interval range of chlorophyll was
of each band reflectivity indicated that the G band exhibited weak 25.92–37.35 mg/L, the chlorophyll mean value was 32.08 mg/L, and the
chlorophyll standard deviation was 2.87 mg/L. For C3 including 75
samples, the interval range of chlorophyll was 27.18–37.31 mg/L, the
chlorophyll average value was 33.12 mg/L, and the chlorophyll stan­
dard deviation was 2.00 mg/L.
CCM and MIC were used to analyze the responses of VI and chloro­
phyll content under different coverages as shown in Fig. 4. Although the
samples were generally in the maize jointing stage, significant

Table 4
Chlorophyll content statistics and sample set division.
Data set Number Maximum Minimum Average Standard
of value value value deviation
samples

Total 216 37.35 20.17 30.60 4.21

sample
Training 162 37.31 20.17 30.56 4.22
set
Validation 54 37.35 20.69 30.70 4.17
set
Fig. 2. Statistics of multi-spectral bands reflectance.

6
L. Qiao et al. Computers and Electronics in Agriculture 196 (2022) 106775

Fig. 3. Statistics of coverage and chlorophyll content.

differences in the response correlation were found between VI and the non-linear methods, the performance of the chlorophyll detection
chlorophyll under different coverages. Considering the linear relation­ model constructed based on RFM was better than that based on MIC.
ship in Fig. 4(a), with the increase in maize canopy coverage, the linear Further comparison of the modeling results of PLS and RF showed that
response relationship decreased between VI and chlorophyll. In detailly, for screening variables, RF was better than PLS in constructing a chlo­
the linear coefficient between VIs and chlorophyll was relatively high rophyll content detection model, and PLS was better than RF in con­
and stable at the coverage level of C1 and C2. However, when coverage structing a chlorophyll content detection model. Comparison results of
level was C3, the linear coefficient between VIs and chlorophyll showed the combination of different variable screening and modeling methods
a sharp decline, and the correlation coefficient was generally lower than are shown in Fig. 5. The chlorophyll content detection model from
0.2. combined PLS and RFM had the best model accuracy, and the R2 and
The combination of linear and non-linear relationships between the RMSE were 0.753 and 2.089 mg/L for the training set, respectively, and
chlorophyll and VIs is shown in Fig. 4(b). Although the MIC coefficient 0.682 and 2.361 mg/L for the validation set, respectively.
of VIs and chlorophyll also has certain differences under different
coverage, the fluctuation range of the difference was smaller than the 4. Discussion
correlation coefficient shown in Fig. 4(a). Compared with that of C1 and
C2, the MIC coefficient showed a decrease and an increase in C3 and In this study, UAV multi-spectral sensors were used to obtain the
fluctuated from 0.1 to 0.2. Among which, the simple vegetation indexes remote sensing images of maize canopy at the jointing stage. Under
represented by DVI and DVIGRE mainly reflected the decreasing trend, different maize coverage conditions, the response law of VI and maize
and the modified vegetation index and functional vegetation index chlorophyll content was analyzed in depth, and the chlorophyll-
represented by EVI, RDVIREG, and MSAVI mainly exhibited the sensitive characteristic variables were optimized. The detection model
increasing trend. And some VIs, such as RVI, GRVI, GR, MSRGRE, NDIg, of canopy chlorophyll content was constructed.
SAVI, GARVI and VARI, showed good correlation with chlorophyll
under three coverage levels. Compared with CCM, the MIC coefficient of 4.1. Influence of different canopy coverages on VI stability
VIs and chlorophyll changed less and showed better stability under three
different coverage level. The stability of VI in different filed environments is the key factor for
the linear model to diagnose crop growth (Yao et al., 2017). Common
3.3. Feature variable screening VIs, such as NDVI, are highly sensitive to changes in field ground spatial
heterogeneity and prone to saturation when the local biological
On the basis of the above analysis, the high-dimensional non-linear coverage reaches medium-to-high conditions (Jinru et al., 2017; Duan
iterative feature variable screening method RFM was further considered. et al., 2019). Fig. 4(a) shows that with the increase in maize canopy
The three types of VI variables were screened using the linear (CCM) and coverage, the linear response relationship between VI and chlorophyll
non-linear (MIC, RFM) methods, and the results are shown in Table 5. was greatly reduced. When the coverage reached C3, the CCM co­
CCM selected 24 VIs, including 7 simple VIs, 8 modified VIs, and 9 efficients were reduced to less than 0.2.
functional VIs. MIC selected 11 VIs, including 5 simple VIs, 2 modified The above viewpoints mostly considered the linear response rela­
VIs, and 4 functional VIs; RFM selected 10 VIs, including 2 simple VIs, 5 tionship between the VI and the target object. However, the VI was
modified VIs, and 3 functional VIs. calculated by the linear or non-linear combination of reflectance be­
tween bands (Jinru et al., 2017). In addition to the linear relationship, a
3.4. Model comparison based on screening of different characteristic non-linear response relationship might exist between VI and the target
variables object. This study extracted the MIC coefficient between VI and chlo­
rophyll content under different coverages based on MIC, and the results
The chlorophyll content detection models based on PLS and RF were are shown Fig. 4(b). Compared with CCM, the difference of MIC coef­
constructed respectively by using the VIs screened by the three variable ficient between VI and chlorophyll was smaller with increasing
screening methods as the input factor and the chlorophyll content as the coverage. Compared with that in C1 and C2, the MIC coefficient
output factor, and the results are shown in Table 6. Comparison of three decreased and increased at C3 and fluctuated from 0.1 to 0.2. The
different variable screening methods showed that the performance of possible reason is that MIC is a method that considers the linear rela­
the chlorophyll detection model constructed based on the non-linear tionship between variables and measures their non-linear relationship
method (MIC, RFM) was better than that of the linear CCM method. In (Shao & Liu, 2021). The linear and non-linear relationships between VI

7
L. Qiao et al. Computers and Electronics in Agriculture 196 (2022) 106775

Fig. 4. Analysis of the correlation and response difference between vegetation index and chlorophyll under different coverage.

Table 5 Table 6
Screening results of VIs variables based on 3 methods. Test results of chlorophyll content.
Method Simple Vegetation Modified Vegetation Functional Vegetation Modeling Variable Number of Training set Validation set
Index Index Index method screening variables
R2 RMSE R2 RMSE
CCM DVI, DVIGRE, NDRE, GNDVI, SAVIGRE, OSAVI,
PLS CCM 24 0.519 2.911 0.311 4.412
DVIRED, GMR, RDVI, RDVIREG, OSAVIREG, OSAVIGRE,
MIC 11 0.650 2.485 0.570 2.778
RVI, CIredege, MSR, MSRREG, MSAVI, ARVI, SARVI,
RFM 10 0.753 2.089 0.682 2.361
CIgreen MTCI, NDIg GARVI, VARI
RF CCM 24 0.918 1.205 0.594 2.665
MIC DVI, DVIGRE, NDRE, RDVIREG SAVIGRE, OSAVI,
MIC 11 0.916 1.215 0.601 2.647
DVIRED, GMR, OSAVIREG, OSAVIGRE
RFM 10 0.896 1.355 0.614 2.601
GDVI
RFM GMR, NRI NDRE, MSRREG, SAVI, OSAVIREG, SARVI
EVI, NDIg, NDIb
interference factors.

and chlorophyll coexisted; hence, the robustness was good, and the 4.2. Comparison of different variable screening methods
response correlation between the two based on MIC evaluation was less
affected by vegetation canopy coverage, soil background, and other Three characteristic variable screening methods (CCM, MIC, and

8
L. Qiao et al. Computers and Electronics in Agriculture 196 (2022) 106775

Fig. 5. Test results of chlorophyll content.

RFM) and two modeling methods (PLS and RF) were used for model vegetation index, MIC screened five VIs including DVI and GMR, and
construction. Table 6 shows that compared with that based on non- RFM screened GMR and NRI. Among them, the DVI, DVIGRE, DVIREG,
linear methods (MIC and RFM), the chlorophyll detection model con­ and GDVI screened by MIC had higher correlations than 0.88 and serious
structed based on linear CCM had lower accuracy. Among which, the multicollinearity as shown in Fig. 6(a). Although RFM had not screened
model verification set R2 constructed by combined CCM and PLS was the above four VIs, the NRI screened by RFM had correlations higher
only 0.311. The main reason is that CCM measures the linear correlation than 0.73 for the above four VIs. This value could reduce the multi­
between VI and chlorophyll (Liu et al., 2020). As shown in Fig. 6, the 24 collinearity between variables during characterization. For the modified
VIs screened by CCM had a serious multicollinear relationship. Taking vegetation index, the correlation between the RDVIREG screened by
the modified vegetation index as an example, among the VIs screened by MIC and the MSRREG screened by RFM reached 0.99 (Fig. 6[b]), indi­
CCM, the correlation coefficients between NDRE, GNDVI, RDVI, cating that the two VIs could be characterized by each other. In addition,
RDVIREG, MSR, MSRREG, and MTCI are within the range [0.83,0.99] NDRE was simultaneously screened by both methods, and RFM addi­
(Fig. 6[b]). The multicollinearity between variables caused variable tionally screened EVI, NDIg, and NDIb. Among them, EVI was improved
uncertainty in the model and thereby reduced the accuracy of the model on the basis of NDVI, that is, the introduced blue band could correct the
(Zhang et al., 2021a; 2021b). influence of soil background and aerosol scattering and enhance the
Table 6 shows that among the non-linear methods, the chlorophyll saturation of medium and high vegetation areas (Ma et al., 2018). NDIb
detection model constructed based on RFM showed better performance had a correlation coefficient of 0.7 with chlorophyll content and a weak
than that based on MIC. For the three types of VI optimization, MIC and correlation with the other VIs. For the functional vegetation index,
RFM screened 11 and 10 VIs, respectively (Table 5). For the simple compared with MIC method, RFM not only selected the type of adjust

Fig. 6. Correlation matrix of different types of vegetation index.

9
L. Qiao et al. Computers and Electronics in Agriculture 196 (2022) 106775

soil vegetation index, but also selected the type of anti-atmospheric with the fertilization level, canopy coverage first increased and then
interference vegetation index (SARVI). Studies had shown that the decreased and reached the maximum value at A5 level. On the basis of
scattering effect of molecules and aerosols in the atmosphere could the real and predicted chlorophyll contents, the field chlorophyll con­
reduce the value of VI; hence, the VI of different types might overlap, tent error distribution map was drawn and was shown in Fig. 7 (c).
and the anti-atmospheric interference vegetation index could effectively Compared with that in the fertilization distribution map, the chlorophyll
diminish this influence (Pôças et al., 2020). The above analysis showed content detection model error was large in the fertilization areas A1, A2,
that compared with MIC, RFM further reduced the multicollinearity and A6. This finding indicated that the error of the chlorophyll content
problem among VI. The optimized VIs included the chlorophyll-sensitive detection model based on UAV multispectral imaging increased when
bands R, REG, and NIR and introduced B band that could suppress the fertilization level was extremely low or high. A possible reason is
environmental interference (Ma et al., 2018). Therefore, the chlorophyll that the growth period of maize is slowed down and the growth trend is
content detection model based on RFM had better accuracy than that poor in low-nitrogen areas. In Fig. 7 (b), it shown that the canopy
based on MIC. coverage was low. Excessive fertilization inhibits the growth of maize,
In the chlorophyll detection modeling, when CCM or MIC was used thus leading to large error in chlorophyll levels in the test areas. In Fig. 7
to screen characteristic variables, the result of RF was better than that of (b), it shown that the canopy coverage at the fertilization area A6 was
PLS. The possible reason is that RF is an integrated learning model lower than that at fertilization area A5. This phenomenon also reflects
(Zhang et al., 2021a; 2021b), which has better generalization ability the potential of the maize canopy chlorophyll content detection for field
compared with PLS. However, considering features screened by RFM, fertilization management decisions in the future.
the modeling result of PLS was better than RF. The possible reason is that Some parameters such as coverage, canopy height and LAI can effi­
RF and RFM are both processed based on instances individually, taking ciently characterize the dynamic changes of crop canopy. Previous study
the one with the majority of votes as the selected prediction, the over- showed that canopy coverage was closely related to LAI and canopy
fitting phenomenon may occur due to overly complex model struc­ height at maize jointing stage (Lang et al., 2022). Compared with LAI
tures, resulting in reduced model accuracy. and canopy height, coverage was easier to calculate in UAV image.
Therefore, this study used the dynamic changes of coverage to charac­
4.3. Application potential and prospects terize the differences of crop canopy. In the future, the comprehensive
growth index which fuse coverage, canopy height and LAI will be con­
The chlorophyll content predicted by the optimal detection model structed to characterize the differences of crop canopy. And this work
constructed based on the combination of PLS and RFM was processed, did not explain how to reduce the model error and improve the
and a field-scale spatial distribution map of the chlorophyll content of robustness of the model when the fertilization level is extremely high or
the maize canopy was obtained to analyze the field application potential low. In the future, this factor will be emphasized, and a general model
of the constructed chlorophyll content detection model. As shown in will be constructed by combining the experimental data of many years
Fig. 7 (a), the chlorophyll content value predicted by the optimal chlo­ and multiple growth periods. In addition, this study only compared two
rophyll content detection model had similar spatial distribution to the regression modeling methods; deep learning has shown good applica­
real chlorophyll content value. Fig. 7 (b) shown that the field distribu­ tion prospects in the classification and monitoring of crop diseases and
tion of maize canopy coverage. As indicated by the field fertilization insect pests (Yang & Xuewei, 2020). In the future, the potential of deep
distribution in Fig. 1(a), the chlorophyll content gradually increased learning methods in crop growth detection based on UAV platform will

Fig. 7. Spatiotemporal distribution map, coverage distribution map and error distribution map of chlorophyll content: a: spatiotemporal distribution map of
chlorophyll content b: coverage distribution map and c: error distribution map of chlorophyll content.

10
L. Qiao et al. Computers and Electronics in Agriculture 196 (2022) 106775

be further considered. Anatoly, A.G., Yuri, G., Mark, N.M., 2003. Relationships between leaf chlorophyll
content and spectral reflectance and algorithms for non-destructive chlorophyll
assessment in higher plant leaves. J. Plant Physiol. 160 (3).
5. Conclusions Chaoyang, W., Li, W., Zheng, N., Shuai, G., Mingquan, W., 2010. Nondestructive
estimation of canopy chlorophyll content using Hyperion and Landsat/TM images.
Int. J. Remote Sens. 31 (8).
The chlorophyll content detection based on UAV was performed by Chaoyang, W., Zheng, N., Quan, T., Wenjiang, H., 2008. Estimating chlorophyll content
analyzing the canopy difference response of VI, and the response law of from hyperspectral vegetation indices: Modeling and validation. Agric. For.
VI and chlorophyll content under different coverage was studied. With Meteorol. 148 (8).
Dash, J., Jeganathan, C., Atkinson, P.M., 2010. The use of MERIS Terrestrial Chlorophyll
the use of three characteristic variable screening methods and two Index to study spatio-temporal variation in vegetation phenology over India. Remote
modeling methods, the detection model of maize canopy chlorophyll Sens. Environ. 114 (7), 1388–1402.
content in the field was constructed, and the visualization distribution Duan, B.o., Liu, Y., Gong, Y., Peng, Y.i., Wu, X., Zhu, R., Fang, S., 2019. Remote
estimation of rice LAI based on Fourier spectrum texture from UAV image. Plant
map of maize chlorophyll and error in the field was finally realized. Methods 15 (1). https://doi.org/10.1186/s13007-019-0507-8.
Results showed that the response relationship between VI and chloro­ Gitelson, A.A., Kaufman, Y.J., Merzlyak, M.N., 1996. Use of a green channel in remote
phyll content was affected by maize canopy coverage. When the canopy sensing of global vegetation from EOS-MODIS. Remote Sens. Environ. 58 (3),
289–298.
coverage was increased, the linear correlation CCM between VI and Gitelson, A.A., Stark, R., Grits, U., Rundquist, D., Kaufman, Y., Derry, D., 2002.
chlorophyll content decreased significantly. The MIC coefficient that Vegetation and soil lines in visible spectral space: a concept and technique for
characterizes the linear and non-linear combination relationship was remote estimation of vegetation fraction. Int. J. Remote Sens. 23 (13), 2537–2562.
Guan, L., Liu, X., Cheng, C., 2009. Research on hyperspectral information parameters of
less affected by vegetation coverage and had high response robustness.
chlorophyll content of rice leaf in Cd-polluted soil environment. Spectroscopy
The 36 VIs of different types were optimized based on the three variable Spectral Analysis 29 (10), 2713–2716.
screening methods of CCM, MIC, and RFM. Among them, RFM screened Haboudane, D., Tremblay, N., Miller, J.R., Vigneault, P., 2008. Remote estimation of
crop chlorophyll content using spectral indices derived from hyperspectral data.
10 VIs with the smallest number of variables and low multicollinearity
IEEE Trans. Geosci. Remote Sens. 46 (2), 423–437. https://doi.org/10.1109/
between variables. The chlorophyll content detection model based on TGRS.2007.904836.
combined PLS and RFM had the highest accuracy. The R2 and RMSE He, Y., Peng, J., Liu, F., Zhang, C.u., Kong, W., 2015. Critical review of fast detection of
were 0.753 and 2.089 mg/L for the model training set, respectively, and crop nutrient and physiological information with spectral and imaging technology.
Trans. Chinese Soc. Agric. Eng. (Transactions of the CSAE) 31 (3), 174–189.
0.682 and 2.361 mg/L for the verification set, respectively. These Hengbiao, Z., Tao, C., Meng, Z., Dong, L., Xia, Y., Yongchao, T., et al., 2019. Improved
findings provide a reference for the rapid and non-destructive acquisi­ estimation of rice aboveground biomass combining textural and spectral analysis of
tion of the chlorophyll content of summer maize by UAV remote sensing UAV imagery. Precis. Agric. 20 (3).
Huete, A.R., 1988. A soil-adjusted vegetation index (SAVI). Elsevier 25 (3), 295–309.
and support for precise agriculture management. Jean-Louis, R., François-Marie, B., 1995. Estimating PAR absorbed by vegetation from
bidirectional reflectance measurements. Remote Sens. Environ. 51 (3).
Jibo, Y., Guijun, Y., Qingjiu, T., Haikuan, F., Kaijian, X., Chengquan, Z., 2019. Estimate
CRediT authorship contribution statement of winter-wheat above-ground biomass based on UAV ultrahigh-ground-resolution
image textures and vegetation indices. ISPRS J. Photogramm. Remote Sens. 150.
Lang Qiao: Conceptualization, Data curation, Formal analysis, Jing, M.C., 2014. Evaluation of vegetation indices and a modified simple ratio for boreal
applications. Canadian J. Remote Sens. 22 (3).
Methodology, Software, Writing – original draft. Weijie Tang: Investi­ Jinru, X., Baofeng, S., & Chenzong, L. (2017). Significant Remote Sensing Vegetation
gation, Methodology, Software. Dehua Gao: Software, Validation. Indices: A Review of Developments and Applications. Journal of Sensors, 2017.
Ruomei Zhao: Investigation, Methodology, Software. Lulu An: Soft­ John, J.S., Dan, S.L., 2016. Spectral considerations for modeling yield of canola. Remote
Sens. Environ. 184.
ware, Validation. Minzan Li: Supervision, Project administration,
Jordan, C.F., 1969. Derivation of leaf-area index from quality of light on the forest floor.
Writing – review & editing. Hong Sun: Conceptualization, Formal Ecology 50 (4), 663–666.
analysis, Supervision, Writing – review & editing. Di Song: Software, Karla, M., Ulf, B., Franziska, M., Henning, K., 2008. Analysis of vegetation indices
Validation. derived from hyperspectral reflection measurements for estimating crop canopy
parameters of oilseed rape (Brassica napus L.). Biosyst. Eng. 101 (2).
Lang, Q., Dehua, G., Junyi, Z., Minzan, L., Hong, S., Junyong, M., 2020. Dynamic
influence elimination and chlorophyll content diagnosis of maize using UAV spectral
imagery. Remote Sensing 12 (16).
Declaration of Competing Interest Lang, Qiao., Dehua, Gao., Ruomei, Zhao., Weijie, Tang., Lulu, An., Minzan, Li., & Hong,
Sun. (2022). Improving estimation of LAI dynamic by fusion of morphological and
The authors declare that they have no known competing financial vegetation indices based on UAV imagery. Computers and Electronics in Agriculture,
192.
interests or personal relationships that could have appeared to influence
Li, F., Miao, Y., Feng, G., Yuan, F., Yue, S., Gao, X., Liu, Y., Liu, B., Ustin, S.L., Chen, X.,
the work reported in this paper. 2014. Improving estimation of summer maize nitrogen status with red edge-based
spectral vegetation indices. Field Crops Research 157, 111–123.
Li, S., Yuan, F., Ata-UI-Karim, S.T., Zheng, H., Cheng, T., Liu, X., Tian, Y., Zhu, Y.,
Acknowledgement Cao, W., Cao, Q., 2019. Combining color indices and textures of UAV-based digital
imagery for rice LAI estimation. Remote Sens. 11 (15), 1763. https://doi.org/
This work was supported by the National Key Research and Devel­ 10.3390/rs11151763.
Liu, H. J., Min-Zan, L. I., Zhang, J. Y., Gao, D. H., Sun, H., & Jing-Zhu, W. U. (2019). A
opment Program (Grant No. 2019YFE0125500), the National Natural Modified Vegetationindex for Spectral Migration During Crop Growth.
Science Fund (Grant No. 31971785), and the Graduate Training Project Liu, Yang., Feng, Haikuan., Huang, Jue., Sun, Qian., Yang, Fuqin. (2020). Estimation of
of China Agricultural University (JG2019004,YW2020007,JG202026, potato biomass based on UAV digital images.Transactions of the Chinese Society of
Agricultural Engineering (Transactions of the CSAE), 36(23):181-192.
QYJC202101,JG202102). We would like to acknowledge the field Liu, N., Xing, Z., Qiao, L., Minzan, L.i., Sun, H., Zhang, Q., 2020b. Discussion on spectral
assistance of Dryland Farming Institute of Hebei Academy of Agriculture variables selection of potato chlorophyll using model population analysis.
and Forestry Sciences. Spectroscopy Spectral Anal. 40 (07), 2259–2266.
Lu, J., Cheng, D., Geng, C., Zhang, Z., Xiang, Y., Hu, T., 2021. Combining plant height,
canopy coverage and vegetation index from UAV-based RGB images to estimate leaf
References nitrogen concentration of summer maize. Biosyst. Eng. 202, 42–54.
Lu, N., Zhou, J., Han, Z., Li, D., Cao, Q., Yao, X., Tian, Y., Zhu, Y., Cao, W., Cheng, T.,
2019. Improved estimation of aboveground biomass in wheat from RGB imagery and
Aasen, H., Honkavaara, E., Lucieer, A., Zarco-Tejada, P., 2018. Quantitative remote
point cloud data acquired with a low-cost unmanned aerial vehicle system. Plant
sensing at ultra-high resolution with UAV spectroscopy: a review of sensor
Methods 15 (1). https://doi.org/10.1186/s13007-019-0402-3.
technology, measurement procedures, and data correction workflows. Remote
Ma, H., Chen, C., Song, Y., et al., 2018. Analysis of vegetation cover change and its
Sensing 10 (7), 1091. https://doi.org/10.3390/rs10071091.
driving factors over the past ten years in Qinghai Province. Res. Soil Water Conserv.
Alessandro, M., Piero, T., Salvatore, D.G., Lorenzo, G., Francesco, V., Jacopo, P., et al.,
25 (06), 137–145.
2015. Intercomparison of UAV, aircraft and satellite remote sensing platforms for
Martina, C., Daniele, C., Giovanni, C., Pietro, M.G., Luca, B., 2018. Does remote and
precision viticulture. Remote Sensing 7 (3).
proximal optical sensing successfully estimate maize variables? A review. Eur. J.
Anatoly, A.G., Yoram, J.K., Mark, N.M., 1996. Use of a green channel in remote sensing
Agron. 99.
of global vegetation from EOS-MODIS. Remote Sens. Environ. 58 (3).

11
L. Qiao et al. Computers and Electronics in Agriculture 196 (2022) 106775

Nguy-Robertson, A.L., Peng, Y.i., Gitelson, A.A., Arkebauer, T.J., Pimstein, A., contents of maize using UVE-PLS based on continuous wavelet transform. Comput.
Herrmann, I., Karnieli, A., Rundquist, D.C., Bonfil, D.J., 2014. Estimating green LAI Electron. Agric. 169, 105160. https://doi.org/10.1016/j.compag.2019.105160.
in four crops: Potential of determining optimal spectral bands for a universal Wei, F., Hai-Yan, Z., Yuan-Shuai, Z., Shuang-Li, Q., Ya-Rong, H., Bin-Bin, G., et al., 2016.
algorithm. Agric. For. Meteorol. 192-193, 140–148. Remote detection of canopy leaf nitrogen concentration in winter wheat by using
Ning, L., Gang, L., Hong, S., 2020. Real-time detection on SPAD value of potato plant water resistance vegetation indices from in-situ hyperspectral data. Field Crops
using an in-field spectral imaging sensor system. Sensors 20 (12). Research 198.
Peng, Y.i., Zhu, T., Li, Y., Dai, C., Fang, S., Gong, Y., Wu, X., Zhu, R., Liu, K., 2019. Wu, W., 2016. Intelligent monitoring and system development of diagnostic index of the
Remote prediction of yield based on LAI estimation in oilseed rape under different wheat seedling growth. YangZhou University.
planting methods and nitrogen fertilizer applications. Agric. For. Meteorol. 271, Yang, L., Xuewei, C., 2020. ANN-Based Continual Classification in Agriculture.
116–125. https://doi.org/10.1016/j.agrformet.2019.02.032. Agriculture 10 (5).
Pôças, I., Calera, A., Campos, I., Cunha, M., 2020. Remote sensing for estimating and Yao, X., Wang, N.i., Liu, Y., Cheng, T., Tian, Y., Chen, Q.i., Zhu, Y., 2017. Estimation of
mapping single and basal crop coefficientes: A review on spectral vegetation indices wheat LAI at middle to high levels using unmanned aerial vehicle narrowband
approaches. Agric. Water Manag. 233, 106081. https://doi.org/10.1016/j. multispectral imagery. Remote Sensing 9 (12), 1304. https://doi.org/10.3390/
agwat.2020.106081. rs9121304.
Schirrmann, M., Giebel, A., Gleiniger, F., Pflanz, M., Lentschke, J., Dammer, K., 2016. Yu, K., Lenz-Wiedemann, V., Chen, X., Bareth, G., 2014. Estimating leaf chlorophyll of
Monitoring Agronomic Parameters of Winter Wheat Crops with Low-Cost UAV barley at different growth stages using spectral indices to reduce soil background
Imagery. Remote Sensing 8 (9), 706. https://doi.org/10.3390/rs8090706. and canopy structure effects. ISPRS J. Photogramm. Remote Sens. 97, 58–77.
Shao, F., Liu, H., 2021. The Theoretical and Experimental Analysis of the Maximal https://doi.org/10.1016/j.isprsjprs.2014.08.005.
Information Coefficient Approximate Algorithm. Journal of Systems Science and Yu, Lei., Zhang, Tao., Zhu, Yaxing., Zhou, Yong., Xia, Tian., Nie, Yan. (2018).
Information 9 (1), 95–104. Determination of soybean leaf SPAD value using characteristic wavelength variables
Silleos, N.G., Alexandridis, T.K., Gitas, I.Z., Perakis, K., 2006. Vegetation Indices: preferably selected by IRIV algorithm.Transactions of the Chinese Society of
Advances Made in Biomass Estimation and Vegetation Monitoring in the Last 30 Agricultural Engineering (Transactions of the CSAE), 34(16):148-154.
Years. Geocarto International 21 (4), 21–28. https://doi.org/10.1080/ Yue, J., Guo, W., Yang, G., Zhou, C., Feng, H., Qiao, H., 2021. Method for accurate multi-
10106040608542399. growth-stage estimation of fractional vegetation cover using unmanned aerial
Sripada, R.P., 2005. Determining in-season nitrogen requirements for corn using aerial vehicle remote sensing. Plant Methods 17 (1). https://doi.org/10.1186/s13007-021-
color-infrared photography. North Carolina State University. 00752-3.
Wan, L., Cen, H., Zhu, J., Li, Y., & He, Y. (2019). Combining UAV-based vegetation Zhang, S.u., Liu, L., 2014. The potential of the MERIS Terrestrial Chlorophyll Index for
indices, canopy height and canopy coverage to improve rice yield prediction under crop yield prediction. Remote Sensing Letters 5 (8), 733–742.
different nitrogen levels. In 2019 Boston, Massachusetts July 7- July 10, 2019. Zhang, J., Cheng, T., Guo, W., Xu, X., Qiao, H., Xie, Y., Ma, X., 2021a. Leaf area index
Wan, L., Cen, H., Zhu, J., Zhang, J., Zhu, Y., Sun, D., Du, X., Zhai, L.i., Weng, H., Li, Y., estimation model for UAV image hyperspectral data based on wavelength variable
Li, X., Bao, Y., Shou, J., He, Y., 2020. Grain yield prediction of rice using multi- selection and machine learning methods. Plant Methods 17 (1). https://doi.org/
temporal UAV-based RGB and multispectral images and model transfer – a case 10.1186/s13007-021-00750-5.
study of small farmlands in the South of China. Agric. For. Meteorol. 291, 108096. Zhang, Y.u., Su, Z., Shen, W., Jia, R., Luan, J., 2016. Remote monitoring of heading rice
https://doi.org/10.1016/j.agrformet.2020.108096. growing and nitrogen content based on UAV Images. Int. J. Smart Home 10 (7),
Wang, W., Yao, X., Tian, Y.-C., Liu, X.-J., Ni, J., Cao, W.-X., Zhu, Y., 2012. Common 103–114.
spectral bands and optimum vegetation indices for monitoring leaf nitrogen Zhang, Y., Xia, C., Zhang, X., Cheng, X., Feng, G., Wang, Y., Gao, Q., 2021b. Estimating
accumulation in rice and wheat. J. Integrat. Agric. 11 (12), 2001–2012. the maize biomass by crop height and narrowband vegetation indices derived from
Wang, Z., Chen, J., Fan, Y., Cheng, Y., Wu, X., Zhang, J., Wang, B., Wang, X., Yong, T., UAV-based hyperspectral images. Ecol. Ind. 129, 107985. https://doi.org/10.1016/
Liu, W., Liu, J., Du, J., Yang, W., Yang, F., 2020. Evaluating photosynthetic pigment j.ecolind.2021.107985.

12