Combining UAV and Sentinel-2 satellite multi-spectral images to diagnose crop growth and N status in winter wheat at the county scale

Abstract

Real-time and non-destructive nitrogen (N) status diagnosis is needed to support in-season N management decision-making for modern wheat production. For this purpose, satellite sensor imaging can act as an effective tool for collecting crop growth information across large areas, but they can be challenging to calibrate with ground reference data. This research aimed to calibrate satellite remote sensing-derived models for crop growth estimation and N status diagnosis based on fne-resolution unmanned aerial vehicle (UAV) images, thus, map wheat growth and N status at the county scale. Seven wheat feld experiments involving multi cultivars and different N applications were conducted at four farms of Xinghua county from 2017 to 2021. A fxed-wing UAV sensing system and the Sentinel 2 (S2) satellite were used to collect wheat canopy multispectral images; three growth variables (plant dry matter (PDM), plant N accumulation (PNA) and N nutrition index (NNI)) and weather data, synchronized with spectral imagery, were obtained at the jointing and booting stages. The farmscale PDM (UAV-PDM) and PNA (UAV-PNA) maps can be derived from the UAV images at the four farms, which were further upscaled to grids to match the S2 image resolution using pixel aggregation method. Then, satellite based prediction models were constructed by ftting four machine learning algorithms to the relationships between satellite spectral indices, upscaled PDM (PNA) and weather data. Amongst the four methods tested, the random forest (RF) achieved the greatest prediction accuracy for PDM (R2 = 0.69–0.93) and PNA (R2 = 0.60–0.77). Meanwhile, an indirect diagnosis method was used to calculate the NNI. The results indicated that the model derived from the S2 imagery performed well for predicting NNI (R2 = 0.46–0.54) at the jointing and booting stages. Thereby, the NNI was used to map winter wheat N nutrition status at the county scale. In summary, this research demonstrated and evaluated an approach to combine UAV and satellite sensor images to diagnose wheat growth and N status across large areas.

Full text

Field Crops Research 294 (2023) 108860
0378-4290/© 2023 Elsevier B.V. All rights reserved.
Combining UAV and Sentinel-2 satellite multi-spectral images to diagnose
crop growth and N status in winter wheat at the county scale
Jie Jiang
a
,
b
,
c
,
d
,
e
, Peter M. Atkinson
f
,
g
,
h
, Chunsheng Chen
i
, Qiang Cao
a
,
b
,
c
,
d
,
e
,
Yongchao Tian
a
,
b
,
c
,
d
,
e
, Yan Zhu
a
,
b
,
c
,
d
,
e
, Xiaojun Liu
a
,
b
,
c
,
d
,
e
,
*
, Weixing Cao
a
,
b
,
c
,
d
,
e
,
*
a
National Engineering and Technology Center for Information Agriculture, Nanjing Agricultural University, Nanjing 210095, China
b
MOE Engineering Research Center of Smart Agricultural, Nanjing Agricultural University, Nanjing 210095, China
c
MARA Key Laboratory for Crop System Analysis and Decision Making, Nanjing Agricultural University, Nanjing 210095, China
d
Jiangsu Key laboratory for Information Agriculture, Nanjing Agricultural University, Nanjing 210095, China
e
Institute of Smart Agriculture, Nanjing Agricultural University, Nanjing 210095, China
f
Faculty of Science and Technology, Lancaster University, Lancaster LA1 4YR, UK
g
Geography and Environmental Science, University of Southampton, Higheld, Southampton SO17 1BJ, UK
h
Institute of Geographic Sciences and Natural Resources Research, Chinese Academy of Sciences, 11A Datun Road, Beijing 100101, China
i
Xinghua Extension Centre for Agricultural Technology, Taizhou 225700, China
ARTICLE INFO
Keywords:
N diagnosis
Pixel aggregation
Vegetation index
Random forest
Large areas
ABSTRACT
Real-time and non-destructive nitrogen (N) status diagnosis is needed to support in-season N management
decision-making for modern wheat production. For this purpose, satellite sensor imaging can act as an effective
tool for collecting crop growth information across large areas, but they can be challenging to calibrate with
ground reference data. This research aimed to calibrate satellite remote sensing-derived models for crop growth
estimation and N status diagnosis based on ne-resolution unmanned aerial vehicle (UAV) images, thus, map
wheat growth and N status at the county scale. Seven wheat eld experiments involving multi cultivars and
different N applications were conducted at four farms of Xinghua county from 2017 to 2021. A xed-wing UAV
sensing system and the Sentinel 2 (S2) satellite were used to collect wheat canopy multispectral images; three
growth variables (plant dry matter (PDM), plant N accumulation (PNA) and N nutrition index (NNI)) and
weather data, synchronized with spectral imagery, were obtained at the jointing and booting stages. The farm-
scale PDM (UAV-PDM) and PNA (UAV-PNA) maps can be derived from the UAV images at the four farms, which
were further upscaled to grids to match the S2 image resolution using pixel aggregation method. Then, satellite-
based prediction models were constructed by tting four machine learning algorithms to the relationships be-
tween satellite spectral indices, upscaled PDM (PNA) and weather data. Amongst the four methods tested, the
random forest (RF) achieved the greatest prediction accuracy for PDM (R
2
=0.69–0.93) and PNA (R
2
=
0.60–0.77). Meanwhile, an indirect diagnosis method was used to calculate the NNI. The results indicated that
the model derived from the S2 imagery performed well for predicting NNI (R
2
=0.46–0.54) at the jointing and
booting stages. Thereby, the NNI was used to map winter wheat N nutrition status at the county scale. In
summary, this research demonstrated and evaluated an approach to combine UAV and satellite sensor images to
diagnose wheat growth and N status across large areas.
1. Introduction
Wheat is one of the important crops that widely cultivated in the
world, which plays a vital role in ensuring the world food security. The
area covered by wheat in China is the fourth with 11% proportion of the
global plantation area, while the wheat yield accounts for appropriate
18% of the total yield in the world (Li et al., 2016; Wu et al., 2022).
Nitrogen (N) has a signicant effect in enhancing crop growth and
improving grain yield formation (Miao et al., 2011). A precision N
management strategy (PNMS) can be used to optimize N fertilizer inputs
* Corresponding authors at: National Engineering and Technology Center for Information Agriculture, Nanjing Agricultural University, Nanjing 210095, China
E-mail addresses: liuxj@njau.edu.cn (X. Liu), caow@njau.edu.cn (W. Cao).
Contents lists available at ScienceDirect
Field Crops Research
journal homepage: www.elsevier.com/locate/fcr
https://doi.org/10.1016/j.fcr.2023.108860
Received 4 August 2022; Received in revised form 24 December 2022; Accepted 9 February 2023

Field Crops Research 294 (2023) 108860
2
and maximize the economic benets for producers. Such a PNMS re-
quires non-destructive and effective tools for crop growth prediction and
N status diagnosis (Diacono et al., 2013).
Traditional N diagnosis methods such as leaf color-based judgement
and chemical analysis based for measured plant sample were relatively
empirical or time-consuming for evaluating the plant N nutrition status,
which may not be sufcient to support making in-season real-time N
management decisions in modern crop production (Padilla et al., 2018).
At the same time, remote sensing technologies provide a non-destructive
approach for real-time diagnosis of crop growth and N status (Moya,
2005; Dong et al., 2019). Ground sensing based on proximal sensors has
been used to predict plant biomass and N accumulation for various
crops, including rice, wheat and maize (Xia et al., 2016; Jiang et al.,
2020; Zhang et al., 2020). However, the small sampling area of ground
sensors makes this approach laborious for crop growth estimation and N
diagnosis at regional scales. Satellite remote sensing can perform spec-
tral sampling for crop N status estimation over large areas, and is likely
to be more suitable for guiding regional crop N management (Magney
et al., 2017). Common satellite missions were classied according to the
spatial resolution: coarse spatial resolution satellite sensors such as
MODIS (spatial resolution ≥250 m) have been used for monitoring
vegetation productivity and mapping foliar N in forests at broad scales
(Guay et al., 2014; Lepine et al., 2016). However, images with coarse
resolution are insufcient to detect eld heterogeneity due to the lack of
pure pixels during the crop growth period (Lepine et al., 2016). Rap-
ideye and IKONOS satellite sensors can produce images with a ne
spatial resolution of 1 m, which have been used for academic research
and agricultural production when combined with easy access to
compute resource (Rinaldi et al., 2010; Magney et al., 2017). Wang et al.
(2019) indicated that vegetation indices such as normalised difference
vegetation index (NDVI) and normalised difference red edge (NDRE)
derived from RapidEye images achieved good precision (R
2
>0.6) for
predicting wheat grain N uptake during the grain lling stage. However,
the cost of ne-resolution images has limited their practical application
in modern crop production. Freely available medium-to-ne resolution
satellite sensor images from Sentinel-2 (S2) have been more popular
with researchers conducting regional studies for crop N management.
Meanwhile, the revisit period of 5 days of S2 makes it suitable for
real-time estimation and diagnosis of crop growth and N status. There-
fore, the advantages of free-access and high revisit rate from the S2
satellite sensor imagery were more benecial for facilitating the prac-
tical agricultural production. Shari (2020) indicated that the simple
ratio red-edge (SRRE) index derived from the S2 satellite sensor image
has a good performance for estimating maize N uptake with R
2
of 0.91
and RMSE of 11.34 kg ha
−1
at the peak greenness date. Additionally, the
normalized difference red edge index (NDRE) and transformed chloro-
phyll absorption ratio index (TCARI) from the S2 images were demon-
strated a good linear estimates for maize NNI (R
2
=0.79) and durum
wheat NNI (R
2
=0.61), respectively (Crema et al., 2020). Therefore, it is
necessary to further evaluate the utility of spectral information derived
from the S2 satellite sensor images with medium-to-ne resolution for
wheat growth estimation and N status diagnosis.
Previous studies calibrated satellite-based models for crop growth
estimation mainly through single or multiple eld measurements, which
are laborious and difcult to upscale to the same spatial resolution as the
satellite sensor images (Huang et al., 2017). UAV-based remote sensing
systems can be operated with ease and have been demonstrated to be
excellent tools for diagnosing crop N status (Zhao et al., 2019).
Furthermore, the ne-resolution images from UAVs can detect eld
heterogeneity and can be aggregated to grids with any desired resolu-
tion. Thus, UAVs offer an opportunity to close the gap between eld
measurements and satellite sensor data. Revill et al. (2020) coupled S2
and UAV observations to bridge the scaling gap between eld data and
satellite sensor images, and the results indicated this method achieved
an accurate retrieval for wheat leaf area index across a large farm.
Similarly, the fractional cover (FCover) of tundra vegetation derived
from a ne-resolution UAV RGB image was aggregated to corresponding
grids with same spatial resolution as Planet (3 m), S2 (10 m, 20 m) and
Landsat 8 (30 m) images. Therefore, the FCover prediction model based
on satellite imagery can be constructed using the relationship between
UAV-FCover and satellite vegetation indices (Deviance explained =89%
at best) over larger extents (Riihim¨
aki et al., 2019). To date, little
research has been performed on the integration of xed-wing UAVs and
satellite sensor images to diagnose the growth and N status of winter
wheat at the county scale. Therefore, the objectives of this study were:
(1) to bridge the scale gap between eld observed wheat growth pa-
rameters and satellite sensor data based on ne-resolution UAV images;
(2) to construct wheat growth estimation and N diagnosis models using
S2 satellite sensor images and (3) to map wheat growth and N status
temporally and spatially at the county scale.
2. Materials and methods
2.1. Experimental design
This study was conducted at the Xinghua experimental station in
Jiangsu Province of East China (Fig. 1). Experiment 1–3 were conducted
using ‘Yangmai 23′and ‘Yangmai 25′cultivar at the Diaoyu farm from
2017 to 2020. Experiment 4 was conducted using ‘Nongmai 88′cultivar
at the Daiyao farm in 2019–2020. Experiment 5 and 6 were conducted
using ‘Yangmai 25′at the Daduo and Zhouzhuang farm, respectively, in
2019–2020. Experiment 7 was conducted based on the local cultivar
such as ‘Yangmai 23′, ‘Yangmai 25′and ‘Nongmai 88′across the Xinghua
county in 2020–2021. The fertilizer treatments of experiment 1–7 fol-
lowed the local farmer’ conventional approach, which were showed in
Table S1 of the Supplymentary le. Wheat plants in experiments 1–7
were grown at a local standard density of 2.25 million seedlings per
hectare. Irrigation application was applied one time to ensure the seeds
germinated securely at the sowing stage if there was no natural rainfall.
The weather data was collected from the local weather station, the Fig. 2
showed the accumulated precipitation, daily average temperature and
accumulated radiation with days after sowing during the whole wheat
growing season from 2017 to 2021. No signicant insects, weeds and
water stress were observed through the whole growing season. Details of
the seven wheat experiments were shown in Table 1.
2.2. Spectral data collection
2.2.1. UAV images collection
The Parrot Sequoia camera (MicaSense, Seattle, WA, USA; Fig. 3)
was mounted on the eBee UAV (senseFly, Cheseaux-Lausanne,
Switzerland; Fig. 3) to collect four multispectral images, including the
green (G, 550 ±40 nm), red (R, 660 ±40 nm), red edge (RE, 735
±10 nm) and near infrared (NIR, 790 ±40 nm) bands. The parameter
setting of the UAV ights and pre-processing method of the multispec-
tral images followed Jiang et al. (2022). UAV ight was conducted at a
speed of 8 m s
−1
under stable low wind, cloudless and sunny-sky con-
ditions from 10:00–14:00. The overlap in the ight direction and sidelap
were set as 75% for each image. The spatial resolution of spectral image
was 10 cm when the ight height was 100 m above the wheat canopy.
The radiation calibration and mosaicking of the acquired images were
performed in the Pix4Dmapper Ag software (Pix4D SA, Prilly,
Switzerland). Several ground control points (GCPs) were located using a
Trimble GeoXH6000, which were then used to geo-rectify the UAV
orthographic image for each farm. Details of UAV image acquisition
times were shown in Table 1.
2.2.2. Sentinel 2 images acquisition
The satellite sensor imagery was acquired as close as possible to the
UAV ights, with a screening criteria of ve days adjacent to the UAV
campaign. Sentinel 2 images can be downloaded from the ofcial
website (https://scihub.copernicus.eu/) as Level-1 C geometrically
J. Jiang et al.

Field Crops Research 294 (2023) 108860
3
corrected, top-of-atmosphere reflectance products. The atmospheric
correction was carried out using the Sen2Cor version 02.08.00 to pro-
duce the Level-2A product. The plug-in of ‘SuperResolution’ in Sentinel
Application Platform (SNAP) version 4.0.2 was used to downscale the
Level-2A image bands with 20 m spatial resolution to 10 m resolution.
Each S2 image include 13 bands (Table 2) with 290 km orbital swath
width: three bands were designed for monitoring atmospheric condi-
tions with 60 m spatial resolution (B1, B10, B11), which were not
considered in this research. Meanwhile, the red edge (RE) band b5 was
selected among the three RE bands, the Narrow NIR band b9 was
selected between the two NIR bands and the SWIR band b12 was
selected between the two SWIR bands. The six selected bands were used
to calculate the spectral indices in Table 3.
2.3. Agronomic and weather data collection
The GPS coordinates of each sampling point was determined by a
Trimble GeoXH6000 (Trimble, CA, USA). Then, 20 plants were selected
randomly and sampled within a range of 10 * 10 m centered around the
sampling point. The plants were separated into the stem and leaf, which
Fig. 1. The four study sites. The green areas and black points indicate the wheat growing area and sampling points, respectively, in each farm.
Fig. 2. Accumulated precipitation, daily average temperature and accumulated radiation with days after sowing during the whole wheat growing season of (a)
2017–2018, (b) 2018–2019, (c) 2019–2020, (d) 2020–2021 in Xinghua experimental station.
J. Jiang et al.

Field Crops Research 294 (2023) 108860
4
were oven dried at 105◦for 30 min and then dried at 70◦to a constant
weight to measure the stem dry matter (SDM) and leaf dry matter
(LDM). The plant dry matter (PDM) was calculated by Eq. (1).
PDM kg •ha−1=SDM kg •ha−1+LDM kg •ha−1(1)
The sub-samples of stem and leaf were later ground into a ne
powder to determine the stem (SNC) and leaf (LNC) N concentration
using the Kjeldahl digestion method (Bremner and Mulvaney, 1982).
The PNA was then calculated by Eq. (2). The plant N concentration (N
a
)
can be calculated as the ratio of PNA and PDM.
PNAkg •ha−1=SDMkg •ha−1×SNC(%) + LDMkg •ha−1
×LNC(%)(2)
The NNI (Eq. (3)) can be calculated using actual plant N concen-
tration (N
a
) divided by critical value, while the critical N concentration
(N
c
) can be calculated using the critical N dilution curve (CNDC; Eq. (4))
developed by Jiang et al. (2020).
NNI =Na/Nc(3)
Nc=4.17∗W-0.39 (4)
where W is the plant biomass.
The grain yield was collected by manually measuring 1 m
2
three
times at each sampling points in the harvest stage, and the observed
value was standardized to 14% grain moisture content.
Previous studies indicated that weather status would inuence crop
growth and the physiological process, and so should be included to in-
crease prediction accuracy during model construction (Wang et al.,
2020; Nonhebel, 1994; Verma et al., 2003). In this study, average daily
temperature (T
ave
), average daily minimum temperature (T
min
), average
daily maximum temperature (T
max
), accumulated daily average
Table 1
Basic information describing the seven eld experiments conducted in this study.
Experiment
No.
Year
Location Cultivar UAV image acquisition
time
S2 image acquisition
time
Plant sampling
date
Sowing
date
Harvest
date
Experiment 1
2017–2018
Diaoyu farm
(33.08◦N, 119.98◦E)
YM 23 22-March (JS)
16-April (BS)
23-March (JS) None 9 Nov. 3 June
Experiment 2
2018–2019
Diaoyu farm
(33.08◦N, 119.98◦E)
YM 25 6-March (JS)
2-April (BS)
6-March (JS)
30-March (BS)
None 2 Nov. 29 May
Experiment 3
2019–2020
Diaoyu farm
(33.08◦N, 119.98◦E)
YM 25 16-March (JS)
2-April (BS)
17-March (JS)
3-April (BS)
None 9 Nov. 2 June
Experiment 4
2019–2020
Daiyao farm
(32.96◦N, 120.17◦E)
NM 88 19-March (JS)
6-April (BS)
24-March (JS)
8-April (BS)
None 5 Nov. 1 June
Experiment 5
2019–2020
Daduo farm
(32.85◦N, 120.02◦E)
YM 25 20-March (JS)
7-April (BS)
24-March (JS)
8-April (BS)
None 4 Nov. 1 June
Experiment 6
2019–2020
Zhouzhuang farm
(32.69◦N, 119.95◦E)
YM 25 20-March (JS)
6-April (BS)
24-March (JS)
8-April (BS)
None 6 Nov. 2 June
Experiment 7
2020–2021
Xinghua county
(32.65–33.25◦N,
119.60–120.32◦E)
YM 23, YM 25,
NM 88
None 14-March (JS)
8-April (BS)
12(14)-March
(JS)
8-April (BS)
31-May to 2-June
(HS)
25 Oct.−10
Nov.
29 May-10
June
Note: the YM23, YM25 and NM 88 represent the Yangmai 23, Yangmai 25 and Nongmai 88 cultivars, respectively. JS, BS and HS represent the jointing, booting and
harvest stages, respectively. The S2 image was not available at the booting stage in experiment 1.
Fig. 3. The eBee xed-wing UAV used in this study.
Table 2
The 13 spectral bands from the S2 satellite sensor image.
Bands Central wavelength
(nm)
Bandwidth
(nm)
Spatial
resolution
(m)
Band 1 Coastal
aerosol
443 20 60
Band 2 Blue 490 65 10
Band 3 Green 560 35 10
Band 4 Red 665 30 10
Band 5 Red edge 705 15 20
Band 6 Red edge 740 15 20
Band 7 Red edge 783 20 20
Band 8 NIR 842 115 10
Band 9 Narrow NIR 865 20 20
Band
10
Water vapor 945 20 60
Band
11
SWIR-Cirrus 1380 30 60
Band
12
SWIR 1610 90 20
Band
13
SWIR 2190 180 20
J. Jiang et al.

Field Crops Research 294 (2023) 108860
5
temperature (T
sum
), accumulated precipitation (Prep
sum
), accumulated
radiation (Rad
sum
) of 30 days before measurement date, and accumu-
lated growing degree day (AGDD) from sowing to measurement date
were used as model inputs to calibrate the growth and N status diagnosis
model.
2.4. Data analysis
The workow for estimation model construction and evaluation was
shown in Fig. 4: when the UAV orthographic images were collected
(Fig. 4a), the PDM and PNA estimation models based on the UAV data
from Jiang et al. (2022) were used to derive the wheat PDM (UAV-PDM)
and PNA (UAV-PNA) maps at each farm for experiments 1–6 (Fig. 4b).
Following the method from Riihim¨
akia et al. (2019), the UAV-PDM and
UAV-PNA maps were upscaled to the same spatial resolution as the S2
images (10 m) using the pixel aggregation function of ArcGIS 10.2
software (Fig. 4c). In order to avoid the inuence of mixing pixels from
the water, road and other objects for model construction, the sampling
points of 57, 52, 53, and 43 that involving the pure pixel inner the wheat
eld were determined randomly at the Diaoyu, Daiyao, Daduo, and
Zhouzhang farm, respectively, to extract the upscaled PDM and PNA
values (Fig. 4c). Meanwhile, the S2 images from each farm of experi-
ment 1–6 were obtained (Fig. 4d), and the vegetation indices at the
corresponding sampling points were extracted from the S2 imagery
(Fig. 4e). Therefore, the data from experiments 1–6 and 10-fold
cross-validation were used to select the optimal machine learning
(ML) modeling method to integrate the S2 spectral indices, weather data
(Fig. 4f), and upscaled PDM (PNA) to construct the estimation models
(Fig. 4g). The methods considered were the Random Forest (RF), Lasso,
articial neural network (ANN) and partial least squares regression
(PLSR). The optimal modelling method with the larger R
2
and smaller
root mean square error (RMSE; Eq. (5)) and relative error (RE; Eq. (6))
was selected to establish the optimal satellite models for PDM and PNA
prediction. Therefore, the satellite prediction models with best modeling
method were established based on the data from experiments 1–6
Table 3
The vegetation indices used in this research.
Index name Formula S2 UAV Reference
Normalised difference
red edge (NDRE)
(NIR - RE)/
(NIR +RE)
√ √ (Barnes et al., 2000)
Red edge soil-adjusted
vegetation index
(RESAVI)
1.5 * (NIR -
RE)/(NIR +RE
+0.5)
√ √ (Sripada et al., 2005)
Red edge chlorophyll
index (CIRE)
(NIR / RE) - 1 √ √ (Sripada et al., 2005)
DATT (NIR - RE) /
(NIR - Red)
√ √ (Datt and B, 2010)
Modied chlorophyll
absorption in
reectance index
(MCARI1)
[(NIR - RE)
−0.2(NIR - G)]
* (NIR / RE)
√ √ (Gitelson et al., 2005)
Ratio water index (RWI) NIR / SWIR √ (Fernandes et al.,
2003)
Normalized difference
water index (NDWI)
(NIR - SWIR) /
(NIR +SWIR)
√ (Gao, 1995)
Ratio blue index (RBI) NIR / B √ (This study, modied
from Pearson and
Miller, 1972)
Normalised difference
blue index (NDBI)
(NIR - B) / (NIR
+B)
√ (This study, modied
from Tucker, 1979)
Green soil adjusted
vegetation index
(GSAVI)
1.5 * (NIR -
G)/(NIR +G +
0.5)
√ √ (Sripada et al., 2005)
Soil adjusted vegetation
index (SAVI)
1.5 * (NIR -
Red)/(NIR +
Red +0.5)
√ (Huete, 1988)
Normalised difference
vegetation index
(NDVI)
(NIR - Red)/
(NIR +Red)
√ √ (Tucker, 1979)
Ratio vegetation index
(RVI)
NIR / Red √ (Pearson and Miller,
1972)
Note: √ represents the vegetation indices calculated based on the satellite and
UAV images.
Fig. 4. The research methodology. NNI: N nutrition index; PDM: plant dry matter; PNA: plant N accumulation; S2: Sentinel 2; ML: machine learning; CNDC: critical N
dilution curve; T
ave
, T
min
, T
max
, T
sum
, Prep
sum
and Rad
sum
represent the average daily temperature, average daily minimum temperature, average daily maximum
temperature, accumulated daily average temperature, accumulated precipitation and accumulated radiation, respectively, of the 30 days before the measurement
date. AGDD represents the accumulated growing degree day from sowing to the measurement date. Fig. 4b and c represent the PNA maps that used as an example.
The x1, x2 in Fig. 4e represent the sampling points for the extraction of vegetation indices from the S2 imagery, while the y1, y2 in Fig. 4c represent the sampling
points for the extraction of upscaled PNA from the upscaled PNA maps.
J. Jiang et al.

Field Crops Research 294 (2023) 108860
6
(Fig. 4h). Independent ground sampling data from the experiment 7 was
used to further validate the optimum PDM and PNA estimation models.
RMSE =
1
n×n
i=1(Pi−Oi)2
(5)
RE(%) = 100 ×
1
n×n
i=1Pi−Oi
Oi2
(6)
where n represents the number of samples, O
i
and P
i
represent the
observed and predicted values, respectively.
An indirect strategy (Fig. 4i) was used to estimate the NNI in this
study: when the PDM and PNA were predicted, the PDM was input to the
CNDC to calculate the N
c
, then the PNA
c
can be calculated as the product
of predicted PDM and N
c
. Therefore, the NNI was calculated as ratio of
predicted PNA and PNA
c
(Eq. 7; Jiang et al., 2022; Zha et al., 2020; Xia
et al., 2016). Three N status categories of N decient, N optimal, and N
excessive can be divided using the predicted NNI according to the
optimal NNI diagnosis interval: 0.92–1.04 and 0.97–1.15 at the jointing
and booting stage (Jiang et al., 2022), respectively. Therefore, the wheat
N diagnosis status at the Xinghua county can be evaluated based on the
indirect NNI diagnosis model and in-season S2 imagery (Fig. 4j).
NNI =Predicted PNA/(Predicted PDM ∗Nc)(7)
The packages of ‘randomForest’, ‘glmnet’, ‘nnet’ and ‘pls’ from R
software were used during the process of model construction and vali-
dation. The determination of wheat growing area at the Xinghua county
referenced the research results from Yang et al. (2021). The ArcGIS 10.2
software was used to generate the growth and N status maps in each
farm and for Xinghua county. The correlation map and scatter diagram
in this study were plotted in the Origin 2021 software.
3. Results
3.1. The correlation of spectral data between UAV and S2 images
The UAV images with 10 cm spatial resolution from experiments 1–6
were resampled (pixel aggregation) to grids that matched the S2 image
resolution. Then, the band reectance derived from the UAV and sat-
ellite sensor images were used to calculate the spectral indices in
Table 3. A correlation analysis (Fig. 5a) between spectral data derived
from the UAV images and the S2 images was performed at the jointing
stage. The results showed that the G band from the UAV images, SWIR1
band and DATT vegetation index from the S2 images produced a
relatively small correlation (r<0.50), while most spectral bands and
vegetation indices achieved a larger correlation between UAV and S2
spectral data (r>0.60). The correlation between the UAV and S2
spectral data across the booting stage was generally smaller than that at
the jointing stage. However, most vegetation indices produced a well
correlation between the UAV and S2 spectral data at the booting stage
(Fig. 5b). As a result, the relatively large correlation between the UAV
and satellite sensor images can be the basis for the integration of UAV
and S2 data.
3.2. Upscaling the PDM and PNA maps based on pixel aggregation
Diaoyu farm in experiment 3 was used as an example. Seven vege-
tation indices (Table 3) were calculated using the reectance from the
UAV multispectral images. Fig. 6(a) shows the NDRE (0.11–0.55) maps
at the UAV image resolution (10 cm), revealing a large variance in
wheat growth across the whole farm. According to the UAV model for
PDM and PNA estimation from Jiang et al. (2022), the PDM (Fig. 6b;
UAV-PDM) and PNA (Fig. 6c; UAV-PNA) maps with 10 cm resolution
were calculated based on the seven vegetation indices extracted from
the UAV images, which had a range of 0.64–7.14 t ha
−1
and
17.71–263.70 kg ha
−1
, respectively, at the jointing stage across Diaoyu
farm. Then, the UAV-PDM (UAV-PNA) maps were upscaled to grids with
the same spatial resolution as S2 images (10 m) based on the pixel ag-
gregation. The upscaled PDM (Fig. 6d) and PNA (Fig. 6e) maps had a
range of 1.19–5.96 t ha
−1
and 28.70–216.41 kg ha
−1
, respectively.
Then, 57 sampling points were determined randomly to extract the
upscaled PDM and PNA values for calibrating the satellite estimation
models.
The statistical analysis was performed for the upscaled PDM and
upscaled PNA at the jointing and booting stages in experiment 1–6. The
results from the Table 4 showed the upscaled PDM and upscaled PNA
varied greatly across six experiments. The upscaled PDM ranged from
1.05 to 5.71 with the coefcient of variation (CV) of 46.05% at the
jointing stage, and from 2.99 to 6.08 with the CV of 18.59% at the
booting stage. Similarly, the upscaled PNA ranged from 24.93 to 136.39
with the CV of 40.18% at the jointing stage, and from 55.32 to 164.89
with the CV of 29.27% at the booting stage. The large variability in the
upscaled PDM and PNA renders the dataset suitable to evaluate the
performance of using satellite remote sensing information to diagnose
winter wheat N status.
Fig. 5. The correlation (r) of spectral data: between UAV and S2 images at the jointing (a) and booting (b) stages across experiments 1–6. Note: * means a signicant
difference at the 0.05 probability level. The ‘Spectral band_UAV’ and ‘Vegetation index_UAV’ represent the spectral band and vegetation index, respectively, derived
from the UAV imagery. The ‘Spectral band_S2′and ‘Vegetation index_S2′represent the spectral band and vegetation index, respectively, derived from the Sentinel-2
sensor imagery.
J. Jiang et al.

Field Crops Research 294 (2023) 108860
7
3.3. PDM and PNA estimation based on the S2 satellite sensor images
The PDM and PNA values were extracted from the upscaled PDM and
PNA maps, respectively. Meanwhile, 13 vegetation indices (Table 3)
were calculated based on the reectance derived from the S2 images.
Four ML methods were tted to the relationship between the upscaled
PDM, weather data and 13 spectral indices in the jointing and booting
stages. The results from 10-fold cross-validation (Table 5) indicated that
the RF method predicted PDM with a high accuracy among the four ML
methods. The RF model based on the S2 images achieved an R
2
of 0.93
and 0.69, RMSE of 0.43 and 0.51 t ha
−1
, and RE of 17.02% and 12.33%
in the jointing and booting stage, respectively. Therefore, the RF method
was selected as the PDM prediction model across experiments 1–6. In-
dependent data from experiment 7 were used to validate the PDM pre-
diction model based on the RF method. The results show that the PDM
model based on the S2 images produced an R
2
of 0.65 and 0.42, RMSE of
0.71 and 0.59 t ha
−1
, and RE of 33.85% and 11.65% in the jointing and
booting stage, respectively (Table 6).
Similarly, the four ML methods were tted to the relationships be-
tween the upscaled PNA, weather data and spectral indices derived from
the S2 images. The 10-fold cross-validation shows that the RF method
achieved accurate prediction of PNA among the four ML methods
(Table 5). The RF model had an R
2
of 0.77 and 0.60, RMSE of
16.35 kg ha
−1
and 20.41 kg ha
−1
, and RE of 24.52% and 20.69% in the
jointing and booting stage, respectively. Therefore, the RF method was
selected as the PNA prediction model across experiments 1–6. Inde-
pendent data from experiment 7 were used to validate the PNA pre-
diction model using the RF algorithm. The results show that PNA model
based on the S2 images had an R
2
of 0.72 and 0.70, RMSE of 13.40 and
19.05 kg ha
−1
, and RE of 31.54% and 15.15% at the jointing and
booting stage, respectively (Table 6).
Based on the criterion of InNodePurity from the RF model, the
relative importance of each input parameter for PDM and PNA estima-
tion can be evaluated (Fig. 7). Generally, the spectral indices have a
relatively higher importance than the weather variables for predicting
Fig. 6. The NDRE map with (a) UAV (10 cm) image resolution; the PDM maps with (b) UAV (10 cm) and (d) S2 (10 m) image resolution; and the PNA maps with (c)
UAV (10 cm) and (e) S2 (10 m) image resolution for Diaoyu farm in experiment 3 at the jointing stage. Note: the green points in Fig. 6(e) represent the randomly
determined sampling points.
Table 4
Descriptive statistics of upscaled plant dry matter (PDM) and plant N accumu-
lation (PNA) at the jointing and booting stages across experiments 1–6.
Parameter Growth stage N Min. Max. SD
a
CV
b
(%)
Upscaled PDM
(t ha
−1
)
Jointing 277 1.05 5.71 1.69 46.05
Booting 240 2.99 6.08 0.92 18.59
Upscaled PNA
(kg ha
−1
)
Jointing 277 24.93 136.39 34.22 40.18
Booting 240 55.32 164.89 32.20 29.27
Note: SD
a
indicates standard deviation of the mean; CV
b
indicates coefcient of
variation (%).
Table 5
The 10-fold cross-validation results across experiments 1–6 using the four ML
algorithms at the jointing and booting stages: for the relationship between PDM,
weather data and 13 vegetation indices from S2 images; and for relationship
between PNA, weather data and 13 vegetation indices from S2 images.
Parameter Method Jointing stage Booting stage
R
2
RMSE RE (%) R
2
RMSE RE (%)
PDM
(t ha
−1
)
RF 0.93 0.43 17.02 0.69 0.51 12.33
Lasso 0.93 0.44 17.1 0.67 0.53 12.64
ANN 0.92 0.51 19.3 0.56 0.71 16.45
PLSR 0.83 0.69 36.85 0.36 0.74 18.64
PNA
(kg ha
−1
)
RF 0.77 16.35 24.52 0.60 20.41 20.69
Lasso 0.72 18.11 27.31 0.52 22.31 22.27
ANN 0.63 23.05 30.74 0.42 23.90 25.95
PLSR 0.61 21.38 35.67 0.34 25.93 26.06
Table 6
Independent validation results of optimal PDM and PNA prediction model using
the eld data from experiment 7 at the jointing and booting stages.
Parameter Jointing stage Booting stage
R
2
RMSE RE (%) R
2
RMSE RE (%)
PDM (t ha
−1
) 0.65 0.71 33.85 0.42 0.59 11.65
PNA (kg ha
−1
) 0.72 13.40 31.54 0.70 19.05 15.15
J. Jiang et al.

Field Crops Research 294 (2023) 108860
8
the PDM and PNA. The Rad
sum
, T
ave
and AGDD were more important
variables for PDM estimation among seven weather parameters at the
jointing and booting stages. The variables of T
max
and AGDD performed
a relatively higher importance than other weather parameters for PNA
estimation at the jointing and booting stages.
The above analysis showed the model based on the S2 satellite im-
ages performed an accurate prediction for wheat PDM and PNA.
Therefore, the S2 estimation model was used to estimate the PDM and
PNA at the county scale. The Fig. 8 showed the wheat PDM value had a
range of 1.51–4.35 and 3.77–5.85 t ha
−1
in the jointing (Fig. 8a) and
booting (Fig. 8b) stage, respectively; while the PNA value had a range of
34.45–120.86 and 68.18–154.33 kg ha
−1
in the jointing (Fig. 8c) and
booting (Fig. 8d) stage, respectively, across the Xinghua county in
2020–2021.
3.4. N nutrition diagnosis based on NNI at the county scale
After the PDM and PNA were predicted, PNA
c
was calculated from
the predicted PDM and N
c
. Then, the NNI was derived as predicted PNA/
PNA
c
. Compared to the observed values of NNI from experiment 7, the
results show the NNI estimation model based on the S2 images had an R
2
of 0.54 and 0.46, RMSE of 0.12 and 0.13, and RE of 11.80% and 11.85%
in the jointing (Fig. 9a) and booting (Fig. 9b) stage, respectively. The
model based on the S2 images was used to predict the NNI at the county
scale. According to the optimal NNI diagnosis interval of 0.92–1.04 and
0.97–1.15 in the jointing and booting stage, respectively. Fig. 10a shows
that the wheat N status had areas of 21.14%, 42.58%, and 36.28% in the
N decient, optimal and excessive categories, respectively, at the
jointing stage; while the proportions of N decient, optimal and
excessive category were 34.35%, 53.87% and 11.78%, respectively, at
the booting stage (Fig. 10b) for Xinghua county in 2020–2021.
To further evaluate the performance of the satellite-based models
and NNI diagnosis maps in experiment 7, a linear relationship was
established between grain yield and predicted NNI derived from the NNI
estimation model and N diagnosis map of Xinghua county (Fig. 11). The
results indicated that the correlation between wheat yield and predicted
NNI has R
2
of 0.43 and 0.47 in the jointing (Fig. 11a) and booting
(Fig. 11b) stage, respectively.
4. Discussion
4.1. Closing the gap between the eld observation and satellite data based
on the ne resolution UAV images
Crop growth observation is relatively simple at the eld scale, but the
laborious plant measurement across large areas is more challenging.
Therefore, satellite remote sensing can play a signicant role for sam-
pling crop information across large areas (Guay et al., 2014). Single or
multiple eld measurement methods are commonly used to calibrate
satellite-based models for crop growth monitoring, which is
time-consuming in terms of non-destructive plant sampling. Meanwhile,
it is difcult to match the satellite sensor images and eld observations
at the same spatial resolution (Huang et al., 2017). Fine-resolution UAV
images offer the possibility for producing crop growth maps at multiple
scales, hence providing a much needed link between eld and satellite
sensor data. Previous studies showed that tundra vegetation can be
classied based on UAV RGB ortho-mosaics in the arctic, which were
further converted to Planet (3 m), S2 (10 m, 20 m) and Landsat 8 (30 m)
Fig. 7. The importance (InNodePurity) value of each input parameter from the RF model for PDM estimation at the (a) jointing and (b) booting stages; for PNA
prediction at the (c) jointing and (d) booting stages.
J. Jiang et al.

Field Crops Research 294 (2023) 108860
9
image grids to train satellite-based models for vegetation cover moni-
toring (Riihim¨
aki et al., 2019). The UAV model of Jiang et al. (2022)
achieved a high accuracy for predicting wheat PDM (R
2
=0.69–0.78)
and PNA (R
2
=0.83–0.84) at the farm scale. Therefore, the PDM
(0.64–7.17 t ha
−1
) and PNA (17.71–263.70 kg ha
−1
) values at Diaoyu
farm of experiment 3 can be derived from the UAV prediction model
(Fig. 6b and c, respectively), showing visually the large variability of
wheat growth across the whole farm. Furthermore, the PDM maps
(0.64 −7.17 t ha
−1
) with UAV image resolution were upscaled to the S2
image resolution, while the range of PDM decreased with an increase in
Fig. 8. Maps of PDM at the (a) jointing and (b) booting stages and PNA at the (c) jointing and (d) booting stages based on S2 images in 2020–2021 across Xing-
hua county.
Fig. 9. Independent validation results of NNI prediction model using the eld data from experiment 7 at the jointing (a) and booting (b) stages. Note: the blue line in
the gure indicates the regression line.
J. Jiang et al.

Field Crops Research 294 (2023) 108860
10
pixel size, as expected. The upscaled PDM maps with 10 m resolution
had values of 1.19–5.96 t ha
−1
. This aggregation effect arises as part of
the well-known Modiable Areal Unit Problem (MAUP). Previous
studies indicated that the phenomenon will arise when ner resolution
data are aggregated to coarser spatial resolution (Dark and Bram, 2007),
and similar results were demonstrated by Riihim¨
aki et al. (2019). The
UAV-PDM (PNA) values with large variability over the farm can be used
as reference data for satellite-based model construction. Revill et al.
(2020) derived wheat LAI maps from the UAV model, which were then
upscaled to S2 grids to train satellite-based estimation models, similar to
this study. The co-registration error caused by the GPS deviation be-
tween the UAV-derived maps and satellite sensor images should be
considered during the analysis process. Therefore, a certain number of
control points were set at each farm to calibrate the UAV ortho-mosaics
to ensure the accuracy of geographic location. Additionally, more UAV
and satellite sensor images from different cultivars and eco-sites should
be collected to construct robust models for crop growth prediction and N
diagnosis.
4.2. Wheat growth prediction and N diagnosis models based on satellite
multi-spectral information and weather variables
Previous studies demonstrated that multi-source information based
on spectral indices can increase the accuracy of crop growth and N status
prediction, while ML has been found highly suitable for integrating
multi-source data (Wang et al., 2021). In this research, four ML methods
were used to combine satellite spectral indices and weather data. The RF
performed most accurately for predicting PDM (R
2
=0.42–0.65) and
PNA (R
2
=0.70–0.72), which are comparable to the results of Jiang et al.
(2022) who also increased the accuracy of PDM (R
2
=0.52–0.68) and
PNA (R
2
=0.67–0.82) prediction with the integration of UAV spectral
indices, weather and eld management data based on the RF method.
Nevertheless, the eld management data such as N application rates
were not considered in this research due to the difculty of determining
it in the different elds over large areas. Although the weather variables
play a relatively low importance for PDM and PNA estimation (Fig. 7).
However, the changes for radiation and precipitation were reported to
affect the crop growth and N nutrition status through adjusting ambient
conditions such as air humidity and temperature, which can inuence
plant stomatal conductance, water status and other physiological func-
tions that control the plant root N absorption and transfer (Naylor et al.,
2020; Nonhebel, 1994). Temperature information like AGDD performed
a relatively high importance for PDM and PNA estimation (Fig. 7), which
may due to the AGDD represent the heat accumulation during the crop
growth period, and directly affect to plant growth rate and phenological
process (Santos et al., 2021; Zhou et al., 2020). Similar research was
Fig. 10. The N diagnosis maps based on S2 images at the (a) jointing and (b) booting stages across the Xinghua county scale in 2020–2021.
Fig. 11. The linear relationship between wheat yield and predicted NNI from the N diagnosis maps of the Xinghua county at the jointing (a) and booting (b) stage
in 2020–2021.
J. Jiang et al.

Field Crops Research 294 (2023) 108860
11
conducted by Wang et al., 2021 who integrate the ground sensing in-
formation and weather variables such as accumulated precipitation and
growing degree day for accurately monitoring the maize NNI and grain
yield at V8-V9 growth stage. Additionally, six bands (B, G, R, RE, NIR
and SWIR) extracted from the S2 images were used to construct the
prediction model in this study. Abundant spectral information was
demonstrated to be more representative for characterizing crop growth
and N nutrition (Verrelst et al., 2012, 2015). Li et al. (2021) also indi-
cated the integration of multi-source information from S2 images
increased the accuracy of chlorophyll prediction across typical lakes in
China, similar with those presented in this study. Several studies indi-
cated RF method exhibits a signicant performance to integrate the
multi variables for predicting plant biomass, leaf area index, and N
concentration in wheat, rice, and soybean crops (Muharam et al., 2021;
Liang et al., 2018; Maimaitijiang et al., 2020). During the RF model
construction, multiple sample sub-sets can be obtained from the original
sample sets using the bootstrap re-sampling method, while each sample
subset was used to construct an independent decision tree for model
prediction. Therefore, the fusion of predictions from multiple decision
trees was used as the nal results of RF models. The specicity of
re-sampling and multiple decision trees was demonstrated to well pro-
cess the outliers during the model construction and improve the model
prediction accuracy (Svetnik et al., 2003).
The PDM and PNA prediction accuracies at the booting stage (R
2
=
0.34–0.69) were lower than at the jointing stage (R
2
=0.61–0.93),
which may be due to the inuence of wheat canopy closure at the later
growth stages (Cao et al., 2015). At the booting stage, the winter wheat
grows strongly and all leaves are grown out from the plant. Meanwhile,
the top leaves shelter the lower leaves and stem, which limits detection
of the whole plant using spectral sensors. Other researchers also
demonstrated this similar phenomenon when estimating aboveground
biomass and N uptake of rice and wheat leaf area index during the later
growth stages (Cao et al., 2013; Zhang et al., 2019). The harvest yield
validation for the NNI prediction model and N diagnosis maps achieved
a comparable result at the jointing (R
2
=0.43) and booting (R
2
=0.47)
stages, which was similar with results from Crema et al. (2020) who
demonstrated a relationship between S2 image-derived NNI and maize
yield with a correlation coefcient r of 0.6. During the practical pro-
duction, more factors affected the crop growth and yield formation,
including fertilizers other than N, water status, soil nutrition, and insect
pests and weeds, etc. These useful information should also be considered
in model construction to increase the diagnosis accuracy in future study.
4.3. The potential for Sentinel 2 to diagnose crop N status across large
areas
Satellite remote sensing has great potential for predicting crop
growth across large areas due to the larger sampling extent than ground-
based and aerial spectral sensing systems (K. Zhang et al., 2020; C.
Zhang et al., 2020). Previous studies used satellite sensor images for
applied predicting crop LAI, aboveground biomass and N content in
wheat at the farm scale (Li et al., 2019; Fabbri et al., 2020), while not for
diagnosing N status across larger areas such as a county. In this research,
S2 images with swath widths of 240 km were demonstrated to provide
large area crop information compared to the small image swath widths
of, for example, the Planet mission (24.6 km; Li et al., 2019). Generally,
the ne spatial resolution images were demonstrated to involve more
crop information than at coarser resolution. However, this also neces-
sitates complex calculations during data analysis when used for practical
applications. On the other hand, coarse spatial resolution images such as
from Landsat (30 m) cannot detect eld heterogeneity, especially as
most of the elds studied here are approximately 40 m wide. Thus,
medium resolution sensors such as Landsat are not suitable for devel-
oping crop management strategies in each eld (Huang et al., 2017). In
this regard, it was implicit that the S2 images with medium-ne reso-
lution were more feasible for characterising crop N nutrition status and
guiding N nutrition management at the county scale. Areas of 36.28%
and 21.14% belong to the N excessive and decient categories, respec-
tively, at the jointing stage, which means less and more N demand,
respectively, compared to the optimal N status (Fig. 10a). However, an
area of 34.35% of the N decient category was found at the booting
stage, which may due to inappropriate topdressing N application by
farmers across the whole county (Fig. 10b). Therefore, a suitable N
regulation algorithm should be developed to adjust N topdressing rates
on the basis of farmers’ N management at the jointing stage. Previous
studies showed the PNMS supported by UAV remote sensing data opti-
mized crop growth and improved the NUE for wheat production
(Argento et al., 2021), while the relevant PNMS based on satellite sensor
images should also be developed and applied at the county scale. The
crop growth stage may vary over such a large area, and it is unwise to
apply the same management strategy to crops under different growth
stages. As a result, the inuence of the growth stage should be consid-
ered for crop N status diagnosis and regulation. Previous studies
demonstrated that the synthetic aperture radar (SAR) and optical
time-series data derived from the satellite and UAV remote sensing
systems have been used for accurately tracking the crop phenological
phrase in rice, winter wheat, maize and soybean (Diao et al., 2021; Liu
et al., 2022; Guo et al., 2022; Zhao et al., 2022). Therefore, the inte-
gration of crop phenology estimation technology and N management
strategy may better facilitate to most precisely diagnose the crop N
status and determine the optimal N recommendation rates at the optimal
growth stages, which would be signicant for improving the crop
growth and increasing the N use efciency. However, the difference in
growth period for winter wheat was not more than 5 days in Xinghua
county according to the survey. Therefore, the growth stages were
regarded as uniform across the whole county and the inuence of
growth stage was not considered during model construction in this
study. Nevertheless, it should be considered for larger areas in future
studies.
5. Conclusion
This research demonstrated the farm-scale PDM (UAV-PDM) and
PNA (UAV-PNA) maps derived from ne-resolution UAV images can be
aggregated to grids that match the S2 satellite image resolution to
calibrate satellite-based models for wheat growth and N status estima-
tion. Meanwhile, the results indicated that the S2 imagery-derived
model based on the RF algorithm produced a high accuracy for pre-
dicting the PDM, PNA and NNI in the jointing and booting stages.
Thereby, wheat growth and N nutrition status were mapped across
Xinghua county, China. We conclude that the combination of UAV and
satellite images can be used to diagnose and map wheat growth and N
status across wide areas (the county scale).
Funding
This work was supported by the National Key Research and Devel-
opment Program of China (No. 2022YFD2301402), the National Natural
Science Foundation of China (No. 32071903), the Jiangsu Provincial
Key Technologies R&D Program of China (No. BE2019386), and the
Guidance Foundationthe Sanya Institute of Nanjing Agricultural Uni-
versity (No. NAUSY-ZD01), China.
CRediT authorship contribution statement
Jie Jiang: Conceptualization, Methodology, Formal analysis, Soft-
ware, Formal analysis, Investigation, Writing −original draft, Visuali-
zation. Peter M. Atkinson: Methodology, Writing −review & editing.
Chunsheng Chen: Methodology, Investigation. Qiang Cao: Methodol-
ogy, Investigation. Yongchao Tian: Methodology, Writing −review &
editing. Yan Zhu: Investigation, Writing −review & editing. Xiaojun
Liu: Conceptualization, Methodology, Supervision, Funding acquisition,
J. Jiang et al.

Field Crops Research 294 (2023) 108860
12
Writing −review & editing. Weixing Cao: Supervision, Funding
acquisition, Writing −review & editing.
Declaration of Competing Interest
The authors declare that they have no known competing nancial
interests or personal relationships that could have appeared to inuence
the work reported in this paper.
Data availability
Data will be made available on request.
Acknowledgments
Thanks to Xue Wang, Jiayi Zhang, Zhihao Zhang, Yang Gao, Xinge Li
for eld management, ground measurement, and chemical analysis of
plant samples.
Appendix A. Supporting information
Supplementary data associated with this article can be found in the
online version at doi:10.1016/j.fcr.2023.108860.
References
Argento, F., Anken, T., Abt, F., Vogelsanger, E., Walter, A., Liebisch, F., 2021. Site-
specic nitrogen management in winter wheat supported by low-altitude remote
sensing and soil data. Precis. Agric. 22, 364–386. https://doi.org/10.1007/s11119-
020-09733-3.
Barnes, E.M., Clarke, T.R., Richards, S.E., Colaizzi, P.D., Thompson, T., 2000. Coincident
detection of crop water stress, nitrogen status, and canopy density using ground
based multispectral data. Proc. 5th Int. Conf. Precis. Agric. Other Resour. Manag.
16–19 (Bloomington, MN, USA).
Bremner, J., Mulvaney, C., 1982. In: Miller, R.H., A.L., Keeney, D.R. (Eds.), Nitrogen
-total. In Methods of Soil Analysis. In Chemical and Microbial Properties. American
Society of Agronomy, and Soil Science Society, Madison, WI, USA, pp. 595–624.
Cao, Q., Miao, Y., Wang, H., Huang, S., Cheng, S., Khosla, R., Jiang, R., 2013. Non-
destructive estimation of rice plant nitrogen status with Crop Circle multispectral
active canopy sensor. Field Crops Res. 154, 133–144. https://doi.org/10.1016/j.
fcr.2013.08.005.
Cao, Q., Miao, Y., Feng, G., Gao, X., Li, F., Liu, B., Yue, S., Cheng, S., Ustin, S.L.,
Khosla, R., 2015. Active canopy sensing of winter wheat nitrogen status: An
evaluation of two sensor systems. Comput. Electron. Agric. 112, 54–67. https://doi.
org/10.1016/j.compag.2014.08.012.
Crema, A., Boschetti, M., Nutini, F., Cillis, D., Casa, R., 2020. Inuence of Soil Properties
on Maize and Wheat Nitrogen Status Assessment from Sentinel-2 Data. Remote Sens.
12, 2175. https://doi.org/10.3390/rs12142175.
Dark, S.J., Bram, D., 2007. The modifable areal unit problem (MAUP) in physical
geography. Prog. Phys. Geogr. 31, 471–479. https://doi.org/10.1177/
0309133307083294.
Datt, B., 2010. Visible/near infrared reectance and chlorophyll content in Eucalyptus
leaves. Int. J. Remote Sens. 20, 2741–2759. https://doi.org/10.1080/
014311699211778.
Diacono, M., Montemurro, F., 2013. Precision nitrogen management of wheat. A review.
Agron. Sustain. Dev. 33, 219–241. https://doi.org/10.1007/s13593-012-0111-z.
Diao, C., Yang, Z., Gao, F., Zhang, X., Yang, Z., 2021. Hybrid phenology matching model
for robust crop phenological retrieval. ISPRS J. Photogramm. Remote Sens. 181,
308–326. https://doi.org/10.1016/j.isprsjprs.2021.09.011.
Dong, T., Shang, J., Chen, J.M., Liu, J., Zhou, G., 2019. Assessment of portable
chlorophyll meters for measuring crop leaf chlorophyll concentration. Remote Sens.
11, 2706. https://doi.org/10.3390/rs11222706.
Fabbri, C., Mancini, M., Marta, A.D., Orlandini, S., Napoli, M., 2020. Integrating satellite
data with a Nitrogen Nutrition Curve for precision top-dress fertilization of durum
wheat. Eur. J. Agron. 120, 126148 https://doi.org/10.1016/j.eja.2020.126148.
Fernandes, R., Butson, C., Leblanc, S., Latifovic, R., 2003. Landsat-5 TM and landsat-7
ETM +based accuracy assessment of leaf area index products for Canada derived
from SPOT-4 VEGETATION data. Can. J. Remote Sens. 29, 241–258. https://doi.org/
10.5589/m02-092.
Gao, B., 1995. NDWI—A normalized difference water index for remote sensing of
vegetation liquid water from space. Remote Sens. Environ. 58, 257–266. https://doi.
org/10.1016/S0034-4257(96)00067-3.
Gitelson, A.A., Vi˜
na, A., Ciganda, V., Rundquist, D.C., Arkebauer, T.J., 2005. Remote
estimation of canopy chlorophyll content in crops. Geophys. Res. Lett. 32, 1–4.
https://doi.org/10.1029/2005GL022688.
Guay, K., Beck, P., Berner, L., Goetz, S., Baccini, A., Buermann, W., 2014. Vegetation
productivity patterns at high northern latitudes: A multi-sensor satellite data
assessment. Glob. Change Biol. 20, 3147–3158. https://doi.org/10.1111/gcb.12647.
Guo, Y., Xiao, Y., Li, M., Hao, F., Zhang, X., Sun, H., Beurs, K., Fu, Y., He, Y., 2022.
Identifying crop phenology using maize height constructed from multi-sources
images. Int. J. Appl. Earth Obs. Geoinf. 115, 103121 https://doi.org/10.1016/j.
jag.2022.103121.
Huang, S., Miao, Y., Yuan, F., Martin, G., Yao, Y., Cao, Q., Wang, H., Victoria, L.W.,
Georg, B., 2017. Potential of RapidEye and WorldView-2 Satellite Data for
Improving Rice Nitrogen Status Monitoring at Different Growth Stages. Remote Sens.
9, 227. https://doi.org/10.3390/rs9030227.
Huete, A., 1988. A soil adjusted vegetation index (SAVI). Remote Sens. Environ. 17,
37–53. https://doi.org/10.1016/0034-4257(88)90106-X.
Jiang, J., Wang, C., Wang, Y., Cao, Q., Liu, X., 2020. Using an active sensor to develop
new critical nitrogen dilution curve for winter wheat. Sensors 20, 1577. https://doi.
org/10.3390/s20061577.
Jiang, J., Peter, M., Zhang, J., Lu, R., Zhou, Y., Cao, Q., Tian, Y., Zhu, Y., Cao, W., Liu, X.,
2022. Combining xed-wing UAV multispectral imagery and machine learning to
diagnose winter wheat nitrogen status at the farm scale. Eur. J. Agron. 138, 126537
https://doi.org/10.1016/j.eja.2022.126537.
Lepine, L., Ollinger, S., Ouimette, A., Martin, M., 2016. Examining spectral reectance
features related to foliar nitrogen in forests: Implications for broad-scale nitrogen
mapping. Remote Sens. Environ. 173, 174–186. https://doi.org/10.1016/j.
rse.2015.11.028.
Li, K., Yang, X., Tian, H., Pan, S., Liu, Z., Lu, S., 2016. Effects of changing climate and
cultivar on the phenology and yield of winter wheat in the North China Plain. Int. J.
Biometeorol. 60 (1), 21–32. https://doi.org/10.1007/s00484-015-1002-1.
Li, S., Song, K., Wang, S., Liu, G., Mu, G., 2021. Quantication of chlorophyll-a in typical
lakes across China using Sentinel-2 MSI imagery with machine learning algorithm.
Sci. Total Environ. 778, 146271 https://doi.org/10.1016/j.scitotenv.2021.146271.
Li, W., Jiang, J., Guo, T., Zhou, M., Yao, X., 2019. Remote sensing Generating Red-Edge
Images at 3 M Spatial Resolution by Fusing Sentinel-2 and Planet Satellite Products.
Remote Sens. 11, 1422. https://doi.org/10.3390/rs11121422.
Liang, L., Di, L., Huang, T., Wang, J., Lin, L., Wang, L., Yang, M., 2018. Estimation of leaf
nitrogen content in wheat using new hyperspectral indices and a random forest
regression algorithm. Remote Sens. 10, 1940.
Liu, L., Cao, R., Chen, J., Shen, M., Wang, S., Zhou, J., He, B., 2022. Detecting crop
phenology from vegetation index time-series data by improved shape model tting
in each phenological stage. Remote Sens. Environ. 277, 277. https://doi.org/
10.1016/j.rse.2022.113060.
Magney, T.S., Eitel, J.U.H., Vierling, L.A., 2017. Mapping wheat nitrogen uptake from
RapidEye vegetation indices. Precis. Agric. 18, 429–451. https://doi.org/10.1007/
s11119-016-9463-8.
Maimaitijiang, M., Sagan, V., Sidike, P., Daloye, A., Erkbol, H., Fritschi, F., 2020. Crop
Monitoring Using Satellite/UAV Data Fusionand Machine Learning. Remote Sens.
12, 1357. https://doi.org/10.3390/rs12091357.
Miao, Y., Stewart, B., Zhang, F., 2011. Long-term experiments for sustainable nutrient
management in China. A review. Agron. Sustain. Dev. 31, 397–414. https://doi.org/
10.1051/agro/2010034.
Moya, A.C.A.Z., 2005. Optically assessed contents of leaf polyphenolics and chlorophyll
as indicators of nitrogen deciency in wheat (Triticum aestivum L.). Field Crops Res.
91, 35–49. https://doi.org/10.1016/j.fcr.2004.05.002.
Muharam, F.M., Nurulhuda, K., Zulkai, Z., Tarmizi, M.A., Abdullah, A., Hashim, M.,
Zad, S., Derraz, R., Ismail, M., R., 2021. Uav- and random-forest-adaboost (rfa)-based
estimation of rice plant traits. Agronomy 11, 915. https://doi.org/10.3390/
agronomy11050915.
Naylor, D., Sadler, N., Bhattacharjee, A., Graham, E.B., Anderton, C.R., McClure, R.,
Lipton, M., Hofmockel, K.S., Jansson, J.K., 2020. Soil microbiomes under climate
change and implications for carbon cycling. Annu. Rev. Environ. Resour. 45, 29–59.
https://doi.org/10.1146/annurev-environ-012320-082720.
Nonhebel, S., 1994. The effects of use of average instead of daily weather data in crop
growth simulation models. Agric. Syst. 44, 377–396. https://doi.org/10.1016/0308-
521X(94)90194-K.
Padilla, F., Gallardo, M., Pe˜
na-Fleitas, M., Souza, R., Thompson, R., 2018. Proximal
Optical Sensors for Nitrogen Management of Vegetable Crops: A Review. Sensors 18,
2083. https://doi.org/10.3390/s18072083.
Pearson, R., Miller, L., 1972. Remote Mapping of Standing Crop Biomass for Estimation
of Productivity of the Shortgrass Prairie. Oct 2–6 Proceedings of the Eighth
International Symposium on Remote Sensing of Environment. Environmental
Research Institute of Michigan,, Ann Arbor, MI, pp. 1357–1381. https://doi.org/
10.1177/002076409904500102. Oct 2–6.
Revill, A., Florence, A., Macarthur, A., Hoad, S., Williams, M., 2020. Quantifying
Uncertainty and Bridging the Scaling Gap in the Retrieval of Leaf Area Index by
Coupling Sentinel-2 and UAV Observations. Remote Sens. 12, 1843. https://doi.org/
10.3390/rs12111843.
Riihim¨
aki, H., Luoto, M., Heiskanen, J., 2019. Estimating fractional cover of tundra
vegetation at multiple scales using unmanned aerial systems and optical satellite
data. Remote Sens. Environ. 224, 119–132. https://doi.org/10.1016/j.
rse.2019.01.030.
Rinaldi, M., Ruggieri, S., Garofalo, P., Vonella, A.V., Satalino, G., Soldo, P., 2010. Leaf
Area Index Retrieval Using High Resolution Remote Sensing Data. Ital. J. Agron. 5,
155–166. https://doi.org/10.4081/ija.2010.155.
Santos, A., Lacerda, L., Rossi, C., Moreno, L., Oliveira, M., Pilon, C., Silva, R., Vellidis, G.,
2021. Using UAV and multispectral images to estimate peanut maturity variability
on irrigated and rainfed elds applying linear models and articial neural networks.
Remote Sens. 14, 93. https://doi.org/10.3390/rs14010093.
Shari, A., 2020. Using Sentinel-2 data to predict nitrogen uptake in maize crop. IEEE J.
Sel. Top. Appl. Earth Obs. Remote Sens. 99, 1. https://doi.org/10.1109/
JSTARS.2020.2998638.
J. Jiang et al.

Field Crops Research 294 (2023) 108860
13
Sripada, R.P., Heiniger, R.W., White, J.G., Weisz, R., 2005. Aerial Color Infrared
Photography for Determining Late-Season Nitrogen Requirements in Corn. Agron. J.
97, 1511–1514. https://doi.org/10.2134/agronj2004.0314.
Svetnik, V., Liaw, A., Tong, C., Culberson, J., Sheridan, R., Feuston, B., 2003. Random
Forest: A Classication and Regression Tool for Compound Classication and QSAR
Modeling. J. Chem. Inf. Comput. Sci. 43, 1947–1958.
Tucker, C.J., 1979. Red and photographic infrared linear combinations for monitoring
vegetation. Remote Sens. Environ. 8, 127–150. https://doi.org/10.1016/0034-4257
(79)90013-0.
Verma, U., Ruhal, D.S., Hooda, R.S., Yadav, M., Hooda, L., 2003. Wheat Yield Modelling
Using Remote Sensing and Agrometeorological Data in Haryana State. Indian Soc.
Agricult. Stat. (India) 56 (2), 190–198.
Verrelst, J., Mu˜
noz, J., Alonso, L., et al., 2012. Machine learning regression algorithms
for biophysical parameter retrieval: Opportunities for Sentinel-2 and -3. Remote
Sens. Environ. 118, 127–139. https://doi.org/10.1016/j.rse.2011.11.002.
Verrelst, J., Rivera Caicedo, J., Veroustraete, F., Mu ˜
noz, J., Clevers, J.G.P.W., Camps-
Valls, G., Moreno, J., 2015. Experimental Sentinel-2 LAI estimation using
parametric, non-parametric and physical retrieval methods – A comparison. ISPRS J.
Photogramm. Remote Sens. 108, 260–272. https://doi.org/10.1016/j.
isprsjprs.2015.04.013.
Wang, K., Huggins, D.R., Tao, H., 2019. Rapid mapping of winter wheat yield, protein,
and nitrogen uptake using remote and proximal sensing. Int. J. Appl. Earth Obs.
Geoinf. 82, 101921 https://doi.org/10.1016/j.jag.2019.101921.
Wang, X., Miao, Y., Dong, R., Hainie, Z., Xia, T., Zhichao, C., Kusnierek, K., Mi, G.,
Sun, H., Li, M., 2021. Machine learning-based in-season nitrogen status diagnosis
and side-dress nitrogen recommendation for corn. Eur. J. Agron. 123, 126193
https://doi.org/10.1016/j.eja.2020.126193.
Wu, J.J., Wang, N., Shen, H.Z., Ma, X.Y., 2022. Spatial–temporal variation of climate and
its impact on winter wheat production in guanzhong plain, china. Comput. Electron.
Agric. 195, 106820 https://doi.org/10.1016/j.compag.2022.106820.
Xia, T., Miao, Y., Wu, D., Hui, S., Khosla, R., Mi, G., 2016. Active optical sensing of spring
maize for in-season diagnosis of nitrogen status based on nitrogen nutrition index.
Remote Sens. 8, 605. https://doi.org/10.3390/rs8070605.
Yang, G., Yu, W., Yao, X., Zheng, H., Cao, Q., Zhu, Y., Cao, W., Cheng, T., 2021. AGTOC:
A novel approach to winter wheat mapping by automatic generation of training
samples and one-class classication on Google Earth Engine. Int. J. Appl. Earth Obs.
Geoinf. 102, 102446 https://doi.org/10.1016/j.jag.2021.102446.
Zha, H., Miao, Y., Wang, T., Li, Y., Kusnierek, K., 2020. Improving unmanned aerial
vehicle remote sensing-based rice nitrogen nutrition index prediction with machine
learning. Remote Sens. 12, 215. https://doi.org/10.3390/rs12020215.
Zhang, C., Marzougui, A., Sankaran, S., 2020. High-resolution satellite imagery
applications in crop phenotyping: An overview. Comput. Electron. Agric. 175,
105584 https://doi.org/10.1016/j.compag.2020.105584.
Zhang, J., Liu, X., Liang, Y., Cao, Q., Tian, Y., Zhu, Y., Cao, W., Liu, X., 2019. Using a
portable active sensor to monitor growth parameters and predict grain yield of
winter wheat. Sensors 19, 1108. https://doi.org/10.3390/s19051108.
Zhang, K., Yuan, Z., Yang, T., Lu, Z., Tian, Y., Zhu, Y., Cao, Q., Liu, X., 2020. Chlorophyll
meter–based nitrogen fertilizer optimization algorithm and nitrogen nutrition index
for in-season fertilization of paddy rice. Agron. J. 112, 288–300. https://doi.org/
10.1002/agj2.20036.
Zhao, J., Zhang, X., Gao, C., Qiu, X., Cao, W., 2019. Rapid Mosaicking of Unmanned
Aerial Vehicle (UAV) Images for Crop Growth Monitoring Using the SIFT Algorithm.
Remote Sens. 11, 1226. https://doi.org/10.3390/rs11101226.
Zhao, W., Qu, Y., Zhang, L., Li, K., 2022. Spatial-aware SAR-optical time-series deep
integration for crop phenology tracking. Remote Sens. Environ. 276, 113046 https://
doi.org/10.1016/j.rse.2022.113046.
Zhou, M., Ma, X., Wang, K., Cheng, T., Wang, J., Zhu, Y., Wu, Z., Niu, Q., Gui, L., Yue, C.,
Yao, X., 2020. Detection of phenology using an improved shape model on time-series
vegetation index in wheat. Comput. Electron. Agric. 173, 105398 https://doi.org/
10.1016/j.compag.2020.105398.
J. Jiang et al.