Computers and Electronics in Agriculture 147 (2018) 109–117

Contents lists available at ScienceDirect

Computers and Electronics in Agriculture

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

Original papers

Vineyard water status estimation using multispectral imagery from an UAV T

platform and machine learning algorithms for irrigation scheduling
management
⁎
Maria Romeroa,b, Yuchen Luoc, Baofeng Sud, Sigfredo Fuentese,
a
Agricultural Engineering University of Montpellier, School of Oenology, Faculty of Wine Science, 34060, France
b
University of Turin, School of Agricultural Science, Faculty of Agriculture, Forest and Food Science, 10124, Italy
c
Hochschule Geisenheim University, D-65366, Germany
d
College of Mechanical and Electronic Engineering, Northwest A&F University, Yangling 712100, Shaanxi, China
e
University of Melbourne, School of Agriculture and Food, Faculty of Veterinary and Agricultural Sciences, VIC 3010, Australia

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

Keywords: Remote sensing can provide a fast and reliable alternative for traditional in situ water status measurement in
Remote sensing vineyards. Several vegetation indices (VIs) derived from aerial multispectral imagery were tested to estimate
Stem water potential midday stem water potential (Ψstem) of grapevines. The experimental trial was carried out in a vineyard in the
Artificial neural networks Shangri-La region, located in Yunnan province in China. Statistical methods and machine learning algorithms
Unmanned aerial vehicle
were used to evaluate the correlations between Ψstem and VIs. Results by simple regression between VIs in-
Water stress
dividually and Ψstem showed no significant relationships, with coefficient of determination (R2) for linear fitting
smaller than 0.3 for almost all the indices studied, except for the Optimal Soil Adjusted Vegetation Index
(OSAVI); R2 = 0.42 with statistical significance (p ≤ 0.001). However, results from a model obtained by fitting
using Artificial Neural Network (ANN), using all VIs calculated as inputs and real Ψstem from plants within the
study site (n = 90) as targets (Model 1), showed high correlation between the estimated water potential through
ANN (Ψstem ANN) and the actual measured Ψstem. Training, validation and testing data sets presented individual
correlations of R = 0.8, 0.72 and 0.62 respectively. The models obtained from the study site were then applied to
a wider area from the vineyard studied and compared to further Ψstem measured obtained from different sites
(n = 23) showing high correlation values between Ψstem ANN and real Ψstem (R2 = 0.83; slope = 1; p ≤ 0.001).
Finally, a pattern recognition ANN model (Model 2) was developed for irrigation scheduling purposes using the
same Ψstem measured in the study site as inputs and with the following thresholds as outputs: Ψstem below
−1.2 MPa considered as severe water stress (SS), Ψstem between −0.8 to −1.2 MPa as moderate stress (MS) and
Ψstem over −0.8 MPa with no water stress (NS). This model can be applied to analyze on a plant by plant basis to
identify sectors of stress within the vineyard for optimal irrigation management and to identify spatial variability
within the vineyards.

1. Introduction
Water is the key factor affecting quality and quantity of wine grapes.
Excessive water applications may result in increases of the vegetative
growth and yield, but other parameters associated to quality traits, such
as sugar content, acidity and pigment synthesis may be adversely af-
fected. In contrast, severe water stress will drive to partial or complete
stomatal closure, causing a reduction on the photosynthetic activity
(Van Leeuwen et al., 2009).
In regions where precipitations are scarce and concentrated in brief
periods along the year, irrigation plays a significant role to achieve
desired yield and quality. To manage irrigation scheduling there are
three questions to answer: (i) when to irrigate, (ii) how much to irrigate
and (iii) where to apply water. The latter question can be only answered
with a spatial assessment of plant water status. Water potential (Ψ) is
one of the in situ methods, that has been traditionally effective to assess
crop water status in agriculture (Smith and Mullins, 2000; Choné, 2001;
Van Leeuwen et al., 2006; Jones, 2007). However, the most accurate
measurements of Ψ can be achieved by using a pressure chamber
(Scholander et al., 1964), which applied to a large scale can be both
labor intensive and time consuming. Hence, adequate sampling and
repeatability are usually limited to a few sentinel plants and assumed to

⁎
Corresponding author.
E-mail address: sfuentes@unimelb.edu.au (S. Fuentes).

https://doi.org/10.1016/j.compag.2018.02.013
Received 12 December 2017; Received in revised form 26 January 2018; Accepted 10 February 2018
0168-1699/ © 2018 Elsevier B.V. All rights reserved.
M. Romero et al.
Fig. 1. Location of the study in the Southern of China and reference plot used for this study. Yellow dots denote the locations where plant measurements were performed in a grid for each
cultivar: Chardonnay (n = 35) in the bottom section and Cabernet sauvignon (n = 55) at the top section. (For interpretation of the references to color in this figure legend, the reader is
referred to the web version of this article.)
be representative of the whole field. These assumptionscomplicate the
implementation of efficient irrigation scheduling practices and any
other management practices focused to control or uniform yield or
quality within vineyards. The latter, mainly due to the heterogeneity of
soils and microclimate found in most of vineyards around the world.
Therefore, more efficient monitoring systems that have both, high
temporal and spatial resolutions are required.
Technological advances in remote sensing have provided non-in-
vasive, time and cost-effective techniques that can detect the spatial
variability of plant water status at a wider range of temporal scales
compared to any manual method (Acevedo-Opazo et al., 2008; Dorigo
et al., 2007). Over the last 30 years, spectral reflectance data have been
used more and more in the studies of vegetation, due to the connection
between the spectral proprieties and their biochemical and biophysical
attributes, e.g., biomass, leaf pigment contents, canopy water content,
crop coefficient and crop evapotranspiration (Penuelas et al., 1997;
Zarco-Tejada et al., 2005; Rodríguez-Pérez et al., 2007;Baluja et al.,
2012; Ferreira et al., 2012; Hunt et al., 2012; Pôças et al., 2015; Costa
et al., 2016). Spectral data are commonly used as mathematical com-
binations of two or more spectral bands, which are named as spectral
indices. The rising availability of sensors with the ability to provide
reflectance information over a spectral range, has contributed to in-
crease the interest in using vegetation indices (VI) based on the visible,
red-edge and near infrared (NIR) regions of the electromagnetic light
spectrum. The VIs based on the visible and red-edge regions have been
reported to be able to detect crop water status at the canopy level
(Rossini et al., 2013). This is based on the knowledge of the effect of
110
Computers and Electronics in Agriculture 147 (2018) 109–117
water deficit in the photosynthesis process, specially, the epoxidation
states of xanthophyll cycle pigments and the chlorophyll fluorescence
emission (Moya et al., 2004). Even though there is not water absorption
in the visible spectra domain, several VIs based on this region have
shown high correlation levels with plant water status and are com-
monly used as its proxy to improve vineyard management (Govaerts
and Verhulst, 2010; Pôças et al., 2015; Rodríguez-Pérez et al., 2007).
The large amount of data collected through remote sensing allows to
describe almost the whole plant physiology and even changes within
plants depending on the resolution of sensors. Nevertheless, the re-
sponse of the plant to different environmental changes not always
corresponds to linear relationships, for this reason sometimes tradi-
tional statistical approach are not powerful enough to obtain accurate
plant water status estimations. In this case, the use of Artificial Neural
Network (ANN) to deal with large and complex task become a good
alternative due to its ability to model both linear and nonlinear systems.
The learning performed by a network runs automatically and it is based
on the selection of appropriate value of weight (Riad et al., 2004;
Samborska et al., 2014). The ANN models have shown to be powerful
prediction tools for support decision making in the area of agriculture,
several examples could be mentioned such as: rainfall-runoff process
(Hsu et al., 1995), prediction of evaporation (Sudheer et al., 2002), soil
erosion and its relationship with leaking of nutrients in plants (Kim and
Gilley, 2008), estimation of plant water uptake (Qiao et al., 2010),
plant classification using leaf recognition (Wu et al., 2007), crop yield
prediction (Khairunniza-Bejo et al., 2014) and vegetation mapping
(Carpenter et al., 1999). Nevertheless, in terms of modelling plant water
M. Romero et al.
Table 1
Vegetation indices considered for this study with their respective equations and references.
Number Indices
1 Difference Vegetation Index
2 Green Index
3 Modified Soil Adjusted Vegetation Index
4 Normalized Difference Vegetation Index
5 Normalized Difference Greenness Vegetation Index
6 Normalized Difference Red Edge Index
7 Optimal Soil Adjusted Vegetation Index
8 Red Green Ratio Index
9 Renormalized Difference Vegetation Index
10 Simple ratio Index
response, the high input-output mapping ability performance by the
ANN has presented higher performance accuracy (R2 = 0.96) than
multiple lineal regression to predict steam water potential from soil
moisture measurements (Martí et al., 2013). Preliminary assessments
on grapevines using multispectral information obtained from UAV
platforms to assess plant water status and ANN modelling techniques
have been recently implemented (Poblete et al., 2017). However, the
latter study included as inputs for ANN modelling, specific multi-
spectral bands and not indices as predictors of stem water potential
(Ψstem), which may hinder accuracy.
The aim of this study was to assess vine water status (Ψstem) through
VIs calculated from multispectral imagery as a possible alternative so-
lution for the traditional pressure chamber measurement. Furthermore,
this study also aimed to implement machine learning modelling using
all the relevant VIs reported in previous literature as linked to plant
water status as inputs.
2. Material and methods
2.1. Vineyard site description and vine material
The research was carried out in a vineyard located in the Yunnan
province (28°33′47.91″N and 98°52′18.05″E; 2300 m.a.s.l.), China
(Fig. 1). Vineyards are dominantly on slopes and terraces, and this
makes it a unique landscape favoring the wine production style from
the region. The vineyard was originally established by local Tibetan
farmers, based on community family units. Thus, plots usually are
smaller than half hectare each. The climate type of Dêqên County can
be described as site-specific. It varies along with the change of altitude,
which contributes to the long vine growing season in the area. In
general, it is a transition between a subtropical highland climate and
humid continental climate. The mean annual total precipitations can
exceed 250 mm and more than 70% concentrates during the months of
June, July and August. The soil texture is mainly sandy clay with clay
content ranging from 10 to 25%.
The main cultivars used for modelling purposes were Cabernet
Sauvignon and Chardonnay both training on vertical shoot positioned
and planted on 2004 on their own roots. The water was supplied using
furrow irrigation with a maximum furrow width not exceeding 150 cm.
The plantation distance is mainly 1.2 m between vines and 1.5 m be-
tween rows. The area considered in this study has a total surface of
2.8 ha. One plot planted with Cabernet Sauvignon and Chardonnay
with no irrigation during the period of study was selected to be used as
a reference area to measure intensively vine water status and to gen-
erate the machine learning models.
Once the machine learning models were developed, they were ap-
plied to the whole valley, which have Cabernet Sauvignon and Merlot
cultivars trained on vertical shoot positioned system (VSP) and planted
on 2004 and 2012 respectively. The whole valley is irrigated using
furrow irrigation system with the same planting distance of 1.2 × 1.5 m
between plants and rows respectively. The whole valley vineyards have
Computers and Electronics in Agriculture 147 (2018) 109–117
Abrev. Equation Reference
DVI NIR-RED Richardson and Everitt (1992)
GI GREEN/RED Chamard et al. (1991)
MSAVI 2NIR+1-(√2(NIR+1)2-√8(NIR-RED))/2 Qi et al. (1994)
NDVI (NIR-RED)/(NIR+RED) Rouse et al., (1973)
NDGI (NIR-GREEN)/(NIR+GREEN) Gamon and Surfus (1999)
NDRE (NIR-REG)/(NIR+REG) Barnes et al. (2000)
OSAVI (NIR-RED)/(NIR+RED+0.16) Rondeaux et al. (1996)
RGRI RED/GREEN Gamon and Surfus (1999)
RDVI (NIR-RED)/√NIR+RED Roujean and Breon (1995)
SR NIR/RED Birth and McVey (1968)
a total area of 4.7 ha.
2.2. Plant water status and UAV-based spectral field measurements
Vine water status was assessed by Ψstem measured using two
Scholander pressure bomb (PSM 600 D. Albany, OR, USA) between
12:00 and 14:00, after two hours after covering the leaves and no
transpiration adaptation inside plastic/aluminum foil bags. The phy-
siological measurements were conducted on 90 vines distributed within
a grid in the measurement plot which were considered as observed data
(Ψstem obs) for modelling purposes. Each measured vine was separated
by three adjacent vines as buffer to attain maximum variability within
the data for modelling purposes. The experiment was carried out at
véraison during the 2016 growing season in three dates: October 7th,
October 22nd and November 4th.
The reflectance data of the whole experimental vineyard were col-
lected using a UAV platform (Phantom 3 professional, DJI Co. Ltd,
Shenzhen, China) equipped with a SEQUOIA multispectral sensors
(Parrot Co. Ltd, Paris, France). The flight height was 30 m above ground
level, which allowed obtaining imagery at 4 cms spatial resolution in
the VIS, GREEN, RED, RED EDGE and NIR region. Imagery was ac-
quired at 13:00. for all the bands in one manual flight per day of
measurement.
2.3. Image processing, data analysis and machine learning modelling
The full size raw multispectral images were processed using Pix4d
mapper 3.0.1 (Pix4D S A, Switzerland) to obtain the ten vegetation
indices calculated (Table 1). The geometric correction and resampling
at 10 cm spatial resolution were performed using Quantum GIS Geo-
graphic Information System 2.14.11 (Qgis, Free Software Foundation,
Boston, USA) using one measurement as a control point for the cali-
bration of the remaining dates. The pure canopy pixel extraction was
performed for each index and date from plants measured for midday
Ψstem using System for Automated Geoscientific Analyses version 3.0.0
(SAGA, Free Software Foundation, Boston, USA). From these data, ten
VIs, which incorporate the GREEN, RED, RED EDGE and NIR region
were calculated (Table 1). To compare Ψstem with the VIs calculated
from data collected from the aerial imagery, linear and curvilineal
models were fitted to the data using Microsoft Excel 2016 (Microsoft,
USA, Washington).
The machine learning modelling method used to process multi-
spectral and physiological data (Model 1) was implemented using the
Matlab Neural Network Toolbox™ 10 (Mathworks Inc., Matick, MA,
USA). All the ten VIs calculated were considered as inputs and midday
Ψstem obs as targets for all plants measured in all measurement dates
(n = 270). Model 1 (Fig. 2A) was an ANN fitting model constructed
using 10 neurons for the hidden layer, data division was performed
automatically by dividing the data into three stages: (i) training, which
is necessary to compute the gradient and to update the weights and
biases, (ii) validation, which is needed to minimize the error and avoid
111
M. Romero et al.
Fig. 2. (A) Model 1, considering a two-layer feedforward network with ten hidden neu-
rons and sigmoid functions to fit regression models for ten vegetation indices (VIs) as
inputs and stem water potential (Ψstem) as target; (B) Model 2 for pattern recognition
considering also ten VIs as inputs and three water stress categories: Non-stressed (NS)
vines (Ψstem ≤ 0.8 MPa), moderate stressed (MS) vines (0.8 MPa ≥ Ψstem ≤ 1.2 MPa) and
severe stressed (SS) vines (Ψstem ≥ 1.2 MPa). For hidden and output layers, w = weights
and b = biases.
overfitting by monitoring the validation error in the training stage, and
(iii) testing, which is necessary to compare different models. Specifi-
cally, a random data division function in Matlab® was used, which
randomly divides the data from samples into the three stages with 60%
(n = 162) for training, 20% (n = 54) for validation with a mean
squared error performance algorithm and 20% (n = 54) for testing with
a default derivative function, the data used in each stage is not used in
the other two, which makes every stage independent from each other.
Model 2 was constructed to train a pattern recognition neural net-
work model (Fig. 2B). The objective of the model was the classification
of water status levels based on ten VIs measured as inputs (Table 1) into
the three levels of vine water stress as targets, namely: Non-stressed
(NS) vines (Ψstem ≤ 0.8 MPa), moderate stressed (MS) vines
(0.8 MPa ≥ Ψstem ≤ 1.2 MPa) and severe stressed (SS) vines
(Ψstem ≥ 1.2 MPa). The ANN are capable of finding patterns in an in-
ductive way through algorithms based on the pre-existing data Beale
et al. (2010). The ANN used in this study was a standard two-layer
feedforward network with a sigmoid transfer function in the hidden
layer and a linear transfer function in the output layer structure con-
formed using the 10 components (VIs) obtained using the UAV
(Table 1) as the input layer and three water stress levels as the targets.
The selected number of neurons in the hidden layer was ten as depicted
in Fig. 2B. The validation and testing data were randomly divided using
a random data division function in 60% (n = 162) for training using a
scale conjugate gradient training function, 20% (n = 54) for validation
using a cross-entropy performance function and 20% (n = 54) for
Fig. 3. Spatial distribution of Stem water potential (MPa) in the study plot during the three dates using kriging interpolation techniques from all Ψstem
(n = 90).
112
Computers and Electronics in Agriculture 147 (2018) 109–117
testing using a default derivative function (Mathworks Inc. Matick, MA.
USA).
2.4. Statistical analysis
To assess and compare the accuracy and performance of the ANN
Model 1 obtained, statistical data such as the determination coefficients
(R2), standard error of estimates (SEE) and root mean square error
(RMSE) were obtained for the linearized model output. There was no
outlier analysis for any of the fitting models presented, hence models
include all the data (n = 90) from replicates for Ψstem measurements
from the three dates, which sum up a total of 270 data. The statistical
significance was calculated using Turkey HDS range test.
2.5. Assessment of model performances at the pixel by pixel, plant by plant
and vineyard scales
Once VIs were calculated for all the study area, multispectral image
processing algorithms were applied to obtain a concatenated matrix
cube of size (M ∗ N ∗ VIL). In which M represent the number or pixel
rows within spectral images, N the number of columns and VIL the
vegetation index image layers considered (n = 10). Models 1 and 2
were then applied to process pixel by pixel from the M,N,VIL cube using
a customized loop code to extract Ψstem per pixel (Model 1) and nu-
merical values corresponding to the three water stress levels considered
(Model 2), as 1 = NS, 2 = MS and 3 = SS. Values below 1 were ob-
tained in sections with no spectral indices available.
The plant by plant assessment was performed after applying Model
2 to the M,N,VIL cube using the method proposed by (Su et al., 2016)
which recognizes the grapevines rows and separate each plant in a box
corresponded to the plantation space, which is 1.2 m between plants
and 1.5 m between rows. This procedure allows the automatic calcu-
lation of data within each region corresponding to each plant. Data
extraction was obtained as the mode water stress threshold value in the
middle section of each plant section. Once having the processed matrix
representing water stress threshold values per plant, a mesh grid spline
interpolation (based on a cubic spline using a not-a-knot end condi-
tions) technique was used to map spatial differences of thresholds. The
interpolated value at specific analysis points is based on a cubic inter-
polation on the values at neighboring grid points for each respective
dimension (Mathworks Inc., Matick. MA. USA).
obs measurement points per date
M. Romero et al.
3. Results and discussion
3.1. Vine water status spatial assessment
The spatial distribution of Ψstem obs of 90 sampled vines for three
measurement dates are shown in Fig. 3. The range of Ψstem obs values
found were from −0.5 MPa to −1.5 MPa, showing that the temporal
and spatial distribution of plant water status were from non-stress to
severe stress levels. In the first measurement date (October 7th), most of
the area was not stressed, largely due to irrigation received just five
days before. Only 20% of measurements showed Ψstem obs values lower
than −1.0 MPa, mainly concentrated in the southwestern edge of this
plot (green to red color). The second measurement date (October 22nd)
showed that vines with Ψstem obs lower than −1.0 MPa reached 30%
and the stressed zone was extended from the edge to the center. In the
last measurement date (November 4th), around 40% of vines showed
Ψstem obs lower than −1.0 MPa, hence the stressed area increased sig-
nificantly.
Several studies (Choné, 2001; Williams and Araujo, 2002; Van
Leeuwen et al., 2009) have successfully demonstrated the accuracy of
midday Ψstem to assess plant water status in vineyards. Besides, similar
results have been obtained in others crops, such as pear, citrus and
apple. Van Beek et al. (2013) showed a decrease of Ψstem under dif-
ferent water status regime in pear orchards. However, there is an im-
portant source of variability of Ψstem among the vines of the same row
and area due to heterogeneity of soil and microclimate, which could
lead to errors in irrigation management decisions when selecting sen-
tinel vines that are not representative for the majority of plants for each
plot. Van Leeuwen et al. (2006) also showed high variability in vine to
vine water status in the same plot. On the other hand, the measurement
could be influenced by several factors such as microclimate changes
which affect evapotranspiration of plants, position and condition of the
sampling leaf, petiole cut, among others (Turner, 1988).
The pressure bomb method to measure plant water status is very
accurate as described before, however it offers low spatial resolution in
practical terms. This method is restricted to sentinel plants for man-
agement purposes and it is still considered an expensive and time-
consuming, since it requires trained personnel to operate a heavy in-
strument (chamber plust N2 gas cylinder and attachments).
3.2. Vine water status assessment by UAV-based multispectral imagery
3.2.1. Analysis of the different predictors for vine water status
Relationships between vine water status and data retrieved from
multispectral imagery and respective VIs calculated from Ψstem obs
monitored plants are presented in Table 2. Linear and polynomial re-
lationships were found between the VIs calculated and Ψstem obs, from
which polynomial fits had a better performance. Nevertheless, the R2
for most of the indices in this study were low and not statistically sig-
nificant, which fluctuated between R2 = 0.09–0.49 and in most cases
Table 2
Statistical analysis showing the correlations and equations between the different vegetation indices studied and Ψstem.
Index Equation (Ψstem=)
NDVI 2 a
20.824x - 40.015x + 26.531
NDGI 27.945x2 - 15.941x + 9.5242
NDRE 11.969x2 - 17.259x + 10.793
GI 1.012x2 - 5.8603x + 14.154
RGRI 12.283x2 - 16.031x + 12.777
SR 0.0308x2 − 0.9932x + 14.446
RDVI −3.6x2 + 6.8144x + 5.9953
OSAVI 52.796x2 - 71.675x + 31.293
MSAVI 14.42x2 - 24.391x + 17.078
DVI 34.37x2 - 35.221x + 15.681
a
Correspond to the respective index value; bSignificance level: *p ≤ 0.05;
Computers and Electronics in Agriculture 147 (2018) 109–117
not higher than R2 = 0.25. The statistical analysis showed significant R
and R2 between Ψstem obs and NDGI, GI and OSAVI indices. In parti-
cular, OSAVI showed the best performance, R2 = 0.42 (p ≤ 0.001).
OSAVI has been described by Rondeaux et al. (1996) as an adjusting
factor for the SAVI family of indices, comparing to some typical indices
like NDVI or DVI. OSAVI has enhanced sensitivity to vegetation cover
and soil brightness. Moreover, it allows to minimize the NDVI sensi-
tivity to variation in soil background; for this reason, it is used in zones
with spare vegetation where soil is visible through the canopy and also
where NDVI saturates due to high vegetation density. In this study,
OSAVI was the index that best correlated with Ψstem obs, one of the
reasons could be related to the high vine density and the narrow leaf
area in the Southwestern zone of the plot, which favors the saturation of
NDVI value. Indexes, such as NDVI, NDRE, NDGI or DVI were expected
to be somehow better correlated with plant water status as many au-
thors have previously reported (Carter, 1993; Penuelas et al., 1997;
Rodríguez-Pérez et al., 2007; Eitel et al., 2011; Baluja et al., 2012;
Ferreira et al., 2012; Calderón et al., 2013; Van Beek et al., 2013; Zarco-
Tejada et al., 2013; Pôças et al., 2015). However, poor correlations and
high SEE for all the indices were in contradiction to these studies, which
limited the possibility of using a standard spectral wavelength for-
mulation or a single VI for real-time prediction of water status.
To verify if there was any improvement of the correlations under
different stress levels, the data collected were divided by measurement
dates. Results in Table 3 showed that most of the VIs are poorly or not
well correlated with Ψstem obs among the three dates. Besides, there was
no significant improvements in correlations. It should be pointed out
that in the last measurement date (November 4th), which corresponded
to severe water stress in considerable parts of the reference plot, a few
VIs showed a significant correlation (p ≤ 0.01) with Ψstem, however as
it was mentioned before, the R2 was considerably lower. According to
these results presented GI, RDVI, OSAVI, MSAVI and DVI corresponded
to the indices that could predict better long-term water stress. More-
over, GI and OSAVI displayed an increasing tendency of the correlation
improvements over time. In general, it is possible to say that Red and
NIR spectral region could serve as long-term predictors of water stress.
Several authors (Ferreira et al., 2012; Jones, 2007; Pôças et al.,
2015; Williams and Araujo, 2002) have suggested the relevance of pre-
dawn water potential (Ψpw) as a better indicator of water status due to
the equilibrium between plant and soil. However, Pôças et al. (2015)
using optimized VI by narrow bands reported poor correlation with Ψpw
and indices that incorporate information between Red and NIR region
(R2 ≤ 0.6). Nevertheless, in the same study, better correlations were
found with indices in the visible domain spectrum like GI, NDGI and
RGRI with R2 ranging from 0.77 to 0.82. Furthermore, evidence has
been presented using night-time transpiration sensors (sap flow) that
have demonstrated that in Mediterranean conditions eucalyptus
(Fuentes et al., 2007), almond (Fuentes et al., 2013) and grapevines
(Fuentes et al., 2014) open stomata at night-time and significant water
loss can be achieved, which makes difficult or even impossible the
R (n = 270) R2 b
(n = 270) SEE RMSE
−0.43 0.20 2.61 2.57
−0.40 0.18* 2.65 2.61
−0.13 0.05 2.87 2.80
−0.37 0.14** 2.69 2.67
0.39 0.20 2.67 2.58
−0.40 0.27 2.65 2.46
−0.16 0.05 2.86 2.81
−0.60 0.42*** 2.32 8.02
−0.43 0.25 2.60 2.50
−0.41 0.23 2.64 2.53
**
p ≤ 0.01; p ≤ 0.001; x denoted the specific index used for the fitting.
***
113
M. Romero et al. Computers and Electronics in Agriculture 147 (2018) 109–117

Table 3 set and calculated VIs of further random plants with respective Ψstem
Statistics analysis of correlations (R) between different vegetation indexes and Ψstem obs measured from another vineyard plot (n = 28). The linear regression
(MPa) for the three dates studied.
between Ψstem obs and Ψstem ANN, as expected had a significant
Index R (n = 90) (p ≤ 0.001) and high correlation (R = 0.83; slope = 1.0). In both cases,
the residual analyses showed independence, normality, and homo-
7-Oct 22-Oct 4-Nov geneity of the variance (data not shown). These results also showed that
more than 74% of the variance in the Ψstem obs can be explained by the
NDVI −0.54 −0.42 −0.35
NDGI −0.42 −0.27 −0.38 ANN model. A possible explanation for the good ANN results could be
NDRE −0.13 −0.07 −0.12 associated to the wavelength range in which each index is calculated.
GI −0.29 −0.36 −0.35** Most of the indices in the present research are built by bands in the Red
RGRI 0.41 0.50 0.46 and NIR spectra, commonly associated to physiological changes in
SR −0.40 −0.55 −0.40
plants (Kazmierski et al., 2011; Rossini et al., 2013; Gamon et al., 2015;
RDVI −0.56 −0.67 −0.60**
OSAVI −0.63 −0.59 −0.62** Pôças et al., 2015). In addition, Green and Red edge bands have also
MSAVI −0.50 −0.45 −0.37** been mentioned as water status indicators in vegetative tissue.
DVI −0.47 −0.43 −0.35** Zygielbaum et al. (2009) demonstrated the strong correlation of
greenness measure with crop water status, mainly because water stress
manifests an increase in green reflectance. Instead, Red edge also has
equilibrium between plant and soil water potentials, hence rendering
been associates to the chlorophyll content and considered as a possible
inaccuracies in the use of Ψpw in these conditions. All the latter studies
predictor of crop water status (Holer et al., 1983; Fillela and Penuelas,
were conducted under pressurized irrigation systems (mainly drip ir-
1994; Pinar and Curran, 1996).
rigation) in which non-uniform wetting patterns are achieved. This can
In fact, the wavelengths and their respectively indices used in this
be the same case with furrow irrigation systems used in this study,
study have the potential to be predictor of water status, but at the same
which result in non-uniform wetting patterns.
time each index presents some problems associated to reflectance and
On the other hand, similar research done in citrus groves by Waldo
radiance, e.g. DVI does not deal with difference between soil and ve-
and Schumann (2009) showed significant correlation between Ψstem
getation in shady area; NDVI could present oversaturation in high
and NDVI (R2 = 0.31). Furthermore, Rallo et al. (2014) reported good
density vineyards and does not eliminate atmospheric effect; SR reduces
prediction of Ψleaf in olive orchards using optimized indices like NDGI
the effects of atmosphere and topography and OSAVI reduces problems
and GI.
of saturation (Carter, 1993; Rondeaux et al., 1996). In this case, it is
possible to associate the ANN results to a compensation effect. That is to
3.2.2. Estimating Ψstem based on artificial neural networks (ANN) say, once the ten VI are combined, using ANN as a tool to evaluate the
The ANN uses a two-layer feedforward network and non-linear influence of each index over the target can significantly improve the
methods to assess patterns between the inputs and targets producing a estimation of Ψstem. Furthermore, even though some VIs indexes did not
linearized model for simplification (Model 1). All the VIs used as inputs correlate well with Ψstem obs it is not required to discard them for the
showed positive and linear/non-linear correlations. Model 1 showed a ANN model construction, since these modelling techniques can identify
strong and statistically significant performance when comparing ob- the pattern of the data and individual weights of inputs and fit the ANN
served stem water potential (Ψstem obs) versus estimated stem water model accordingly.
potential (Ψstem ANN). Specifically, the ANN regressions (Table 4), Nowadays, due to great computing power advancements, modelling
showed high correlations between Ψstem obs and the Ψstem ANN using ANN requires ubiquitous personal computers. The modelling
(R = 0.88) for the training data (n = 176). Additionally, the random technique can be performed by consultancy services and research in-
data used for the validation and testing also presented high correlations stitutions that have access to these software platforms, such as Matlab®.
with Ψstem (R = 0.73 and 0.62 respectively). The final linearized neural The product obtained by this technique for growers would be a software
fit regression combines the results obtained from the training, valida- package containing the ANN modelling algorithms to process auto-
tion and testing data. As expected, a high correlation was observed matically information from UAV multispectral imagery for irrigation
between Ψstem obs and Ψstem ANN (R = 0.84; slope = 1.0). It is important management purposes. This will require normal computer skills level
to note that the final model is not linear, since the linearization is a for users.
simplification of the complex model procedures to assess accuracy only.
The final model considers sigmoidal functions as presented in Fig. 2, 3.2.3. Using Ψstem estimation to manage vineyard irrigation
which represent most of processes found in nature. Overall, the analysis of canopy reflectance showed high potential
Model 1 was tested a second time, besides the model performance for Ψstem estimation. As an illustration on how Model 1 could be used
assessment presented before, to predict Ψstem using a multispectral data for irrigation management, Fig. 4A shows the vineyard map using Ψstem
ANN from Model 1 (RMSE = 0.15 MPa) compared to Ψstem obs (Fig. 4B).

Table 4 Model 1 rendered adequate accuracy, which is important for irrigation

Neural fitting using Leveng-Marquard algorithm (Model 1) with VIs as inputs and Ψstem as scheduling management, considering the criteria to determine levels of
target. Pattern recognition model (Model 2) using VIs as inputs and three water status stress. As previously presented, the criteria has as minimum 0.2 MPa
classification levels as targets: Non-stressed (NS) vines (Ψstem ≤ 0.8 MPa), moderate (Non-stressed (NS) vines (Ψstem ≤ 0.8 MPa), moderate stressed (MS)
stressed (MS) vines (0.8 MPa ≥ Ψstem ≤ 1.2 MPa) and severe stressed (SS) vines
(Ψstem ≥ 1.2 MPa). Na = non-applicable.
vines (0.8 MPa ≥ Ψstem ≤ 1.2 MPa) and severe stressed (SS) vines
(Ψstem ≥ 1.2 MPa). Hence the error of Model 1 is lower than this cri-
Stage Samples MSE R Accuracy (%) terion (0.15). The map created through the ANN Model 1 estimations
shows several zones in the vineyard requiring different irrigation
Training 162 2.68 0.88 Na
Validation 54 3.35 0.73 Na
management strategies. According to literature review, Ψstem below
Testing 54 3.68 0.62 Na −1.5 MPa should be considered as a cut-off between moderate and
Model 1 270 3.24 0.84 Na severe water stress (Choné, 2001; Van Leeuwen et al., 2009). However,
Training 162 Na Na 83.5 these thresholds could be considered as more appropriate to cultivars
Validation 54 Na Na 82.9
such as Shiraz (Acevedo-Opazo et al., 2010).
Testing 54 Na Na 82.9
Model 2 270 Na Na 83.3 For the present study, Ψstem below −1.2 MPa (orange to red colors)
were considered as cut off between moderate and severe water stress,

114
M. Romero et al.
Fig. 4. Interpolated map of stem water potential (MPa) for reference plot on 22nd November: (A) Ψstem
Fig. 5. Stem water potential (MPa) estimate on a pixel by pixel basis through the ANN
Model 2 for the Adong vineyard. Non-stressed (NS) pixels (Ψstem ≤ 0.8 MPa), moderate
pixels (MS) vines (0.8 MPa ≥ Ψstem ≤ 1.2 MPa) and severe stress (SS) pixels
(Ψstem ≥ 1.2 MPa). The yellow broken lined oval represents the Cabernet sauvignon and
Merlot plots considered in Fig. 6. (For interpretation of the references to color in this
figure legend, the reader is referred to the web version of this article.)
115
Computers and Electronics in Agriculture 147 (2018) 109–117
ANN map estimated through ANN model 1 and (B) observed Ψstem obs.
which have been shown to be more appropriated for cultivars such as
Chardonnay (Williams and Araujo, 2002). Furthermore, Ψstem between
−0.8 and −1.2 MPa (light blue to green color) corresponded to mod-
erate stress and finally Ψstem over −0.8 MPa (blue color) was con-
sidered with no water stress. Some studies done in Cabernet sauvignon
and Chardonnay considered plants without water stress when Ψstem was
over −0.6 MPa, instead when Ψstem fell to −1.2 MPa vines are con-
sidered under water deficit, which had a positive effect in the devel-
opment of aroma in cultivars such a Chardonnay (Deluc et al., 2009).
Using the previous classification and the ANN Model 2, it was possible
to create an irrigation management map for the whole vineyard. Fig. 5,
shows a vineyard water status classification at the pixel by pixel scale
using Model 2. Three main groups were used: NS vines (blue color), MS
(green to yellow color), SS (red color). By obtaining regular Ψstem maps
throughout the growing irrigation season using the methodology pro-
posed in this paper, irrigation scheduling using UAV-based remote
sensing becomes more practical for the assessment of spatial variability
of water status.
Finally, Fig. 6 shows an example of Model 2 applied to Cabernet
Sauvignon and Merlot plots (yellow oval from Fig. 5) using a plant by
plant data extraction procedure (Fig. 6A) and spline interpolation
techniques (Fig. 6B). The latter figure shows two defined spatial sectors
of NS, and SS with some plants surrounding SS at MS level. From
Fig. 6B, it can be seen a high spatial variability of vine water status,
which can give information to growers for site specific management
strategies. These management strategies can be focused to either
manage variability or to maintain a certain wine style, which many
times defines a region.
4. Conclusion
The use of multispectral sensors onboard of a UAV platform allowed
obtaining high-resolution images to monitor vine water status, which
has the potential to replace the traditional pressure chamber method,
especially for vineyards in extreme water stress conditions. Machine
learning methods (ANN) demonstrated to be a powerful tool to mining
data from multi-dimensional maps. As it was shown in this study, once
the ANN models were trained using different VIs, high correlations
were found between the estimates (Ψstem ANN) and actual Ψstem obs
(Model 1). A pixel by pixel data information availability per plant may
be useful for physiological studies and more in-depth irrigation prac-
tices, such as identification of localized irrigation problems (broken
emitters or flooded furrows). However, this level of information offers
M. Romero et al.
A B
Fig. 6. Stem water potential (MPa) estimated on a plant by plant basis (A) using Model 2 and interpolated by spline interpolation techniques (B) for sections of cultivars Cabernet
Sauvignon (top block) and Merlot (bottom block) for the Adong vineyard. Map B show mainly two sections of non-stressed (NS) vines (Ψstem ≤ 0.8 MPa, blue), and severe stress (SS) vines
(Ψstem ≥ 1.2 MPa, yellow) with some of the latter section surrounded with vines at moderate stress (MS) vines (0.8 MPa ≥ Ψstem ≤ t1.2 MPa, green). (For interpretation of the references
to color in this figure legend, the reader is referred to the web version of this article.)
little help for more practical irrigation scheduling strategies based on
irrigation blocks. To address the latter problem, Model 2 showed to
perform accurately in the identification of regions within vineyards of
low, medium and high-water stress, which offer spatial information to
growers for optimal irrigation scheduling practices and to identify
strategies to manage this variability more efficiently. This study con-
firmed that an UAV platform equipped with multispectral sensors and
machine learning modelling strategies could be an efficient and af-
fordable option to assess vine water status and to map vineyard spatial
variability to optimize irrigation management. Other applications such
as biomass calculation, weeds monitoring and disease detection using
these techniques are worthy of further investigation. Nevertheless,
problems associated to irradiance and reflectance still need to be in-
vestigated for other potential applications.
Appendix A. Supplementary material
Supplementary data associated with this article can be found, in the
online version, at http://dx.doi.org/10.1016/j.compag.2018.02.013.
References
Acevedo-Opazo, C., Tisseyre, B., Ojeda, H., Guillaume, S., 2010. Spatial extrapolation of
the vine (Vitis vinifera L.) water status: a first step towards a spatial prediction model.
Irrig. Sci. 28, 143.
Acevedo-Opazo, C., Tisseyre, B., Ojeda, H., Ortega-Farias, S., Guillaume, S., 2008. Is it
possible to assess the spatial variability of vine water status? J. Int. des Sci. la Vigne
du Vin. 42, 203–220.
Baluja, J., Diago, M.P., Balda, P., Zorer, R., Meggio, F., Morales, F., Tardaguila, J., 2012.
Assessment of vineyard water status variability by thermal and multispectral imagery
using an unmanned aerial vehicle (UAV). Irrig. Sci. 30, 511–522. http://dx.doi.org/
10.1007/s00271-012-0382-9.
Barnes, E., Clarke, T., Richards, S., Colaizzi, P.D., Haberland, J., Kostrzewski, M., Waller,
P., Choi, C., Riley, E., Thompson, T., Lascano, R.J., Li, H., Moran, M.S., 2000.
Coincident Detection of Crop Water Stress, Nitrogen Status and Canopy Density Using
Ground-Based Multispectral Data. Proc. Fifth Int. Conf. Precis. Agric. [CD Rom].
Beale, M.H., Hagan, M.T., Demuth, H.B., 2010. Neural Network ToolboxTM User’s Guide.
MathWorks, Inc., MathWorks.
Birth, G.S., McVey, G.R., 1968. Measuring the Color of Growing Turf with a Reflectance
Spectrophotometer.
Calderón, R., Navas-Cortés, J.A., Lucena, C., Zarco-Tejada, P.J., 2013. High-resolution
airborne hyperspectral and thermal imagery for early detection of Verticillium wilt of
olive using fluorescence, temperature and narrow-band spectral indices. Remote
Sens. Environ. 139, 231–245. http://dx.doi.org/10.1016/j.rse.2013.07.031.
Carpenter, G.A., Gopal, S., Macomber, S., Martens, S., Woodcock, C.E., Franklin, J., 1999.
A neural network method for efficient vegetation mapping. Remote Sens. Environ.
70, 326–338. http://dx.doi.org/10.1016/S0034-4257(99)00051-6.
Carter, G.A., 1993. Responses of leaf spectral reflectance to plant stress. Am. J. Bot. 80,
239–243. http://dx.doi.org/10.2307/2445346.
116
Computers and Electronics in Agriculture 147 (2018) 109–117
Stress Level
3 (SS)
2.5
2.0 (MS)
1.5
1 (NS)
0.5
0
Chamard, P., Courel, M.F., Docousso, M., Guénégou, M.C., LeRhun, J., Levasseur, J.,
Togola, M., 1991. Utilisation des bandes spectrales du vert et du rouge pour une
meilleure évaluation des formations végétales actives In Télédétection et
Cartographie. AUPELF-UREF, Sherbrooke, QC pp. 203–209.
Choné, X., 2001. Stem water potential is a sensitive indicator of grapevine water status.
Ann. Bot. 87, 477–483. http://dx.doi.org/10.1006/anbo.2000.1361.
Costa, J.M., Vaz, M., Escalona, J., Egipto, R., Lopes, C., Medrano, H., Chaves, M.M., 2016.
Modern viticulture in southern Europe: Vulnerabilities and strategies for adaptation
to water scarcity. Agric. Water Manag. 164, 5–18. http://dx.doi.org/10.1016/j.
agwat.2015.08.021.
Deluc, L.G., Quilici, D.R., Decendit, A., Grimplet, J., Wheatley, M.D., Schlauch, K.A.,
Mérillon, J.-M., Cushman, J.C., Cramer, G.R., 2009. Water deficit alters differentially
metabolic pathways affecting important flavor and quality traits in grape berries of
Cabernet Sauvignon and Chardonnay. BMC Genomics 10, 212. http://dx.doi.org/10.
1186/1471-2164-10-212.
Dorigo, W.A., Zurita-Milla, R., de Wit, A.J.W., Brazile, J., Singh, R., Schaepman, M.E.,
2007. A review on reflective remote sensing and data assimilation techniques for
enhanced agroecosystem modeling. Int. J. Appl. Earth Obs. Geoinf. 9, 165–193.
http://dx.doi.org/10.1016/j.jag.2006.05.003.
Eitel, J.U.H., Vierling, L.A., Litvak, M.E., Long, D.S., Schulthess, U., Ager, A.A., Krofcheck,
D.J., Stoscheck, L., 2011. Broadband, red-edge information from satellites improves
early stress detection in a New Mexico conifer woodland. Remote Sens. Environ. 115,
3640–3646. http://dx.doi.org/10.1016/j.rse.2011.09.002.
Ferreira, M.I., Conceição, N., Silvestre, J., Fabião, M., 2012. Transpiration and water
stress effects on water use, in relation to estimations from NDVI: application in a
vineyard in SE Portugal. Options Méditerranéennes. Ser. B Stud. Res. 67, 203–208.
Fillela, I., Penuelas, J., 1994. The red edge position and shape as indicators of plant
chlorophyll content, biomass and hydric status. Int. J. Remote Sens. http://dx.doi.
org/10.1080/01431169408954177.
Fuentes, S., De Bei, R., Collins, M.J., Escalona, J.M., Medrano, H., Tyerman, S., 2014.
Night-time responses to water supply in grapevines (Vitis vinifera L.) under deficit
irrigation and partial root-zone drying. Agric. Water Manag. 138, 1–9. http://dx.doi.
org/10.1016/j.agwat.2014.02.015.
Fuentes, S., Mahadevan, M., Bonada, M., Skewes, M.A., Cox, J.W., 2013. Night-time sap
flow is parabolically linked to midday water potential for field-grown almond trees.
Irrig. Sci. 31, 1265–1276. http://dx.doi.org/10.1007/s00271-013-0403-3.
Fuentes, S., Palmer, A.R., Taylor, D., Zeppel, M., Whitley, R., Eamus, D., 2007. An au-
tomated procedure for estimating the leaf area index (LAI) of woodland ecosystems
using digital imagery, MATLAB programming and its application to an examination
of the relationship between remotely sensed and field measurements of LAI. 5th Int.
Work. Funct. Struct. Plant Model 1070–1079. http://dx.doi.org/10.1071/fp08045.
Gamon, J.A., Kovalchuck, O., Wong, C.Y.S., Harris, A., Garrity, S.R., 2015. Monitoring
seasonal and diurnal changes in photosynthetic pigments with automated PRI and
NDVI sensors. Biogeosciences 12, 4149–4159. http://dx.doi.org/10.5194/bg-12-
4149-2015.
Gamon, J.A., Surfus, J.S., 1999. Assessing leaf pigment content and activity with a re-
flectometer. New Phytol. 143, 105–117. http://dx.doi.org/10.1046/j.1469-8137.
1999.00424.x.
Govaerts, B., Verhulst, N., 2010. The normalized difference vegetation index (NDVI)
Greenseeker (TM) handheld sensor: toward the integrated evaluation of crop man-
agement. Part A-Concepts and case studies. Int. Maize Wheat Improv. Cent. 1–12.
Holer, D.N.H., Dockray, M., Barber, J., 1983. The red edge of plant leaf reflectance. Int. J.
Remote Sens. 4, 273–288. http://dx.doi.org/10.1080/01431168308948546.
Hsu, K., Gupta, H.V., Sorooshian, S., 1995. Artificial neural network modeling of the
rainfall-runoff process. Water Resour. Res.
Hunt, E.R., Doraiswamy, P.C., McMurtrey, J.E., Daughtry, C.S.T., Perry, E.M., Akhmedov,
M. Romero et al.
B., 2012. A visible band index for remote sensing leaf chlorophyll content at the
Canopy scale. Int. J. Appl. Earth Obs. Geoinf. 21, 103–112. http://dx.doi.org/10.
1016/j.jag.2012.07.020.
Jones, H.G., 2007. Monitoring plant and soil water status: established and novel methods
revisited and their relevance to studies of drought tolerance. J. Exp. Bot. 58,
119–130. http://dx.doi.org/10.1093/jxb/erl118.
Kazmierski, M., Glemas, P., Rousseau, J., Tisseyre, B., 2011. Temporal stability of within-
field patterns of ndvi in non irrigated mediterranean vineyards. J. Int. des Sci. la
Vigne du Vin. 45, 61–73.
Khairunniza-Bejo, S., Mustaffha, S., Wan Ismail, I., 2014. Application of artificial neural
network in predicting crop yield: a review. J. Food Sci. Eng. 705, 283–291. http://dx.
doi.org/10.1016/j.aca.2011.06.033.
Kim, M., Gilley, J.E., 2008. Artificial Neural Network estimation of soil erosion and nu-
trient concentrations in runoff from land application areas. Comput. Electron. Agric.
64, 268–275.
Martí, P., Gasque, M., González-Altozano, P., 2013. An artificial neural network approach
to the estimation of stem water potential from frequency domain reflectometry soil
moisture measurements and meteorological data. Comput. Electron. Agric. 91,
75–86. http://dx.doi.org/10.1016/j.compag.2012.12.001.
Moya, I., Camenen, L., Evain, S., Goulas, Y., Cerovic, Z.G., Latouche, G., Flexas, J., Ounis,
A., 2004. A new instrument for passive remote sensing. Remote Sens. Environ. 91,
186–197. http://dx.doi.org/10.1016/j.rse.2004.02.012.
Penuelas, J., Pinol, J., Ogaya, R., Filella, I., 1997. Estimation of plant water concentration
by the reflectance Water Index WI (R900/R970). Int. J. Remote Sens. 18, 2869–2875.
http://dx.doi.org/10.1080/014311697217396.
Pinar, A., Curran, P.J., 1996. Technical Note Grass chlorophyll and the reflectance red
edge. Int. J. Remote Sens. 17, 351–357. http://dx.doi.org/10.1080/
01431169608949010.
Poblete, T., Ortega-Farías, S., Moreno, M., Bardeen, M., 2017. Artificial neural network to
predict vine water status spatial variability using multispectral information obtained
from an unmanned aerial vehicle (UAV). Sensors 17, 2488. http://dx.doi.org/10.
3390/s17112488.
Pôças, I., Rodrigues, A., Gonçalves, S., Costa, P.M., Gonçalves, I., Pereira, L.S., Cunha, M.,
2015. Predicting grapevine water status based on hyperspectral reflectance vegeta-
tion indices. Remote Sens. 7, 16460–16479. http://dx.doi.org/10.3390/rs71215835.
Qi, J., Chehbouni, A., Huete, A.R., Kerr, Y.H., Sorooshian, S., 1994. A Modified Soil
Adjusted Vegetation Index 126, 119–126. http://dx.doi.org/10.1016/0034-4257(94)
90134-1.
Qiao, D.M., Shi, H.B., Pang, H.B., Qi, X.B., Plauborg, F., 2010. Estimating plant root water
uptake using a neural network approach. Agric. Water Manag. 98, 251–260.
Rallo, G., Minacapilli, M., Ciraolo, G., Provenzano, G., 2014. Detecting crop water status
in mature olive groves using vegetation spectral measurements. Biosyst. Eng. 128,
52–68. http://dx.doi.org/10.1016/j.biosystemseng.2014.08.012.
Riad, S., Mania, J., Bouchaou, L., Najjar, Y., 2004. Rainfall-runoff model usingan artificial
neural network approach. Math. Comput. Model. 40, 839–846. http://dx.doi.org/10.
1016/j.mcm.2004.10.012.
Richardson, A.J., Everitt, J.H., 1992. Using spectral vegetation indices to estimate ran-
geland productivity. Geocarto Int. 7, 63–69. http://dx.doi.org/10.1080/
10106049209354353.
Rodríguez-Pérez, J.R., Riaño, D., Carlisle, E., Ustin, S., Smart, D.R., 2007. Evaluation of
hyperspectral reflectance indexes to detect grapevine water status in vineyards. Am.
J. Enol. Vitic. 58, 302–317.
Rondeaux, G., Steven, M., Baret, F., 1996. Optimization of Soil-Adjusted Vegetaion
Indices.
Rossini, M., Fava, F., Cogliati, S., Meroni, M., Marchesi, A., Panigada, C., Giardino, C.,
Busetto, L., Migliavacca, M., Amaducci, S., Colombo, R., 2013. Assessing canopy PRI
from airborne imagery to map water stress in maize. ISPRS J. Photogramm. Remote
Sens. 86, 168–177. http://dx.doi.org/10.1016/j.isprsjprs.2013.10.002.
117
Computers and Electronics in Agriculture 147 (2018) 109–117
Roujean, J.L., Breon, F.M., 1995. Estimating PAR absorbed by vegetation from bidirec-
tional reflectance measurements. Remote Sens. Environ. 51, 375–384. http://dx.doi.
org/10.1016/0034-4257(94)00114-3.
Rouse, J.W., Hass, R.H., Schell, J.A., Deering, D.W., 1973. Monitoring vegetation systems
in the great plains with ERTS. Third Earth Resour. Technol. Satell. Symp. 1, 309–317.
https://doi.org/citeulike-article-id:12009708.
Samborska, I.A., Alexandrov, V., Sieczko, L., Kornatowska, B., Goltsev, V., Cetner, M.D.,
Kalaji, H.M., 2014. Artificial neural networks and their application in biological and
agricultural research. J. NanoPhotoBioSciences 2, 14–30.
Scholander, P.F., Hammel, H.T., Hemmingsen, E.A., Bradstreet, E.D., 1964. Hydrostatic
pressure and osmotic potential in leaves of mangroves and some other plants. Proc.
Natl. Acad. Sci. 52, 119–125.
Smith, K. a, Mullins, C.E., 2000. Soil and Environmental Analysis. doi: 10.1201/
9780203908600.
Su, B.F., Xue, J.R., Xie, C.Y., Fang, Y.L., Song, Y.Y., Fuentes, S., 2016. Digital surface
model applied to unmanned aerial vehicle based photogrammetry to assess potential
biotic or abiotic effects on grapevine canopies. Int. J. Agric. Biol. Eng. 9, 119–130.
http://dx.doi.org/10.3965/j.ijabe.20160906.2908.
Sudheer, K.P., Gosain, A.K., Mohana Rangan, D., Saheb, S.M., 2002. Modelling eva-
poration using an artificial neural network algorithm. Hydrol. Process. 16,
3189–3202.
Turner, N.C., 1988. Measurement of plant water status by the pressure chamber tech-
nique. Irrig. Sci. 9, 289–308. http://dx.doi.org/10.1007/BF00296704.
Van Beek, J., Tits, L., Somers, B., Coppin, P., 2013. Stem water potential monitoring in
pear orchards through worldview-2 Multispectral Imagery. Remote Sens. 5,
6647–6666. http://dx.doi.org/10.3390/rs5126647.
Van Leeuwen, C., Goutouly, J., Azaïs, C., Marguerit, E., Roby, J., Chone, X., Germain, C.,
2006. Intra-block variations of vine water status in time and space Variations intra-
parcellaires temporelles et spatiales du régime hydrique de la vigne. Vine 64–69.
Van Leeuwen, C., Tregoat, O., Chon??, X., Bois, B., Pernet, D., Gaudill??re, J.P., 2009.
Vine water status is a key factor in grape ripening and vintage quality for red bor-
deaux wine. How can it be assessed for vineyard management purposes? J. Int. des
Sci. la Vigne du Vin 43, 121–134. http://dx.doi.org/10.20870/oeno-one.2009.43.3.
798.
Waldo, L., Schumann, A., 2009. Alternative methods for determining crop water status for
irrigation of citrus groves. Soil Water 63–71.
Williams, L.E., Araujo, F.J., 2002. Correlations among Predawn Leaf, midday leaf, and
midday stem water potential and their correlations with other measures of soil and
plant water status in Vitis vinifera. J. Am. Soc. Hortic. Sci. 127, 448–454. http://dx.
doi.org/10.1016/j.yqres.2009.07.010.
Wu, S.G., Bao, F.S., Xu, E.Y., Wang, Y.-X., Chang, Y.-F., Xiang, Q.-L., 2007. A leaf re-
cognition algorithm for plant classification using probabilistic neural network, in:
Signal Processing and Information Technology, 2007 IEEE International Symposium
on. IEEE, pp. 11–16.
Zarco-Tejada, P.J., Berjón, A., López-Lozano, R., Miller, J.R., Martín, P., Cachorro, V.,
González, M.R., de Frutos, A., 2005. Assessing vineyard condition with hyperspectral
indices: Leaf and canopy reflectance simulation in a row-structured discontinuous
canopy. Remote Sens. Environ. 99, 271–287. http://dx.doi.org/10.1016/j.rse.2005.
09.002.
Zarco-Tejada, P.J., González-Dugo, V., Williams, L.E., Suárez, L., Berni, J.A.J.,
Goldhamer, D., Fereres, E., 2013. A PRI-based water stress index combining struc-
tural and chlorophyll effects: assessment using diurnal narrow-band airborne imagery
and the CWSI thermal index. Remote Sens. Environ. 138, 38–50. http://dx.doi.org/
10.1016/j.rse.2013.07.024.
Zygielbaum, A.I., Gitelson, A.A., Arkebauer, T.J., Rundquist, D.C., 2009. Non-destructive
detection of water stress and estimation of relative water content in maize. Geophys.
Res. Lett. 36. http://dx.doi.org/10.1029/2009GL038906.