Agricultural and Forest Meteorology 226–227 (2016) 119–131

Contents lists available at ScienceDirect

Agricultural and Forest Meteorology

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

A new wet reference target method for continuous infrared

thermography of vegetations
Wouter H. Maes a,b,∗ , Annelies Baert a , Alfredo R. Huete b , Peter E.H. Minchin c ,
William P. Snelgar c , Kathy Steppe a
a
Laboratory of Plant Ecology, Department of Applied Ecology and Environmental Biology, Ghent University, Coupure Links 653 – Bl. A, BE-9000 Ghent,
Belgium
b
University of Technology, Sydney (UTS), Climate Change Cluster, Remote Sensing, 745 Harris Street, Broadway, NSW 2007, Australia
c
Plant and Food Research Institute, 412 No 1 Road, RD 2, Te Puke, 3182, New Zealand

a r t i c l e i n f o a b s t r a c t

Article history: Although infrared thermography for stress detection in plants is popular in scientific research, it is rarely
Received 10 August 2015 used in continuous and automated applications. One of the main reasons for this is that the most precise
Received in revised form 4 May 2016 method for generating wet reference targets, used for normalizing the leaf or canopy surface temperature
Accepted 29 May 2016
for microclimatic conditions, requires manual wetting before each image capture. In this article, we
present and evaluate a new type of wet reference target that remains wet while having an energy balance
Keywords:
as similar as possible to that of the canopy. This reference target consists of a cloth knitted around a solid
Infrared thermography
frame whose shape and dimensions mimic those of the leaves. The cloth remains wet by constantly
Grapevine
Kiwifruit
absorbing water from a reservoir.
Ground-based thermal remote sensing The new reference target was evaluated on grapevine and kiwifruit plants in greenhouse and orchard
Stomatal conductance conditions. In greenhouse conditions, measured stomatal conductance was consistently more highly
CWSI correlated with the stomatal conductance index Ig when Ig was calculated with the new wet reference
rather than the manually wetted reference target. Furthermore, the temperature difference between
leaves and the new reference target remained stable for as long as measured, in contrast with the manually
wetted leaves. Ig obtained with the new reference target method was also highly correlated with stomatal
conductance (gs ) of both crops in orchard conditions.
A new empirical regression model to estimate gs from Ig in greenhouse conditions was introduced and
evaluated. This regression model incorporates the background temperature, a parameter that needs to
be included in thermographic measurements for obtaining correct surface temperatures, thus avoiding
the need for any additional measurements. The same regression model can be applied on different days
with differing conditions. The model performed better than other tested empirical models and provided
unbiased estimates of gs on days with different conditions, resulting in a root mean square error of 22–25%
of gs . Thus, it provides a promising method for continuous remote sensing of stomatal conductance or
drought stress detection of plants and vegetations.
© 2016 Elsevier B.V. All rights reserved.

1. Introduction plants and terrestrial vegetations, with applications in irrigation

Ground-based thermal remote sensing is an established method
for estimating evapotranspiration and assessing drought stress of
∗ Corresponding author at: Laboratory of Plant Ecology, Department of Applied
Ecology and Environmental Biology, Ghent University, Coupure Links 653 – Bl. A,
Ghent BE-9000, Belgium.
E-mail addresses: Wh.maes@UGent.be (W.H. Maes),
Annelies baert@hotmail.com (A. Baert), Alfredo.huete@uts.edu.au
(A.R. Huete), Peter.Minchin@plantandfood.co.nz (P.E.H. Minchin),
Bill.Snelgar@plantandfood.co.nz (W.P. Snelgar), Kathy.steppe@UGent.be
(K. Steppe).
scheduling, plant breeding and disease detection (see reviews of
Maes and Steppe, 2012 and Costa et al., 2013). These methods are
based on the linear relationship between leaf or canopy surface
temperature (Tl or Tc ) and evapotranspiration.
For most applications, a mathematical normalisation of Tl or Tc
is necessary because of the large influence of microclimatic con-
ditions on Tl or Tc (Maes and Steppe, 2012). The most successful
normalisation makes use of a minimum and maximum tempera-
ture for given conditions. In the crop water stress index (CWSI),
where this approach was first introduced, the upper and lower
temperatures corresponded to the measured canopy that is either
not transpiring (maximum temperature) or that is transpiring at

http://dx.doi.org/10.1016/j.agrformet.2016.05.021
0168-1923/© 2016 Elsevier B.V. All rights reserved.
120 W.H. Maes et al. / Agricultural and Forest Meteorology 226–227 (2016) 119–131
the maximal rate (minimum temperature). CWSI has mainly been
applied in an empirical form, in which the minimum tempera-
ture is calculated as a linear function of vapour pressure deficit,
the so-called non-water stressed baseline (NWSB). Although this
method is widely applied, particularly at orchard or field scale (e.g.
Gonzalez-Dugo et al., 2014; Sezen et al., 2014; Taghvaeian et al.,
2014; Gonzalez-Dugo et al., 2015), it requires very stable weather
conditions and depends on the availability of the NWSB for the spe-
cific crop and crop stage (Maes and Steppe, 2012; Prashar and Jones,
2014).
Thermal cameras have facilitated a new approach, in which the
minimum and maximum temperatures are estimated directly from
reference targets included within the image. This direct approach
requires no external microclimatic measurements or plant prop-
erty data, but still normalizes the thermal data for radiation,
humidity and, to a large extent, wind speed (Maes and Steppe,
2012). Different from the empirical approach, the bias of the ther-
mal sensor does not need to be corrected either (thus only requiring
corrections for emissivity and background radiation), which fur-
ther reduces errors (Prashar and Jones, 2014). The direct approach
has become a popular method for assessing drought stress or esti-
mating evapotranspiration (Maes et al., 2011; Fuentes et al., 2012;
Grant et al., 2012; Maes and Steppe, 2012; Agam et al., 2013;
Ballester et al., 2013; Agam et al., 2014; Maes et al., 2014; Pou et al.,
2014; Rud et al., 2014).
The direct approach allows calculating the direct version of
CWSI (CWSId ) or the stomatal conductance index Ig (Jones, 1999).
Although CWSId is most commonly used, it is not linearly related
to evapotranspiration or stomatal conductance (Maes and Steppe,
2012). Ig , on the other hand, is linearly related with leaf stomatal
conductance, at least for isolateral, hyper- or hypostomatous leaves
and when the correct wet reference target is used (Guilioni et al.,
2008; see Section 2.1):
 
Tdry − Tl
Ig = = Ggs (1)
(Tl − Twet )
with Tdry and Twet the surface temperature of the maximum and
minimum reference targets (K or ◦ C) and gs the stomatal conduc-
tance (mmol m−2 s−1 ). G (m2 s mmol −1 ) is a function of the air
temperature and of the boundary layer resistance to moisture (raV )
and to heat (raH ) transfer. G is independent of radiation or humidity
but changes with air temperature and wind speed. The linear rela-
tion between Ig and gs has been confirmed empirically in several
studies (Maes and Steppe, 2012).
Despite its potential, the direct approach has so far not
been applied outside scientific research, for instance in irrigation
scheduling or automated stress detection in agriculture. Probably
the most important reason for this is related to the practical aspects
of the creation of the reference targets (Jones et al., 2009; Maes
et al., 2011). Indeed, to obtain precise results, it is vital that the ref-
erence targets have leaf properties and experience microclimatic
conditions as similar to those of the leaf or canopy as possible
(Kaukoranta et al., 2005; Leinonen et al., 2006; Maes et al., 2011;
Maes and Steppe, 2012).
Developed at plant scale, Jones and his co-workers recom-
mended creating dry reference targets by covering leaves of the
plant with petroleum jelly on the stomatal side(s) and wet refer-
ence surfaces by spraying one (hypo-/hyperstomatous leaves) or
two (isolateral leaves) leaf sides with water before image capture
(Jones, 1999; Jones et al., 2002; Leinonen and Jones, 2004).
The practical problems with dry reference target leaves are
restricted to the distinction of the leaves in the image, particularly
at larger scale (Jones et al., 2009; Maes et al., 2011). In contrast,
the creation of the wet reference target leaves poses a range of
issues, which is why the focus of this article is on these wet ref-
erence target leaves. These issues are: (1) the manual spraying
hinders the whole process from being fully automated, an impor-
tant requirement for continuous application; (2) manual spraying
is labour-intensive and unpractical. Plants can be damaged and
their temperature affected by disturbance; (3) wet leaves should
be sprayed about one minute before image capture. However, it is
unclear whether this prevails in all conditions and to what extent
the method is still reliable when images are taken earlier or later;
and (4) this method cannot be applied at larger scale, where single
leaves are not distinguishable on the thermal images.
A few alternative reference targets have been used. At small
scale, Pou et al. (2014) used the temperatures of an evaporime-
ter placed at the leaf angle. Tdry is estimated as the temperature
of a thin black metal plate (5 × 1 cm), Twet as the temperature of a
black cotton wick of the same dimensions absorbing water from
a small reservoir filled with water. At larger scale, Meron et al.
(2003) presented the wet artificial reference surface (WARS) to
derive Twet , which has been used in a number of studies (Maes and
Steppe, 2012; Agam et al., 2013, 2014; Rud et al., 2014). WARS is
a water-absorbent cloth floating on a polystyrene foam board in a
water-filled tray.
As pointed out by Prashar and Jones (2014), these two alter-
natives are not fully satisfactory. As to the evaporimeter, black
surfaces may heat up many degrees above the temperatures of non-
transpiring leaves. The energy balance of both the wet and the dry
evaporimeter parts are unlikely to be representative of the energy
balance of the leaves, because of the very different boundary layer
conditions. Similarly, the energy balance of WARS is unlikely to
correspond to that of the canopy (Maes and Steppe, 2012; Prashar
and Jones, 2014). In addition, the large thermal mass of the WARS
implies that it will also have a longer thermal time constant (larger
thermal inertia) and might respond more slowly to environmental
changes than the actual canopy (Prashar and Jones, 2014).
In this article, we present and evaluate a new type of wet ref-
erence target that remains wet, thus allowing automated canopy
measurements, while having an energy balance as similar as possi-
ble to that of the canopy. This new reference target is evaluated on
two orchard crops, grapevine (Vitis vinifera L.) and kiwifruit (Actini-
dia chinensis) in indoor greenhouse and outdoor orchard conditions,
focusing on the relation between Ig and gs . For the indoor green-
house measurements, we also compare the new method with the
standard manual wetting method. Finally, we propose an empirical
method to estimate gs with Ig .
2. Material and methods
2.1. Theoretical background and development of empirical model
The most practical way to create a wet reference target with
properties similar to those of the leaves is to maintain the target
wet on both sides. The temperature of a reference target wetted on
both sides (Twet 2sides ) will be different from one that is wetted on
one side (Twet 1side ). Consequently, Ig will be different when calcu-
lated with Twet 2sides instead of Twet 1side (see Eq. (1)). In Appendix
A1, we evaluated the differences between Twet 1side and Twet 2sides
and their influence on Ig using the leaf surface temperature model
developed by Maes and Steppe (2012). This revealed that Twet 2sides
is always lower than Twet 1side and therefore that Ig 2sides is smaller
than Ig 1side (Fig. A1).
Differences in leaf properties (length, shape, spectral properties,
orientation) between the reference targets and the actual leaves
can be an important source of error in the estimation of Ig that can
(Maes et al., 2011; Maes and Steppe, 2012). Although Ig 2sides is still
influenced by differences in leaf properties between the reference
target and the actual leaf, the model simulations indicate that this
 W.H. Maes et al. / Agricultural and Forest Meteorology 226–227 (2016) 119–131
influence is much reduced when compared to Ig 1side , particularly
in the case of deviations in orientation (Fig. A1.1, second row; Fig
A1.2). Consequently, using a reference target wetted on both sides
is not only practical, but can reduce errors.
However, the relation between Ig and gs in Eq. (1) is only valid if
the reference target is only wetted on the side(s) where the stomata
are located. Else, the relation between Ig and gs is less straightfor-
ward. When reference targets wetted on both sides are used to
measure plants with hypo- or hyperstomatous leaves, as are a lot
of the economically interesting crops, the relation between Ig and
gs is given by (Guilioni et al., 2008):
 s raH rR

gs−1 = 0.92raH + I−1
g − 0.92raH
 raH + rR
In Eq. (2), gs is in units of [s m−1 ], raH is the leaf’s total boundary
layer resistance to heat transfer [s m−1 ], rR is the (virtual) leaf resis-
tance to radiative heat transfer [s m−1 ], s is the slope of saturation
vapour pressure curve at the air temperature (Pa K−1 ) and ␥ the
psychrometric constant [Pa K−1 ].
Our model simulations indicate that this Ig 2sides −gs relationship
very closely resembles a quadratic function passing through the
origin (Figs. A1.2, A1.3) or
2 within-row distances of about 1.5 m and between-row distances of
g s = aIg 2sides + bIg 2sides
The simulations furthermore show that, similar to the relation
between Ig 1side and gs , the Ig 2sides −gs relationship is influenced
by air temperature and wind speed, but very little by radiation or
relative humidity (Fig. A1.3). For a broad range of air temperatures
or wind speeds, the parameters a and b of Eq. (2) are nearly linearly
related to the air temperature or wind speed (Fig. A1.4).
In greenhouse conditions, wind speed is low and its spatial vari-
ation is hard to measure accurately. Vermeulen et al. (2012) found
that energy balance calculations inside greenhouses gave more reli-
able results when constant wind speeds were assumed than when
wind speed was measured. If constant wind speeds are indeed
assumed, the parameters a and b of Eq. (3) can be approximated
as linear functions of Ta , or
a = a0 + a1 Ta
b = b0 + b1 Ta
Combining Eqs. (3), (4) and (5) results in
2 Microclimate in both greenhouses was monitored (every minute
g s = (a0 + a1 Ta )Ig 2sides + (b0 + b1 Ta )Ig 2sides
The close linear (or quadratic) relation between gs and Ig has
often been observed for measurements taken in the course of
a few hours or one single day. However, this relation tends to
change on a consecutive day when conditions are different, which
reduces the applicability of the direct method to detect drought
stress or for irrigation steering. One way to overcome this is by
measuring all microclimatic variables and applying energy balance
models to relate gs with Ig (e.g. Leinonen et al., 2006) or even to
relate Tl directly with gs and dropping the reference target mea-
surements (e.g. Vermeulen et al., 2012). However, this requires
precise microclimatic measurements, which makes the method
error-prone in greenhouse conditions, where microclimatic con-
ditions can be highly variable (Grant et al., 2006; Kaukoranta et al.,
2005; Vermeulen et al., 2012).
Maes et al. (2011) showed that the large variability in micro-
climatic conditions in greenhouses can be largely overcome by
creating a sufficiently large number of reference target sets. Based
on the assumption of constant wind speed, Eq. (6) provides a way
to empirically relate Ig 2sides with gs for hypo- or hyperstoma-
tous leaves in greenhouse conditions. This approach only requires
Ig 2sides and Ta as input and a limited dataset of gs -measurements
to estimate the parameters a0 , a1 , b0 and b1 . Although it is inspired
121
by the model simulations of the leaf temperature model, no energy
balance model or other microclimatic measurements are needed.
We will test this empirical model and compare its performance
with linear or quadratic models not incorporating Ta (See Section
2.3).
In the rest of the article, we will work with wet reference targets
wetted on both sides and Ig calculated from this. To increase the
readability, we will therefore use the terms Twet and Ig instead of
Twet 2sides and Ig 2sides .
2.2. Greenhouse measurements
(2) 2.2.1. Experimental set-up
Greenhouse measurements were performed on grapevine
plants (Vitis vinifera L. cv. Chardonnay) (8 and 14 June 2012) and on
kiwifruit seedlings (Actinidia chinensis) (7 and 21 December 2011).
The grapevine plants were growing in the greenhouse facilities
of the Faculty of Bioscience Engineering, Ghent University, Belgium.
They were 2 years old, approximately 1.5 m tall and were grow-
ing in 50L containers (0.4 m diameter; 0.4 m height) filled with
CDM Mediterra compost. They were trained according to the single
Guyot system. The grapevines were positioned in three rows with
(3)
1 m. Measurements were done on seven plants, three plants of the
first row, two of the second and two of the third row.
The kiwifruit plants were growing in greenhouse facilities of
the Plant and Food Research Orchard in Te Puke, New Zealand. They
were one-year old Bruno seedlings grafted with Hort16A. The vines
were about 2.3 m tall, with a trunk diameter of 13.7 ± 1.3 mm and an
average leaf number of 16.7 ± 3.2 on 7 Dec 2011 and 27.0 ± 9.4 on 19
Dec 2011. Measurements were performed on six plants, positioned
next to each other in one row, with a seventh plant in the centre on
which the reference targets were created.
In both set-ups, irrigation was fully automated and none of the
plants experienced drought stress before or at the time of the mea-
surements. The kiwifruit plants were part of an experimental set-up
investigating potential infection of Pseudomonas syringae pv. actini-
diae (Psa-V), the causal agent of bacterial canker in kiwifruit, in
(4) grafts (Snelgar et al., 2012). Although the plants used in the experi-
(5) ment belonged to two different infection treatments, no differences
in plant growth or properties, gs or Ts between the treatments were
observed and treatments are not further discussed.
(6)
in the grapevine set-up and every five minutes in the kiwifruit set-
up) with a weather station located at 2 m above ground level which
included measurements of air temperature, relative humidity and
photosynthetic active radiation. No microclimatic data were avail-
able for the kiwifruit set-up on 21 December 2011 because of a
datalogger problem.
One set of reference target leaves was created in the canopy of
the central plant of the kiwifruit set-up. In the grapevine set-up,
plants were not as homogeneous in location and shape, and there-
fore a separate set of reference target leaves was created for each
of the seven plants. Each set consisted of two to three individual
leaves. The leaf size (length and width) and leaf angle of ten random
leaves of each crop were measured to derive the average and range
of leaf sizes and orientations. Each reference target leaf was fabri-
cated by knitting a cloth around a steel wire frame. The dimensions,
shapes and orientations of the frames were based on the range of
the leaf sizes and orientations of the grapevine or kiwifruit leaves
(Fig. 1). A light green, 100% cotton cloth with a thickness of 0.50 mm
and a density of 15.3 10−3 g cm−2 was used. The bottom part of the
cloth was put in a bottle filled with tap water. Acting as a wick, this
kept the imitation leaves wet for several days. The bottles were
wrapped with aluminium foil to avoid algal growth or excessive
heating by sunlight. They were filled with water the day before
122 W.H. Maes et al. / Agricultural and Forest Meteorology 226–227 (2016) 119–131
Fig. 1. Images of the wet artificial reference target leaves on grapevine plants. a: One set consisting of three artificial reference target leaves in a grapevine orchard. b,c:
Visual and thermal image of the wet artificial reference target leaves (black oval), a manually wetted reference target (light blue oval) and dry reference target leaf (red oval)
in the grapevine greenhouse. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
the measurements to make sure the temperature of the water was
equal to the greenhouse air temperature.
The manually wetted targets were created by spraying a canopy
leaf on both sides with water to which a small quantity of standard
dishwashing detergent was added, to make sure the liquid spread
evenly over the entire leaf surface (Jones, 1999; Jones et al., 2002).
The water sprayer was also filled the day before the measurements
and was wrapped in aluminium foil. The same leaves were used
for all reference target measurements. In the kiwifruit set-up, two
leaves of the central vine were used, in the grapevine set-up, one
leaf of each plant was used. Dry reference target leaves were created
by covering the lower leaf side with petroleum jelly, again two of
the central plant in the kiwifruit set-up and one leaf of each plant of
the grapevine set-up. The dry reference target leaves were created
in the close vicinity of the wet and artificial reference target leaves
and had comparable leaf properties. Cardboard crosses wrapped
with aluminium foil were used to extract background temperature
(Tbg , see further).
A dynamic porometer (AP4, DeltaT Devices, UK) was used for
all stomatal conductance measurements. The porometer was cali-
brated prior to the measurements and when required during the
measurements. Three and eight rounds of porometer measure-
ments were carried out on the grapevine plants on 8 and 14 June
2012, respectively, and six and four rounds on the kiwifruit plants
on 7 and 21 December 2011, respectively (Table 1). Each round, gs
was measured on four leaves per plant for the kiwifruit and three
leaves per plant for the grapevine set-up. The kiwifruit leaves were
taken randomly at the same height as the reference target leaves;
the grapevine leaves were marked with a black marker at the base of
the leaf before the measurements and measurements were always
taken on the same leaves. The average gs per plant and per round
was calculated and used for the analyses.
Thermal images were taken immediately prior to and after each
round of gs -measurements. All thermal measurements were made
with a FLIR SC305 thermal camera (Flir Systems, Inc.), which has an
accuracy of ±2 ◦ C and a thermal sensitivity of <0.05 ◦ C. The camera
has a built-in 18 mm lens with a field of view (FOV) of 25 × 18.8◦
and a spatial resolution of 1.38 mrad (320 × 240 pixels) which was
used for the grapevine set-up. For the kiwifruit set-up, a 10 mm
lens was mounted on the thermal camera, amplifying the field of
view to 45 × 34◦ . An in-house developed (Laboratory of Plant Ecol-
ogy, Ghent University, Belgium) robotic pan-tilt system, described
by Maes et al. (2014), was used for taking all images. Software run
on a remote laptop enabled scanning a set of pre-defined pan and
tilt angles, focusing, recording and collecting images from the ther-
mal camera and triggering a visual camera (Nikon Coolpix L23). For
the grapevine set-up, a separate image was taken automatically of
each plant using the pre-programmed pan-tilt system, which was
installed on the greenhouse frame about 3 m in front of the first row
of plants and at a height of 3.5 m. For the kiwifruit set-up, all plants
could be imaged with one shot, and the thermal and visual cam-
eras were installed on a tripod at a fixed position at 2.5 m height and
3.5 m in front of the plants. The manually wetted reference targets
were wetted about 1 min before the images were taken.
In addition to these scans, the temperature evolution of the wet
reference target created with the manually wetted and with the
new method, and the effect on Ig , were studied in the grapevine
set-up on 14 June 2012. Three trials were set up. In each trial, the
manually wetted target leaf was sprayed, after which images were
taken every 10 s during six minutes, recording the temperature of
the new and the manually wetted wet reference targets and of the
other leaves of the plant. For these measurements, the pan-tilt head
remained in a fixed position. All trials were performed between
rounds 6 and 7 (15:10-15:50).
2.2.2. Image analysis and data processing
Thermal images were processed with ThermaCam Researcher
Professional 2.10 (Flir Systems, Inc., USA) software. Settings for air
temperature (input for Reflected temperature, Atmospheric tem-
perature, Temperature of External optics) and relative humidity
were extracted from the microclimatic measurements. By setting
the emissivity to 1, the brightness temperatures (Tbr ) was cal-
culated. Using the Polygon tool of the software, the aluminium
crosses, the dry, manually wet and new wet target leaves and four
leaves (kiwifruit set-up) or three leaves (grapevine set-up) were
marked and the Tbr of each polygon was calculated, copied and
pasted to Excel 2010 (Microsoft, Ca, USA).
The surface temperature (Ts , K) of all leaves and reference tar-
gets was calculated as (Maes and Steppe, 2012):
 1/4
T4br − (1 − ε) T4bg
Ts = (7)
ε
with ␧ the emissivity and Tbg the background temperature. Tbg was
estimated as the brightness temperature of the aluminium crosses
(Maes and Steppe, 2012). A value of ␧ = 0.98 was used for all leaves
and reference targets. For the leaves and dry reference target, 0.98
was used as a good average value for a range of crops (López et al.,
2012). For the new wet target, the emissivity estimate was based
on the emissivity of dark cloth (␧ = 0.98) (Bramson, 1968, cited in
Pou et al., 2014) and the fact that (a thin layer of) water has an
emissivity of 0.98 (Pou et al., 2014). Similarly, the emissivity of the
 W.H. Maes et al. / Agricultural and Forest Meteorology 226–227 (2016) 119–131 123

Table 1
Overview of microclimatic conditions and average stomatal conductance for each measurement round on grapevine and on kiwifruit inside the greenhouse and in orchards.
For the kiwifruit measurements on 21 December 2011, no microclimatic data were available.

Round Start End PAR (mmol m−2 s−1 ) Ta (◦ C) VPD (kPa) gs (mmol m−2 s−1 )

1. Greenhouse measurements
1.1. Grapevine
8 June 2012
1 12:09 12:23 554 26.7 2.1 317 ± 151
2 13:34 13:50 358 23.5 1.6 206 ± 92
3 14:07 14:27 203 19.1 0.9 123 ± 61
14 June 2012
1 11:52 12:04 960 27.9 2.9 227 ± 99
2 12:52 13:02 1012 29.6 3.2 169 ± 65
3 13:16 13:26 656 29.4 3.2 176 ± 55
4 13:50 13:59 756 28.9 3.1 242 ± 53
5 14:29 14:39 846 30.6 3.2 152 ± 50
6 14:54 15:04 398 30.2 3.1 140 ± 41
7 16:10 16:20 329 26.2 2.4 82±27
8 16:58 17:08 406 25.4 2.6 90 ± 28
1.2. Kiwifruit
7 Dec 2011
1 10:25 10:36 614 24.3 1.8 508 ± 42
2 12:03 12:16 1289 26.4 1.9 527 ± 182
3 12:53 13:05 1327 26.7 2 584 ± 101
4 14:41 14:53 1038 27.6 2.2 617 ± 123
5 16:08 16:23 762 27.8 2.3 436 ± 57
6 17:54 18:08 296 22.6 1.2 288 ± 57
21 Dec 2011
1 15:15 15:35 – – – 392 ± 51
2 15:48 16:02 – – – 330 ± 82
3 16:20 16:38 – – – 262 ± 47
4 16:44 16:58 – – – 232 ± 32

2. Orchard measurements
2.1. Grapevine (28 June 2012)
1 13:59 14:14 1218 30.5 1.8 250 ± 54
2 14:21 14:36 1818 29.5 1.6 271 ± 67
3 16:03 16:17 1303 30.5 1.6 217 ± 73
4 16:18 16:31 1372 31.4 1.7 135 ± 21
2.2. Kiwifruit (8 December 2011)
1 13:50 14:30 2190 22.7 1.3 182 ± 36 (Inf)
364 ± 39 (Not Inf)
2 15:00 15:30 1942 22.4 1.3 204 ± 29 (Inf)
305 ± 24 (Not Inf)

Inf = Canes with Psa-infected leaves; Not Inf = canes not having any leaves with visible Psa-spotting.

manually wetted target was estimated as 0.98 because of the thin
layer of water and the emissivity of regular leaves being both 0.98.
The stomatal conductance index Ig was calculated per plant with
Eq. (1). In the grapevine set-up, Tbg , Tdry , and the surface tempera-
tures of the new target leaves (Twet new , resulting in Ig new ) or of the
manually wetted reference target leaves (Twet manual , resulting in
Ig manual ) were calculated for each plant individually. The average
leaf temperature Tl of each plant was calculated as the average Ts
of the three marked leaves. On 14 June 2012, the measured leaves
of one of the grapevine plants (plant 7) in the third row could not
be identified and Ig of this plant was not calculated. In the kiwifruit
set-up, Tbg , Tdry , Twet new and Twet manual were the same value for
all six plants, and Tl per plant was the average Ts of the four leaves.
2.3. Regression models for estimating gs with Ig,new
To evaluate the new regression model for estimating gs (See Sec-
tion 2.1), we compared its performance with that of other empirical
models. Empirical models were constructed for the greenhouse
grapevine and the greenhouse kiwifruit set-ups, aiming to assess
whether gs is best predicted by a linear or by a quadratic model and
with or without incorporating Ta . The tested models were:
- Linear models not incorporating Ta passing through the origin
(gs = a0 Ig new ) or not (gs = a0 Ig new + c0 ).
- Quadratic models not incorporating Ta passing through the origin
(gs = a0 Ig new + b0 Ig new 2 ) or not (gs = a0 Ig new + b0 Ig new 2 + c0 ).
- Linear models incorporating Ta passing through the origin
(gs = (a0 + a1 Ta ) Ig new ) or not (gs = (a0 + a1 Ta ) Ig new + c0 ).
- Quadratic models incorporating Ta passing through the origin
(gs = (a0 + a1 Ta ) Ig new + (b0 + b1 Ta ) Ig new 2 , see Eq. (6)) or not
(gs = (a0 + a1 Ta ) Ig new + (b0 + b1 Ta ) Ig new 2 + c0 ).
Regression models were constructed using the grapefruit and
kiwifruit greenhouse datasets separately. A part of the datasets was
used as input for the models (rounds 1 and 3 of 8/06, rounds 1, 3,
5, 7 and 8 of 14/06 for the grapevine dataset; rounds 1, 2, 4, 6 on
7/12 and round 2 on 21/12 for the kiwifruit dataset). The remaining
records (4 rounds of the grapevine dataset, three of the kiwifruit
dataset) were used as a dataset for model validation. Models were
built for all plants together.
The regression models were fit in Matlab R2015b (MathWorks,
Natick, USA) using the fitlm-command and occurred in two steps.
First, the regression was fit through all the points of the dataset
and studentized residuals and Cook’s distances were calculated of
all points. Data points with both studentized residuals > 2.0 (outly-
ing values) and with a Cook’s distance of > 1.0 (points with a large
influence) were removed from the dataset and the linear regres-
sion was repeated on the reduced dataset. To check whether the
intercept needed to be retained, the models were first fit by includ-
ing the intercept c0 . If c0 was not a significant factor, the analyses
124 W.H. Maes et al. / Agricultural and Forest Meteorology 226–227 (2016) 119–131
were repeated forcing the regressions through the origin. The per-
formance of the regression models was evaluated with the root
mean square error (RMSE), the adjusted R2 (R2 adj ) and by looking
at Akaike’s Final Prediction Error (FPE).
Finally, the gs values were estimated of the validation datasets
for each regression model. The performance of the models to esti-
mate the gs of the validation dataset correctly was evaluated by
looking at the correlation between the measured and the estimated
gs , by looking at the mean ± standard deviation of the residuals and
by plotting the residuals versus the measured gs .
As input for air temperature, we tested the measured air tem-
perature (only in the grapevine greenhouse experiment) and the
background temperature Tbg . The rationale behind using Tbg is that
in greenhouse conditions, Tbg is often very similar to air tempera-
ture. Furthermore, Tbg is measured close to each plant (Maes and
Steppe, 2012).
2.4. Outdoor measurements
2.4.1. Grapevine orchard measurements
Measurements were performed in a grapevine orchard on 28
June 2012 in Ghent, Belgium. The grapevine plants (Vitis vinifera
L. cv. Müller thurgau) were between 5 and 30 years old and were
trained according to the single Guyot system. Planting distance was
about 2 m between the rows and 1.5 m within the rows. The vines
were about 1.8 m tall. Sets of wet artificial leaves were installed
on five plants growing in the same canopy row. Each set consisted
of three to four artificial leaves, each with a different orientation.
A dry reference target leaf was created adjacent to or above the
wet reference target leaf. Crumpled aluminium foil was placed in
the canopy close to the plants to derive Tbg . Stomatal conductance
measurements were performed on three randomly selected leaves
(different per round) close to the reference target leaves per plant.
Four measurement rounds were carried out in the afternoon. The
average gs was calculated per plant and per round (Table 1).
The robotic pan-tilt system with thermal and visual cameras
was mounted on a tripod (Manfrotto 269HDB-3U) at a height of
five meter and standing six meter in front of the rows. Before and
after each stomatal conductance measurement round, a thermal
scan was completed.
Tl was calculated as the average temperature of five leaves of
the plant. All other temperatures were calculated as described in
Section 2.1. Per round, Ig of the scans directly before and after the
scans was averaged and compared with gs per plant and per round.
Because not all dry reference target leaves were clearly detectable
amongst the canopy, Tdry used in Eq. (1) was the average Tdry of all
dry reference target leaves.
2.4.2. Kiwifruit orchard measurements
The data presented are recalculated data originally presented in
Maes et al. (2014), where a more detailed description of the set-up,
measurements and processing can be found. Thermal and stomatal
measurements were performed in an orchard block of HORT-16A
kiwifruit in the Plant and Food Research Orchard in Te Puke, New
Zealand, on 8 December 2011. The thermal and visual cameras were
installed on the robotic pan-tilt head mounted on the 7 m-tall tripod
and the head was programmed to scan two rows of kiwifruit plants
on each side of the tripod. Vines in this orchard block were infected
with Psa-V. Six canes in the vicinity of the tripod showing typical
signs of visual Psa-V-infection and six canes without any signs of
infection were identified.
One set of new artificial wet reference target leaves was installed
on each side of the blocks. Each set consisted of four joint imitation
leaves, differing in leaf angle and size. Similarly, two sets of dry
reference target leaves were created by covering all leaves of a cane
with petroleum jelly. These canes were marked with a carbon board
frame covered with aluminium foil. In two measurement rounds,
stomatal conductance (gs ) was measured with the AP4 dynamic
porometer on three leaves of each of the Psa-infected and non-
infected canes (Table 1).
The image processing and calculation of Tl , Tdry and Twet new
were described in Maes et al. (2014). Ig was calculated using Eq. (1)
and compared to the average gs for each of the marked areas and
for each measurement round.
2.5. Statistical processing
Pearson correlations between the variables were calculated. For
greenhouse and outdoor measurements of the two crops, differ-
ences in Tl and gs between the rounds were analysed with repeated
measures ANOVA, with the measurement rounds as between-
subject factors and, in the greenhouse grapevine set-up, the plants
as within-subject factor. These statistical tests were performed
with IBM SPSS Statistics 22.0 (IBM, Armonk, NY, USA).
3. Results
3.1. Greenhouse grapevine measurements
3.1.1. Dynamic measurements: influence of time on Twet new and
Twet manual
(Tl − Twet manual ) had a similar pattern for the three trials. It
took about 30–60 s for the reference targets to cool down, indi-
cated by the blue areas in Fig. 2. This was followed by a period
in which (Tl − Twet manual ) remained more or less constant. The
length of this period differed between the trials, ranging from less
than 100 s in trial two to about 240 s in trial one (Fig. 2a–c). After
this period, (Tl − Twet manual ) decreased rapidly as the reference
target started to dry, hence warming up (red areas in Fig. 2). In
contrast, (Tl − Twet new ) remained more or less constant over the
entire period. Even during the period in which (Tl − Twet manual ) was
largely constant, Tl was slightly higher correlated with Twet new than
with Twet manual (r = 0.87 vs 0.81 in trial one, 0.82 vs 0.79 in trial two
and 0.99 vs 0.98 in trial three).
Twet manual was slightly lower than Twet new in the first trial, sim-
ilar in trial two but a few degrees higher in trial three (Fig. 2d–f).
Ig manual was lower than Ig new when Twet manual was lower than
Twet new and vice-versa (Fig. 2a–c and g–i). The pattern of Ig manual
and Ig new plotted against the time after wetting is similar during the
period when Twet manual remained constant, with Ig new following a
slightly smoother pattern than Ig manual .
3.1.2. Correlations with stomatal conductance
On both days, Twet new and Twet manual or Ig new and Ig manual did
not differ significantly and were very significantly correlated with
each other. Correlations were lower on 14 June (r = 0.94 on 8 June
and r = 0.63 on 14 June for Twet new vs Twet manual ) and the scat-
terplots of Twet manual vs Twet new or Ig manual vs Ig new show large
variation on this day (Fig. 3a,b). A more detailed overview of the
relation between Twet new vs Twet manual can be found in Appendix
A2.
If the data of all plants and all rounds are pooled, stomatal con-
ductance was more highly correlated with Ig new than with Ig manual ,
on both days (Table 2; Fig. 3c, d). On 14 June, Ig new and Ig manual of
one of the plants (plant 3) were slightly higher for a given gs than
those of the other plants (Fig. 4d). Without this plant, the overall
correlation for this day further increases to 0.75 (Ig new vs gs ) and
0.66 (Ig manual vs gs ).
If the data were pooled per measurement round, gs was more
highly correlated with Ig new than with Ig manual for two of the three
measurement rounds on 8 June (Table 2). On 14 June, correlations
were lower and only significant for one of the eight rounds, with
 W.H. Maes et al. / Agricultural and Forest Meteorology 226–227 (2016) 119–131 125

Fig. 2. a–c. The temperature difference between the average leaf temperature (Tl ) of the greenhouse grapevine plants and Twet new (black circles) and Twet manual (white
triangles) as a function of the time after spraying the manually wetted reference target. Each trial corresponds with one column. The blue area corresponds with the time
when the manually wetted reference target is still cooling down; the red area with the time when the manually wetted reference target starts warming up. d–f. Time effect
on the Twet new (black circles) Twet manual (white triangles) and on (Twet new − Twet manual ) (secondary axis; grey diamonds). g–i. The resulting Ig new (black circles) and Ig manual
(white triangles). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Table 2
Overview of correlations (r-value) between Ig manual -gs and Ig new -gs per round (data of all plants pooled) and per plant (data of all rounds pooled, only for 14 June) of the
greenhouse measurements on grapevine on 8 and 14 June 2012.

Ig manual -gs Ig new -gs Ig manual -gs Ig new -gs Ig manual -gs Ig new -gs

8 Jun 8 Jun 14 Jun 14 Jun 14 Jun 14 Jun

Round 1 0.72 0.79* 0.89* 0.38 Plant 1 0.44 0.81*

Round 2 0.85* 0.82* 0.08 −0.28 Plant 2 0.56 0.74*
Round 3 0.91** 0.92** 0.52 −0.49 Plant 3 0.03 0.81*
Round 4 0.31 0.64 Plant 4 0.75* 0.84*
Round 5 0.75 0.37 Plant 5 0.76* 0.81*
Round 6 0.80 0.81* Plant 6 0.85** 0.88**
Round 7 0.46 0.60
Round 8 −0.65 0.34

Overall 0.77** 0.85** 0.58** 0.65** Overall 0.58** 0.65**

* **
Indicates significant correlations at P <0.05. At P <0.01. To compare the new and manually wetted reference target methods, the most positive of the two correlationsis
underlined. For the correlations per round, n = 7 on 8 June and n = 6 on 14 June (n = 5 for Round 8), for the measurements per plant, n = 8 (n = 7 for Plant 1), for the overall
measurements, n = 21 on 8 June and n = 47 on 14 June.

Ig new and Ig manual performing equally well (Table 2). If the data are
pooled per plant (only on 14 June), gs was more highly correlated
with Ig new than with Ig manual for each plant (Table 2).
3.2. Greenhouse kiwifruit measurements
Twet manual and Twet new were very significantly correlated
(r = 0.97 on 7 December and r = 0.74 on 21 December) (Fig. 4a), with
Twet manual 0.8 ± 0.4 ◦ C higher than Twet new . Consequently, Ig manual
was slightly higher than Ig new (Fig. 4b).
Although correlations were not always statistically significant,
gs was more highly correlated with Ig new than with Ig manual for the
data of all plants pooled, for each round on both days (except round
2 on 21 December) and for each plant individually (only calculated
on 7 December) (Table 3 – Fig. 4c,d).
126 W.H. Maes et al. / Agricultural and Forest Meteorology 226–227 (2016) 119–131

a 24 1.5
b

1.2
22

Twet_manual (°C)
0.9

Ig_manual (-)
20
0.6

18
R²=0.89**; n=69 0.3 R²=0.66**; n=69
R²=0.39**; n=107 R²=0.40**; n=107

16 0.0
16 18 20 22 24 0.0 0.3 0.6 0.9 1.2 1.5
Twet_new (°C) Ig_new (-)
1.5
c 1.5 d

1.0
1.0

Ig (-)
Ig (-)

0.5 0.5

R²=0.71**; n=21 R²=0.42**; n=47

R²=0.62**; n=21 R²=0.34**; n=47

0.0 0.0
0 200 400 600 0 100 200 300 400
gs (mmol m-2 s-1) gs (mmol m-2 s-1)

Fig. 3. Scatterplots and linear regressions of the greenhouse grapevine set-up. Figures a and b shows the scatterplots and linear regressions of Twet manual vs Twet new and
Ig manual vs Ig new , respectively, for the two measurements days (black triangles, full lines: 8 June 2012; white triangles, dotted lines: 14 June 2012). The dashed line indicates
the 1:1 line. In figures c and d, the scatterplots of Ig new vs gs (black dots; full lines) and Ig manual vs gs (white dots; dashed lines) are given for each day separately (c 8 June
2012; d 14 June 2012). In Fig. d, the data of the (deviating) plant 3 are indicated in red (Ig new vs gs ) and orange (Ig manual vs gs ); plotted regressions include these data. For
plots a and b, all data were used. In plot c, the average Ig and gs per round and per plant were plotted. (For interpretation of the references to colour in this figure legend, the
reader is referred to the web version of this article.)

Table 3
Comparison of the correlations between Ig manualc -gs and Ig new -gs for the greenhouse kiwifruit measurements on 7 and 21 December 2011.

Ig manual -gs Ig new -gs Ig manual -gs Ig new -gs Ig manual -gs Ig new -gs

7 Dec 7 Dec 21 Dec 21 Dec 7 Dec 7 Dec

Round 1 0.32 0.43 0.44 0.70 Plant 1 0.40 0.68

Round 2 0.77 0.90** −0.06 −0.06 Plant 2 0.49 0.69
Round 3 0.40 0.48 0.05 0.41 Plant 3 0.53 0.77
Round 4 0.64 0.66 −0.02 0.75 Plant 4 0.33 0.89*
Round 5 0.33 0.34 Plant 5 0.92** 0.96**
Round 6 0.58 0.68 Plant 6 0.65 0.80

Overall 0.53** 0.76** 0.65** 0.81** 0.53** 0.76**

For the correlations per round, n = 6 (n = 5 for Round 1 on 21 Dec), for the correlations per plant, n = 6; for the overall correlations, n = 36 on 7 Dec and n = 23 on 21 Dec).
*
Indicates significant correlations at P <0.05.
**
At P <0.01. To compare the new and manually wetted reference target methods, the most positive of the two correlations with gs were underlined.

3.3. Modelling gs with Ig
For both datasets, the quadratic models incorporating Tbg fitted
the input dataset best, as they had the lowest RMSE, lowest FPE
and the highest adjusted R2 (Table 4). The modelled gs of the data
of the validation sets were most highly correlated with the mea-
sured gs when fitted with the linear model incorporating Tbg , closely
followed by the quadratic model incorporating Tbg (Table 4). The
standard deviations of the residuals of the validation model were
lowest in the models incorporating Tbg in both datasets.
Most models, in both datasets, slightly underestimated gs of
the validation sets, but this underestimation was never significant
(Table 4, right column). Nevertheless, in the grapevine dataset, the
residuals of the validation set were slightly but significantly neg-
atively correlated with the measured gs (not shown), indicating a
small overestimation of low gs values and small underestimation
of high gs values.
In the kiwifruit dataset, the models not incorporating Tbg per-
formed poorly (Fig. 5c). Incorporating Tbg in the quadratic models
improved the regression significantly, although gs was still slightly
overestimated at low values and slightly underestimated at high
values (Fig. 5d).
3.4. Outdoor measurements
In the grapevine measurements, Ig new and gs were very signif-
icantly correlated with each other (r = 0.81, P < 0.01) (Fig. 6a). The
gs -Ig regression line crossed the X-axis at 69 mmol m−2 s−1 .
 W.H. Maes et al. / Agricultural and Forest Meteorology 226–227 (2016) 119–131 127

26 3.5
b
3.0
24
2.5

Twet_manual (°C)

Ig_manual (-)
2.0
22
1.5

1.0
20
R²=0.94**; n=13 R²=0.68**; n=36
R²=0.55**; n=11 0.5 R²=0.71**; n=23

18 0.0
18 20 22 24 26 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5
Twet_new (°C) Ig_new (-)
3.5 3.5
c d
3.0 3.0

2.5 2.5

2.0 2.0
Ig (-)

Ig (-)
1.5 1.5

1.0 1.0

0.5 R²=0.58**; n=36 0.5 R²=0.66**; n=23

R²=0.28**; n=36
R²=0.43**; n=23
0.0 0.0
200 400 600 800 1000 0 200 400 600
gs (mmol m-2 s-1) gs (mmol m-2 s-1)

Fig. 4. Scatterplots and linear regressions of the greenhouse kiwifruit measurements. Figures a and b show the scatterplots and linear regressions of Twet manual vs Twet new and
Ig manual vs Ig new , respectively, for the two measurements days (black triangles, full lines: 07 December 2011; white triangles, dotted lines: 21 December 2011). The dashed
line in plots a and b indicates the 1:1 line. In figures c and d, the scatterplots of Ig new vs gs (black dots; full lines) and Ig manual vs gs (white dots; dashed lines) are given for
each day separately (c 07 December; d 21 December).

Table 4
Comparison of the empirical regression models estimating gs of the grapevine and kiwifruit datasets. The root mean square error (RMSE; mmol m−2 s−1 ), adjusted R2 (R2 adj )
and Final Prediction Error (FPE) of the regression models is given. Of the validation datasets, the correlation between the measured and the modelled gs of the validation set
(r(gs,meas ; gs,mod )) and the mean ± standard deviation of the residuals (gs,meas -gs,mod ) are given.

RMSE R2 adj FPE r (gs,meas ; gs,mod ) (gs,meas -gs,mod )

[mmol m−2 s−1 ] [mmol m−2 s−1 ]

Grapevine
Linear without Ta or Tbg 52 0.63 11.8 104 0.62** −12 ± 42
Quadratic without Ta or Tbg 47 0.49 9.9 104 0.61** −15 ± 38
Linear with Tbg 48 0.65 10.7 104 0.79** −18 ± 29
Quadratic with Tbg 44 0.74 9.6 104 0.75** −20 ± 31
Linear with Ta 51 0.66 11.5 104 0.63** −14 ± 41
Quadratic with Ta 48 0.68 11.2 104 0.56** −19 ± 39

Kiwifruit
Linear without Ta or Tbg 119 0.04 6.0 105 0.46 −18 ± 141
Quadratic without Ta or Tbg 127 0.06 7.2 105 0.41 6 ± 139
Linear with Tbg 117 0.35 6.2 105 0.86** −29 ± 99
Quadratic with Tbg 107 0.46 5.4 105 0.86** −28 ± 86
*
Indicates significant correlations at P < 0.05.
**
At P <0.01. To compare the different regression models, the best results were underlined.

In the kiwifruit measurements, Ig new was significantly lower on 4. Discussion

canes that showed symptoms of Psa-V-infection, similar to obser-
vations for CWSI or gs (Maes et al., 2014). For both measurements
rounds, the correlation between gs and Ig new was highly significant
(0.77 in round 1; 0.79 in round 2) (Fig. 6b). The overall correlation
was lower (r = 0.68), but still highly significant.
4.1. The new reference target has the potential to become a
standard method
The new wet reference target has the potential to replace the
manually wetted reference leaves at small scale or the WARS ref-
128 W.H. Maes et al. / Agricultural and Forest Meteorology 226–227 (2016) 119–131

a 500 b 500

Modelled gs (mmol m-2 s-1)

400 400

Modelled gs (mmol m-2 s-1)

300 300

200 R² = 0.37 200 R² = 0.55

n = 21
n = 21
100 100

0 0
0 100 200 300 400 500 0 100 200 300 400 500
Measured gs (mmol m-2 s-1) Measured gs (mmol m-2 s-1)
c 900 d 900
Modelled gs (mmol m-2 s-1)

Modelled gs (mmol m-2 s-1)

600 600

R² = 0.71
n = 16
300 R² = 0.12 300
n = 16

0 0
0 300 600 900 0 300 600 900
Measured gs (mmol m-2 s-1) Measured gs (mmol m-2 s-1)

Fig. 5. Measured versus modelled stomatal conductance for the grapevine greenhouse dataset (a,b) and the kiwifruit dataset (c,d), using (a,c) the linear model not incorporating
Ta or Tbg and (b,d) the quadratic model incorporating Tbg . The grey diamonds are the data of the dataset used to develop the models, the black triangles are the values of the
validation dataset. The regression line was fitted through the validation, and the number of data points and R2 of this validation is given. The dashed line represents the 1:1
line.

a 1.4 b 6
R²=0.65**; n=20 R²=0.48**; n=21
R²=0.60**; n=10
5 R²=0.63**; n=11
1.0
4
Ig (-)

Ig (-)

3
0.6
2

0.2 1

0
0 100 200 300 400 0 100 200 300 400
-0.2 gs (mmol m-2 s-1) gs (mmol m-2 s-1)

Fig. 6. Ig new -gs scatterplots of orchard measurements. Figure a: Ig new -gs scatterplot measurements in a grapevine orchard in Ghent, Belgium, on 18 June 2012. The data
shown are measurements on five different plants and four measurement rounds. Figure b: Ig new -gs scatterplot and regression lines of two measurement rounds (black circles,
dotted line: round 1; white circles, dashed line: round 2; full line = regression through all data) of a H16A kiwifruit canopy area with and without Psa-infection obtained at 8
December 2011 over an experimental orchard in Te Puke, NZ. Figure b recalculated from data of Maes et al. (2014).

erence at larger scale. We justify this with the fact that in nearly the manual wetting method, as measurements should only be per-
all situations, gs was more highly correlated with Ig new than with formed for the limited period when (Twet manual − Tleaf ) remains
Ig manual . One reason for this improvement is that Twet new is an stable (Fig. 2a–c). Otherwise, Twet and hence Ig are overestimated.
average Twet of leaves representing different leaf angles, sizes In outdoor conditions with higher wind speeds and quicker drying
and orientations, which is not easy to achieve with the manu- of the leaves, this period might even be more restricted. It appears
ally wetted or other methods. The dynamic measurements (Section that the only way to verify whether Twet manual is estimating Twet
3.1.1) showed that even during the period when (Tl − Twet manual ) correctly is by taking dynamic measurements.
was stable, Twet new was still more highly correlated with Tl and In clear contrast, the new reference method has the important
(Tl − Twet new ) was less noisy than (Tl − Twet manual ) (Fig. 2d–f). benefit of remaining continuously wet, even in demanding green-
Furthermore, the dynamic measurements revealed an impor- house conditions, avoiding overestimation of Ig . This follows from
tant risk related to Twet manual , which has received very little the dynamic measurements, where (Twet new − Tleaf ) remained sta-
attention in literature. Although our dataset was limited to three ble during the entire measurement period, from the consistent
trials, it clearly illustrates that errors may be hard to avoid with performance of Ig new in all trials, from the absence of a system-
 W.H. Maes et al. / Agricultural and Forest Meteorology 226–227 (2016) 119–131
atic bias between Twet manual and Twet new and from the fact that
(Twet new − Twet manual ) was not correlated with any of the micro-
climatic variables for any plant (See Appendix A3). Thus, the new
method can be applied for automated and continuous thermal mea-
surements. Nevertheless, it deserves further testing to see whether
the reference targets remain wet in outdoor conditions in (semi-
)arid or tropical conditions.
Another advantage of the new method is its low cost, as it is
produced from low-cost and readily available materials. The ref-
erence target leaves can easily be hand-made, and adjusted to the
required leaf shape and dimensions. Still, it is clear that a more pro-
fessional approach (finer yet solid frame, optimized cloth) could
further increase the precision and make the method available for a
wide range of crops (e.g. including grain crops).
Correlations between Ig and gs were higher when data were
aggregated per plant and per measurement day than when data
were aggregated per measurement round (Tables 2, 3). This might
seem surprising, given the larger differences during the day in Ta ,
affecting the Ig -gs relationship. However, whereas the large con-
trast in gs during the day was obvious, there was little contrast in gs
or Tl between the plants during the same round, in the absence of
drought treatments. As a consequence, within-plant standard devi-
ations in gs and Tl were of the same order (Tl ) or even higher (gs )
than between-plant standard deviations. This was the case in both
grapevine and kiwifruit set-ups, thus causing an almost unavoid-
able ‘noise’ in the dataset. The large within-plant variation is a likely
consequence of the typically high variability in irradiance and air
temperature conditions within greenhouse structures (Kaukoranta
et al., 2005; Grant et al., 2006; Vermeulen et al., 2012).
4.2. Evaluation of the empirical method estimating gs with Ig and
Tbg
The quadratic model incorporating Tbg (see Eq. (6)) as an estima-
tor of (local) air temperature clearly performed better than linear
or quadratic models not incorporating Tbg (Table 4). This confirmed
the findings of the theoretical modelling and provides an appeal-
ing new empirical method for estimating gs . Measurements of Tbg
can be done with thermal cameras and are required for accurate
estimates of Tl or Tc (Eq. (3)). This implies that no additional mea-
surements need to be made to apply this empirical method. The
other advantage of Tbg is that it is measured by the same sen-
sor as the leaves and reference targets so any bias in temperature
measurement is eliminated.
In the grapevine greenhouse measurements, the empirical mod-
els incorporating Tbg also performed much better than those
incorporating the measured air temperature, despite the high fre-
quency of the air temperature measurements (once per minute).
This again is likely due to the high variability of air temperature in
greenhouse structures. Air temperature was measured at one spot
in the middle of the experimental set-up, whereas Tbg reflects the
air temperature the individual plant is experiencing.
The RMSEs of the kiwifruit measurements were 2.4 times
higher than those of the grapevine measurements (107 vs
44 mmol m−2 s−1, Table 4). However, the overall average gs of the
kiwifruit measurements (418 mmol m−2 s−1 ) was 2.1 times higher
than that of the grapevine measurements (197 mmol m−2 s−1 ), so
the overall RMSE were comparable, about 22.3% in the grapevine
greenhouse and 25.5% in the kiwifruit greenhouse.
With more data, regression models could be developed for indi-
vidual plants. This should further decrease the RMSE of the models,
as the four parameters of Eq. (6) are a function of the boundary layer
resistance to heat transfer (raH ), a variable that is plant- (and even
leaf-) specific (Jones, 1992). Indeed, if no validation set was used
and the regression models were fit for each plant individually, the
overall RMSE of the quadratic model incorporating Tbg (still the best
129
performing model) dropped to about 20% in both datasets − at the
cost of higher FPE (not shown).
An aspect that needs to be elaborated further is the amount
of input data and the number of measurement days which are
required for establishing empirical models that can be applied to
all days within the same season, as well as the optimal conditions
of these measurement days. When we constructed the empirical
models with the data of just one day, using the data of the other
day as validation set, we obtained biased estimates for this valida-
tion set, particularly for the kiwifruit dataset (data not shown). It
is clear that measurements made over several days and in differing
conditions are required to establish reliable empirical models.
In this article, the focus was on the relation between Ig 2sides and
gs of hypo- or hyperstomatous leaves–leaves with stomata on one
side. In the case of isolateral leaves -leaves with equal stomatal con-
ductance on each leaf side– the relation between Ig 2sides and gs is
in theory linear. By analogy with the case of hypo- or hyperstoma-
tous leaves, it can be shown that the slope of this linear regression
is a function of wind speed and air temperature, and that, in case of
constant wind speed, gs can be modelled as gs = (a0 + a1 Tbg ) Ig 2sides .
In the (rare) case of plants with leaves in which the stomatal con-
ductance of both sides are not zero and are significantly different,
the relation between Ig 2sides and gs – like that between Ig 1sides and
gs – is more complex (Guilioni et al., 2008).
4.3. Remaining challenges for outdoor measurements
The relatively high correlations between gs and Ig new obtained
in the outdoor measurements illustrate that the new method devel-
oped here is also applicable at larger scales, where it can be used
as an alternative for WARS. Still, the gs -Ig new relationship for the
grapevine outdoor measurements was not as expected, as it did
not pass through the origin. This is probably caused by the difficul-
ties in distinguishing Tdry from the rest of the canopy, rather than
by issues with Twet new . Because of the large variation in canopy
leaf temperatures and the limited image resolution in the outdoor
measurements, we found that dry reference surfaces were often
very hard to distinguish from the canopy, in line with observations
in other studies (Jones et al., 2009; Maes and Steppe, 2012). Fur-
ther investigation is needed to determine whether the application
of a ‘dry version’ of the wet reference target leaf could overcome
this. This dry version would consist of the same synthetic leaves
as the wet version, but which are not absorbing any water. This
dry reference method would show similarities to the dry refer-
ence target leaves proposed by Loveys et al. (2008), although they
suggested using paper leaves and measuring with contact thermo-
couples. Using cotton, or a similar cloth material, would have the
benefit of allowing one to leave the reference surface continuously
in the field.
We did not apply an empirical model to our outdoor measure-
ments, because of the limited number of measurements made, but
also because wind speed will influence the quadratic gs -Ig relation-
ship outdoors, in a similar way as air temperature. Hence, wind
speed needs to be additionally incorporated and more parame-
ters would need to be assessed for establishing empirical models
outdoors.
However, it is uncertain whether an empirical modelling
approach can generate equivalent results for the outdoor measure-
ments as for the greenhouse datasets. Wind speeds can be highly
variable within the canopy and more research is needed to assess if
wind speed measured at one central location can accurately correct
for wind speeds at the individual plant level. In addition, for outdoor
conditions, Tbg is not a good estimator of Ta (unless in cloudy condi-
tions), and is normally much lower (Maes and Steppe, 2012). Hence,
for outdoor conditions, incorporating Tbg will probably not give
satisfactory results. On the other hand, as air temperature can be
130 W.H. Maes et al. / Agricultural and Forest Meteorology 226–227 (2016) 119–131
expected to be less variable than in greenhouse conditions, incor-
porating Ta will probably give better results for outdoor than for
indoor measurements.
4.4. Towards continuous thermal measurements
Measuring transpiration or detecting drought stress with ther-
mal imaging is not as precise as direct plant-based measurements
such as sap flow or stem diameter variation sensors, but thermal
imaging offers the important advantage that several plants can be
monitored simultaneously. To fully benefit from this advantage,
an automated, computer-controlled pan-tilt scanning system, like
the one used in this study, seems indispensable. Still, Maes et al.
(2014) discussed limitations of this method for outdoor conditions,
related to differences in resolution, viewing angle and illumination
angle between the images. In indoor (greenhouse) conditions, simi-
lar issues can be expected. However, our study illustrates that these
effects can be minimized by using a separate set of reference target
leaves for each viewing angle, in line with the recommendations
by Maes et al. (2011). The relatively good results of the regression
models showed that the aspects of viewing and illumination angle
are well corrected for. Installing the pan-tilt system on a rail system
could be a very interesting expansion of the scanning method and
would allow covering the entire greenhouse facility.
Automation of the processing routine for the thermal images
is an absolute prerequisite for successful continuous thermal
remote sensing and remains an important challenge. This includes
automatic co-registration of the visual and thermal images, iden-
tification of all reference target leaves and the canopy leaves,
conversion of brightness temperature into canopy temperature
and, if possible, separation of shaded and sunlit leaf canopy.
Despite recent progress in automated visual and thermal image
co-registration for fixed (Raza et al., 2014) or unfixed (Raza et al.,
2015; Yahyanejad and Rinner, 2015; Wang et al., 2010) visual and
thermal cameras, further research is still needed before this can be
applied commercially.
There are alternatives for our regression models in case of
continuous measurements. For instance, if the main purpose of
the thermal measurements is to detect and quantify drought
stress, advanced statistical techniques can be employed that do
not require time-consuming stomatal conductance (or other plant-
based) measurements. For instance, Baert et al. used microclimatic
data and stem diameter variation data (Baert et al., 2012) or sap
flow data (Baert et al., 2013) of grapevine plants that were not
experiencing drought stress to calibrate two statistical methods,
one an unfold principal component analysis (UPCA) and the other,
a functional unfold principal component analysis (FUPCA). When
the plants were later deprived of water, both statistical techniques
were able to detect drought stress at an early stage by signalling
when the stem diameter or sap flow was statistically deviating
from the trends projected based on the microclimatic conditions.
A similar approach could be used to link Ig to drought stress, for
both indoor and outdoor conditions, provided measurements of the
unstressed canopy can be guaranteed.
5. Conclusions
The new wet reference method can become a standard reference
method at small and large scale and can be used for continuous
measurements of Ig , because (Tl − Twet new ) remained stable during
the entire measurement period and because Ig new was more highly
correlated with gs than Ig manual . The new quadratic model incorpo-
rating Tbg as an estimator of (local) air temperature was superior in
estimating gs than other regression methods, yielding models with
a RMSE of about 22–25% of gs . Thus, this method can be applied for
continuous measurements of gs . Future research needs to focus on
the amount of input data that is required for establishing reliable
regression models.
Acknowledgements
We would like to thank two anonymous reviewers for their
useful suggestions.
WHM was previously funded by the Special Research Fund (BOF)
from Ghent University, and currently by a Marie Curie International
Outgoing Fellowship (IOF) within the 7th European Community
Framework Programme (PIOF-GA-2012-331934). WHM obtained
travel support from the Plant & Food Research (NZ) internally
funded research funds.
The thermal camera was funded by the Flemish agency for
Innovation by Science and Technology (IWT) on the Agricultural
Research (LO) project TipRelet. We would like to thank Erik Moer-
man, Geert Favyts and Philip Deman for their technical assistance
with the grapevine greenhouse set-up.
Appendix A. Supplementary data
Supplementary data associated with this article can be found, in
the online version, at http://dx.doi.org/10.1016/j.agrformet.2016.
05.021.
References
Agam, N., Cohen, Y., Berni, J.A.J., Alchanatis, V., Kool, D., Dag, A., Yermiyahu, U.,
Ben-Gal, A., 2013. An insight to the performance of crop water stress index for
olive trees. Agric. Water Manage. 118, 79–86.
Agam, N., Segal, E., Peeters, A., Levi, A., Dag, A., Yermiyahu, U., Ben-Gal, A., 2014.
Spatial distribution of water status in irrigated olive orchards by thermal
imaging. Precis. Agric. 15, 346–359.
Baert, A., Villez, K., Steppe, K., 2012. Functional unfold principal component
analysis for automatic plant-based stress detection in grapevine. Funct. Plant
Biol. 39, 519–530.
Baert, A., Villez, K., Steppe, K., 2013. Automatic drought stress detection in
grapevines without using conventional threshold values. Plant Soil 369,
439–452.
Ballester, C., Jiménez-Bello, M.A., Castel, J.R., Intrigliolo, D.S., 2013. Usefulness of
thermography for plant water stress detection in citrus and persimmon trees.
Agric. For. Meteorol. 168, 120–129.
Costa, J.M., Grant, O.M., Chaves, M.M., 2013. Thermography to explore
plant–environment interactions. J. Exp. Bot. 64, 3937–3949.
Fuentes, S., De Bei, R., Pech, J., Tyerman, S., 2012. Computational water stress
indices obtained from thermal image analysis of grapevine canopies. Irrig. Sci.
30, 523–536.
Gonzalez-Dugo, V., Zarco-Tejada, P.J., Fereres, E., 2014. Applicability and
limitations of using the crop water stress index as an indicator of water deficits
in citrus orchards. Agric. Forest Meteorol. 198–199, 94–104.
Gonzalez-Dugo, V., Goldhamer, D., Zarco-Tejada, P.J., Fereres, E., 2015. Improving
the precision of irrigation in a pistachio farm using an unmanned airborne
thermal system. Irrig. Sci. 33, 43–52.
Grant, O.M., Chaves, M.M., Jones, H.G., 2006. Optimizing thermal imaging as a
technique for detecting stomatal closure induced by drought stress under
greenhouse conditions. Physiol. Plant. 127, 507–518.
Grant, O.M., Davies, M.J., James, C.M., Johnson, A.W., Leinonen, I., Simpson, D.W.,
2012. Thermal imaging and carbon isotope composition indicate variation
amongst strawberry (Fragaria×ananassa) cultivars in stomatal conductance
and water use efficiency. Environ. Exp. Bot. 76, 7–15.
Guilioni, L., Jones, H.G., Leinonen, I., Lhomme, J.P., 2008. On the relationships
between stomatal resistance and leaf temperatures in thermography. Agric.
For. Meteorol. 148, 1908–1912.
Jones, H.G., Stoll, M., Santos, T., de Sousa, C., Chaves, M.M., Grant, O.M., 2002. Use of
infrared thermography for monitoring stomatal closure in the field:
application to grapevine. J. Exp. Bot. 53, 2249–2260.
Jones, H.G., Serraj, R., Loveys, B.R., Xiong, L.Z., Wheaton, A., Price, A.H., 2009.
Thermal infrared imaging of crop canopies for the remote diagnosis and
quantification of plant responses to water stress in the field. Funct. Plant Biol.
36, 978–989.
Jones, H.G., 1992. Plants and microclimate. In: A Quantitative Approach to
Environmental Plant Physiology. Cambridge University Press, Cambridge, UK.
Jones, H.G., 1999. Use of thermography for quantitative studies of spatial and
temporal variation of stomatal conductance over leaf surfaces. Plant Cell
Environ. 22, 1043–1055.
 W.H. Maes et al. / Agricultural and Forest Meteorology 226–227 (2016) 119–131
Kaukoranta, T., Murto, J., Takala, J., Tahvonen, R., 2005. Detection of water deficit in
greenhouse cucumber by infrared thermography and reference surfaces. Sci.
Hortic. 106, 447–463.
López, A., Molina-Aiz, F.D., Valera, D.L., Peña, A., 2012. Determining the emissivity
of the leaves of nine horticultural crops by means of infrared thermography.
Sci. Hortic. 137, 49–58.
Leinonen, I., Jones, H.G., 2004. Combining thermal and visible imagery for
estimating canopy temperature and identifying plant stress. J. Exp. Bot. 55,
1423–1431.
Leinonen, I., Grant, O.M., Tagliavia, C.P.P., Chaves, M.M., Jones, H.G., 2006.
Estimating stomatal conductance with thermal imagery. Plant Cell Environ. 29,
1508–1518.
Loveys, B.R., Jones, H.G., Theobald, J.C., McCarthy, M.G., 2008. An assessment of
plant-based measures of grapevine performance as irrigation scheduling tools.
Proceedings of the Fifth International Symposium on Irrigation of Horticultural
Crops vol. 792, 421–427.
Maes, W.H., Steppe, K., 2012. Estimating evapotranspiration and drought stress
with ground-based thermal remote sensing in agriculture: a review. J. Exp. Bot.
63, 4671–4712.
Maes, W.H., Achten, W.M.J., Reubens, B., Muys, B., 2011. Monitoring stomatal
conductance of Jatropha curcas seedlings under different levels of water
shortage with infrared thermography. Agric. For. Meteorol. 151, 554–564.
Maes, W.H., Minchin, P.E.H., Snelgar, W.P., Steppe, K., 2014. Early detection of Psa
infection in kiwifruit by means of infrared thermography at leaf and orchard
scale. Funct. Plant Biol. 41, 1207–1220.
Meron, M., Tsipris, J., Charitt, D., 2003. Remote mapping of crop water status to
assess spatial variability of crop stress. In: Stafford, J., Werner, A. (Eds.),
Precision Agriculture. Proceedings of the 4th European Conference on Precision
Agriculture, Berlin, Germany. Academic Publishers, Wageningen, pp. 405–410.
Pou, A., Diago, M.P., Medrano, H., Baluja, J., Tardaguila, J., 2014. Validation of
thermal indices for water status identification in grapevine. Agric. Water
Manage. 134, 60–72.
131
Prashar, A., Jones, H.G., 2014. Infra-red thermography as a high-throughput tool for
field phenology. Agronomy 4, 397–417.Raza, S.E.A., Smith, H.K., Clarkson, G.J.J.,
Taylor, G., Thompson, A.J., Clarkson, J., Rajpoot, N.M., 2014. Automatic
detection of regions in spinach canopies responding to soil moisture deficit
using combined visible and thermal imagery. PLoS One 9, 10.
Raza, S.E.A., Sanchez, V., Prince, G., Clarkson, J.P., Rajpoot, N.M., 2015. Registration
of thermal and visible light images of diseased plants using silhouette
extraction in the wavelet domain. Pattern Recognit. 48, 2119–2128.
Rud, R., Cohen, Y., Alchanatis, V., Levi, A., Brikman, R., Shenderey, C., Heuer, B.,
Markovitch, T., Dar, Z., Rosen, C., et al., 2014. Crop water stress index derived
from multi-year ground and aerial thermal images as an indicator of potato
water status. Precis. Agric. 15, 273–289.
Sezen, S.M., Yazar, A., Daşgan, Y., Yucel, S., Akyıldız, A., Tekin, S., Akhoundnejad, Y.,
2014. Evaluation of crop water stress index (CWSI) for red pepper with drip
and furrow irrigation under varying irrigation regimes. Agric. Water Manage.
143, 59–70.
Snelgar, W.P., Blattmann, P., Tyson, J., Curtis, C., Manning, M., 2012. Girdles can Be
Infected with Psa-V. Zespri Group Limited, Kiwifruit Journal Mount Manganui,
NZ.
Taghvaeian, S., Comas, L., DeJonge, K.C., Trout, T.J., 2014. Conventional and
simplified canopy temperature indices predict water stress in sunflower. Agric.
Water Manage. 144, 69–80.
Vermeulen, K., Aerts, J.-M., Dekock, J., Bleyaert, P., Berckmans, D., Steppe, K., 2012.
Automated leaf temperature monitoring of glasshouse tomato plants by using
a leaf energy balance model. Comput. Electron. Agric. 87, 19–31.
Wang, X.Z., Yang, W.P., Wheaton, A., Cooley, N., Moran, B., 2010. Efficient
registration of optical and IR images for automatic plant water stress
assessment. Comput. Electron. Agric. 74, 230–237.
Yahyanejad, S., Rinner, B., 2015. A fast and mobile system for registration of
low-altitude visual and thermal aerial images using multiple small-scale UAVs.
ISPRS J. Photogramm. Remote Sens. 104, 189–202.