Content uploaded by Kiryn Dobbie
Author content
All content in this area was uploaded by Kiryn Dobbie on Nov 22, 2023
Content may be subject to copyright.
Phytopathology®r2023 r113:1405-1416 rhttps://doi.org/10.1094/PHYTO-02-23-0078-R
Bioinformatics and Computational Plant Pathology
Previsual and Early Detection of Myrtle Rust on Rose Apple Using Indices
Derived from Thermal Imagery and Visible-to-Short-Infrared Spectroscopy
Michael S. Watt,1,†Michael Bartlett,2Julia Soewarto,2Dilshan de Silva,2Honey Jane C. Estarija,2Peter Massam,2
David Cajes,2Warren Yorston,2Elizaveta Graevskaya,2Kiryn Dobbie,2Stuart Fraser,2
Heidi S. Dungey,2and Henning Buddenbaum3
1Scion, 10 Kyle St., Christchurch 8011, New Zealand
2Scion, P.O. Box 3020, Rotorua 3010, New Zealand
3Environmental Remote Sensing and Geoinformatics, Trier University, 54286 Trier, Germany
Accepted for publication 5 April 2023.
Abstract
Myrtle rust, caused by the fungus Austropuccinia psidii, is a serious
disease, which affects many Myrtaceae species. Commercial nurseries
that propagate Myrtaceae species are prone to myrtle rust and require a
reliable method that allows previsual and early detection of the disease.
This study uses time-series thermal imagery and visible-to-short-infrared
spectroscopy measurements acquired over 10 days from 81 rose apple
plants (Syzygium jambos) that were either inoculated with myrtle rust or
maintained disease-free. Using these data, the objectives were to (i) quantify
the accuracy of models using thermal indices and narrowband hyperspectral
indices (NBHI) for previsual and early detection of myrtle rust using data
from older resistant green leaves and young susceptible red leaves and
(ii) identify the most important NBHI and thermal indices for disease
detection. Using predictions made on a validation dataset, models using
indices derived from thermal imagery were able to perfectly (F1 score =
1.0; accuracy =100%) distinguish control from infected plants previsually
one day before symptoms appeared (1 DBS) and for all stages after early
symptoms appeared. Compared with control plants, plants with myrtle
rust had lower and more variable normalized canopy temperature, which
was associated with higher stomatal conductance and transpiration. Using
NBHI derived from green leaves, excellent previsual classification was
achieved3DBS,2DBS,and1DBS(F1scorerange=0.89 to 0.94). The
accurate characterization of myrtle rust during previsual and early stages
of disease development suggests that a robust detection methodology could
be developed within a nursery setting.
Keywords: biosecurity, disease screening, nursery, regularized discriminant
analysis
The fungus Austropuccinia psidii (basionym Puccinia psidii,
Sphaerophragmaciaceae, Pucciniales), is native to South and
Central America (Beenken 2017; Glen et al. 2007). Commonly
known as myrtle rust (MR), eucalyptus rust, or ohi’a rust, the dis-
ease affects a wide range of plants in the family Myrtaceae. Spores
of A. psidii can be transported over long distances, particularly via
wind and anthropogenic transport (e.g., the nursery trade and gar-
den plants), which has allowed the pathogen to be widely spread.
As reviewed by Carnegie and Pegg (2018), it has been reported
over the past two decades in North America (Florida, California,
and Hawaii), East and Southeast Asia (Japan, China, Indonesia, and
Singapore), Australia, New Caledonia, South Africa, and most re-
cently New Zealand. Combining observations of symptoms in the
field with laboratory inoculations, known hosts of A. psidii comprise
at least 480 Myrtaceae species worldwide (Soewarto et al. 2019),
which makes this pathogen a major threat to native forest ecosys-
tems dominated by Myrtaceae, as well as myrtle-related industries.
ThepandemicstrainofA. psidii was first detected at a nurs-
ery in mainland New Zealand in 2017 (Ho et al. 2019), and the
pathogen rapidly spread across urban and forest environments on the
North Island of New Zealand (Toome-Heller et al. 2020). Symptoms
†Corresponding author: M. S. Watt; michael.watt@scionresearch.com
Funding: Support was provided by Scion Strategic Science Investment Fund (SSIF).
e-Xtra: Supplementary material is available online.
The author(s) declare no conflict of interest.
Copyright © 2023 The Author(s). This is an open access article
distributed under the CC BY 4.0 International license.
have been recorded in at least 30 native and nonnative Myrtaceae
taxa. Despite concerted responses to manage MR, the pathogen has
spread through almost all of its predicted climatic range (Beresford
et al. 2018), including regions of the South Island (Tasman, Marl-
borough, and Canterbury), and has recently reached the Chatham
Islands, which lie approximately 800 km east of the South Island.
A. psidii is considered an “unwanted organism” in New Zealand
under the Biosecurity Act of 1993. The act requires commercial
nurseries to follow approved hygiene measures, including use of
fungicides and visual assessment of plants by nursery staff for the
appearance of disease symptoms, to manage any risk of spreading
the pathogen. Visual inspection to identify MR symptoms requires
training, is time-consuming, may be unreliable or prone to biases,
and relies on disease occurring at the stage where it is visually de-
tectable. Treating plants prior to the emergence of any symptoms
is highly desirable to minimize plant and financial losses, as plants
with symptoms are unable to be dispatched. However, such a treat-
ment program requires an automated and highly reliable detection
method that can be used for previsual and early identification of
infected plants.
The description of visible signs and symptoms and the disease
incubation period has been documented for a range of suscepti-
ble hosts (Coutinho et al. 1998; Lee et al. 2015; Pegg et al. 2014;
Rayachhetry et al. 1997; Roux et al. 2016; Soewarto et al. 2021).
The descriptions have been supplemented by the cytological work
of Xavier et al. (2001) and Yong et al. (2019), which describes
in more detail the process by which A. psidii infects and devel-
ops inside eucalypt hosts. The pathogen infects younger actively
growing tissues of susceptible hosts, including new leaves, stems,
flowers, and fruits, while older tissues develop an ontogenic resis-
tance to infection (Beresford et al. 2020). The infection process first
Vol. 113, No. 8, 2023 1405
involves adhesion of urediniospores to the surfaces of susceptible
tissue. Under optimum conditions (15 to 25°C, darkness, and >90%
humidity), urediniospores swell and develop a germ tube between
6 and 24 h after inoculation. Within 48 h, the germ tube forms a
specialized structure called an appressorium, which produces an in-
fection peg that directly penetrates the host through the cuticle and
epidermis. Within the host tissue, intercellular colonization by the
pathogen starts with the formation of hyphae, followed by haustoria
to draw nutrients directly from the mesophyll cells. Externally, the
first visible symptoms of infection can appear as distortion, blister-
ing, and small purple/red lesions and necrotic and chlorotic spots.
Eventually the hyphae growing in susceptible tissue give rise to ure-
dinia or telia that erupt from the leaf epidermis. The urediniospores
or teliospores impart the characteristic “rusty and powdery” aspect
of MR. The length of time for development of symptoms and signs
is influenced by temperature and host (Beresford et al. 2020). Un-
der optimal conditions, initial visual symptoms typically occur from
4 to 7 days after inoculation, and the time to eruption of the first
uredinia or telia (latent period) ranges from 5 to 10 days. As the in-
fection progresses, pustules enlarge and merge, often causing leaf
distortion. At 12 days postinoculation, heavy infection can result in
leaf and stem dieback.
Conventional diagnosis of A. psidii relies on host identification,
observation of signs and symptoms, and microscopic examination
of rust pustules and spores (Roux et al. 2013). However, because
morphological identification relies on the presence of pustules and
spores, morphological identification of A. psidii basedonearly
symptoms prior to pustule development is not possible. In ad-
dition, close visual inspection of large volumes of Myrtaceae in
a nursery setting may not always be practical or cost-effective.
DNA-based detection techniques involve extracting and amplify-
ing target DNA from infected plant tissue to identify the presence
of the fungus from diverse tissue types, including symptomless or
cryptically contaminated plant tissue. A nested polymerase chain
reaction (PCR) (Langrell et al. 2008) and quantitative PCR (qPCR)
assays (Baskarathevan et al. 2016; Bini et al. 2018) have been de-
veloped. The nested PCR approach was specific to A. psidii,butis
time-consuming and requires two rounds of PCR. The qPCR assays
enabled faster detection of the pathogen and high-throughput test-
ing and had higher accuracy than the conventional PCR method.
However, while it is possible to detect the pathogen in asymp-
tomatic tissue using these methods, they can be expensive. As
these methods sample a relatively small amount of tissue, it is
likely that the pathogen may not be detected when sampling from
asymptomatic plants. Consequently, these methods are much better
applied following visual detection of symptoms to guide sampling.
Data from hyperspectral and thermal imagery provide an alter-
native method of previsual and early disease detection that can
be spatially scaled for rapid screening. Plants that are subject to
disease react using a variety of protection mechanisms that can
include changes in photosynthesis, stomatal conductance, transpi-
ration, and photosynthetic pigment content (Hernández-Clemente
et al. 2019). The changes may be detected by shifts in leaf tem-
perature, or spectrally through changes in key pigments related to
the photosynthetic process, such as chlorophyll, carotenoids, an-
thocyanins, and xanthophylls (Hernández-Clemente et al. 2019).
Hyperspectral and thermal imagery have been used to character-
ize many of these traits at the plant level, and research using these
methods has demonstrated previsual or early detection of a range
of diseases in agricultural and forestry crops (Calderón et al. 2015;
Hornero et al. 2021; López-López et al. 2016; Mahlein et al. 2012;
Zarco-Tejada et al. 2018).
Despite progress, few studies have used hyperspectral and ther-
mal imagery for previsual and early detection of MR. A recent study
was undertaken to detect MR on a plantation of lemon myrtle (Back-
housia citriodara) in Australia (Heim et al. 2019). Using a novel
spectral index obtained from spectroradiometer data (ranging from
350 to 2,500 nm), classification of fungicide-treated and untreated
(diseased) leaves had an overall accuracy of 90%. However, the
study focused on detection of relatively advanced symptoms, rang-
ing from small purple spots to large necrotic lesions and yellow
pustules, rather than previsual and early detection of the disease.
Rose apple (Syzygium jambos (L.) Alston) is known to be one
of the hosts most susceptible to A. psidii and is severely damaged
by the disease (Tessmann et al. 2001). Originating from Southeast
Asia, it is planted as an ornamental or fruit tree in the tropics and in
New Zealand, but is regarded as invasive in some regions outside
of its native range. Rose apple develops into a small to medium-
sized tree, and leaves are opposite, oblong-lanceolate, 7.5 to 20 cm
long and 3.4 to 7.8 cm wide, shiny red/purple when growing, but
turn dark green as they age. Young leaf tissues can become heavily
diseased with MR, producing large amounts of inoculum before
becoming blighted and deformed. Therefore, rose apple is an ideal
model host species for this pathosystem and is often used to mass-
produce inoculum and as a positive control for infection in artificial
inoculation studies.
This study examines the accuracy of indices derived from thermal
imagery and visible-to-short-infrared spectroscopy to detect MR
on juvenile rose apple plants growing in a controlled environment.
The objectives of this research were to (i) quantify the accuracy of
models using thermal indices and narrowband hyperspectral indices
(NBHI) for previsual (prior to symptom development) and early
detection (following symptom development) of MR using data from
both older resistant green leaves and young susceptible red leaves
and (ii) identify the most important NBHI and thermal indices for
disease detection.
Materials and Methods
Experimental procedure
The experiment was undertaken in a temperature-regulated dark
room with artificial lighting used for growing plants. A total of 84
healthy rose apple plants were selected that were visually free from
any signs of infection. The plants were originally sourced from seed
collected in the Auckland area and propagated as cuttings during
April 2021 in a nursery at Scion (Rotorua, New Zealand) under
containment. Plants were grown in 2 liter polybags, ranged in height
from 20 to 60 cm, and were watered as needed by hand over the
course of the experiment.
On 27 October 2022, plants were randomly allocated to either an
untreated control group or a MR treatment group. The groups were
separated into plastic enclosures with misting capability. The 32
control plants were placed in a single enclosure arranged in a 3 ×
11 grid. Plants in the MR treatment group were randomly assigned
to one of three identical enclosures and arranged in 3 ×6grids
within each enclosure (18 plants in two enclosures and 16 in the
third). The positions of plants within each enclosure were random-
ized. No differences between the three enclosures were noted in
the timing of symptom development among MR plants. All plants
were maintained at 22°C, 70 to 80% relative humidity, and a 16 h
photoperiod under 30 W LED 3,000 K grow lights. Plants were
only removed from enclosures briefly to take measurements. Prior
to inoculation, all plants were visually assessed for MR symptoms
on 31 October and 2 and 4 November. Two plants in the control
treatment developed minor symptoms on 4 and 6 November and
were removed from the experiment. Another plant was removed
from the MR treatment group when it developed minor symptoms
on 6 November, which was too soon for symptom development fol-
lowing inoculation. Thus, there were 51 plants in the MR treatment
group and 30 in the control group.
Plants in the MR treatment were inoculated on 4 November. In-
oculum used in the experiment was mass produced on rose apple
from a single pustule isolate maintained in the New Zealand Na-
tional Forestry Fungarium and Culture Collection (NZFRIM 6037).
The isolate was originally collected from Lophomyrtus ×ralphii in
Rotorua, in 2019. Urediniospores were harvested using a portable
1406 PHYTOPATHOLOGY®
vacuum pump (Mini Cyclone Spore Collector, Tallgrass Solutions,
KS) into 00 gelatin capsules, desiccated for 48 h using silica beads,
andthenstoredat−80°C. Prior to inoculation, spores were allowed
to warm to room temperature and then suspended in 0.05% Tween
20 in sterile distilled water at a concentration of 1 ×105spores/ml,
determined using a hemocytometer. Inoculum was applied evenly
to both sides of young leaves and stems using a gravity-fed air-
brush powered by a compressor at 2 bar pressure. The plants were
returned to the enclosure and maintained in the dark at 18°C and
>90% relative humidity for the following 24 h to maintain opti-
mal conditions for infection. Control plants were mock inoculated
with sterile 0.05% Tween 20solution, and identical conditions were
maintained to ensure that the environmental change did not differ-
entially affect plant physiology between treatments. Conditions for
both the control and MR treatment groups were returned to those
previously described for the duration of the experiment.
Measurements were taken from both treatments between 2 and
11 November 2022. A baseline preinoculation set of measurements
was taken on 2 November. Starting on 6 November, postinoc-
ulation measurements were repeated daily for 6 days, until 11
November. The measurements encompass the period from 2 days
after inoculation (DAI) to 7 DAI. The first symptoms of infec-
tion within the MR treatment were noted four DAI, and all 51
plants within the treatment were symptomatic by 6 DAI. No MR
symptoms were observed on control plants during the measure-
ments, or 7 days after measurements ended, so they were considered
disease-free.
Visual assessment of symptoms
At each assessment, all new flush leaves and stems were exam-
ined on each plant. Both sides of the leaves (adaxial and abaxial)
were assessed. New growth is easily distinguishable from mature
growth, as it is softer and differs in color. While several assess-
ment scales have been developed for MR (Junghans et al. 2003;
Smith et al. 2020), these scales are not designed to evaluate very
early symptoms. We developed the following method. For each new
leaf and stem, the presence or absence of the following symptoms
and signs was recorded; red/purple spots, slight bumps on the leaf
surface, yellow discoloration, and developing pustules containing
spores (yet to break through the epidermis) or fully emerged pus-
tules. In addition, the percentage of area affected by each symptom
on both the adaxial and abaxial leaf surfaces was estimated using
the following ordinal scale: 0 =not present, 1 =1 to 10%, 2 =11
to 25%, 3 =26 to 50%, 4 =51 to 75%, 5 =76 to 100%. By the last
day of the regular monitoring (11 November 2022), 41 MR plants
had symptoms affecting 1 to 10% of the leaf area (category 1), while
the remaining 10 MR plants had symptoms affecting 11 to 25% of
the leaf area (category 2).
Hyperspectral measurements
Data acquisition. Leaf reflectance was measured from 350 to
2,500 nm using a RS-5400 spectroradiometer (Spectral Evolution,
MA). The spectroradiometer was mounted on a backpack and was
equipped with a leaf clip with a dedicated light source and white ref-
erence. Reflectance measurements were taken from single green and
red leaves located in the uppermost part of each plant canopy. During
all sample dates, measurements were taken from leaves that did not
show any visual symptoms, as the focus was on detection of physi-
ological changes associated with MR rather than the identification
of symptoms. For each measured leaf, three reflectance measure-
ments were recorded; each measurement consisted of an average of
40 spectral readings. The spectroradiometer was calibrated against
a white reference in the device at the start of measurements and
every 10 min thereafter. Measurements were taken from the control
plants before the treated plants, and the instrument was cleaned af-
ter all measurements, to minimize transmission of the pathogen to
control plants.
Processing. Leaf reflectance for each plant was extracted from
the spectroradiometer as .sed files. Spectra collected for each plant
and leaf type over all capture dates were visually examined, and
eight spectra with abnormal values or shapes were excluded from
further analysis. NBHI that were related to pigments, xanthophyll,
red/green/blue, disease, water, and plant structure were derived from
bands in the visible and near infrared (VNIR) and short-wave in-
frared (SWIR) spectral regions. Equations and references for these
NBHI are presented (Supplementary Table S1).
Thermal measurements. Canopy-level thermal measurements
were acquired using a thermal camera (FLIR A655SC, Teledyne
FLIR, U.S.A.). The camera has a spectral response in the range
of 7.5 to 14 µm and can detect temperature differences as small
as 0.3°C. The image resolution is 640 ×480 pixels, and images
were captured in 16-bit TIFF format. The camera was mounted
on a tripod 1.2 m from the target plant (Supplementary Fig. S1)
and controlled using Teledyne FLIR’s ResearchIR Max software
(Version 4.40.11.35). The camera was turned on for 1 h before any
images were captured to avoid inaccuracies due to automatic in-
ternal camera calibration. The accuracy of the thermal camera was
improved by inputting values for emissivity, distance to the target,
air temperature, and relative humidity within the ResearchIR soft-
ware (Supplementary Fig. S1). Following Calderón et al. (2015),
emissivity was set to 0.98. The distance to the target was set to
1.2 m. Room air temperature and relative humidity were continu-
ously measured using an air temperature and relative humidity probe
(HMP155A, Vaisala, Oyi, Finland), and average values for these
variables were added to the object properties within the software for
each captured image during the processing stage (Supplementary
Fig. S1).
Two separate containers, one with a black cotton cloth and an-
other filled with damp soil, were placed within the field of view
(Supplementary Fig. S1). The temperatures of the containers were
recorded by the thermal camera and a handheld thermometer that
was set to an emissivity of 0.98 (FLIR TG165, Teledyne FLIR LLC,
U.S.A.). Using data pooled by capture day, a linear regression be-
tween temperatures from the two sources (handheld thermometer
and camera) was constructed and used during the postprocess-
ing step to calibrate measurements of mean canopy temperature
recorded by the camera.
Following the method described in Calderón et al. (2015), the
canopy of individual plants was accurately delineated using object-
based image segmentation algorithms in the Fiji package of ImageJ
software (Rasband 2012). Although the segmentation mainly con-
sisted of red leaves, which made up the majority of the upper canopy,
some green leaves in the lower canopy were also included in the
canopy delineation. For each captured image, canopy temperature,
Tc, was determined using the mean temperature of the delineated
crown which was adjusted using the regression equation obtained
from the calibration plots described above (see example in Sup-
plementary Fig. S1). Normalized canopy temperatures, Tnorm,were
derived for each plant as canopy temperature less air temperature
(i.e., Tc–Ta) (Calderón et al. 2015). The standard deviation of nor-
malized canopy temperature, TSD, was determined at the plant level
andalsousedinanalyses.
Physiological measurements. Physiological measurements
were taken, where possible from symptom-free leaves, on
15 November, which was 4 days after the last set of measurements.
Net photosynthetic rates (A), stomatal conductance (gs), and
transpiration rate (E) were measured after a 2-min pre-illumination
period, at 400 ppm CO2. The measurements were taken using
a GFS-3000 system coupled with an Imaging-PAM chlorophyll
fluorometer (M-Series, Walz, Effeltrich, Germany) that was
equipped with a CO2cartridge to maintain CO2at a constant level.
A one-way analysis of variance was used to test for treatment
differences between the physiological variables. An additional
thermal capture was taken on 14 November from all plants, and
estimates of Tnorm and TSD were regressed against gsand E.
Vol. 113, No. 8, 2023 1407
Data analysis
Data preparation. As MR symptoms developed over the 3-day
period, the measurements on the MR plants were categorized ac-
cording to the infection stage to standardize further analyses. Within
the dataset, the number of days before symptoms (DBS) ranged
from 4 to 1, while the number of days after symptoms (DAS) ranged
from 1 to 3. There were only two samples with DBS =4, so these
were discarded from the analyses, and the infection stages used in-
cluded DBS =3, 2, 1 and DAS =1, 2, 3. Measurements made on the
day of symptoms (Day Sym) and the pretreatment measurements
taken before inoculation (pretreat) were included in the analyses.
These data were matched with measurements from control plants
for each of the eight infection stages. The majority of the MR
plants (30/51) developed symptoms 4 days after inoculation. Thus,
the control group measurements for the eight stages were selected
from the dates that aligned with this largest MR group. The seven
postinoculation stages were categorized for analyses as previsual
detection (3 DBS, 2 DBS, and 1 DBS) and early detection (Day
Sym, 1 DAS, 2 DAS, and 3 DAS) of the disease.
Model fitting. Regularized discriminant analysis (RDA) was
used to classify the two treatments (Control, MR) from thermal
indices and NBHI for the two sets of leaves (green, red) during the
seven previsual and early detection stages. These models used pre-
dictor variables derived from (i) NBHI extracted from red leaves,
(ii) NBHI extracted from green leaves, and (iii) the two thermal
indices. The thermal indices were not separated by leaf type, as the
thermal imagery mainly comprised leaves from the upper canopy,
which were predominantly red leaves, but also included some green
leaves. Thus, a total of 21 models were developedthat included these
three combinations of predictor variables ×the seven previsual and
early detection stages.
All classification models were developed in R (R Development
Core Team 2011). Recursive feature elimination (RFE) was used to
subset the 101 NBHI (Supplementary Table S1) for each of the 14
models with hyperspectral data (categories i and ii in the previous
paragraph) to the most important features. This method undertakes a
backward selection of predictors to a predefined number of predictor
variables and can be used to identify the number and identity of
features that maximize the accuracy on the training dataset using
internal cross-validation. The reduction of the number of variables
to a core smaller set was undertaken to remove highly collinear
variables, avoid overfitting in the RDA classification models, and
reduce training time.
RFE was implemented with the random forest algorithm to sub-
set the 101 variables to the most important features. The RFE used a
fivefold cross validation, with five repeats, that identified the num-
ber of variables that produced the most accurate models from 31
different-sized predictor subsets ranging from 1 to 101 NBHI (full
dataset). Each variable subset was further reduced by removing
collinear variables with a correlation coefficient (R)>0.9, with
eliminated variables constituting the less important of the collinear
pair, as denoted by variable order from the RFE analysis. The final
number of variables included in the seven NBHI models that used
green leaves ranged from 3 to 14 (mean =9), while the number of
variables in the seven NBHI models with red leaves ranged from
2to12(mean=7). A list of all variables used in the 14 NBHI
models is presented (Table 1). As the two thermal indices (Tnorm
and TSD) were only weakly correlated, both were included in the
seven thermal models.
Regularized discriminant analysis (RDA) (Friedman 1989) was
used for all 21 models to classify plants from the control and MR
treatments. RDA is particularly useful for small datasets with corre-
lated features and has been used previously for analyses of spectral
data (Wu et al. 1996). Two other classification methods were used
as a cross check to determine if RDA was the most suitable method.
These methods included random forest, which has been widely
used, as it is easy to implement, is capable of accounting for nonlin-
ear relationships, and handles collinear and high-dimensional data
very well (Breiman 2001). Support vector machines was also imple-
mented, as this method has been widely used for analyzing remotely
sensed data and is very effective at maximizing the margin be-
tween classes while minimizing misclassification error (Cortes and
Vapnik 1995). However, as analyses showed that the accuracy of
RDA models was slightly superior to that of models constructed
using both random forests and support vector machines (data not
shown), further methods and results will focus on the RDA.
The RDA classification models were fitted to the data using a
fivefold cross validation with five repeats. Cross-fold validation
was used, as it is the most efficient method for small datasets, such
as those used here, when there are insufficient data to set aside a
completely independent test dataset (James et al. 2021). Fivefold
cross validation divides the dataset randomly into five equal groups.
During each round, four groups are used for model training, while
the fifth group is set aside for model testing. This process is then
repeated five times until each of the five groups has been used for
validation (James et al. 2021). As there were five repeats, this entire
process was repeated five times and the model evaluation statistics
were averaged across all 25 subsampled (five folds ×five repeats)
validation datasets.
The two parameters that require tuning in RDA are gamma and
lambda, which range between 0 and 1. During the model training, a
grid search with a fine step size (0.01) was implemented to optimize
TABLE 1. Variables that were included in the 21 classification modelsa
Data Leaf type Infection stagebVariables in the model
NBHI Green 3 DBS Ratio975,HI, NDWI1, DRI_NDWI, SRPI, SIPI, WI, NPQI, CLS, BF4, BF2, MSI2
2DBS Ratio975 ,HI, NDWI1, DRI_NDWI, NPCI, SIPI, WI, NPQI, CLS, PSRI, BF1, BF2, MSI, RGI
1 DBS PRI528, Ratio975, DRI_PRI, WBI, PRI570, PRIn, RDVI, PRI550, PM1, OSAVI, MTVI2, MCARI, NDWI2130, DDI
Day Sym CLS, Ratio975,HI, MCARI2, MCARI3, MSI1, SBRI, RARS, MCARI, RDVI
1 DAS PRI528, HI, PRI570
2DAS HI, MSAVI, PSRI, WI_NDVI
3 DAS PRI528, HI, DRI_PRI, SIPI, PRI570, CLS, Ratio975,WI
Red 3 DBS G, HI, Ratio975, DRI_PRI, PRI550, DCabCxc
2DBS BGI2, PRIm3, PRIn, Ratio975 ,PRIm2
1 DBS RGI, GI, PSRI, BGI1, B, WI, RR, CARI, BF3, DCabCxc, DRI_MSI, Ratio975
Day Sym PRIn, GI, DCabCxc,BGI2
1 DAS DRI_PRI, BGI2,DCabCxc, PRI570, G, PM1
2 DAS PRI_CRI, PRIn
3 DAS PRI570, RGI, PM1, PRI550, DCabCxc, GM4, CI1, MSI1, PSRI, BGI2
Thermal Both All seven Tnorm,TSD
aThe variables are categorized according to data type (narrow band hyperspectral indices [NBHI] and thermal indices) and leaf type (green and red). Full names,
equations, and references for all indices are presented in Supplementary Table S1. For both leaf types, the two NBHI that were most frequently used in the models
are bolded.
b3 DBS, 3 days before symptoms; 2 DBS, 2 days before symptoms; 1 DBS, 1 day before symptoms; Day Sym, the day symptoms were expressed; 1 DAS, 1 day
after symptoms; 2 DAS, 2 days after symptoms; and 3 DAS, 3 days after symptoms.
1408 PHYTOPATHOLOGY®
the accuracy of each fitted model. Variable importance within the
final fitted RDA models was identified using the variableimportance
function in the Caret package (Kuhn 2008).
Model performance. Based on predictions for each model made
on the validation dataset, a confusion matrix was constructed, with
frequencies that were averaged across all 25 cross validation folds
and repeats. The MR plants were designated positive and values
in the confusion matrix, which quantified the percentage of true
positives (TP), true negatives (TN), false positives (FP), and false
negatives (FN), were determined.
As the dataset was unbalanced model performance was predom-
inantly assessed using precision, recall, and F1 score, which were
calculated as follows:
Precision =TP
TP +FP
Recall =TP
TP +FN
F1 score =2×Precision ×Recall
Precision +Recall
Precision measures the proportion of the positive predictions that
were correct, while recall measures the proportion of actual posi-
tives that were correctly identified. The F1 score is the harmonic
mean of precision and recall and is more useful than accuracy for
unbalanced datasets, as it accounts for both false positives and false
negatives. Values for the F1 score range from 0 to 1 and categories
of 0.5 to 0.7, 0.7 to 0.8, 0.8 to 0.9, and >0.9 can be classified as poor,
acceptable, excellent, and outstanding discrimination, respectively.
Accuracy measures the proportion of correct predictions among
the total number of cases examined and is given by
Accuracy =TP +TN
TP +FP +TN +FN
Although accuracy is not as widely used for unbalanced datasets,
this statistic was included for completeness.
Results
Visualization of disease
The first symptoms developed in the MR treatment 4 DAI, and
by 6 DAI all plants were symptomatic. All control plants were
asymptomatic over the course of the experiment. The time series
RGB images contrast a control plant with the two MR plants with
the most severe symptoms (Fig. 1). Considering that these were the
most severely affected plants, the symptoms that were expressed in
the early detection phase (1 to 3 DAS) were often subtle and would
be difficult to detect by visual inspection of plants. Images of the
two MR plants are also shown at 5 and 6 DAS (i.e., 10 DAI), which
show the rapid development of symptoms beyond the previsual and
early detection phases, which were studied here. As expected, older
green leaves on MR plants did not exhibit any visual symptoms over
the course of measurements.
Hyperspectral spectra
Mean reflectance of the green leaves was very similar across the
spectral range prior to inoculation. Following inoculation, variation
in reflectance increased within the MR treatment 2 DBS, 1 DBS,
and on the day of symptoms (Fig. 2). The increased variation was
marked within all spectral regions for 2 DBS but more confined to
the 750- to 1,300-nm range for data collected 1 DBS. Mean values
for MR plants within this region declined below those of the control,
both 1 DBS and on the day of symptoms, but after this stage, mean
values were very similar between treatments.
Prior to treatment, the mean reflectance of the red leaves was
higher in the control than the MR treatment, particularly between
700 and 1,800 nm. Following treatment, mean reflectance in MR
plants increased markedly and exceeded mean values in the con-
trol plants across most of the spectral range for all posttreatment
measurements. Changes were particularly marked within the green
peak (550 nm), where the ratio of reflectance between treatments
(MR/Control) increased from 1.00 for pretreatment measurements
to 2.33 for measurements taken 3 DAS. Values of reflectance from
Day 2
aer inoculaon
Day 4
aer inoculaon
Day 5
aer inoculaon
Day 6
aer inoculaon
Day 7
aer inoculaon
Day 10
aer inoculaon
Inoculated with
Myrtle Rust
Control
2 DBS
Day
Sym. 1 DAS 2 DAS 3 DAS 6 DAS
3 DBS
Day
Sym. SAD 1SBD 1 2 DAS 5 DAS
Fig. 1. RGB images showing the progression of myrtle rust on the two inoculated plants with the most severe symptoms, which first displayed symptoms on the fourth
(top row) and fifth (middle row) days after inoculation. A control plant is shown for reference (bottom row). The days before symptoms (DBS), day of symptoms
(Day Sym), and days after symptoms (DAS) are shown in white text within each photo for the two inoculated plants. Although data taken 5 and 6 DAS were not
used in analyses, the photos are shown for reference.
Vol. 113, No. 8, 2023 1409
ca. 750 to 2500 nm for MR plants were also elevated above the
control for most posttreatment measurements, and the differences
were most marked at 3 DBS, the day of symptoms, 2 DAS, and 3
DAS (Fig. 2).
Analysis of variance
Thermal indices. The two thermal indices did not differ sig-
nificantly between treatments either prior to inoculation or for
measurements taken 3 DBS. Treatment differences for both indices
were significant (P<0.05) at 2 DBS and highly significant (P=
8.16 ×10–16 for Tnorm and 1.98 ×10–23 for TSD) by 1 DBS, at which
stage the MR plants had lower (Fig. 3A) and more variable (Fig.
3B) Tnorm than the control plants. As the disease progressed, both
indices continued to diverge between treatments as Tnorm declined
and TSD increased in the MR plants, and treatment differences re-
mained highly significant for all four of the early detection stages
(Prange =1.03 ×10–20 to 1.43 ×10–28). A scatterplot of the two
thermal indices through each of the eight infection stages showed
treatment overlap for the first three measurements (pretreat, 3 DBS,
2 DBS), but complete treatment separation by 1 DBS, which was
maintained for the remaining four infection stages from the day of
symptoms to 3 DAS (Fig. 4).
Hyperspectral indices—green leaves. Significant treatment
differences were detected in 18% of the 101 NBHI during the three
previsual measurements taken on green leaves. Among all cate-
gories, the disease indices had the highest proportion, with 27%
of variables exhibiting significant differences between treatments
(Table 2). Within this category, healthy index (HI) demonstrated
the most highly significant treatment differences (P<0.001) for
both 3 and 2 DBS (Supplementary Table S2). Values of HI in MR
Fig. 2. Plots of reflectance against wavelength for control plants (teal lines) and plants inoculated with Austropuccinia psidii (cause of myrtle rust [MR], red lines)
with respect to the infection stage and the leaf type. Values are shown for each plant (thin lines) and means for the treatment (bold lines). Displayed values represent
measurements taken prior to inoculation (pretreatment), for the 3 days prior to the expression of symptoms (day before symptoms [DBS]), on the day that symptoms
occurred (Day Sym), and for 3 days following expression of symptoms (day after symptoms [DAS]).
1410 PHYTOPATHOLOGY®
plants were lower than those of control plants and diverged over
the previsual detection phase until 1 DBS (Fig. 3C). Water indices
were also important for previsual detection and 25% of the variables
in this category exhibited significant treatment differences during
this stage (Table 2). Among water indices, Ratio975 was the vari-
able that differed most significantly between treatments, with P<
0.001 over all three previsual detection stages (Supplementary Ta-
ble S2). Compared with the control, values of Ratio975 were reduced
in MR plants 3 and 2 DBS, but increased above the control 1 DBS
(Fig. 3D).
In total, 20% of all examined NBHI from green leaves were sig-
nificantly different between treatments during the early detection
stage, which included the day of symptoms and three postsymp-
tomatic measurements (DAS 1 to 3). The disease indices were
most sensitive among the six categories, with 35% of the indices
showing significant treatment differences. The HI differed most be-
tween treatments during these four early detection measurements
(Prange =0.013 to 1.58 ×10–11), and values in the MR treatment
were consistently lower than those in the control during this phase
(Fig. 3C). Both water and structural indices were also important,
with 27% of variables exhibiting significant treatment differences
(Table 2). However, differences within these two categories were
not maintained for any variable over the four infection stages, and,
of all indices, HI was the only variable that maintained significant
differences for all four infection stages during the early detection
period (Supplementary Table S2).
Hyperspectral indices—red leaves. In contrast to the green
leaves, a markedly higher proportion of NBHI derived from red
leaf data differed significantly between treatments during the pre-
visual detection phase (37 versus 18%). The significant differences
were most prevalent in the xanthophyll indices, in which 94% of the
variables exhibited significant treatment differences (Table 2). The
Fig. 3. Box plots of indices for control (teal boxes) and plants inoculated with Austropuccinia psidii (cause of myrtle rust [MR], red boxes) with respect to the
infection stage for A, normalized canopy temperature (Tnorm), B, standard deviation of Tnorm (TSD), C, healthy index (HI), D, Ratio975 ,E, chlorophyll/carotenoid
index (DCabCxc ), and F, blue/green index (BGI2). Values were taken prior to inoculation (pretreatment), for the 3 days prior to the expression of symptoms (day
before symptoms [DBS]), on the day that symptoms occurred (Day Sym), and for 3 days following expression of symptoms (day after symptoms [DAS]). Lines
inside the boxes represent medians, and the top and bottom lines in each box represent the 75th and 25th quartiles, respectively. Whiskers represent ±1.5×the
interquartile range and dots represent outliers.
Vol. 113, No. 8, 2023 1411
variable PRI515 had the most significant differences between treat-
ments (Supplementary Table S2) for the three previsual infection
stages (Prange =0.00028 to 2.37 ×10–7). Disease indices also
demonstrated strong treatment differences, with 80% of variables
showing significant differences. HI differed most significantly be-
tween treatments, with Pvalues ranging from 0.003 to 2.33 ×10–8
(Supplementary Table S2).
The number of variables that differed significantly between treat-
ments increased from the previsual to the early detection stage with
red leaves (65 versus 37%). As with previsual detection in red
leaves, the categories with the highest percentage of variables that
differed significantly between treatments for early detection were
xanthophyll (95%) and disease indices (80%) (Table 2). Within
these two categories, PRIm2 and HI were the NBHI that differed
Fig. 4. Normalized canopy temperature (Tnorm) plotted against the standard deviation of normalized canopy temperature (TSD) for the control plants (teal solid
circles) and plants inoculated with Austropuccinia psidii (cause of myrtle rust [MR], red solid circles). Measurements that are shown were taken prior to inoculation
(pretreatment), for the 3 days prior to the expression of symptoms, on the day that symptoms occurred, and for 3 days following expression of symptoms.
1412 PHYTOPATHOLOGY®
most between treatments, and treatment differences were highly
significant (P<0.001) for both variables over all four infection
stages (Supplementary Table S2). The percentage of NBHI with
significant treatment differences increased from previsual to early
detection for both water indices (63 versus 23%) and structural
indices (62 versus 4%).
Model predictions
Models using thermal indices. The models using indices de-
rived from thermal imagery had low to moderate predictive power
for measurements taken at 3 DBS (F1 score =0.33) and 2 DBS (F1
score =0.75; Table 3). However, by 1 DBS, the model using indices
derived from thermal imagery was able to previsually distinguish
control from infected plants perfectly (F1 score =1.0, accuracy =
100%), and this accuracy was maintained for all four early detec-
tion stages after the appearance of symptoms (Fig. 4). The perfect
discrimination was consistently repeated on all folds and repeats of
the cross-validation for each of the five infection stages (standard
deviation =0) for all the classification metrics (data not shown).
Models using hyperspectral indices. Using indices obtained
from the green leaves, excellent or outstanding treatment classifica-
tion was achieved using data for 3 DBS (F1 score =0.89; accuracy=
91%), 2 DBS (F1 score =0.94; accuracy =93%), and 1 DBS (F1
score =0.91; accuracy =89%). The three most important indices
for these models were Ratio975, HI, and NDWI1. The agreement di-
minished slightly for Day Sym, but the F1 score was relatively high
between 0.87 and 0.94 for the last three infection stages (1 DAS, 2
DAS, 3 DAS; Table 3). Precision and recall were generally very sim-
ilar for most infection stages. The largest differences occurred on
the Day Sym and 1 DAS, during which times recall was higher than
precision, as there were slightly more false positives than incorrect
classifications predicted by these models during the corresponding
infection stages.
With two exceptions (Day Sym, 3 DAS), models that used in-
dices derived from red leaves had poorer performance than those
that used green leaves (Table 3). During the previsual detection pe-
riod, the red leaf model using data obtained from 1 DBS was the
TABLE 2. The percentage of variables in each narrow band hyperspectral index
(NBHI) category in which there were significant differences between treatments
for green and red leavesa
Infection NBHI category
Leaf type stage Pigment Xanthophyll R/G/B Disease Water Structural
Green Pretreat 0 0 0 40 4 0
3 DBS 17 18 17 20 17 7
2DBS 10 0 28 40 35 0
1 DBS 3 55 0 20 22 53
Day Sym 0 0 6 40 4 47
1 DAS 0 27 0 20 61 0
2 DAS 41 27 50 40 43 60
3DAS 0 27 0 40 0 0
Previsual 10 24 15 27 25 20
Early 10 20 14 35 27 27
Red Pretreat 31 45 39 20 4 7
3 DBS 14 82 33 80 4 0
2 DBS 34 100 33 80 17 7
1 DBS 59 100 44 80 48 7
Day Sym 62 100 44 80 78 87
1 DAS 55 100 50 80 30 67
2 DAS 45 91 56 80 74 40
3 DAS 69 91 78 80 70 53
Previsual 36 94 37 80 23 4
Early 58 95 57 80 63 62
aResults are shown for each infection stage and are also averaged over the
previsual (3, 2, and 1 day before symptoms [DBS]) and early disease expression
stages (day of symptoms [Day Sym], 1, 2, 3 and days after symptoms [DAS]).
The categories shown for NBHI correspond to those listed in Supplementary
Table S1, and data for this table were derived from Supplementary Table S2.
most accurate (F1 score =0.90). In general, models using indices
derived from red leaves were more accurate at early detection than
at previsual detection, while F1 scores were relatively high for the
last two infection stages (Table 3). Recall and precision were very
similar for most infection stages and only varied substantially for 2
DBS and 3 DAS, when recall was higher than precision, reflecting
the relatively large number of false positive predictions, compared
with incorrect classifications.
The most frequently used indices in the seven previsual and
early detection models that utilized data from green leaves were
HI and Ratio975, which were used in six and five models, re-
spectively (Table 1). For red leaves the most frequently used
indices were DCabCxc and BGI2, which were used in five and
four models, respectively. Compared with the control, values for
DCabCxc and BGI2 in the MR treatment declined after inoculation
(Fig. 3E and F).
Tree physiology and relationships with thermal indices
All three physiological variables differed significantly between
treatments at the end of the experiment (Prange =2.6 ×10–5 to
8.6 ×10–11). Compared with the control, values in MR plants were,
respectively, 79 and 73% higher for stomatal conductance (68 vs. 38
mmol m−2s−1) and transpiration rate (0.88 versus 0.51 mmol m−2
s−1). In contrast, assimilation rates in MR plants were only 7.1% of
those recorded in the control plants (0.13 versus 1.81 µmol m−2s−1).
Significant negative relationships were observed between Tnorm and
both stomatal conductance (P<0.001; R2=0.16) and transpiration
rate (P<0.001; R2=0.17; Fig. 5A). In contrast, positive relation-
ships were observed between TSD and both stomatal conductance
(P<0.001; R2=0.22) and transpiration rate (P<0.001; R2=0.23;
Fig. 5B).
Fig. 5. Relationship between transpiration rate and A, normalized canopy tem-
perature, Tnorm,andB, the standard deviation of normalized canopy temperature,
TSD. Data were recorded at the end of the experiment for the control plants (teal
solid circles) and plants inoculated with Austropuccinia psidii (cause of myrtle
rust [MR], red solid circles).
Vol. 113, No. 8, 2023 1413
Discussion
Normalized canopy temperature and TSD were the most robust
indices for detecting MR on rose apple during both the previsual
and the early infection stages. Many studies have used normal-
ized canopy temperature and derived indices to characterize disease
mainly in agricultural and orchard crops (Calderón et al. 2015;
López-López et al. 2016; Nilsson 1991; Zarco-Tejada et al. 2018),
compared with far fewer studies on forestry related tree species
(Hornero et al. 2021). Increases in canopy temperature, used in
combination with plant functional traits extracted from hyperspec-
tral data, have been used for previsual detection of Xylella fastidiosa
in olives (Zarco-Tejada et al. 2018) and Phytophthora cinnamomi in
holm oak (Hornero et al. 2021) and early detection of Verticillium
wilt in olives (Calderón et al. 2015).
However, the linking of canopy temperature to disease impacts is
often complex and depends on the pathogen mode of operation, the
physiological response of the plant to infection, and the stage of the
infection. According to energy balance theory, the temperature of
leaves is linked to the transpiration rate, which is a function of stom-
atal conductance (Still et al. 2019). Although stomatal conductance
is mainly affected by abiotic factors, foliar pathogens have been
found to influence it during infection (Smith et al. 1986). Rates
of transpiration often decline in plants infected by necrotrophic
pathogens, which occurs through stomatal closure, reduction of air
space by hyphae, obstruction of conducting tissue and stomata, and
defoliation (Bassanezi et al. 2002).
In contrast, plant infection by biotrophic pathogens, such as rusts
or mildews, often has the opposite effect on transpiration. During
the early stages, infection by biotrophic pathogens often results in
higher water loss through increasing permeability of leaf cell mem-
branes (Chaerle et al. 2001), establishment of infection structures
on the plant surface that result in damage to the cuticle (Anderson
and Frank 2003; Bassanezi et al. 2002), or inhibited stomatal closure
(Lindenthal et al. 2005; Smith et al. 1986). The increased rates of
water loss reduce the leaf temperature (Anderson and Frank 2003;
Lindenthal et al. 2005; Oerke et al. 2011), and as early stage in-
fection is often localized, cause greater temperature variation in
diseased leaves than in unaffected foliage (Lindenthal et al. 2005;
Oerke et al. 2011). As infection transitions to later phases, and
necrotic tissue becomes more widespread, the leaf temperature typ-
ically increases and becomes less variable (Smith et al. 1986).
Results presented here that characterize infection during the pre-
visual and early stages are consistent with general observations
describing impacts of biotrophic pathogens such as rusts. Infection
by A. psidii increased both stomatal conductance and transpira-
tion in rose apple, which was detected through reduced and more
variable leaf temperature. As has been noted for other rusts, it is
likely that these initial physiological changes occurred through the
rust pustules rupturing the cuticle and preventing stomatal closure
(Smith et al. 1986).
Although the older green leaves were, as expected, visually unaf-
fected by MR (Beresford et al. 2020), hyperspectral indices derived
from these leaves were more accurate at detecting plant previsual
symptoms than indices derived from red leaves. Excellent classifi-
cation of MR plants was achieved as early as 3 DBS on green leaves,
and the key indices used for detection were Ratio975 and healthy in-
dex (HI). HI was developed by Mahlein et al. (2013) for detection of
a range of diseases on sugar beet that included Cercospora leaf spot,
sugar beet rust, and powdery mildew and in this context was found
to outperform many other commonly used indices. Consistent with
results presented here, increased disease severity has been found to
be associated with significant reductions in HI (López-López et al.
2016; Mahlein et al. 2013). HI has a sound physiological basis,
as reflectance at 534 nm denotes changes in photosynthetic func-
tion through interconversion of chlorophyll to xanthophyll cycle
pigments, which facilitates dissipation of unused energy as heat,
avoiding damage to the photosystem (Niyogi 1999). The two bands
on either side of 700 nm, within HI, occur within the red edge, and
a blue shift of reflectance in this region allows early detection of
many forms of plant stress (Watt et al. 2020). Ratio975 is a water
stress index that is based on a prominent water absorption band in
vegetation centered at 970 nm. Ratio975 has been used to predict
variation in foliage water content for a variety of tree species (Pu
et al. 2003), including Pinus patula trees infested with sirex wood
wasp (Sirex noctilio; Mutanga and Ismail 2010).
TABLE 3. Confusion matrix and classification statistics categorized into groupings for models that use thermal indices, narrow band hyperspectral indices (NBHI)
extracted from green leaves, and NBHI using red leavesa
Infection Confusion matrix (%) Classification statistics
Index stagebTP FP FN TN F1 Precision Recall Accuracy (%)
Thermal indices 3 DBS 12.2 20.8 29.0 38 0.33 0.37 0.30 50
2 DBS 54.3 26.9 8.6 10.1 0.75 0.67 0.86 64
1 DBS 63 0 0 37 1.00 1.00 1.00 100
Day Sym 63 0 0 37 1.00 1.00 1.00 100
1 DAS 63 0 0 37 1.00 1.00 1.00 100
2 DAS 62 0 0 38 1.00 1.00 1.00 100
3 DAS 50 0 0 50 1.00 1.00 1.00 100
Hyperspec. indices, green leaves 3 DBS 36.9 4.7 4.3 54.1 0.89 0.89 0.90 91
2 DBS 58.8 3.0 4.2 34.1 0.94 0.95 0.93 93
1 DBS 57.0 4.9 5.9 32.1 0.91 0.92 0.91 89
Day Sym 52.1 13.6 10.9 23.5 0.81 0.79 0.83 76
1 DAS 56.3 10.1 6.7 26.9 0.87 0.85 0.89 83
2 DAS 58.2 3.0 3.8 34.9 0.94 0.95 0.94 93
3 DAS 45.3 6.0 4.7 44.0 0.89 0.88 0.91 89
Hyperspec. indices, red leaves 3 DBS 29.4 10.2 11.8 48.6 0.73 0.74 0.71 78
2 DBS 57.5 13.3 5.4 23.7 0.86 0.81 0.91 81
1 DBS 57.2 5.8 7.3 29.6 0.90 0.91 0.89 87
Day Sym 57.5 8.2 6.2 28.0 0.89 0.88 0.90 86
1 DAS 54.0 8.5 8.5 29.0 0.86 0.86 0.86 83
2 DAS 57.0 5.1 5.1 32.9 0.92 0.92 0.92 90
3 DAS 48.0 6.3 2.0 43.7 0.92 0.88 0.96 92
aAll results were compiled from predictions made on a validation dataset and all values for the confusion matrix and classification statistics were averaged across
the five folds and five repeats obtained from the validation. Abbreviations for categories determined from the confusion matrix are as follows: true positive (TP),
false positive (FP), false negative (FN), and true negative (TN). Models with an F1 score (F1) >0.9, which is rated as outstanding classification, are in bold.
b3 DBS, 3 days before symptoms; 2 DBS, 2 days before symptoms; 1 DBS, 1 day before symptoms; Day Sym, the day symptoms were expressed; 1 DAS, 1 day
after symptoms; 2 DAS, 2 days after symptoms; and 3 DAS, 3 days after symptoms.
1414 PHYTOPATHOLOGY®
Although HI and Ratio975 varied significantly between treat-
ments in green leaves prior to inoculation, evidence suggests the
differences did not negatively impact the use of these indices for
previsual detection. Compared with pretreatment differences, treat-
ment differences at 3 DBS and 2 DBS were far more significant
for both indices (Supplementary Table S2), demonstrating a sub-
stantial impact of inoculation on these indices. Furthermore, even
when HI and Ratio975 were excluded from the analysis, RDA mod-
els with excellent classification capability were developed using
data from 3 DBS (F1 score =0.91) and 2 DBS (F1 score =0.92).
The development of these models demonstrates that alternative
NBHI from green leaves are also sufficiently robust for previsual
detection.
Although treatment classification was generally less precise using
NBHI from red leaves than from green leaves, the most commonly
used indices within these models provided insight into processes oc-
curring in the leaves. The blue/green index (BGI2), which appeared
in four models, was originally developed as an accurate proxy for
chlorophyll content using measurements taken from grapevines
(Zarco-Tejada et al. 2005). Reductions in BGI2 that occurred in
MR plants most likely reflect lower chlorophyll content, as these
changes were mainly driven by substantial increases in reflectance
within MR plants at 550 nm, which is a change associated with
reductions in chlorophyll content (Gitelson and Merzlyak 1996).
Similarly, the reflectance band ratio index, DCabCxc , is a NBHI that
is positively related to chlorophyll and carotenoid content (Datt
1998) and has been used for previsual detection of Xyella fastid-
iosa in olive trees (Poblete et al. 2020). Changes in DCabCxc also
suggest that infection by A. psidii induces reductions in chlorophyll
within the susceptible leaves, which is consistent with the impacts
of several other diseases on plant species (Kabir et al. 2015; Poblete
et al. 2020).
Despite being specifically developed for this pathogen, the lemon
myrtle–MR index (LMMR; Heim et al. 2019) was not used within
the models and did not differ significantly between treatments at
any of the infection stages for either leaf type. In contrast to this
study, which focussed on previsual and early detection, the LMMR
was developed to detect more advanced symptoms on a different
host (Heim et al. 2019), which most likely limited the utility of the
index. This result reinforces the need to test indices on a variety of
Myrtaceae covering a range of stages from previsual detection to
advanced symptom expression for this pathosystem.
The dataset collected here was relatively small, which is likely to
limit the generality of the findings. Analyses were undertaken using
repeated cross validation, as this is the most efficient strategy for
determining the robustness of models using small datasets where
there is insufficient data to set aside a completely independent test
dataset (James et al. 2021). Despite the sample size limitation, ther-
mal indices perfectly discriminated MR from 1 DBS onward. The
validation showed that this result was repeatable, as the perfect clas-
sification was consistent for all 25 validation folds and repeats at 1
DBS and all four subsequent early detection stages. The utility of
thermal data for detecting MR is evident in the separation of the two
groups by the two thermal indices from 1 DBS onward in Figure 4.
Further research using a larger set of plants is needed to confirm
these findings and develop a more generally applicable model.
The accurate characterization of MR during the previsual and
early stages suggests that a robust detection methodology could be
developed within a nursery setting. The most practical thermal in-
dicator for MR may be TSD, as in contrast to Tnorm the TSD does
not require calibration to account for changes in climatic factors
and the physiological status of the plant (Lindenthal et al. 2005).
Accurate and relatively repeatable previsual and early detection of
infected plants were shown to occur above a TSD threshold of ca.
0.2°C (Fig. 4), and measurements should focus on young expand-
ing leaves that are directly infected by the rust, as they showed the
greatest changes in TSD to disease. Results presented here highlight
marked variation in reflectance and indices between the green and
red leaves, which may have implications for presymptomatic and
early detection using canopy-level hyperspectral acquisitions. As
has been found previously (Chavana-Bryant et al. 2017), leaf age
had significant impacts on both the shape and scale of reflectance
across the spectrum. The most age-sensitive spectral domains were
found to be the green peak (550 nm) and the short-wave infrared
region. Reflectance shifts resulting from infection also differed be-
tween leaf ages in these regions (Fig. 2), and differential responses
to infection between green and red leaves occurred for many key
NBHI (data not shown). Although this disparity may make it chal-
lenging to predict infection from canopy-level hyperspectral data, a
greater understanding of the differences could also be used to select
predictors that improve the sensitivity of detection systems.
In conclusion, all of the rose apple plants that were inoculated
with A. psidii developed minor MR symptoms over the monitoring
period that would be difficult to detect through visual observation in
a nursery setting. Using predictions made on the validation dataset,
models using indices derived from thermal imagery were able to
distinguish control from MR plants perfectly at 1 DBS (previsual
detection) and during the entire early detection phase (Day Sym
to 3 DAS). Changes in the indices show that MR plants had lower
and more variable values of Tnorm, which were most likely an in-
dicator of higher values of stomatal conductance and transpiration
within infected tissue. Classification models using NBHI derived
from older green leaves with ontogenic resistance to the disease
were almost always more accurate than models that used NBHI
from susceptible younger red leaves that developed MR symptoms.
Using NBHI derived from green leaves, excellent previsual clas-
sication was achieved 3 DBS, 2 DBS, and 1 DBS, and the most
important three indices in these models characterized changes in
disease (HI) and water stress (Ratio975, NDWI). Models that used
NBHI derived from red leaves were able to accurately classify treat-
ments during 1 DBS, 2 DAS, and 3 DAS. Further research should
use a larger dataset to develop more generally applicable models and
investigate the feasibility of implementing a system for detecting
MR in a nursery setting on a range of Myrtaceae.
Acknowledgments
We thank the Auckland Council for providing access to collect Syzy-
gium jambos seed; Kwasi Adusei-Fosu, Darryl Herron, Te Whaeoranga
Smallman, Katherine Wardhaugh, Maria Zhulanov, and Ricki Hasson for
assistance with disease assessments and maintaining plants during the ex-
periment; Scion nursery staff for propagation and care of plants used in this
experiment; and two anonymous reviewers who provided useful comments
that improved the quality of the manuscript.
Literature Cited
Anderson, M. T., and Frank, D. A. 2003. Defoliation effects on reproductive
biomass: Importance of scale and timing. J. Range Manage. 56:501-516.
Baskarathevan, J., Taylor, R. K., Ho, W., McDougal, R. L., Shivas, R. G., and
Alexander, B. J. R. 2016. Real-time PCR assays for the detection of Puccinia
psidii. Plant Dis. 100:617-624.
Bassanezi, R. B., Amorim, L., Filho, A. B., and Berger, R. D. 2002. Gas ex-
change and emission of chlorophyll fluorescence during the monocycle of
rust, angular leaf spot and anthracnose on bean leaves as a function of their
trophic characteristics. J. Phytopathol. 150:37-47.
Beenken, L. 2017. Austropuccinia: A new genus name for the myrtle rust
Puccinia psidii placed within the redefined family Sphaerophragmiaceae
(Pucciniales). Phytotaxa 297:53-61.
Beresford, R. M., Shuey, L. S., and Pegg, G. S. 2020. Symptom development
and latent period of Austropuccinia psidii (myrtle rust) in relation to host
species, temperature, and ontogenic resistance. Plant Pathol. 69:484-494.
Beresford, R. M., Turner, R., Tait, A., Paul, V., Macara, G., Zhidong, D. Y.,
Lima, L., and Martin, R. 2018. Predicting the climatic risk of myrtle rust
during its first year in New Zealand. N.Z. Plant Prot. 71:332-347.
Bini, A. P., Quecine, M. C., da Silva, T. M., Silva, L. D., and Labate, C. A.
2018. Development of a quantitative real-time PCR assay using SYBR Green
for early detection and quantification of Austropuccinia psidii in Eucalyptus
grandis. Eur. J. Plant Pathol. 150:735-746.
Breiman, L. 2001. Random forests. Mach. Learn. 45:5-32.
Vol. 113, No. 8, 2023 1415
Calderón, R., Navas-Cortés, J. A., and Zarco-Tejada, P. J. 2015. Early detection
and quantification of verticillium wilt in olive using hyperspectral and thermal
imagery over large areas. Remote Sens. 7:5584-5610.
Carnegie, A. J., and Pegg, G. S. 2018. Lessons from the incursion of myrtle rust
in Australia. Annu. Rev. Phytopathol. 56:457-478.
Chaerle, L., De Boever, F., Montagu, M. V., and Straeten, D. V. D. 2001. Ther-
mographic visualization of cell death in tobacco and Arabidopsis. Plant Cell
Environ. 24:15-25.
Chavana-Bryant, C., Malhi, Y., Wu, J., Asner, G. P., Anastasiou, A., Enquist,
B. J., Cosio Caravasi, E. G., Doughty, C. E., Saleska, S. R., and Martin, R. E.
2017. Leaf aging of Amazonian canopy trees as revealed by spectral and
physiochemical measurements. New Phytol. 214:1049-1063.
Cortes, C., and Vapnik, V. 1995. Support-vector networks. Mach. Learn. 20:
273-297.
Coutinho, T. A., Wingfield, M. J., Alfenas, A. C., and Crous, P. W. 1998. Euca-
lyptus rust: A disease with the potential for serious international implications.
Plant Dis. 82:819-825.
Datt, B. 1998. Remote sensing of chlorophyll a, chlorophyll b, chlorophyll
a+b, and total carotenoid content in eucalyptus leaves. Remote Sens. Environ.
66:111-121.
Friedman, J. H. 1989. Regularized discriminant analysis. J. Am. Stat. Assoc.
84:165-175.
Gitelson, A. A., and Merzlyak, M. N. 1996. Signature analysis of leaf reflectance
spectra: Algorithm development for remote sensing of chlorophyll. J. Plant
Physiol. 148:494-500.
Glen, M., Alfenas, A. C., Zauza, E. A. V., Wingfield, M. J., and Mohammed, C.
2007. Puccinia psidii: A threat to the Australian environment and economy—
A review. Australas. Plant Pathol. 36:1-16.
Heim, R. H. J., Wright, I. J., Allen, A. P., Geedicke, I., and Oldeland, J. 2019.
Developing a spectral disease index for myrtle rust (Austropuccinia psidii).
Plant Pathol. 68:738-745.
Hernández-Clemente, R., Hornero, A., Mottus, M., Penuelas, J., González-
Dugo, V., Jiménez, J. C., Suárez, L., Alonso, L., and Zarco-Tejada, P. J. 2019.
Early diagnosis of vegetation health from high-resolution hyperspectral and
thermal imagery: Lessons learned from empirical relationships and radiative
transfer modelling. Curr. For. Rep. 5:169-183.
Ho, W. H., Baskarathevan, J., Griffin, R. L., Quinn, B. D., Alexander, B. J. R.,
Havell, D., Ward, N. A., and Pathan, A. K. 2019. First report of myrtle rust
caused by Austropuccinia psidii on Metrosideros kermadecensis on Raoul
Island and on M. excelsa in Kerikeri, New Zealand. Plant Dis. 103:2128.
Hornero, A., Zarco-Tejada, P. J., Quero, J. L., North, P. R. J., Ruiz-Gómez,
F. J., Sánchez-Cuesta, R., and Hernandez-Clemente, R. 2021. Modelling
hyperspectral- and thermal-based plant traits for the early detection of
Phytophthora-induced symptoms in oak decline. Remote Sens. Environ.
263:112570.
James, G., Witten, D., Hastie, T., and Tibshirani, R. 2021. An introduction to
statistical learning: with applications in R. Springer Science+Business Media.
Junghans, D. T., Alfenas, A. C., and Maffia, L. A. 2003. Rating scale to eucalypts
rust severity evaluation. Fitopatol. Bras. 28:184-188.
Kabir, M. S., Ganley, R. J., and Bradshaw, R. E. 2015. Dothistromin toxin
is a virulence factor in Dothistroma needle blight of pines. Plant Pathol.
64:225-234.
Kuhn, M. 2008. Building predictive models in R using the caret package. J. Stat.
Softw. 28:1-26.
Langrell, S. R. H., Glen, M., and Alfenas, A. C. 2008. Molecular diagnosis of
Puccinia psidii (guava rust)—A quarantine threat to Australian eucalypt and
Myrtaceae biodiversity. Plant Pathol. 57:687-701.
Lee, D. J., Brawner, J. T., and Pegg, G. S. 2015. Screening Eucalyptus cloeziana
and E. argophloia populations for resistance to Puccinia psidii. Plant Dis.
99:71-79.
Lindenthal, M., Steiner, U., Dehne, H. W., and Oerke, E. C. 2005. Effect of
downy mildew development on transpiration of cucumber leaves visualized
by digital infrared thermography. Phytopathology 95:233-240.
López-López, M., Calderón, R., González-Dugo, V., Zarco-Tejada, P. J., and
Fereres, E. 2016. Early detection and quantification of almond red leaf
blotch using high-resolution hyperspectral and thermal imagery. Remote
Sens. 8:276.
Mahlein, A. K., Rumpf, T., Welke, P., Dehne, H. W., Plümer, L., Steiner, U.,
and Oerke, E. C. 2013. Development of spectral indices for detecting and
identifying plant diseases. Remote Sens. Environ. 128:21-30.
Mahlein, A.-K., Steiner, U., Hillnhütter, C., Dehne, H.-W., and Oerke, E.-C.
2012. Hyperspectral imaging for small-scale analysis of symptoms caused
by different sugar beet diseases. Plant Methods 8:1-13.
Mutanga, O., and Ismail, R. 2010. Variation in foliar water content and hyper-
spectral reflectance of Pinus patula trees infested by Sirex noctilio. South.
Forests 72:1-7.
Nilsson, H. E. 1991. Hand-held radiometry and IR-thermography of plant dis-
eases in field plot experiments. Int. J. Remote Sens. 12:545-557.
Niyogi, K. K. 1999. Photoprotection revisited: Genetic and molecular ap-
proaches. Annu. Rev. Plant Biol. 50:333-359.
Oerke, E. C., Fröhling, P., and Steiner, U. 2011. Thermographic assessment of
scab disease on apple leaves. Precis. Agric. 12:699-715.
Pegg, G. S., Brawner, J. T., and Lee, D. J. 2014. Screening Corymbia populations
for resistance to Puccinia psidii. Plant Pathol. 63:425-436.
Poblete, T., Camino, C., Beck, P. S. A., Hornero, A., Kattenborn, T., Saponari,
M., Boscia, D., Navas-Cortes, J. A., and Zarco-Tejada, P. J. 2020. Detection
of Xylella fastidiosa infection symptoms with airborne multispectral and ther-
mal imagery: Assessing bandset reduction performance from hyperspectral
analysis. ISPRS J. Photogramm. 162:27-40.
Pu, R., Ge, S., Kelly, N. M., and Gong, P. 2003. Spectral absorption features as
indicators of water status in coast live oak (Quercus agrifolia) leaves. Int. J.
Remote Sens. 24:1799-1810.
R Development Core Team. 2011. R: A language and environment for statistical
computing. R Foundation for Statistical Computing, Vienna.
Rasband, W. S.2012. ImageJ, U.S. National Institutes of Health, Bethesda, MD.
Rayachhetry, M. B., Elliott, M. L., and Van, T. K. 1997. Natural epiphytotic of
the rust Puccinia psidii on Melaleuca quin-quenervia in Florida. Plant Dis.
81:831-831.
Roux, J., Granados, G. M., Shuey, L., Barnes, I., Wingfield, M. J., and
McTaggart, A. R. 2016. A unique genotype of the rust pathogen, Puccinia
psidii, on Myrtaceae in South Africa. Australas. Plant Pathol. 45:645-652.
Roux, J., Greyling, I., Coutinho, T. A., Verleur, M., and Wingfeld, M. J. 2013.
The myrtle rust pathogen, Puccinia psidii, discovered in Africa. IMA Fungus
4:155-159.
Smith, G. R., Ganley, B. J., Chagné, D., Nadarajan, J., Pathirana, R. N., Ryan,
J., Arnst, E. A., Sutherland, R., Soewarto, J., and Houliston, G. 2020. Resis-
tance of New Zealand provenanceLeptospermum scoparium,Ku nzea ro busta,
Kunzea linearis,andMetrosiderosexcelsa to Austropuccinia psidii. Plant Dis.
104:1771-1780.
Smith, R. C. G., Heritage, A. D., Stapper, M., and Barrs, H. D. 1986. Effect of
stripe rust (puccinia striiformis West.) and irrigation on the yield and foliage
temperature of wheat. Field Crops Res. 14:39-51.
Soewarto, J., Giblin, F., and Carnegie, A. J. 2019. Austropuccinia psidii (myr-
tle rust) global host list, version 4. Australian Network for Plant Conser-
vation, Canberra, ACT. https://www.anpc.asn.au/wp-content/uploads/2019/
09/MyrtleRust_GlobalHostList_final_4.xlsx
Soewarto, J., Somchit, C., Du Plessis, E., Barnes, I., Granados, G. M., Wingfield,
M. J., Shuey, L., Bartlett, M., Fraser, S., and Scott, P. 2021. Susceptibility of
native New Zealand Myrtaceae to the South African strain of Austropuccinia
psidii: A biosecurity threat. Plant Pathol. 70:667-675.
Still, C., Powell, R., Aubrecht, D., Kim, Y., Helliker, B., Roberts, D., Richard-
son, A. D., and Goulden, M. 2019. Thermal imaging in plant and ecosystem
ecology: Applications and challenges. Ecosphere 10:e02768.
Tessmann, D. J., Dianese, J. C., Miranda, A. C., and Castro, L. H. R. 2001.
Epidemiology of a Neotropical rust (Puccinia psidii): Periodical analysis of
the temporal progress in a perennial host (Syzygium jambos). Plant Pathol.
50:725-731.
Toome-Heller, M., Ho, W. W. H., Ganley, R. J., Elliott, C. E. A., Quinn, B.,
Pearson, H. G., and Alexander, B. J. R. 2020. Chasing myrtle rust in New
Zealand: Host range and distribution over the first year after invasion. Aus-
tralas. Plant Pathol. 49:221-230.
Watt, M. S., Buddenbaum, H., Leonardo, E. M. C., Estarija, H. J., Bown, H. E.,
Gomez-Gallego, M., Hartley, R. J. L., Pearse, G. D., Massam, P., and Wright,
L. 2020. Monitoring biochemical limitations to photosynthesis in N and P-
limited radiata pine using plant functional traits quantified from hyperspectral
imagery. Remote Sens. Environ. 248:112003.
Wu, W., Mallet, Y., Walczak, B., Penninckx, W., Massart, D. L., Heuerding,
S., and Erni, F. 1996. Comparison of regularized discriminant analysis linear
discriminant analysis and quadratic discriminant analysis applied to NIR data.
Anal. Chim. Acta 329:257-265.
Xavier, A. A., Alfenas, A. C., Matsuoka, K., and Hodges, C. S. 2001. Infection
of resistant and susceptible Eucalyptus grandis genotypes by urediniospores
of Puccinia psidii. Australas. Plant Pathol. 30:277-281.
Yong, W. T. L., Ades, P. K., Tibbits, J. F. G., Bossinger, G., Runa, F. A., Sandhu,
K. S., and Taylor, P. W. J. 2019. Disease cycle of Austropuccinia psidii on
Eucalyptus globulus and Eucalyptus obliqua leaves of different rust response
phenotypes. Plant Pathol. 68:547-556.
Zarco-Tejada, P. J., Berjón, A., Lopez-Lozano, R., Miller, J. R., Martín, P.,
Cachorro, V., González, M. R., and 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.
Zarco-Tejada, P. J., Camino, C., Beck, P. S. A., Calderon, R., Hornero, A.,
Hernández-Clemente, R., Kattenborn, T., Montes-Borrego, M., Susca, L.,
and Morelli, M. 2018. Previsual symptoms of Xylella fastidiosa infection
revealed in spectral plant-trait alterations. Nat. Plants 4:432-439.
1416 PHYTOPATHOLOGY®
