Content uploaded by Panagiotis Christakakis
Author content
All content in this area was uploaded by Panagiotis Christakakis on Oct 02, 2024
Content may be subject to copyright.
Multi-spectral image transformer descriptor classication combined with
molecular tools for early detection of tomato grey mould
Dimitrios Kapetas
a
, Eleni Kalogeropoulou
b
, Panagiotis Christakakis
a
, Christos Klaridopoulos
c
,
Eleftheria Maria Pechlivani
a,*
a
Centre for Research and Technology Hellas, Information Technologies Institute, Thessaloniki 57001, Greece
b
Benaki Phytopathological Institute, Scientic Directorate of Phytopathology, Laboratory of Mycology, 145 61 Kissia, Attica, Greece
c
iKnowHow S.A., Athens 15451, Greece
ARTICLE INFO
Keywords:
Deep learning
Image classication
Botrytis cinerea
Grey mould
Tomato
Precision agriculture
qRT-PCR
Molecular diagnosis
ABSTRACT
Accurate early detection of plant diseases is a critical milestone in the development of self-navigating robots for
precision agriculture and still remains one of the most signicant challenges. Plant pathogens, like Botrytis cinerea
causing grey mould disease, pose signicant threats to agriculture and food safety. This study focuses on the early
detection of B. cinerea in tomato crops using Articial Intelligence. Specically, the Deep Learning (DL) YOLOv8
architecture was employed to apply single class segmentation on plant captures to extract the tomato leaves
locational information. Each plant capture includes ve images from hyperspectral wavelengths (at 460, 540,
640, 775 and 875 nm) and one RGB images. The leaf segmentation achieved 81.7 % Mean Average Precision
(mAP) at Intersection over Union (IoU) threshold 0.5 (mAP50). Then the for each leaf segment, and array of
descriptors is extracted through various Transformer models. Finally, the descriptors are classied through a
KNN or an LSTM solution and the results of all the descriptors for each leaf from each of the Transformer models,
for each of the six images of the capture, are ensembled to yield the class of each leaf. The best ensemble results
were produced by only accumulating results from some of the Transformer and specically the MaxViT-L, the
ViT-B (P:16 ×16 – C), the VOLO (D:5) and the XCIT-L (L:24 – P:16 ×16), by only using the results provided by
the images at the hyperspectral wavelengths at 540, 640 nm and the RGB image, and by applying the LSTM
solution. The process achieved a 79.41 % classication accuracy and 71.12 % F1-Score for all classes. Addi-
tionally, early, rapid and accurate detection of grey mould at the early stages of the infection or at latent in-
fections when symptoms are not visible was assessed by comparative qPCR-based methods (qRT-PCR) for fungal
biomass estimation in plant tissues and the rest of this study’s methodology. The ndings of this study represent a
signicant advancement in crop disease detection and management, and highlight the potential for integrating
these methods into on-site digital systems (robots, mobile apps, etc.) under real-world settings. The results,
demonstrate the effectiveness of combining segmentation and transformer-based classication models for early
and accurate detection of grey mould disease.
Introduction
Grey mould caused by Botrytis cinerea Pers. is one of the most
destructive diseases of high-value crops throughout the world. B. cinerea
has a disastrous impact on various economically important crops
including grape, strawberry, and tomato [1]. The disease occurs more
often in excessively humid environments attacking both outdoor and
greenhouse plantings as well as plant products during storage and transit
[2,3,4]. The most efcient way to block the propagation of pathogens is
by applying fungicides. However, excessive use of fungicides can lead to
high costs, increased pathogen resistance, and potential environmental
and health risks [5,6]. Therefore, early diagnosis of pathogenic fungi can
potentially allow disease management with minimal fungicide applica-
tions. Consequently, early and accurate detection of B. cinerea is crucial.
Traditional detection methods primarily rely on visual inspection, which
lacks precision, and laboratory-based diagnostic testing [7], which re-
quires specialized equipment, expertise and considerable time. Howev-
er, these methods are tedious and often cannot discriminate between
* Corresponding author.
E-mail address: riapechl@iti.gr (E.M. Pechlivani).
Contents lists available at ScienceDirect
Smart Agricultural Technology
journal homepage: www.journals.elsevier.com/smart-agricultural-technology
https://doi.org/10.1016/j.atech.2024.100580
Received 3 September 2024; Received in revised form 19 September 2024; Accepted 19 September 2024
Smart Agricultural Technology 9 (2024) 100580
Available online 26 September 2024
2772-3755/© 2024 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC license ( http://creativecommons.org/licenses/by-
nc/4.0/ ).
slightly different levels of resistance. Besides, they are not well suited to
assess fungal development in the early phases of the infection, before
macroscopical symptoms are visible. qPCR based methods have been
employed in recent years for fungal biomass quantication for the
discrimination of quantitative resistance to plant pathogens [8]. Addi-
tionally, quantifying fungal biomass based on relative expression of
fungal reference genes via qRT-PCR can further validate and conrm
fungal colonization at the early stages of the infection or at latent in-
fections when symptoms are not visible [9].
To achieve this, a Deep Learning (DL) approach that utilizes image
detection and classication techniques, combined with multi-spectral
imaging, offers a signicant advantage [10]. DL has revolutionized
plant pathology by providing more advanced and precise methods for
disease detection [10,11]. DL systems can process and analyze large
volumes of image data, identifying patterns and symptoms indicative of
plant diseases [12]. These models continually learn and improve their
diagnostic accuracy over time, adapting to new disease patterns and
variations [13], which is crucial for managing the diverse and evolving
nature of plant pathogens. Additionally, applications utilizing such
technologies extend to predicting disease spread and severity, offering
valuable insights for the crop management and the decision-making
processes [14], while also contributing to the overall efciency and
sustainability of agricultural practices [15].
In recent years, advancements have been made towards the appli-
cation of computer vision and DL in crop disease detection as their value
is recognized by a growing body of literature [16]. In combination with
hyperspectral imaging disease detection reached higher precision, while
implemented classication techniques include Convolutional Neural
Networks (CNNs), Support Vector Machine (SVM) models, the K-nearest
neighbor (KNN) algorithm and decision tree-based classiers [16]. Also,
ensemble techniques are gaining traction as an approach to improve
performance of singular models. Mohammed et al. (2023) [17] per-
formed a literature review on ensemble techniques, going over the
various structural elements of an ensemble methodology.
In 2020, Dosovitskiy et al. [18] proposed that the Transformer
methodology, which had already become a standardized approach for
Natural Language Processing tasks, is also applicable to computer vision
tasks. Up until then, the computer vision space was dominated by CNNs,
but Vision Transformer (ViT), when trained on a sufcient dataset,
outperforms convolutional networks. Future research lead to various
permutations of the ViT, such as: a) the Swin Transformer proposed by
Liu et al. (2021) [19], which attempts to combat limitation of ViT like
the squared computational complexity relative to the image resolution
and the xed scale; b) The DeiT model, proposed by Tourvron et al.
(2022) [20], which also focuses on optimizing the computational
complexity of the simple ViT by utilizing a pretrained-ViT at lower
resolution; c) The VOLO model, proposed by Yuan et al. (2021) [21],
which attempts to encode ner-level features and contexts, as opposed
to the ViT that focuses on global dependency modeling at a coarse level;
d) The XCiT model, proposed by El-Nouby et al. (2021) [22], which
mitigates the quadratic complexity of ViT by implementing a
cross-covariance attention layer of linear complexity; e) The MaxViT
model, proposed by Tu et al. (2022) [23], where the solution scales
linearly with image resolution and attempts to enable the model to see
globally throughout the entire network.
In relevant research, Giakoumoglou et al. (2024) [24], explored
early detection of B. cinerea in cucumber plants using multi-spectral
image segmentation. Their approach involves the integration of CNNs
with ViT encoders, achieving a Dice Coefcient of 0.677 with an Inter-
section over Union (IoU) of 0.656 and an overall accuracy of 90.1 %. In
their later research, Christakakis et al. (2024) [25], they improved upon
their previous work, reaching a Dice Coefcient of 0.792, an IoU of
0.816 and an overall accuracy of 92 % using the augmentation technique
Cut-and-Paste to address dataset imbalance. Scarboro et al. (2021) [26],
focused on detection of B. cinerea in lettuce through bispectral imaging
applying single pixel classication. They reached a true-positive rate of
95.25 % with a false positive rate of 9.316 % under laboratory condi-
tions. In addition, Chung et al. (2021) [27] investigated Fusarium fuji-
kuroi detection in rice seedlings, a disease known to reduce crop yield.
They utilized machine vision to distinguish infected seedlings from
healthy ones at early developmental stages using SVM classiers. A
genetic algorithm was employed to optimize model parameters and
select key morphological and color traits. Their method achieved an
accuracy of 87.9 % with a positive predictive value of 91.8 %, demon-
strating the potential of automated disease detection in agricultural
settings. Qasrawi et al. (2021) [28] applied machine learning techniques
to detect and classify diseases affecting tomato plants using smartphone
images. Their study focused on ve common tomato diseases, including
Botrytis cinerea, and utilized a dataset of 3000 images collected from
various regions. The models employed included neural networks, lo-
gistic regression, and clustering techniques. The clustering method
achieved a 70 % accuracy in grouping the diseases, while the neural
network and logistic regression models reported classication accu-
racies of 70.3 % and 68.9 %, respectively. For the image classication
task, Lorente et al. (2021) [29] employed standardized descriptor
extraction methods like SIFT and SURF [30] and then applied Bag of
Visual Words (BoVW) [31] classication with multiple algorithms,
reaching accuracy levels ranging from 0.6 to 0.96 depending on the
model and conguration. Hyperspectral imaging has also been applied
effectively in other plant disease detection studies. For instance, Naga-
subramanian et al. (2019) [32] utilized a novel 3D deep convolutional
neural network (DCNN) architecture to classify charcoal rot-infected
soybean stems. Their model achieved a classication accuracy of
95.73 %, with an F1-score of 87 % for the infected class. A signicant
aspect of their study was the use of saliency maps to identify the most
sensitive wavelengths contributing to classication accuracy. The re-
sults indicated that wavelengths in the near-infrared (NIR) region,
particularly around 733 nm, were critical for identifying infected sam-
ples, while wavelengths in the visible spectrum (400–700 nm) were
found to be more sensitive for infected samples compared to healthy
ones, demonstrating the physiological relevance of these spectral re-
gions for disease detection. In a more recent research, Nguyen et al.
(2021) [33] applied hyperspectral imaging and DL techniques for the
early detection of grapevine viral diseases, specically focusing on the
grapevine vein-clearing virus (GVCV). Their work demonstrated the
effectiveness of hyperspectral imagery in distinguishing between
healthy and GVCV-infected grapevines at early asymptomatic stages.
The study identied several key wavelength regions that were particu-
larly useful for classication, for both early and later infection stages,
including the 900–940 nm range in the NIR region and the 449–461 nm
range in the visible spectrum (VIS). These wavelengths were found to be
strongly associated with physiological changes in the vines, such as
reduced leaf water content and pigment degradation, which affected the
reectance patterns. These studies demonstrate the potential of hyper-
spectral imaging for the early detection of plant diseases, a concept that
aligns closely with our approach in detecting Botrytis cinerea.
In this study, the main objective is the effective detection of
B. cinerea, especially during the early stages of fungal infection. The
bioassays consist of non- and B. cinerea-inoculated tomato plants culti-
vating in two separate controlled environment plant growth chambers
simulating the environment typically prevailing in greenhouses. Disease
severity was assessed over 42 days post-inoculation for each leaet and
leaf using a linear disease progression model. The dataset consists of
images of healthy, articially and naturally inoculated with B. cinerea
tomato plants captured on 5 hyperspectral wavelengths and in RGB
mode. Each leaet is segmented and extracted using YOLOv8 [34], and a
suite of Transformer models is employed to extract a descriptor for each
leaet. Finally, two descriptor classication solutions – one based on the
KNN algorithm and one based on an LSTM model – are applied for
classifying the leaves into three distinct classes (healthy,
invisible-botrytis, visible-botrytis). To further improve the classication
performance and ensure robustness, the classication results of the
D. Kapetas et al.
Smart Agricultural Technology 9 (2024) 100580
2
multiple images for each capture (ve hyperspectral and one RGB) and
of the multiple descriptors from each Transformer model, of either
classication solution, are ensembled, applying a weighted voting al-
gorithm to extract the best result, effectively combining the classica-
tion predictions of the multiple image wavelengths and the multiple
Transformer models. Additionally, grey mould detection was assessed
by comparative qPCR-based methods (qRT-PCR) for fungal biomass
estimation in plant tissues at the early stages of the infection when
symptoms are not visible and the rest of this study’s methodology.
Reverse transcription quantitative PCR (RT-qPCR) is widely recog-
nized as a valuable standard for precise, sensitive, and rapid gene
expression measurement. The method involves converting mRNA to
complementary DNA (cDNA) using the enzyme reverse transcriptase,
followed by the PCR amplication of the cDNA. It is considered as the
most reliable technique for conducting simultaneous measurements of
the relative levels of gene transcripts in many different samples because
of its efciency and sensitivity [35,36]. Compared to conventional
methods, RT-qPCR is the only method available for detecting low copy
number mRNA of selected genes [37]. It has been used for pathogen
detection like bacteria, viruses and fungi and gene expression analyses
in plant tissues and soil [38,39]. In many fungi, the above method has
become a frequent rst choice for gene expression studies; the general
molecular strategy for studying fungal biology involves evaluation of
gene expression levels in order to establish correlations between the
transcript levels of specic genes and the fungal responses and adapta-
tions to environmental conditions [40].
In contrast to previous works by Giakoumoglou et al. (2024) [24]
and Christakakis et al. (2024) [25], that focused on a single model for
the semantic segmentation task, this study approaches the problem by
using a segmentation model to extract the leaf information and then an
ensemble of Transformer models to classify those segments. Further-
more, qPCR-based methods have been employed, quantifying fungal
biomass in plant tissues based on relative expression of fungal reference
genes via qRT-PCR to further validate and conrm grey mould detection
at early stages or at latent infections when symptoms are not visible. So,
by comparing qPCR-based methods and segmentation model results, this
current study aims to provide a more robust benchmark for DL models,
yielding valuable insights and improvements over previous research.
The rest of the paper is structured as follows: Section 2 outlines the
materials and methods employed in this study Section 3 presents the
results of the study, while Section 4 concludes the study, summarizing
key ndings and implications.
Materials and methods
Articial inoculation processes and experimental setup
Tomato (Solanum lycopersicum L.) cv. Darren (ВЕХ−24) F1 plants
Fig. 1. High-level owchart, presenting the proposed methodology. The RGB and the ve hyperspectral images are fed to the YOLOv8 model to extract the single leaf
images, which are then fed to the Transformer models to extract the descriptors of each leaf image. Then the descriptors pass through the classier to extract each
leaf’s class on a per wavelength, per model basis, and then a weighted voting ensemble is applied to those classications to produce the nal class of each leaf.
Finally, the information for all the leaves is collected back and applied on the original image.
D. Kapetas et al.
Smart Agricultural Technology 9 (2024) 100580
3
were used. At the rst-true-leaf stage, when the rst set has fully
expanded, and the second set of true leaves are visible, seedlings were
transplanted into plastic pots (diameter 24 cm, height 24 cm, capacity 7
L) containing sterile compost (Potground; Klasmann). The plants were
maintained in controlled environment growth chambers at 21 ±1 ◦C, 16
h light at 352.81
μ
mol/sec photosynthetic photon ux and 85 %–90 %
relative humidity, irrigated twice a week and kept under these condi-
tions resembling those of greenhouse-grown plants [24].
А Botrytis cinerea isolation from naturally infected cucumber plants
causing grey mould disease was used. For inoculum production,
B. cinerea was cultured on PDA in petri dishes at 21 ◦C for 10–15 days in
darkness. Spore suspensions were prepared in sterile distilled water
(SDW) containing 2 % sucrose and 0.01 % Tween® 20 (Sigma-Aldrich,
United States) [41]. Conidial concentration was adjusted to 10
5
conidia
ml
-1
[24].
Fig. 1.
Following a completely randomized design, the bioassays were
conducted at Benaki Phytopathological Institute (BPI), in two controlled
environment plant growth chambers with the same set of environmental
conditions (21 ±1 ◦C, 16 h light/8 h dark photoperiod and 85 %–90 %
relative humidity); one for B. cinerea-inoculated plants and the other for
mock-inoculated (water treated) plants (Fig. 2).
Plants at the stage of four fully expanded leaves were articially
inoculated with the pathogen by spraying the adaxial surface of leaves
with the conidial suspension to the stage of run-off using a low-pressure
hand sprayer (approximately 500
μ
L per leaf of spore suspension) [24].
Initially, the adaxial surface of the rst and second true leaves was
sprayed with the conidial suspension. Since plant defense responses
resulted in limitation of the infection on these two rst leaves, a second
articial inoculation was decided to be carried out on the 3rd, 5th and
13th leaf of each plant. For mock inoculated plants, the same procedure
was followed without the addition of B. cinerea inoculum.
In planta assays for disease assessment and image acquisition
The bioassay comprised of six tomato plants inoculated with the
pathogen numbered from 1 to 6, and three mock-inoculated tomato
plants numbered from 1 to 3. The rst thirteen leaves of each plant and
all the fully expanded leaets of each leaf were also numbered. Disease
severity (%) estimated as the percentage of leaf area infected with
visible disease symptoms, was assessed twice a week until the end of
plant’s growing season (ca two months). For each leaet and leaf, dis-
ease severity was assessed over 42 days post-inoculation (dpi), quanti-
fying the percentage of leaf area showing visible disease symptoms. The
data were analyzed using a logit transformation to linearize disease
progression in relation to time [42]. Linear regression was employed to
Fig. 2. In planta assays. (a) Mock- and (b) B. cinerea-inoculation of young tomato plants. Grey mould symptoms at early (c) and late stage (d) of the disease.
Fig. 3. Plant Image captured at distinct wavelengths: (a) RGB, (b) 460 nm, (c) 540 nm, (d) 640 nm, (e) 775 nm, (f) 875 nm.
D. Kapetas et al.
Smart Agricultural Technology 9 (2024) 100580
4
calculate the rate of disease progression and to estimate the infection
onset time on each leaf.
Multi-spectral imaging was used to capture both early and late
symptoms of B. cinerea. The rst set of images was taken on the 3rd dpi,
followed by subsequent captures on 4, 5, 11, 18, 25, 32, and 39 dpi. The
captured images include individual leaets and leaves, and whole
plants, clearly depicting all of their leaves. A total of 174 and 104 cap-
tures were collected from B. cinerea - and mock-inoculated plants,
respectively.
The utilized equipment for the multi-spectral image captures was a
customized Qcell [35] Phenocheck camera. This camera was tuned to
capture six images from the visible and NIR spectrum at a 3000 by 1900
pixels resolution. Specically, the captured ve wavelengths at 460 nm,
540 nm, 640 nm, 775 nm and 875 nm. And one RGB image. An example
of the individual spectra of a B. cinerea-inoculated plant can be seen in
Fig. 3.
In planta assays for gene expression measurements and image acquisition
at early stages of inoculation
The bioassay initially comprised of three plants (biological repli-
cates) per treatment (mock- or B. cinerea -inoculated by spraying the 1st,
2nd, 3rd, 5th and 13th leaf of each plant) and per time point (3, 4 and 5
dpi).
For multi-spectral imaging and gene expression measurements, the
mock- and B. cinerea -inoculated leaves were collected from each tomato
plant, treatment (mock- and B. cinerea -inoculated leaves) and each time
point (3, 4 and 5 dpi), used immediately for multi-spectral imaging, ash
frozen and stored at −80 ◦C.
The gene expression was performed with reverse-transcription
quantitative PCR (RT-qPCR) in the above samples. Initially, leaf sam-
ples were pulverized under liquid nitrogen and total RNA was isolated
using NucleoSpin® RNA Plant and Fungi Kit (Macherey-Nagel GmbH &
Co) according to the manufacturer’s instructions. Puried RNA was
quantied and checked using a NanoDrop® ND-1000 Spectrophotom-
eter while the RNA integrity was also assessed by agarose gel electro-
phoresis [43]. cDNA synthesis with elimination of traces of
contaminating genomic DNA was performed using PrimeScript™ RT
reagent Kit with gDNA Eraser for Perfect Real Time (TAKARA) according
to the manufacturer’s instructions.
Quantitative RT-qPCRs were run on the Applied Biosystems StepO-
nePlus Real-Time PCR thermocycler using the master mix KAPA SYBR®
FAST qPCR Master Mix (2X) (KAPABIOSYSTEMS). Each reaction con-
tained 5
μ
L of FAST qPCR Master Mix (2X), 0.2
μ
L of 10
μ
M of each gene-
specic primer pair, 0.2
μ
L of 50X ROX High and 1
μ
L of cDNA template
to a nal volume of 10
μ
L.
Fungal biomass estimated by the relative expression of B. cinerea
reference gene BcRPL5 (Bcin01g09620) [9]. The relative expression of
the disease responsive tomato gene SlWRKY33 (Solyc09g014990) was
also estimated in samples inoculated with the pathogen and in the
mock-inoculated control [9]. Relative gene expression levels were
calculated by using the Comparative CT (2
-ΔΔCT
) method [43,44]. All
primers used are described in Supplementary Table S1. All reactions
were performed in duplicate. The absence of non-specic products and
primer dimers was conrmed by the analysis of melting curves. As
reference gene to normalize transcript level for each sample, the
housekeeping tomato gene SlUBQ was used [9].
Image annotation
A total of 278 images were annotated using Roboow [45]. The
annotations encapsule each whole leaf area on the images with poly-
gons. Leaves were categorized into distinct classes based on the pro-
gression of B. cinerea disease. Leaves from mock-inoculated plants were
labeled as "healthy". Leaves from B. cinerea-inoculated plants that
showed no disease symptoms were also classied as "healthy," while
those with 0.1–5 % disease severity were labeled as "botrytis-invisible,"
indicating that the disease is in a stage with no visible symptoms. Leaves
from inoculated plants that showed 5 % to 100 % infected leaf area with
visible disease symptoms were marked as “botrytis-visible”. The
“healthy”, “botrytis-invisible” and the “botrytis-visible” leaves represent
Class 0, Class 1 and Class 2, respectively. Fig. 4 depicts an RGB image
with its true annotations mask overlaid.
Leaf segmentation
In computer vision tasks, image segmentation is a crucial step. It is
the process of splitting an image into distinct segments-objects [46].
This process involves a DL model capable of processing a dataset of
images and understanding image features. An example of such a model
is YOLO, which learns a generalizable representation of objects on an
image and presents a solution with a good balance between accuracy
and inference speed.
In this study, the state-of-the-art YOLOv8 model was employed.
Specically, the pretrained YOLOv8-Small model was used for the seg-
mentation task. The training on the multispectral annotated dataset was
conducted for 100 epochs with batch sizes of 2, 4 and 8 and image
resolutions of 1024 ×1024 pixels and 1600 ×1600 pixels. The SGD
optimizer was used with an initial learning rate of 0.01. Other specied
parameters included a cosine learning rate schedule, disabling the
mosaic augmentation after 20 epochs, a weight decay of 0.0005, and a
maximum detection threshold of 150.
Dataset preprocessing and augmentation
To classify individual leaves using manual annotations, each leaf was
isolated from its original image and placed into a new image. In these
new images, all areas outside the leaf were made completely transparent
(dataset-a). The images were then resized to 384 ×384 pixels to ensure
compatibility with the Transformer models described in Section 2.7.
However, after resizing from their original resolution of 3200 ×1900
pixels to 384 ×384, the leaf became a small portion of the image,
resulting in too few pixels to capture its details effectively.
To address this, two zooming techniques were applied. The rst
method involved cropping the images to half of their original width and
height, effectively zooming in by a factor of four (dataset-b). This
approach was chosen because it preserved more detail while still tting
even the largest leaves in the dataset. The second technique involved
cropping the images directly around the leaf area (dataset-c). This
method maximized the information retained for each leaf but had the
disadvantage of losing leaf size context, as every leaf lled the available
space in the image. Fig. 5 shows an example of a leaf image with no
zoom, four-times zoom, and cropping around the leaf.
Extracting all leaves from the original images resulted in a dataset
Fig. 4. Ground truth annotations overlaid on an RGB image. Green leaves
indicate “healthy” class, blue leaves indicate “botrytis-invisible” class, and red
leaves indicate “botrytis-visible” class.
D. Kapetas et al.
Smart Agricultural Technology 9 (2024) 100580
5
consisting of 33,360 images of the “healthy” class, 1242 images from the
“botrytis-invisible” class and 1398 images from the “botrytis-visible”
class. This dataset includes images captured from all ve hyperspectral
wavelengths (460 nm, 540 nm, 640 nm, 775 nm and 875 nm) as well as
the RGB images, with each leaf covering at least 0.35 % of the image
area.
For the classication process, maintaining a relative balance among
the count of images of each class is essential. For the case of this study,
dataset augmentation techniques were considered, to increase the
number of images in the “botrytis-invisible” and the “botrytis-visible”
classes. The case of applying rotational and scaling augmentations was
rejected on the basis that the gap between the number of images in those
two classes and the number of images in the “healthy” class was too big,
and the coverage would lead the classication models presented in
Section 2.8 to overt. Also, brightness, blurring and other color-based
augmentations were rejected because applying them would defeat the
purpose of having the unique grayscale color values that are provided by
the hyperspectral imaging. Therefore, the opted solution was to reduce
the number of the “healthy” class images. In the newly created dataset, it
was also important to ensure representation from each day and to pre-
serve all six images of each capture. To achieve this, approximately 5 %
of the unique leaves for each day were retained, where a unique leaf is
dened by the total set of all six images. This resulted in the "healthy"
class being reduced to 1662 images. Finally, the images were split into
training and testing sets for all classes, with approximately 80 % of the
images allocated for training and 20 % for testing. The split was per-
formed on a per-day basis to ensure proper representation of each day in
the nal result. Additionally, different wavelength variations of the
same image were grouped together in either the training or the testing
set, to avoid biased results. Table 1 provides details on the distribution of
classes in the training and testing sets for each day.
Descriptor extraction
To extract a descriptor from each individual leaf, a suite of Trans-
former models was utilized. A descriptor is a numerical array that
characterizes the content of an image with the output of each machine
learning model being a descriptor and the global descriptor describing
the overall features of an image [47]. More specically for this study,
pretrained models from the timm [48] library were employed. These
models, pretrained on ImageNet-1 K [49], initially produce a descriptor
of length 1000, reecting the model’s condence in classifying the
image into each of the 1000 ImageNet-1 K classes. Therefore, to obtain
the actual global descriptor for each image, the nal classication layer
was removed from each model. This adjustment allowed the models to
provide the comprehensive feature representation of the images. Table 2
details the transformer models used and their respective descriptor sizes.
Descriptor classication
Since there are six different descriptors to describe each leaf image
(ve from the hyperspectral wavelengths and one from the RGB image),
a total of six separate classication actions will be done for each leaf
image. To classify the aforementioned extracted descriptors, two
different approaches were employed.
The rst approach involved KNN classication [50]. This method
constructs a BoVW, essentially a digital library that stores all the
available descriptors from the training set, thereby forming a collection
of visual features, with a separate BoVW created for each distinct
wavelength. Descriptors from the testing set are compared against each
element in the corresponding BoVW for the same wavelength using a
distance function, and the resulting distances are recorded. The classi-
cation result is determined by averaging the K number of smallest
distances. In this study, the distance functions employed were the
Euclidian distance, the Bray-Curtis distance and the Cosine distance.
The second approach utilized a Long Short-Term Memory (LSTM)
Sequential model implemented with Keras [51]. This approach involved
training an LSTM network on the descriptors to learn relevant features
from the descriptors through an LSTM layer, followed by a Batch
Normalization layer, and then another LSTM layer. The output was
mapped to the corresponding image label through a Dense layer with a
linear activation function. The LSTM model was trained iteratively,
separately for each wavelength, while a separate model was trained for
each descriptor extraction model used. The initial learning rate of the
Fig. 5. Example leaf image from datasets (a): with no zoom, (b): with four-times zoom, (c): with crop around the leaf area. (a) also depicts visually the extraction of a
single leaf from the original image.
Table 1
Distribution of leaf dataset images per class and set across different dpi values.
Training Set Testing Set
Dpi Class 0 Class 1 Class 2 Class 0 Class 1 Class 2
3 150 66 0 36 30 0
4 144 90 0 36 42 0
5 132 42 0 36 60 0
11 174 168 48 42 36 18
18 168 234 138 48 12 42
25 168 204 240 48 18 66
32 168 114 312 48 24 102
39 210 72 366 54 30 66
Total Images 1314 990 1104 348 252 294
Table 2
Transformer models used and their descriptor size.
Transformer Model Descriptor Size
ViT-L (P:16 ×16) 1024
ViT-L (P:32 ×32) 1024
ViT-B (P:16 ×16 - C) 1024
ViT-B (P:32 ×32 – C) 1024
Swin-L (P:4 ×4 – W:12 ×12) 1536
VOLO (D:5) 768
XCIT-L (L:24 – P16 ×16) 768
DEIT-L (P16 ×16) 1024
MaxViT-L 1024
D. Kapetas et al.
Smart Agricultural Technology 9 (2024) 100580
6
models was set to 0.001 and a learning scheduler was applied. The
training was conducted for 12 epochs per wavelength with a batch size
of 128.
Furthermore, a Per Class Condence Multiplier (PCCM) was applied
to the classication results, to address class imbalance due to varying
image counts across classes. As shown in Table 1, “healthy” class has
1314 training images, “botrytis-invisible” class has 990 training images
and “botrytis-visible” class has 1104 training images. The PCCM is
calculated as the inverse of the number of training images in a given
class, multiplied by the number of images of the class with the most
images (“healthy” class with 1314 training images). The formula for this
calculation is shown in Eq. (1), while the resulting PCCM is shown in
Table 3.
PCCMi=Nmax
Ni
,i∈ [0,2](1)
Where Ni is the number of training images in class i and Nmax is the
number of training images in the class with the most images.
Evaluation metrics
This study involves classication of the images into three distinct
classes, where each image is assigned to exactly one class, and the
models always return a prediction for one class. Therefore, the relevant
metrics extracted based on the predictions are True Positives (TP), False
Positives (FP), True Negatives (TN) and False Negatives (FN)
The formulas for calculating Accuracy, Precision, Recall and F1-
Score [52] are as follows:
Accuracy =TN +TP
TN +FP +TP +FN
Precision =TP
TP +FP
Recall =TP
TP +FN
F1Score =2×Precision ×Recall
Precision +Recall
The testing set of descriptors was processed through both the KNN
and the LSTM models to generate predictions for each descriptor/image.
Each prediction was then compared to the ground truth class derived
from the manual annotation and the TP, FP, TN and FN were calculated
separately for each class on each day, allowing a precise evaluation of
the results per class and per day. Subsequently, the Accuracy, the Pre-
cision and the F1-Score were calculated using the TP, FP, TN and FN
counts. These metrics were computed on various abstraction levels,
including for each class on each day, for each class on all days and for all
classes on all days totally.
Ensemble techniques
Up to this point, predictions were made on a per model per wave-
length basis, resulting in different predictions for each model and six
separate predictions for each leaf image (one for each of the ve
hyperspectral wavelengths and one for the RGB image).
To improve performance and achieve robustness, an ensemble
methodology was employed. This approach aggregates multiple pre-
dictions to produce the nal result. For each nal prediction, multiple
models and multiple wavelengths are used iteratively, and their results
are stored separately. This setup allows for further ensemble results to be
calculated from other ensembled results. For each ensembled result, the
prediction is determined by calculating the condence of each class and
then selecting the class with the highest condence. The condence
score for each class is determined using the formula in Eq. (2), where the
nal condence is computed as the sum of the predictions that identied
the class, with each prediction weighted by the predicted class F1-Score
of the corresponding model and wavelength. An iterative procedure
applies this methodology to all possible ensemble combinations of
descriptor sets across multiple combinations of multi-wavelength com-
putations, in order to identify the optimal combination for the nal
prediction.
Ci=∑
W
k=1∑
M
j=1(pj,k⋅Aj,k)(2)
where Ci denotes the condence score for class i across all models and
wavelengths, pj,k indicates whether the prediction made by model j for
wavelength k corresponds to class i (1 if it does, 0 otherwise), Aj,k refers
to the F1-Score of model j for wavelength k, M represents the total
number of models, and W represents the total number of wavelengths.
Results and discussion
Disease severity
Disease severity was assessed per leaet of all thirteen leaves for each
tomato plant as the percentage of leaf area with visible symptoms of grey
mould to the total leaf area and was periodically recorded for 42 dpi.
Disease progress was assessed at 3 to 5-day intervals.
In the majority of plants, no disease progress was recorded after
articially inoculation on the 1st leaf. Additionally, most of the 2nd, 3rd
and 5th articially inoculated leaves were progressively infected natu-
rally later in time (Fig. 6). On the contrary, in all plants, the 13th arti-
cially inoculated leaf was successfully infected and showed the rst
symptoms after 7 dpi (Fig. 6). Οn each plant, non-inoculated leaves were
infected later in time and started having visible symptoms of the disease
approximately in a period of 14 to 26 dpi (Fig. 5, 9th and 11th leaf). The
above observed variations in the disease progress were due to leaf age
and position on stems, leaf content in nutrients, the plant defense re-
sponses against the pathogen and the secondary cycles of the disease
that were initiated by airborne conidia released from the rst grey
mould lesions developing on the inoculated leaf tissues. Conidia of
B. cinerea play a major role in the rapid colonization of neighboring
plants or plant parts when conditions are favorable. They are released
during variation of air humidity or by other mechanic forces related to
wind or splash, or human activities and are easily disseminated by air
currents, while their survival is highly dependent on temperature,
moisture and exposure to sunlight/light [53]. Also, B. cinerea is most
destructive on mature or senescent tissues of dicotyledonous hosts, but it
usually gains entry to such tissues at a much earlier stage in crop
development and remains quiescent for a considerable period before
rapidly rotting tissues when the environment is conducive and the host
physiology changes [54].
Tomato susceptibility to Botrytis cinerea infection
In the present study, the progression of B. cinerea infection in tomato
leaves at the early stages of the infection onset was studied at 3, 4 and 5
dpi, a period when grey mould symptoms are not visible. This was
accomplished by the study of the relative expression of the B. cinerea
reference gene BcRPL5 (Bcin01g09620) via RT-qPCR [9].
The presence of the fungus was detected in plant tissues of the 3rd,
5th and 13th leaves (Fig. 7a-c). In the plant tissues of the 3rd and 13th
Table 3
Condence multiplier per class.
Class Train Images Condence Multiplier
0-healthy 1314 1
1-botrytis-invisible 990 1.327
2-botrytis-visible 1104 1.193
D. Kapetas et al.
Smart Agricultural Technology 9 (2024) 100580
7
Fig. 6. Disease progress in terms of disease severity (% leaf area showing grey mould symptoms), is presented for a period of 42 dpi. Initially, plants were articially
inoculated with a conidial suspension on the 1st and 2nd leaf, while a second inoculation was conducted on the 3rd, 5th and 13th leaf. Accidental spread of the
disease was recorded on the other leaves (e.g., 9th and 11th leaf).
D. Kapetas et al.
Smart Agricultural Technology 9 (2024) 100580
8
leaf, the fungal biomass showed a signicant increase over time (from 3
to 5 dpi) (Fig. 7a, c), while there was an increase in fungal biomass that it
was not signicant in the 5th leaf, (Fig. 7b). These ndings coincide with
the phenotypic observations according to which the 5th leaf did not
show any symptoms of the disease, indicating either the latent phase of
the pathogen or the host’s response to the pathogen’s attack. The above
variations in fungal biomass may be due to the fact that B. cinerea
infection can proceed by several different routes which vary according
to the plant species, tissue type, and external conditions. Additionally,
the disease development is multilayered and regulated by multiple
factors, with subtle contributions from virulence factors. On the other
hand, plant defense is activated early on; despite the lack of complete
resistance against B. cinerea, plant immunity has the potential to reduce
and even prevent disease development [55].
Although there were no visible symptoms in tomato leaves ve days
after articial inoculation with B. cinerea, some differences in fungal
biomass and tomato responses were observed. To provide initial support
that the pathogen is capable of inducing disease responses in the host,
and does not merely behave as a saprotroph (i.e., feeding on dead tis-
sue), the expression of the host gene SlWRKY33 (Solyc09g014990) was
evaluated, which is well-known to be pathogen-responsive but is not
induced by abiotic stresses [56,57]. To test that the induction of this
gene occurred only as a result of inoculation and not wounding, a
mock-inoculated control was included in the analyses. The expression
patterns of SlWRKY33 measured by qRT-PCR reected the accumulation
of fungal biomass in each of the treatments (Fig. 8). The comparison
between B. cinerea– and mock-inoculated leaves showed that the
expression of SlWRKY33 gene was signicantly reduced by B. cinerea
inoculation in the 3rd and especially the 13th leaf at 3dpi, while the gene
was overexpressed in all tested leaves at 5dpi and especially in 13th leaf
at 4 and 5 dpi (Fig. 8). Also, the host’s responses to the pathogen inoc-
ulation seem to reduce over the time in all tested leaves (Fig. 8). Ac-
cording to the above, the pathogen is capable of inducing disease
responses in the host that are variable depending on the leaf age and the
time point after inoculation. It was recently reported that the over-
expression of SlWRKY33 regulates cold tolerance in tomatoes in a pos-
itive manner [58].
The 3rd and the 13th leaf showed an increase of B. cinerea biomass in
their tissues (Fig. 7). Additionally, the 13th articially inoculated leaves
were the most susceptible to the pathogen infection and showed the rst
symptoms after 7 dpi (Fig. 6). So, the variations in B. cinerea biomass in
tomato tissues and the host’s responses to the pathogen’s attack seem to
be in line with the variations observed in the disease progress. Specif-
ically, the 13th true leaves were the most susceptible to B. cinerea
infection as they showed the highest intensity of grey mould symptoms
over a period of 42 dpi and had increasing fungal biomass in their tissues
Fig. 7. Quantication of fungal biomass expressed as the relative transcript abundance of the fungal reference gene BcRPL5 (Bcin01g09620) in mock- and Botrytis
cinerea-inoculated (A) third, (B) fth and (C) thirteenth tomato leaves at 3, 4 and 5 dpi via RT-qPCR, using the Comparative CT (2 −ΔΔCt) method. The columns
represent the means of three independent leaf samples per treatment (mock- and B. cinerea-inoculated). The vertical bars indicate standard errors and columns with
different letters are statistically different according to Fisher’s LSD multiple range test at P ≤0.05.
Fig. 8. Relative expression of the disease responsive tomato gene SlWRKY33 (Solyc09g014990) in mock- and Botrytis cinerea-inoculated (A) third, (B) fth and (C)
thirteenth tomato leaves at 3, 4 and 5 days post-inoculation (dpi) via RT-qPCR, using the Comparative CT (2 −ΔΔCt) method. The columns represent the means of
three independent leaf samples per treatment (mock- and B. cinerea-inoculated). The vertical bars indicate standard errors and columns with different letters are
statistically different according to Fisher’s LSD multiple range test at P ≤0.05.
D. Kapetas et al.
Smart Agricultural Technology 9 (2024) 100580
9
and decreased immune responses at the early stages of the infection.
Leaf segmentation results
Using the YOLOv8 model, the best results were achieved with the
YOLOv8-Small variant, using an input resolution of 1600 ×1600 pixels
and a batch size set of 4. The metric results of the various runs are
presented in Table 4, while Fig. 9 depicts a preview of the best model’s
leaf segmentation on a sample image.
Table 4
Performance evaluation (mAP50, recall and precision) of the YOLOv8 model.
Model Image Size mAP
50
mAP
50–95
Recall Precision
YOLOv8s-seg 1024 0.776 0.489 0.684 0.781
YOLOv8l-seg 1024 0.795 0.499 0.703 0.780
YOLOv8s-seg 1600 0.817 0.547 0.745 0.781
Fig. 9. (a) original images and (b) the YOLOv8 segmentation results on images. The images were captured at a wavelength of 875 nm. The red marks represent the
shapes that the model found as leaves.
Table 5
Accuracy (Acc) and F1-Score (F1) metrics on a per Transformer model, per solution (KNN/LSTM), per dataset basis.
(a) No Zoom Dataset (b) 4 times Zoom Dataset (c) Leaf Zoom Dataset
KNN LSTM KNN LSTM KNN LSTM
Models Acc % F1 % Acc % F1 % Acc % F1 % Acc % F1 % Acc % F1 % Acc % F1 %
ViT-L (P:16 ×16) 66.23 49.35 67.10 50.65 70.92 56.38 69.13 53.69 68.60 52.90 69.13 53.69
ViT-L (P:32 ×32) 67.10 50.65 63.64 45.45 69.13 53.69 68.23 52.35 64.73 47.10 68.23 52.35
ViT-B (P:16 ×16 - C) 63.64 45.45 64.94 47.40 71.81 57.72 71.36 57.05 68.60 52.90 65.16 47.74
ViT-B (P:32 ×32 – C) 65.37 48.05 65.80 48.70 72.26 58.39 71.36 57.05 70.75 56.13 65.16 47.74
Swin-L (P:4 ×4 – W:12 ×12) 59.31 38.96 61.04 41.56 67.79 51.68 66.89 50.34 68.60 52.90 65.16 47.74
VOLO (D:5) 64.94 47.40 66.23 49.35 69.57 54.36 69.57 54.36 66.45 49.68 69.89 54.84
XCIT-L (L:24 – P16 ×16) 66.23 49.35 67.10 50.65 66.00 48.99 66.00 48.99 69.03 53.55 63.87 45.81
DEIT-L (P16 ×16) 63.20 44.81 60.17 40.26 70.47 55.70 69.57 54.36 64.30 46.45 63.44 45.16
MaxViT-L 64.50 46.75 61.04 41.56 69.57 54.36 66.89 50.34 69.03 53.55 64.30 46.45
Table 6
Accuracy (Acc) and F1-Score (f1) metrics on a per Transformer model, per solution (KNN/LSTM), per dataset basis, when applying the PCCM.
(b) 4 times Zoom Dataset (c) Leaf Zoom Dataset
KNN LSTM KNN LSTM
Models Acc % F1 % Acc % F1 % Acc % F1 % Acc % F1 %
ViT-L (P:16 ×16) 70.20 54.52 74.74 61.34 69.67 53.72 66.17 48.48
ViT-L (P:32 ×32) 70.20 54.52 69.29 53.16 67.04 49.79 67.48 50.45
ViT-B (P:16 ×16 - C) 71.11 55.89 70.20 54.52 69.23 53.07 68.79 52.41
ViT-B (P:32 ×32 – C) 73.83 59.98 67.02 49.75 70.98 55.69 67.48 50.45
Swin-L (P:4 ×4 – W:12 ×12) 70.20 54.52 68.84 52.48 67.48 50.45 64.42 45.86
VOLO (D:5) 67.02 49.75 71.56 56.57 67.48 50.45 64.42 45.86
XCIT-L (L:24 – P16 ×16) 67.47 50.43 67.93 51.12 69.23 53.07 66.17 48.48
DEIT-L (P16 ×16) 70.65 55.21 67.47 50.43 62.24 42.59 67.48 50.45
MaxViT-L 71.11 55.89 66.11 48.39 70.10 54.38 63.11 43.90
D. Kapetas et al.
Smart Agricultural Technology 9 (2024) 100580
10
Classication performance evaluation
The KNN and LSTM solutions were applied to dataset-a, dataset-b
and dataset-c and all the descriptors were classied. The classication
results from the ve hyperspectral wavelengths and from the RGB image
for each separate leaf were then ensembled. The extracted performance
metrics from the classication include Accuracy, Precision, Recall and
F1-Score. To compare models with one another, the F1-Score was used
as the comparative factor, as it is the harmonic mean of precision and
recall and can present a better evaluation metric for each model’s
quality.
The resulting F1-Scores ranged from 38.96 % to 58.39 %, as shown in
Table 5. The data indicates that dataset-a performed signicantly worse
than the other two datasets, with an average F1-Score of 12.45 % and
7.55 % less than the average F1-Score of dataset-b and dataset-c,
respectively. This is expected since the pixels that are covered by the
leaf area in dataset-a are considerably less than those in dataset-b and
dataset-c. Consequently, further experiments with this dataset were not
conducted. Also, dataset-b seems to be outperforming dataset-c, indi-
cating that leaf size and shape are more important for the classication
process than preserving the maximum number of pixels. For the KNN
solution, the Euclidian distance function was used with a K value of 3.
Applying the PCCM lead to 3.5 % average improvement in dataset-b
for the KNN solution, and to 0.1 % average improvement for the LSTM
solution. As a result, all following experimentations incorporate PCCM.
The full results are shown in Table 6.
For the KNN solution, no training was required, and a separate BoVW
was employed for each individual Transformer model and for each of the
six distinct images corresponding to a single leaf. Therefore, for the KNN
solution, each classication of a leaf from an image of one wavelength,
was not at all affected by the existence of the other ve leaf images.
Table 7 depicts the F1-Score classication percentage of the KNN solu-
tion separately for each of the six image wavelengths for Transformer
model.
In contrast, for the LSTM solution, a separate model was trained for
each descriptor set (i.e., each of the descriptor extraction models).
Consequently, for each classication task, the accumulated knowledge
from all six images (one RGB and ve hyperspectral images) contributed
to the nal result. However, as observed in the average F1-Score values
across all models in Table 7, there were signicant uctuations in per-
formance across the various wavelengths. These uctuations may
negatively impact the LSTM model’s training process.
Therefore, in order to further improve the LSTM models’ perfor-
mance, multiple runs were performed, selectively removing certain
wavelengths from the calculations, specically iterating on removing
every possible combination of the wavelengths of the 460 nm, 775 nm
and 875 nm wavelengths, which are the 3 worst performing ones. The
best results were achieved by excluding the wavelengths 775 nm and
875 nm, retaining the remaining four. This approach improved the per
model average F1-Score by 3.21 % and 0.37 % for dataset-b and dataset-
c, respectively. Thus, any further results involving the LSTM approach
excluded the 775 nm and 875 nm wavelengths. The results suggest that
these two NIR wavelengths were less effective in capturing meaningful
information compared to the visible and lower NIR wavelengths for the
specic task. By retaining only the more informative wavelengths, the
model’s ability to differentiate between classes was improved. The nal
classication accuracies and F1-Scores for the four remaining wave-
lengths are presented in Table 8.
The metrics presented in Tables 5 and 6 for the KNN solution on
Tables 5 and 6, were all extracted using the Euclidian distance function,
that proved to provide the best results. On the dataset-b, the Euclidean
distance function outperformed the Bray-Curtis and Cosine distance
functions by a per model average of 0.98 % and 1.54 %, respectively.
Detailed performance metrics are shown in Table 9. Similarly, the K
value of 3 was chosen, since it was 5.20 %, 0.89 % and 0.07 % more
performant than the K value of 1, 5 and 7 respectively. The full metrics
for the various K values are shown in Table 9.
Table 7
F1-Score (%) of the KNN solution on a per Transformer model, per wavelength
basis.
Models RGB 460
nm
540
nm
640
nm
775
nm
875
nm
ViT-L (P:16 ×16) 43.85 45.19 46.09 53.24 42.95 43.40
ViT-L (P:32 ×32) 48.10 42.51 49.44 51.45 40.27 42.06
ViT-B (P:16 ×16 - C) 46.98 44.74 46.31 48.32 41.16 41.61
ViT-B (P:32 ×32 – C) 47.87 41.83 43.85 50.11 46.09 45.19
Swin-L (P:4 ×4 –
W:12 ×12)
43.62 41.83 41.39 46.98 42.28 41.61
VOLO (D:5) 40.94 42.06 39.60 45.41 40.49 38.93
XCIT-L (L:24 – P16 ×
16)
45.86 38.70 41.39 44.74 37.81 38.26
DEIT-L (P16 ×16) 41.61 38.93 46.09 45.19 44.07 40.04
MaxViT-L 45.86 43.62 40.72 47.87 40.94 40.94
Average Value
Across all Models
44.97 42.16 43.87 48.15 41.78 41.34
Table 8
Per model, per dataset Accuracy (Acc) and F1-Scores (F1), using wavelengths:
RGB, 460 nm, 540 nm, 640 nm, for the LSTM solution.
(b) 4 times Zoom Dataset (c) Leaf Zoom Dataset
Models Acc (%) F1 (%) Acc (%) F1 (%)
ViT-L (P:16 ×16) 67.34 51.01 66.02 49.03
ViT-L (P:32 ×32) 64.65 46.98 68.60 52.90
ViT-B (P:16 ×16 - C) 74.05 61.07 67.31 50.97
ViT-B (P:32 ×32 – C) 67.79 51.68 67.74 51.61
Swin-L (P:4 ×4 – W:12 ×12) 67.79 51.68 70.75 56.13
VOLO (D:5) 72.71 59.06 68.17 52.26
XCIT-L (L:24 – P16 ×16) 72.71 59.06 66.45 49.68
DEIT-L (P16 ×16) 70.92 56.38 64.73 47.10
MaxViT-L 64.65 46.98 63.87 45.81
Table 9
Per Transformer model F1-Score performance comparison of: (a): Euclidian,
Bray-Curtis, Cosine distance functions for the K value of 3, (b) 1, 3, 5, 7 K values
for the Euclidian distance function.
(a) Distance Function (b) K Value
Models Bray-
Curtis
Cosine Euclidian 1 3 5 7
ViT-L
(P:16 ×
16)
53.69 53.69 54.52 56.38 54.52 55.15 59.13
ViT-L
(P:32 ×
32)
58.39 58.39 54.52 53.02 54.52 59.80 59.80
ViT-B
(P:16 ×
16 - C)
56.38 58.39 55.98 53.02 55.98 53.15 55.15
ViT-B
(P:32 ×
32 – C)
54.36 55.03 54.52 50.34 54.52 54.48 52.49
Swin-L
(P:4 ×4
– W:12
×12)
54.36 55.03 54.52 50.34 54.52 54.48 52.49
VOLO
(D:5)
54.36 50.34 49.75 46.98 49.75 51.16 51.83
XCIT-L
(L:24 –
P16 ×
16)
44.97 42.95 50.43 46.98 50.43 49.17 47.17
DEIT-L
(P16 ×
16)
54.36 52.35 55.21 55.03 55.21 51.16 51.16
MaxViT-L 55.70 56.38 55.89 53.02 55.89 55.81 55.81
D. Kapetas et al.
Smart Agricultural Technology 9 (2024) 100580
11
Ensemble
Having approached the classication problem with two separate
solutions, using multiple models and multiple wavelengths per image,
the best F1-Score results remained at 59.98 % for the KNN method by
the VOLO (D:5) model, and at 61.07 % for the LSTM method by the ViT-
B (P:16 ×16 – C) model.
Applying the ensemble solution described in Section 2.10, led to the
highest overall F1-Score, with dataset-b and the LSTM approach. The
combination of models that produced this outcome include the MaxViT-
L, the ViT-B (P:16 ×16 – C), the VOLO (D:5), and the XCIT-L (L:24 – P:16
×16). This ensemble achieved an overall classication accuracy of
79.41 % with an F1-Score of 71.12 %. On a per-class basis, the classi-
cation accuracy was 74.49 %, 80.53 % and 83.21 % for the “healthy”,
“botrytis-invisible” and “botrytis-visible” classes respectively, while the
F1-Score was at 68.33 %, 63.71 % and 77.47 %. These results were
obtained by applying the classication to the images from the 540 nm
and 640 nm hyperspectral wavelengths, as well as the RGB images while
a reduction in performance was observed when other wavelength
combinations were used. The results of all wavelength combinations
employed are presented in Table 10.
Finally, Table 11shows the Accuracy (Acc) and F1-Score (F1) metrics
for the best ensemble model on a per day total and on a per day per class
level. The F1-Score for Class 1 (“botrytis-invisible”) and its consistency
across all dpi, even as early as day 3, demonstrates the approach’s
capability for early detection.
Qualitative results
Precisely identifying and classifying leaves of tomato plants in regard
to B. cinerea infection is crucial for plant disease diagnosis and crop
management. A side-by-side comparison of manually annotated results
and those produced using the above-described methodology for three
test images is presented in Fig. 10, where the generated results for
manual-predicted annotations range from very similar, to mildly similar
and to very dissimilar.
Table 10
Best ensemble model F1-Score (F1) metrics for multiple wavelength(s) combi-
nations for the LSTM approach for dataset-b.
F1 (%)
Wavelength(s) Total Class 0 Class 1 Class 2
RGB 64.42 58.41 59.31 71.67
460 50.99 57.50 47.58 37.68
540 60.39 56.57 48.24 67.65
640 67.77 68.57 53.15 75.27
RGB, 460, 540, 640 69.11 65.63 59.31 78.10
460, 540, 640 67.10 67.65 50.08 76.29
RGB, 460, 640 69.11 68.66 59.94 75.00
RGB, 540, 640 71.12 68.33 63.71 77.48
540, 640 67.77 66.09 57.75 74.14
Table 11
Accuracy (Acc) and F1-Score (F1) metrics for the best ensemble model (640 nm) on a per day total and on a per day per class level, for the LSTM approach for dataset-b.
Acc (%) F1 (%)
DPI Total Class 0 Class 1 Class 2 Total Class 0 Class 1 Class 2
3 79.49 100.00 68.42 84.62 65.64 100.00 57.00 –
4 77.78 81.25 70.83 85.71 62.00 80.00 53.15 –
5 86.96 95.24 80.00 88.89 66.00 85.71 69.50 –
11 79.22 72.97 85.19 84.62 54.94 58.33 67.33 66.67
18 72.73 62.07 90.48 68.75 70.75 52.17 68.28 66.67
25 82.98 62.50 100.00 90.48 90.48 70.00 100.00 91.67
32 78.43 66.67 72.73 83.87 85.87 66.67 64.14 86.49
39 77.27 71.43 85.71 78.26 80.26 71.43 73.67 78.26
Total 79.42 74.50 80.54 83.22 71.12 68.33 63.72 77.48
Fig. 10. Qualitative assessment of the proposed methodology’s performance: (a)-(b)-(c) depict three images overlayed with the manual annotations. (d)-(e)-(f) depict
the same three images, overlayed with the predicted annotations from the proposed methodology of this study. Specically, (d) represents a case where the proposed
methodology had very dissimilar results compared to the manual annotations. Similarly, (e) presents a case where the results are mildly similar and (f) presents a case
where the results are very similar. Green leaves indicate “healthy” class, blue leaves indicate “botrytis-invisible” class, and red leaves indicate “botrytis-visible” class.
D. Kapetas et al.
Smart Agricultural Technology 9 (2024) 100580
12
Early-stage capabilities validation through comparison with qPCR-based
results
To further evaluate the early detection of B. cinerea at 3, 4 and 5 dpi
when grey mould symptoms are not visible, the best performing archi-
tecture from Section 3.4.1 was tested on the detached leaves used for
estimation of the fungal biomass via RT-qPCR (Section 3.2). The original
images (captured at 5 hyperspectral wavelengths and 1 RGB) were fed
into the YOLO model. Then, single leaets were extracted from these
images and processed separately. The new images were passed through
the Transformer models for the extraction of descriptors, that were
nally classied to give the nal results. According to them, the model
detected the pathogen as early as 3dpi in the 3rd and 5th leaves, but did
not detect it in the 13th leaves before 5 dpi. The results are promising,
demonstrating the model’s excellent segmentation capabilities on a
different background from that it was originally trained on, and
providing valuable insights into early detection of B. cinerea.
In Fig. 11, four different detached leaves are indicatively depicted
with the predictions from the presented methodology overlaid on each
image. The rst leaf was mock-inoculated (Fig. 11a), while the
remaining leaves were B. cinerea-inoculated with relative transcript
abundance of the fungal reference gene BcRPL5: 0.074 (Fig. 11b), 0.264
(Fig. 11c), and 10.498 (Fig. 11d).
Conclusions and future work
Botrytis cinerea is a widespread plant pathogen with a necrotrophic
lifestyle causing massive crop losses both pre- and post-harvest. No plant
shows complete resistance against B. cinerea, but plant immune re-
sponses have the potential to signicantly reduce disease progression
[55]. Following the pathogen’s initial contact with the host, two distinct
phases have been described: an early phase characterized by the for-
mation of local infection foci without spreading, and a late stage char-
acterized by the production of abundant fungal biomass and lesion
spread [55]. The balance between the fungus and the host-plant is
delicate, and even subtle changes can signicantly affect disease pro-
gression (e.g., changes in the fungal inoculum, timing of defense acti-
vation or external conditions). In the coming years, new technologies,
such as advanced imaging, are expected to provide detailed information
on specic processes and molecules that affect disease development.
In this study, the focus is the early and late detection of B. cinerea on
images captured on multiple wavelengths (RGB, 460 nm, 540 nm, 640
nm, 775 nm, 875 nm) through a two-step process. The rst step included
the training of a YOLOv8 model for the leaf segmentation task. The
YOLO model reached a mAP
50
of 81.7 % demonstrating its strong
capability in accurately detecting tomato leaves within images. The
second step was the classication of the grey mould disease on tomato
leaves. To that end, three leaf datasets were generated and then a suite of
transformers was employed to extract the leaf descriptors. Finally, a
KNN and a LSTM approach was introduced to classify the extracted
descriptors.
Among the three leaf datasets, dataset-b emerged as the best per-
forming. For the KNN approach, a K value of 3 with the Euclidian dis-
tance function returned the best performing results reaching a maximum
of 73.83 % accuracy and 59,98 % F1-Score across all classes among all
models. For the LSTM method, the best performing model was achieved
by excluding images from the 775 nm and 875 nm wavelengths from the
training process, resulting in an accuracy of 74.05 % and an F1-Score of
61.07 % across all classes among all models. The removal of these two
wavelengths proved necessary, since these two wavelengths were the
two lowest performing ones and they were negatively affecting the
LSTM model’s training.
A subsequent step was to employ an ensemble technique that com-
bined the predictions across multiple transformer models and multiple
wavelengths. Using the LSTM approach, the best ensemble model ach-
ieved a 79.41 % classication accuracy across all classes, with an F1-
Score of 71.12 %, leading to a 16.45 % performance increase
compared to the best individual model. Given the results yielded by the
application of the ensemble technique, the approach proved valid and its
increased performance overhead is deemed acceptable to gain such a
signicant performance boost. Based on the highest overall F1-Score,
the optimal solution was found using the LSTM approach. On a per
class basis, this best model achieved an accuracy of 74.49 %, 80.53 %
and 83.21 % respectively for each of the three classes. Similarly, the F1-
Score was at 68.33 %, 63.71 % and 77.47 % for each class respectively.
Evaluating these per class results based on the accuracy metric, it is
shown that the approach classies correctly a signicant portion of the
Fig. 11. Detached leaves classied as mock-inoculated (a) and B. cinerea-inoculated with relative transcript abundance of the fungal reference gene BcRPL5. 0.074
(b), 0.264 (c) and 10.498 (d). Green leaves indicate “healthy” class, blue leaves indicate “botrytis-invisible” class, and red leaves indicate “botrytis-visible” class.
D. Kapetas et al.
Smart Agricultural Technology 9 (2024) 100580
13
data. However, comparing each class accuracy with its corresponding
F1-Score, highlights that the “healthy” and the “botrytis-visible” classes
have relatively good balance between precision and recall, due to the
small percentage difference of the two values, while the bigger differ-
ence in the “botrytis-invisible” class, suggests that the model may
benet from further renement to better distinguish this class from
others. This indicates a potential area of improvement, likely related to
class distribution in the training data, which could be addressed to
further enhance model performance.
Additionally, the comparison of the qRT-PCR results with the output
of the presented architecture for validation purposes highlights the early
detection of grey mould disease through this approach. It demonstrates a
gradual detection of infected leaves corresponding to the increasing
relative expression of the B. cinerea reference gene BcRPL5. The results
contribute to the obtaining of new tools towards the early and late
detection of grey mould disease in widely used crops such as tomato
(Solanum lycopersicum), since this study is one of the rst attempts in
which DL techniques were extrapolated to tomato plants grown under
controlled conditions that were similar to those prevailing in a
greenhouse.
Future work could expand in different directions, but a promising
start would be to implement and experiment with other model archi-
tectures for the classication approach like CNNs and other variations
like ResNet based models, SVMs, Transformer or other attention-based
models. Also, further research could emphasize more into data
augmentation or manipulation techniques or in additional data
gathering.
Funding
This work supported by the Green Deal PestNu project, funded by
European Union’s Horizon 2020 research and innovation programme
under the grant agreement No 101,037,128.
Ethics statement
Not applicable: This manuscript does not include human or animal
research.
CRediT authorship contribution statement
Dimitrios Kapetas: Writing – original draft, Software, Methodology,
Investigation, Formal analysis. Eleni Kalogeropoulou: Writing – orig-
inal draft, Visualization, Methodology, Investigation, Formal analysis.
Panagiotis Christakakis: Writing – review & editing, Validation,
Methodology, Investigation. Christos Klaridopoulos: Writing – review
& editing, Data curation. Eleftheria Maria Pechlivani: Writing – re-
view & editing, Supervision, Resources, Project administration, Funding
acquisition, Conceptualization.
Declaration of competing interest
The authors declare that they have no known competing nancial
interests or personal relationships that could have appeared to inuence
the work reported in this paper.
Data availability
Data will be made available on request.
Acknowledgements
We thank Mrs. E. Fotopoulou and Dr. Emilia Markelou from Benaki
Phytopathological Institute for their invaluable technical support in
molecular experiments. We extend our sincere thanks to Mr. Nikolaos
Gaikoumoglou for his assistance in capturing the multi-spectral images
and data annotation.
Supplementary materials
Supplementary material associated with this article can be found, in
the online version, at doi:10.1016/j.atech.2024.100580.
References
[1] B. Williamson, B. Tudzynski, P. Tudzynski, J.A.L. Van Kan, Botrytis cinerea : the
cause of grey mould disease, Mol. Plant Pathol. 8 (5) (2007) 561–580, https://doi.
org/10.1111/j.1364-3703.2007.00417.x. Sep.
[2] B.A. Latorre, K. Elfar, E.E. Ferrada, Gray mold caused by Botrytis cinerea limits
grape production in Chile, Cienc. Inv. Agr. 42 (3) (2015) 1, https://doi.org/
10.4067/S0718-16202015000300001, 1.
[3] M. Al-Sarayreh, M.M. Reis, W.Q. Yan, R. Klette, Potential of deep learning and
snapshot hyperspectral imaging for classication of species in meat, Food Control
117 (2020) 107332, https://doi.org/10.1016/j.foodcont.2020.107332. Nov.
[4] G. Romanazzi, J.L. Smilanick, E. Feliziani, S. Droby, Integrated management of
postharvest gray mold on fruit crops, Postharvest Biol. Technol. 113 (2016) 69–76,
https://doi.org/10.1016/j.postharvbio.2015.11.003. Mar.
[5] P. Leroux, et al., Mechanisms of resistance to fungicides in eld strains of Botrytis
cinerea, Pest Manag. Sci. 58 (9) (2002) 876–888, https://doi.org/10.1002/ps.566.
Sep.
[6] M. Bilkiss, M.J.A. Shiddiky, R. Ford, Advanced diagnostic approaches for
necrotrophic fungal pathogens of temperate legumes with a focus on Botrytis spp,
Front. Microbiol. 10 (2019) 1889, https://doi.org/10.3389/fmicb.2019.01889.
Aug.
[7] H.J. Rosslenbroich, D. Stuebler, Botrytis cinerea — history of chemical control and
novel fungicides for its management, Crop Protection 19 (8–10) (2000) 557–561,
https://doi.org/10.1016/S0261-2194(00)00072-7. Sep.
[8] C. Gachon, P. Saindrenan, Real-time PCR monitoring of fungal development in
Arabidopsis thaliana infected by Alternaria Brassicicola and Botrytis cinerea, Plant
Physiol. Biochem. 42 (5) (2004) 367–371. doi.org/10.1016/j.plaphy.2004.04.001.
[9] S. Petrasch, C.J. Silva, S.D. Mesquida-Pesci, K. Gallegos, C. van den Abeele,
V. Papin, F.J. Fernandez-Acero, S.J. Knapp, B. Blanco-Ulate, Infection strategies
deployed by Botrytis cinerea, Fusarium acuminatum, and Rhizopus stolonifer as a
function of tomato fruit ripening stage, Front. Plant Sci. 10 (2019) 223, https://doi.
org/10.3389/fpls.2019.00223.
[10] J. W¨
aldchen, P. M¨
ader, Machine learning for image based species identication,
Methods Ecol. Evol. 9 (11) (2018) 2216–2225, https://doi.org/10.1111/2041-
210X.13075. Nov.
[11] K. Tsiakas, et al., An autonomous navigation framework for holonomic mobile
robots in conned agricultural environments, Robotics 12 (6) (2023) 146, https://
doi.org/10.3390/robotics12060146. Oct.
[12] E.M. Pechlivani, G. Gkogkos, N. Giakoumoglou, I. Hadjigeorgiou, D. Tzovaras,
Towards sustainable farming: a robust decision support system’s architecture for
agriculture 4.0, in: 2023 24th International Conference on Digital Signal
Processing (DSP), Rhodes (Rodos), Greece, IEEE, 2023, pp. 1–5, https://doi.org/
10.1109/DSP58604.2023.10167922. Jun.
[13] A. Voulodimos, N. Doulamis, A. Doulamis, E. Protopapadakis, Deep learning for
computer vision: a brief review, Comput. Intell. Neurosci. 2018 (2018) 1–13,
https://doi.org/10.1155/2018/7068349.
[14] Potgieter Saleem, Mahmood Arif, Plant disease detection and classication by deep
learning, Plants 8 (11) (2019) 468, https://doi.org/10.3390/plants8110468. Oct.
[15] M. Shoaib, et al., An advanced deep learning models-based plant disease detection:
a review of recent research, Front. Plant Sci. 14 (2023) 1158933, https://doi.org/
10.3389/fpls.2023.1158933. Mar.
[16] M. Ouhami, A. Haane, Y. Es-Saady, M. El Hajji, R. Canals, Computer vision, IoT
and data fusion for crop disease detection using machine learning: a survey and
ongoing research, Remote Sens. (Basel) 13 (13) (2021), https://doi.org/10.3390/
rs13132486. Art. no. 13Jan.
[17] A. Mohammed, R. Kora, A comprehensive review on ensemble deep learning:
opportunities and challenges, J. King Saud Univ. - Comput. Inf. Sci. 35 (2) (2023)
757–774, https://doi.org/10.1016/j.jksuci.2023.01.014. Feb.
[18] A. Dosovitskiy, et al., An image is worth 16x16 words: transformers for image
recognition at scale, arXiv (2021), https://doi.org/10.48550/arXiv.2010.11929.
Jun. 03.
[19] Z. Liu, et al., Swin transformer: hierarchical vision transformer using shifted
windows, arXiv (2021), https://doi.org/10.48550/arXiv.2103.14030. Aug. 17.
[20] H. Touvron, M. Cord, H. J´
egou, DeiT III: revenge of the ViT, arXiv 14 (2022),
https://doi.org/10.48550/arXiv.2204.07118. Apr.
[21] L. Yuan, Q. Hou, Z. Jiang, J. Feng, S. Yan, VOLO: vision outlooker for visual
recognition, arXiv (2021), https://doi.org/10.48550/arXiv.2106.13112. Jun. 28.
[22] A. El-Nouby, et al., XCiT: cross-covariance image transformers, arXiv (2021),
https://doi.org/10.48550/arXiv.2106.09681. Jun. 18.
[23] Z. Tu, et al., MaxViT: multi-axis vision transformer, arXiv (2022), https://doi.org/
10.48550/arXiv.2204.01697. Sep. 09.
[24] N. Giakoumoglou, et al., Early detection of Botrytis cinerea symptoms using deep
learning multi-spectral image segmentation, Smart Agricul. Technol. 8 (2024)
100481, https://doi.org/10.1016/j.atech.2024.100481. Aug.
D. Kapetas et al.
Smart Agricultural Technology 9 (2024) 100580
14
[25] P. Christakakis, N. Giakoumoglou, D. Kapetas, D. Tzovaras, E.M. Pechlivani, Vision
transformers in optimization of AI-based early detection of Botrytis cinerea, AI 5 (3)
(2024), https://doi.org/10.3390/ai5030063 volnoArt. no. 3, Sep.
[26] C.G. Scarboro, S.M. Ruzsa, C.J. Doherty, M.W. Kudenov, Quantication of gray
mold infection in lettuce using a bispectral imaging system under laboratory
conditions, Plant Direct 5 (3) (2021) e00317, https://doi.org/10.1002/pld3.317.
Mar.
[27] C.L. Chung, K.J. Huang, S.Y. Chen, M.H. Lai, Y.C. Chen, Y.F. Kuo, Detecting
Bakanae disease in rice seedlings by machine vision, Comput. Electr. Agricult. 121
(2016) 404–411, https://doi.org/10.1016/j.compag.2016.01.008. Feb.
[28] R. Qasrawi, M. Amro, R. Zaghal, M. Sawafteh, S.V. Polo, Machine learning
techniques for tomato plant diseases clustering, prediction and classication, in:
2021 International Conference on Promising Electronic Technologies (ICPET),
2021, pp. 40–45, https://doi.org/10.1109/ICPET53277.2021.00014. Nov.
[29] `
O. Lorente, I. Riera, and A. Rana, “Image classication with classic and deep
learning techniques.” arXiv, May 11, 2021. Available: http://arxiv.org/abs/2
105.04895.
[30] D. Mistry, A. Banerjee, Comparison of feature detection and matching approaches:
SIFT and SURF, GRD J.- Global Res. Develop. J. Eng. 2 (2017) 7–13. Mar.
[31] N. Mansoori, M. Nejati, P. Razzaghi, and S. Samavi, “Bag of visual words approach
for image retrieval using color information,” May 2013, pp. 1–6. doi: 10.1109/Iran
ianCEE.2013.6599562.
[32] K. Nagasubramanian, S. Jones, A.K. Singh, S. Sarkar, A. Singh,
B. Ganapathysubramanian, Plant disease identication using explainable 3D deep
learning on hyperspectral images, Plant Methods 15 (1) (2019) 98, https://doi.
org/10.1186/s13007-019-0479-8.
[33] “Early detection of plant viral disease using hyperspectral imaging and deep
learning.”. Available: https://www.mdpi.com/1424-8220/21/3/742.
[34] J. Redmon, S. Divvala, R. Girshick, A. Farhadi, You only look once: unied, real-
time object detection, arXiv (2016), https://doi.org/10.48550/arXiv.1506.02640.
May 09.
[35] “Qcell – spectral vision camera systems.”. Available: https://qcell.tech/.
[36] R.D. Machado, A.P. Christoff, G. Loss-Morais, M. Margis-Pinheiro, R. Margis, A.
P. K¨
orbes, Comprehensive selection of reference genes for quantitative gene
expression analysis during seed development in Brassica napus, Plant Cell Rep. 34
(2015) 1139–1149.
[37] J. Huggett, K. Dheda, S. Bustin, A. Zumla, Real-time RT-PCR normalisation;
strategies and considerations, Genes Immun. 6 (2005) 279–284.
[38] M.M. L´
opez, E. Bertolini, A. Olmos, P. Caruso, M.T. Gorris, P. Llop, R. Penyalver,
M. Cambra, Innovative tools for detection of plant pathogenic viruses and bacteria,
Int. Microbiol. 6 (2003) 233–243, 6.
[39] S. Schneider, F. Widmer, K. Jacot, R. Koelliker, J. Enkerli, Spatial distribution of
Metarhizium clade 1 in agricultural landscapes with arable land and different semi-
natural habitats, Appl. Soil Ecol. 52 (2012) 20–28.
[40] Y. Song, Y. Wang, D. Guo, L. Jing, Selection of reference genes for quantitative real-
time PCR normalization in the plant pathogen Puccinia helianthi Schw, BMC Plant
Biol. 19 (1) (2019) 20, https://doi.org/10.1186/s12870-019-1629-x. Jan 11PMID:
30634896; PMCID: PMC6329156.
[41] “The glutaredoxin ATGRXS13 is required to facilitate Botrytis cinerea infection of
Arabidopsis thaliana plants - La Camera - 2011 - The Plant Journal - Wiley Online
Library.”. Available: https://onlinelibrary.wiley.com/doi/10.1111/j.1365-313X.
2011.04706.x.
[42] C.L. Campbell, L.V. Madden, Introduction to Plant Disease Epidemiology, 1st
edition, Wiley-Interscience, New York, 1990.
[43] E. Kalogeropoulou, D. Beris, S.E. Tjamos, I. Vloutoglou, E.J. Paplomatas,
Arabidopsis β-amylase 3 affects cell wall architecture and resistance against
Fusarium oxysporum, Physiol. Mol. Plant Pathol. (2022) 101945, https://doi.org/
10.2139/ssrn.4251657.
[44] N. Liu, Y. Sun, Y. Pei, X. Zhang, P. Wang, X. Li, et al., A pectin Methylesterase
inhibitor enhances resistance to verticillium wilt, Plant Physiol. 176 (2018)
2202–2220, https://doi.org/10.1104/pp.17.01399.
[45] “Roboow: computer vision tools for developers and enterprises.”. Available: htt
ps://roboow.com/.
[46] S. Minaee, Y. Boykov, F. Porikli, A. Plaza, N. Kehtarnavaz, D. Terzopoulos, Image
segmentation using deep learning: a survey, arXiv (2020), https://doi.org/
10.48550/arXiv.2001.05566. Nov. 14.
[47] P. Zhang, Y. Wu, B. Liu, Leveraging local and global descriptors in parallel to
search correspondences for visual localization, arXiv 22 (2020), https://doi.org/
10.48550/arXiv.2009.10891. Sep.
[48] “timm (PyTorch image models).”. Available: https://huggingface.co/timm.
[49] “Papers with Code - ImageNet-1K dataset.”. Available: https://paperswithcode.
com/dataset/imagenet-1k-1.
[50] P. Cunningham, S.J. Delany, K-Nearest Neighbour Classiers 54 (6) (2022) 1–25,
https://doi.org/10.1145/3459665, 2nd Edition (with Python examples)Jul.
[51] “Keras: deep learning for humans.”. Available: https://keras.io/.
[52] H N.B., “Confusion matrix, accuracy, precision, recall, f1 score,” Analytics Vidhya.
Available: https://medium.com/analytics-vidhya/confusion-matrix-accuracy-preci
sion-recall-f1-score-ade299cf63cd.
[53] P.J. Pinter, R.D. Jackson, C. Elaine Ezra, H.W. Gausman, Sun-angle and canopy-
architecture effects on the spectral reectance of six wheat cultivars, Int. J. Remote
Sens. 6 (12) (1985) 1813–1825, https://doi.org/10.1080/01431168508948330.
[54] B. Williamson, B. Tudzynski, P. Tudzynski, J.A.L. Van Kan, Botrytis cinerea: the
cause of grey mould disease, Mol. Plant Pathol. 8 (5) (2007) 561–580, https://doi.
org/10.1111/j.1364-3703.2007.00417.x.
[55] K. Bi, Y. Liang, T. Mengiste, A. Sharon, Killing softly: a roadmap of Botrytis cinerea
pathogenicity, Trends Plant Sci. 28 (2) (2023) 211–222. Issue, doi.org/10.1016/j.
tplants.2022.08.024.
[56] J. Zhou, J. Wang, Z. Zheng, B. Fan, J.Q. Yu, Z Chen, Characterization of the
promoter and extended C-terminal domain of Arabidopsis WRKY33 and functional
analysis of tomato WRKY33 homologues in plant stress responses, J. Exp. Bot. 66
(2015) 4567–4583, https://doi.org/10.1093/jxb/erv221.
[57] J. Zhou, J.Q. Yu, Z Chen, The perplexing role of autophagy in plant innate immune
responses, Mol. Plant Pathol. 15 (2014) 637–645, https://doi.org/10.1111/
mpp.12118.
[58] M.F. Guo, Y.C. Liu, J. Zou, Z. Qi, V. Fotopoulos, G. Lu, J. Yu, J. Zhou, A single-
nucleotide polymorphism in WRKY33 promoter is associated with the cold
sensitivity in cultivated tomato, New Phytol. 236 (2022) 989–1005, https://doi.
org/10.1111/nph.18403.
D. Kapetas et al.
Smart Agricultural Technology 9 (2024) 100580
15
