Traitement du Signal

Vol., No., Month, Year, pp. **-**

Journal homepage: http://iieta.org/journals/ts

Early Detection of Grapevine Viral Disease Using Hyper spectral Imaging and Optimized
Rajalakshmi S1, Alagumani S2, Ananthi R.K3 and Surendran Rajendran4*
Neural Networks
1
Department of Department of Artificial Intelligence and Data Science, Dr. N.G.P. Institute of Technology,
Coimbatore, 641048, India. mrajislm@gmail.com
2
Department of Computer Science and Engineering, SRM Institute of Science and Technology,
Ramapuram Campus, Chennai – 600 089, India. alagumas@srmist.edu.in
3
Department of Artificial Intelligence and Data Science, K.Ramakrishnan College of Engineering, Tiruchirapalli,
621112, India. anuu.rk@gmail.com
4
Department of Computer Science and Engineering, Saveetha School of Engineering, Saveetha Institute of Medical
and Technical Sciences, Chennai, 602105, India. surendran.phd.it@gmail.com

Corresponding Author Email: surendran.phd.it@gmail.com

https://doi.org/10.18280/ts.xxxxxx ABSTRACT
Received: Early identification of viral infections in grapevines is necessary for prompt intervention to
Accepted: prevent disease dissemination across the field. Remote sensing using hyper spectral technology
may be able to non - destructively identify and measure viral infections. The present research
Keywords: used hyper spectral photography at the plant level to identify and categories grapevines infected
Chlorophyll content, grapevine viral disease, with the newly found DNA virus grapevine vein-clearing virus (GVCV), mostly during the first
hyper spectral images, neural network asymptomatic Here, only grapevine pixels were kept in the hyper spectral images after
optimization, Artificial Humming Bird calibration and preprocessing. In Addition, a statistical technique was used to distinguish
Optimization and ZfNet+VGG19 between two reflectance spectrums patterns from vines that are both healthy and GVCV. After
preprocessing the data, an artificial hummingbird optimization technique is used to extract the
features, ensuring that only the most pertinent features are chosen and that the overall model
classification is improved. Furthermore, a nondestructive method is chosen to calculate the total
chlorophyll (Chl) content of grape leaves. Chl concentration and the red-edge chlorophyll index
were found to be correlated with reflectance measurements in the near-infrared (755-765 nm)
and red-edge (710-720 nm) spectral ranges. Following this assessment, a hybrid of
ZfNet+VGG19 is employed for both pixel-wise and image-wise classification of disease
severity. The suggested approach, known as the Artificial Humming Bird Optimized
ZfNet+VGG19 neural network (AHB_ ZfNet+VGG19), which has the advantage of achieving
large gain in speed and an increase in accuracy. The incorporation of prior training and the
deepening of the model were mostly responsible for this. It is evaluated against standard
approaches and hence the suggested method achieves98.27% of accuracy, 97.67% of precision,
97.41% of recall and 97.74% of F1score for Salinas dataset and 98.45% of accuracy, 97.1% of
precision, 97.41% of recall and 97.6% of F1-score for Indian pine

1. INTRODUCTION pathogen relationship [3]. Viruses, fungi, and bacteria are

typical pathogens. The collection, preparation, and analysis
Plant diseases have a significant negative economic of samples using these techniques takes at least a few days.
impact on the worldwide agricultural industry. To avoid Indirect (proxy) techniques may identify plant diseases
the spread of infection and make effective management using a range of factors, such as morphological changes,
methods possible, precision agriculture relies heavily on temperature change, evaporation rate reduction, and
plant health monitoring and pathogen detection [1]. Early volatile compounds released by affected crops [4]. These
detection of plant pathogens and other agricultural approaches mostly depend on optical imaging
disturbances may be a useful source of information for technologies. The optical imaging sensors give in-depth
improving crop management strategies and disease control data based on several electromagnetic spectra, allowing for
approaches to prevent the establishment and spreading of the prediction of the health of the plant. Among the most
diseases. There are two primary categories for automated popular indirect techniques for identifying plant diseases
plant disease detection techniques: direct and indirect include thermography, fluorescence, and hyper spectral
(proxy) techniques. Examples of direct detection imaging. The characterization, detection, modelling, and
techniques include molecular and serological techniques, categorization of plant diseases have all been accomplished
which encourage high-throughput examination of several with the use of hyper-spectral imaging [5]. Through the
samples [2]. By identifying the disease-causing pathogens use of a hypercube, hyper-spectral imaging analyses the
directly, these techniques accurately identify the disease- specular reflections from crops over the electromagnetic
1
 spectrum in thousands of small groups. The interaction of a processing technology. Thus, it is appropriate and relevant
plant with various electromagnetic spectrum regions is to analyse recent events. This section examines
influenced by the biochemical elements and anatomical interdisciplinary approaches to hyper spectral imaging
composition of its leaves. Due to the photosynthetic techniques, data processing, and disease classification
pigments in their leaves, healthy plants normally absorb models. According to Narayanan et al. [12], a hybrid
light in the visible spectrum (VIS 400–700 nm). Near- convolution neural network (HCNN) allows the
infrared light scattering in the 700–1000 nm region is very identification of banana illness, and the classification is
sensitive to leaf cell structure [6]. Leaf water and chemical suggested to get around these problems and help the
contents are the main determinants of leaf reflectance in farmers by allowing the use of fertilisers necessary for
the short-wave infrared range about SWIR 1000-2500 nm. preventing the disease in its early stages. As a result, it
Plants respond to diverse pressures by changing provide better accuracy with more computational time. The
biophysically and biochemically, such as by depleting the held-out dataset used by Yakkundimath et al. [13], it was
amount of chlorophyll in the leaves or altering the cell analysed through a threefold cross-validation procedure
structures of the leaves. The benefit of hyper-spectral employing pretrained VGG-16 as well as Google Net
imaging is that it can detect these minute variations in plant convolutional neural network models. The average
spectral reflectance [7]. Based on these spectral reflectance categorization accuracy for the tested VGG-16 and Google
values, a variety of machine learning techniques have been Net convolutional neural networks was 92.24% and
created [8] for the automated categorization of plant 91.28%, accordingly.
diseases. Typically, the process for predicting sick leaves The sensitivity of various wavelengths was explained by
involves extracting characteristics from spectral Cao et al. [14] using a Spectral Dilated Convolution three-
reflectance, building a classifier model using images of dimensional Convolutional Neural Network (SDC-
sick and normal plants, and then applying the model to 3DCNN) with thresholding display. The results revealed
new data [9]. A common feature extraction method is the that the SDC-3DCNN model's accuracy is 95.4427% when
estimation of spectral Vegetation Indices (VIs) linked to the cumulative number of inputs is 50 characteristic
specific physiological parameters [10]. These VIs, wavelengths and the dilation ratio is fixed at
however, often aren't made to distinguish between healthy 5.Convolutional neural networks (CNNs) and other
and ill plants [11]. The amount of data in real-world learning frameworks, according to Shankarnarayanan
scenarios makes analysis difficult. Attributes and Nalini et al [15], have made substantial strides in the fine-
characteristics exist in datasets. Not all features are tuning of image processing to meet a database of a plant's
necessary for data extraction. A model's efficiency could leaves that was separately produced for various plant
suffer from redundant features. Each dataset's size is illnesses. It provides around 78.9% of accuracy.
decreased by feature reduction while retaining accuracy. A combined convolutional neural network and support
Feature selection is a component of feature reduction. vector machine (CNN-SVM) technique is suggested by
While feature selection chooses necessary features, feature Gui et al. [16] for identification of SMV. According to the
extraction adds additional features from already-existing experimental findings, the CNN-SVM model's training set
datasets. As a result, AHB is a cutting-edge meta-heuristic accuracy rate was 96.67% and its training and testing
method that enhances accuracy. It mimics the accuracy rate was 94.17%. Deep convolutional neural
extraordinary flight prowess and cunning feeding strategies network (DCNN) is a method suggested by Kukreja et al.
of wild hummingbird. A flexible adversarial approach is [17] that can quickly diagnose wheat rust infections
provided to enable the first strategy to generate more automatically without human inspection. Furthermore, our
precise responses with greater obstacles involving two DCNN training and testing procedure yields accurate
employees. diagnosis and classification results for wheat rust illnesses.
This research makes the following contributions: To This method achieves 89.4% of accuracy with less
achieve this, a comparative analysis is conducted between computational time. According to Jiang et al. [18], six
the reflectance spectra of GVCV-infected vines and normal machine learning-assisted techniques were created based
vines, considering the datasets from each collection date. on the chosen spectral fingerprint characteristics for the
To enhance the reliability and performance of the analysis, early detection of anthracnose and grey mould in
the artificial hummingbird (AHB) feature extraction strawberries. According to Wang et al. [19], a novel deep
method is employed. This method strikes an improved learning architecture called outliers removal auxiliary
balance between exploration and exploitation, making the classifier generative adversarial network (OR-AC-GAN) is
newly designed algorithm more robust and efficient presented. In addition to incorporating the classification
compared to its predecessor. function into the model, it can also uncover the inherent
The remainder of the paper is organized as follows: The data characteristics and lessen the consequences of data
related efforts for the categorization of plant diseases using outlier side effects.
hyper spectral images are included in Section 2. The The experiment uses the Tomato Spotted Wilt Virus
proposed neural network with classification layer is (TSWV), a widespread pathogen, to validate the concept.
described in Section 3. The experimental analysis is In [21] employed hyperspectral remote sensing to identify
presented with graphs and a comparison with two state-of- grapevine viral infections early. During asymptomatic
the-art techniques in Section 4. Section 5 is the conclusion phases, it effectively recognised and categorised
and recommendations for further research conclude the grapevines infected with the newly found DNA virus
essay. GVCV. Specific vegetation indicators demonstrated strong
discriminating power. SVM and RF classifiers performed
2. RELATED WORK well in classification, while the 3D-CNN feature extractor
outperformed the 2D-CNN. This method achieves 78.4%
Rapid viral detecting techniques based on distant and of accuracy. In [22] employed hyperspectral imaging to
nearby optical sensors have been developed more quickly identify grapevine leafroll disease (GLD) in white and red
as a result of recent advancements in image and data grapevines. In greenhouse and outdoor conditions, models
2
 distinguished between sick, asymptomatic, and healthy
plants with great accuracy. Here, the complexity is more. In this research, assume N low-altitude nodes, each with
In [23] used Random Forest (RF) and 3D-Convolutional an indexing set of X ={1 , 2 ,… x }.Nodes are used in the
Neural Network (CNN) models outperformed ocular sensing of a planar area for surveillance tasks fitted with
evaluation of symptoms in distinguishing infected vs. non- GPS, inertial measurement (IMU), webcams, detectors,
infected leaves. While differentiating co-infected plants and a wireless transmission interface. It is assumed that all
was more difficult, both models demonstrated good results nodes are randomly dispersed in a 3D space. With a
across all infection categories. In [24] four machine standard constant communication range R at each place,
learning algorithms were used. In 10-fold cross-validation, each node may perceive a specific region. Using GPS, each
the boosted regression tree (BRT) model with SPA- node is conscious of its precise position(a ,b ,c ).
selected wavelengths produced the best results, with 85.2%
The BS, which is regarded as the target of the data
accuracy and an AUC of 0.932.The approach effectively
packets, receives the data from UAVs that monitor the
identified TSWV at the presymptomatic stage prior to
region and collect photos and video from the surveillance
molecular identification, suggesting its potential for early uav uav uav
detection of the virus in tobacco plants. In [25] use of area. The values of t j=(a j ,b j , c j ) and Q i,
methodologies based on machine learning and modelling accordingly, provide the positional data and transmitting
has shown to be beneficial in reducing the complexity of power of node i∈ X .
data analysis To specify the network model as G= ( X , R ) , where X
The training of a high number of spectral inputs and
is the node set and R={r 1 , r 2 , … r n } is the set of node
establishing their subjects is the fundamental difficulty
faced by neural networks while processing hyper spectral locations, taking into account all node’s placements and
data [20]. The use of NN classifiers for the categorization transmission strengths.
of VIs and SDIs makes this even more difficult. Overall, To take into account the forwarding route of N number
the Hughes effect, sometimes known as "the curse of of nodes for collision-free pathways. Assume that the
dimensionality," is a significant challenging issue with r i ( t )=¿ location coordinates of node i at time t ,
hyper-spectral data since it involves spectral band variety incorporate the forwarding route of N number of nodes for
and distortions. The NN modelling may be impacted by the collision-free pathways. Assume that the position
Hughes phenomenon. It often occurs when the ratio of coordinates of node i at time t are r i ( t )=¿, with
training pixels is higher than the required minimal to
determining the quantitative fit.
∀ i∈ {1 ,2 , … , N }, t ≥ 0.

3. PROPOSED METHODOLOGY 3.2 Hyper spectral image preprocessing

The Artificial Humming Bird Optimized ZfNet+VGG19 The computational complexity brought on by
neural networks (AHB_ ZfNet+VGG19) used in this processing the enormous volume of data is one of the
research are developed as seen in figure 1. The hyper difficult issues in processing high dimensional data with
spectral photos of grapes are first captured using a wireless improved spectral and temporal resolution. In particular,
sensor network, then preprocessed. In order to determine this is valid for hyperspectral photographs with a wide
the reflectance spectra of healthy and diseased leaves, the range of spectral bands. Hyperspectral photography must
preprocessed data were subjected to a spectral signal be preprocessed in order to decrease the dimensions and
discrimination approach. To get the best results, the computing complexities of the data, as well as for display
artificial hummingbird (AHB) optimization approach for and optimal band picking. When compared to pixels that
dimensionality reduction is used, followed by are far apart from one another, those with comparable
ZfNet+VGG19-based classification. spatial placements are more likely to be part of the same
sort of thing. The distance between comparable pixels is
nearer and more probable to convergence in the feature
space. The pixels are no longer separate data points when
viewed from a location within the data field. Instead, they
stand for a variety of radioactive particles. Any given spot
emits energy into the whole region that the picture covers.
With greater distance, the energy's intensity diminishes.
Every pixel collects energy from the points around it and
radiates energy outward to other points. In this formula, the
potential energy function is determined (1).
( x−n y )
k

(1)
φ=m× e

where k , which in this case is set to two, indicates the

Euclidean distance, k ∈ N signifies the distance index,
and m ≥0 indicates the grey value of a pixel.
The impact factor n, is a constant in the data packet that
expresses the possible interaction between the pixels.
Insufficient effect between the pixels results in low
Figure 1. Block diagram of disease classification model clustering when this factor is minimal.
In these conditions, separate pixel-centric energy zones
3.1 Network model are also described by the lines of equal potential. The

3
 interactions among individual data pixels rise as the impact amazing capacity to find the best solutions with these
factor rises, and the line features become closer together. remarkable abilities. The modelling description of the
AHB is demonstrated by establishing the initial population
3.3 Spectral Signal Discrimination of X hummingbirds from N individuals, as seen in Eq. (4)

In preprocessed hyperspectral images spatial and X i =L+ r × ( U−L ) , i=1 ,2 , 3 … N (4)

spectral information are available which has to be separate.
Because, need only spectral signals for analysis these can
Where L∧U denote for a D dimension's upper and
be done by t-. Each independent band sample t-tests were
used to analyse spectral alterations specifically attributable lower limits, correspondingly. r is a random vector with a
to that infection phase in every dataset matching to every [0, 1] range. Eq. (5) is also used to produce a visited table
collection date, as well as in the aggregated dataset, to of food sources:
determine the distinction between the reflectance spectra of
normal and GVCV pathogen vines. Three t-test statistical
presumptions were examined. The two groups were once
thought to be independent of each other. The reflectance
{
VT ij = 0 if i ≠ j , i=1 , … . N , j=1 , …. N
null if i= j
(5)
values of every band, which made up the dependent
variable, were assumed to have a normal distribution, Where VT ij represents the food consumed by a
which was supported by evaluating the distribution's
hummingbird at a particular food source and becomes null
skewness (symmetry) and kurtosis (peakness).The second
presumption was that the dependent variable's variance if i= j. Moreover, they represent a hummingbird visiting a
was about equal between the two groups. As a result of food source when i≠ j and VT ij reach zero.
Levene's test returning no significant results for all bands, Guided Foraging- During foraging at this stage, three
pooled or equal variance t-tests were employed, as shown flying capabilities—omnidirectional, diagonal, and axial
in Equation (2) flight—are used. Eq. (6) is used to describe the axial flight

SD =2
p
( n1−1 ) SD21 + ( n2−1 ) SD22
( n1 −1 ) + ( n2−1 )
(2) {
D(i )= 1, if i=randi ( [ 1 , d ] } i=1 , 2 , … .d
0 , else
(6)

The t values were determined by substituting Eq. (1) Eq. (7) may be used to represent the diagonal flight
into Eq. (2):

{
1 , if i=Pp ( j ) , jϵ [ 1 , k ]
D(i )= , Pp=randperm ( Kp )

√
M 1−M 2 SD 2p SD 2p (7)
where , SE( M −M )= + (3) , Kp ∈[2 , [ r 1 . ( d−2 ) ] +1]
SE(M − M )
1 2
1 2
n1 n2 0 , else i=1 , … .d

where n1 is the number of specimens taken from healthy Eq. (8) represents the omni-directional flying.
vines, M 1 is the mean reflectance of band i of normal
(i )
vines, M 2 is the meanreflectance of band i of pathogens D =1 i=1 ,2 , … d (8)
2
SD
vines, and
p
is the number of specimens taken from whererand i([1 , d ]) stands for a randomly generated
n2 integer between 1∧d , randperm(k ) for a randomly
2
diseased vines is n2 , and the pooled variance is SD p in Eq. generated permutations of the values between 1∧k , and
(3). r 1∈[0 , 1] for a random number between 0 and 1. Eq. (9)
is used to generate the directed foraging behavior
3.4 Artificial hummingbird optimization based feature
extraction V i ( t+1 ) =X i ,t ( t ) +a × D × ¿(9)
Here, the Artificial Hummingbird (AHB) optimization
algorithm's feature extraction procedure is described. The Where X i , t ( t )indicates the food source i for t iteration.
AHB replicates the remarkable flying prowess and cunning The hummingbirds' preferred food source is X i , t ( t ).
feeding strategies of hummingbirds in the wild. Axial, Territorial Foraging- A hummingbird is highly possible
diagonal, and omnidirectional foraging methods are used to look for a new food source rather than to visit other
in this method. The goal is to simulate the hummingbird's existing food sources when flower nectar runs out. A
ability to remember where food is located, directed, hummingbird may therefore effortlessly fly to a nearby
territorial, and migratory foraging methods are also used, location inside its region where it can discover a
along with a visiting table. The method is simple and has potentially superior food source. Eq. (10) is used to
just a few fixed parameters that may be changed. Each represent the situation
hummingbird in the AHB is given a distinct food source
from which to feed. Hummingbirds are able to retain the
V i ( t+1 ) =X i ,t ( t ) +b × D × X i (t ) , bϵN (0 ,1)
location and frequency of nectar replenishment for this
specific food source. It can remember the intervals (10)
between trips to each food source. The AHB has an
4
 Migration Foraging-A hummingbird will travel to a
different eating area if its favorite spot runs out of food. This section introduces the categorization procedure
The visit table will change when this hummingbird utilizing a pretrained CNN. Three layers make up a CNN:
switches from its prior food source to the new one. The a convolutional layer, a pooling layer, and a fully
migratory of a hummingbird from a nectar source with the connected layer. Computer vision activities including
fewest nectar replenishments to the one with a randomized picture creation, image classification, image captioning,
rate of nectar generation is described Eq. (11). and many more may be performed using four pretrained
networks. VGG19 and ZfNet were two of these pretrained
X w ( t +1 )=L+r ×(U−L) (11) networks utilized in this research. For layer-by-layer
convolutional network visualization and comprehension,
The food supply with the least fitness value is ZfNet+VGG19 is utilized. The network used batch
stochastic gradient descent for training and ReLUs for
represented by X w in this case. The visiting table is a key
activation. ZfNet+VGG19 architecture considerably
part of the AHA technique. Equ. (12) used to update the outperforms AlexNet by dissecting the convolutional
visiting table for every hummingbird. network layer by layer, changing the layer hyper-
parameters like filter size or stride, and successfully
VT i , k =VT i ,k +1 ,if k ≠i∧k ≠target , k=1 , 2 … . hn reducing the error rates. The model architecture of a deep
(12) convolutional neural network is shown in Figure 2. The
depth of the network depends on the number of hidden
The time that the same hummingbirds visited each food layers. Hidden layers are those layers that exist between
source is shown in this visiting table. A high number of the input and output layers. It has 3 completely connected
visits is indicated by a lengthy time between visits. layers, max-pooling layers, and 5 shared convolutional
layers.
3.5 Chlorophyll estimation The dataset was separated into a training and testing
dataset for every class after just a feature extraction
A conceptual approach was created to quantify the process. It comprises of training and testing pictures for
concentration of plant pigments such total chlorophyll, DS1, DS2, DS3, and DS4 of 2452, 4238, 4011, and 10292
anthocyanins, and carotenoids using three distinct spectral photos and 584, 1971, 1357, and 3912 images,
bands. In three spectral bandsτi , the model establishes a accordingly. The complete dataset that serves as the input
relationship between the target pigment and leaf for the model was downsized to 224224 pixels. Based on
reflectance ρ τi in Equ. (13): the number of categories, the output of the final completely
linked layer was divided into 5 groups. It calculated and
−1 −1 −1 recorded the average value that had exponentially
pigment contnet ∝ a pigment =( ρτin −ρτ 2n )× ρτ 3n depreciated in the two instances before it and calculated
the averaged of the prior gradients that had depreciated
(13) exponentially.

wherea pigment is the relevant pigment's absorption

coefficient. In the τi band of the spectrum, reflectance is
most sensitive to the absorption of the color of interest;
however, other pigments' absorbing and leaf scatters also
have an impact.
Depending on which λ1 was chosen, there were two
different approaches to estimate Chl using the three-band
model of Eq. 14. As a result, the following types of
chlorophyll indices (CI) have been proposed.

P NIR
CI green = −1 (14)
P green

P NIR
CI red edge= −1 (15) Figure 2. Architecture of ZfNet+VGG19
Pr 4 ed edge
Pooling layer: This is employed to reduce the spatial
It was discovered that CI green is only a reliable indicator domain and hence the network's calculation after the
of chlorophyll concentration in leaves lacking of convolution layer. Typically, the kernel size in
anthocyanin IN Eq. (15). Anthocyanin consumes in situ at ZfNet+VGG19 is 2 × 2 with stride 2. In this case, the
around 550 nm; as a result, if ρλ1 is near 550 nm in the pooling layer executes the maximum operation across the
green band, the index will be significantly impacted by the restricted spatial area R, yielding a feature map in Eq. (16):
absorption of both anthocyanin and Chl. The amount of l l
p =maxi ∈R α i (16)
Chl was overestimated. In order to estimate Chl in leaves
Fully connected (FC) layer: In ZfNet+VGG19, FC are
that contain anthocyanins, it was recommended to utilize
emulated employing a convolution with a size of n1, n2,
theCI red edge. where n1 × n2are the sizes of the input and output tensors,
accordingly. In most cases, n1 is an integer and n2 is a
3.6 ZfNet+VGG19 based classification triplet (7 × 7 × 512).
5
 Dropout: This layer, which is also known as "Drop," is Precision=True Positives /(True Positives+ False Positives)
often used to reduce the input fit and enhance the DL
algorithm's hypothesis. Typically, it gives the network  Recall, also known as sensitivity or true positive rate,
nodes weights (in PDCNN the percentage of 0.5 is is another performance metric used in statistics and
assigned to the two drop layers). machine learning to assess the effectiveness of a
Softmax: A ReLU layer accompanies the DL model classification model. Recall measures the proportion
with several layers and a convolution layer, establishing of correctly predicted positive instances out of the
the nonlinearity in the ZfNet+VGG19 model, and is often total actual positive instances.
represented as “σ”. Recall=True Positives/(True Positives+ False Negatives)
The categorization of the grape picture to 3-D groups
has been performed along with the borders. The grape
image has been coupled in groups with size b and then sent  F1-Score of the classification algorithm is the
weighted measure whose value ranges between 0 and
to the In ZfNet+VGG19.The d ×d ×n groups are then fed
1, where value 1 denotes the better performance of the
into the first layer of convolution (c1), which is composed
classification algorithm and the value 0 denotes the
of k c1 filters of the form l c 1 ×l c 1 × q c1 , where q c1 =n, poor performance of the classification algorithm.
the stride is constant at 1, and padding is not present. F 1 score=2∗( precision∗Sensitivity )/( precision+Sen
After applying the ReLU function, the feature maps for
c1
k were generated using c1, and these were then directed 3.8 Data collection
to MaxPool's first layer (mp1) using a l mp 1 × l mp 1kernel,
stride of 2, and padding. A total of 200 hyperspectral images were collected
The volume of the simulation analysis for during the trial period in summer 2019. The data collection
mp 1 mp 1 mp 1 c1 spanned over several days, with images captured at
p =d ×d ×k has been targeted for the
different time points to account for variations in
subsequent convolution layer (c2) with k c2 filters of size
c2 c2 c2 c2 c1
environmental conditions and disease progression. The
l ×l ×q , where q =k and that has a comparable rationale behind choosing the University of Missouri South
beginning convolution stride in addition to without Farm Research Center in Columbia, Missouri (38.92 N,
padding. 92.28 W) as the data collection location was based on its
known high vulnerability to Grapevine Vein Clearing
3.7 Performance analysis Virus (GVCV). This location provided an ideal setting to
study the early detection of GVCV in the Chardonel
The effectiveness of our proposed Artificial Humming cultivar, as it is known to be affected by the virus. To
Bird Optimized ZfNet+VGG19 neural network (AHB_ ensure robust data collection and experimental control, the
ZfNet+VGG19), utilising metrics including accuracy, grapevines were divided into two groups: one with normal
precision, recall and F1-score. Two baseline methods such vines and the other with vines infected with GVCV
as Spectral Dilated Convolution 3-Dimensional pathogens. This division allowed for a comparative study
Convolutional Neural Network (SDC-3DCNN), VGG-16 under comparable circumstances, minimizing potential
+GoogleNetare evaluated. confounding factors. The hyperspectral imaging system
 Accuracy describes how closely a specific value used for data collection had a fixed image size of 512×512
matches cases that have been categorized. Accuracy is pixels, capturing information across 204 bands ranging
the representation of systematic mistakes and from 397 to 1004 nm with a spectral resolution of 3 nm.
statistical bias. Additionally, it is the recognition The viewing area was 0.55 x 0.55 m, and the system
(combined TP and TN values) among the count of the claimed to provide a spatial resolution of 1.07 mm at a
assessed classes as well as the estimation's adequacy distance of 1 meter from the object. For this investigation,
to the genuine value. It's calculated as, the grapevines were photographed at the canopy level from
TP+True Negative (TN) a distance of 1-2 meters, ensuring that the imaging system
Accuracy= captured detailed and representative information from the
TP+TN + FP+ FN
vines. Overall, the careful data collection process and the
choice of location and grapevine cultivar ensured a
 Precision is a performance metric commonly used in
comprehensive dataset for studying the early detection of
statistics and machine learning to evaluate the
GVCV using hyperspectral imaging and optimized neural
accuracy of a classification model. It measures the
networks Table 1 shows the performance analysis of
proportion of correctly predicted positive instances
accuracy for different methods.
out of the total instances predicted as positive.
Precision can be defined as:
Table 1. Performance analysis of accuracy for different methods

Salinas dataset Indian pine dataset

Class SDC- VGG- AHB_ Class label SDC-3DCNN VGG- AHB_
label 3DCN 16+GoogleNe ZfNet+VGG19 16+GoogleNet ZfNet+VGG19
N t
Sable 85.6 83.6 94.6 Autumn Royal 84.6 85.3 94.3
Seedless
Alphonse 83.7 86.4 97.3 Crimson 82.6 84.6 97.8
Lavallée
Lival 84.9 84.6 95.7 Itum4 87.1 85.4 96.3
Black 83.6 89.6 94.6 Itum5 85.6 86.4 98.5
6
 Magic
Sugarone 81.6 84.3 92.8 Itum9 87.3 86.2 92.6
Superior
Seedless
Thompson 82.6 87.6 95.8 Vinyard_untrained 82.4 86.7 94.7
Seedless
Victoria 83.4 84.9 92.6 Vinyard_vertical_trelli 87.3 87.9 96.8
s

Figure 4. Analysis of accuracy for Indian pine dataset

Figures 3 and 4 depict the accuracy evaluation for Salinas

and Indian pine dataset. The comparison is done between
existing SDC-3DCNN, VGG-16 +GoogleNet with the
proposed AHB_ ZfNet+VGG19. X axis and Y axis show that
various class labels and the values obtained in percentage,
respectively. When contrasted with existing SDC-3DCNN
and v-gg-16 +GoogleNet methods achieve 85.2% and 86.4%
of accuracy, respectively, while the proposed AHB_
ZfNet+VGG19method achieves98.27% of accuracy for
Salinas datasets , which is13.07% and 12.27% better than
SDC-3DCNN and VGG-16 +GoogleNet. While analyzing
Indian pine dataset, SDC-3DCNN and VGG-16 +GoogleNet
method achieves 86.5% and 87%, while the proposed AHB_
ZfNet+VGG19 achieves 98.45% of accuracy, which is
12.15% and 11.45% better than existing methods. Table 2
Figure 3. Analysis of accuracy for Salinas’s dataset
shows the performance analysis of precision for different
methods.

Table 2. Performance analysis of precision for different methods

Salinas dataset Indian pine dataset
Class SDC- VGG- AHB_ Class label SDC-3DCNN VGG- AHB_
label 3DCN 16+GoogleNe ZfNet+VGG1 16+GoogleNe ZfNet+VGG1
N t 9 t 9
Sable 82.6 87.4 98.3 Autumn Royal 84.6 86.5 98.6
Seedless
Alphonse 84.9 86.5 97.4 Crimson 87.5 89.7 98.7
Lavallée
Lival 87.3 87.2 98.5 Itum4 83.5 83.4 96.6
Black 82.5 86.3 97.8 Itum5 84.6 87.5 97.9
Magic
Sugarone 84.6 87.9 94.7 Itum9 87.4 84.9 97.5
Superior
Seedless
Thompso 87.3 85.3 96.3 Vinyard_untrained 84.2 82.6 97.8
n
Seedless
Victoria 85.9 84.6 94.8 Vinyard_vertical_trelli 86.7 84.9 98.4
s

7
 Figures 5 and 6 depict the precision evaluation for Salinas and
Indian pine dataset. The comparison is done between existing
SDC-3DCNN, VGG-16 +GoogleNet with the proposed
AHB_ ZfNet+VGG19. X axis and Y axis show that various
class labels and the values obtained in percentage,
respectively. When contrasted with existing SDC-3DCNN
and v-gg-16 +GoogleNet methods achieve 82.1% and 84.3%
of precision, respectively, while the proposed AHB_
ZfNet+VGG19method achieves97.67% of precision for
Salinas dataset, which is 15.57% and 13.37% better than
SDC-3DCNN and VGG-16 +GoogleNet. While analyzing
Indian pine dataset, SDC-3DCNN and VGG-16 +GoogleNet
method achieves 84% and 82%, while the proposed AHB_
ZfNet+VGG19 achieves 97.1% of precision, which is 13.1%
and 15% better than the existing methods. Table 3 shows the
performance analysis of recall for different methods.
Figure 5. Analysis of precision for Salinas’s dataset

Figure 6. Analysis of precision for Indian pine dataset

Figure 7. Analysis of recall for Salinas dataset

Table 3. Performance analysis of recall for different methods

Salinas dataset Indian pine dataset

Class label SDC- VGG- AHB_ Class label SDC-3DCNN VGG- AHB_
3DCNN 16+GoogleNet ZfNet+VGG1 16+GoogleNet ZfNet+VGG19
9
Sable 87.4 86.4 97.45 Autumn 86.4 84.6 96.53
Seedless Royal
Alphonse 89.3 87.4 96.2 Crimson 84.5 87.5 97.56
Lavallée
Lival 84.4 86.5 98.4 Itum4 87.4 84.6 98.4
Black Magic 87.6 85.3 96.9 Itum5 82.6 87.5 97.6
Sugarone 85.1 82.6 94.8 Itum9 84.6 86.2 98.2
Superior
Seedless
Thompson 86.3 84.7 97.5 Vinyard_untr 87.4 86.5 96.8
Seedless ained
Victoria 84.3 84.6 96.3 Vinyard_vert 89.6 86.3 94.5
ical_trellis

8
 Figure 9. Analysis of F1-score for Salinas dataset

Figure 8. Analysis of recall for Indian pine dataset

The recall assessment for the Salinas and Indian Pine

datasets is shown in Figures 7 and 8. The suggested AHB_
ZfNet+VGG19 is compared with the already existing SDC-
3DCNN VGG-16 + GoogleNet. Different class names and
percentage values are shown on the X and Y axes,
correspondingly. When contrasted with existing SDC-
3DCNN and VGG-16 +GoogleNet methods achieve 87.3%
and 85.2% of recall, respectively, while the proposed AHB_
ZfNet+VGG19method achieves97.41% of recall for Salinas
datasets, which is 20.11% and 12.21% better than SDC-
3DCNN and VGG-16 +GoogleNet. While analyzing Indian
pine dataset, SDC-3DCNN and VGG-16 16 +GoogleNet
methods achieves 89% and 86% while the proposed AHB_
ZfNet+VGG19 achieves 97.41% of recall, which is 6.5% and
11% better than the existing methods. Table 4 shows the
performance analysis of F1_score for different methods.
Figure 10. Analysis of F1-score for Indian pine dataset

Table 4. Performance analysis of F1_score for different methods

Salinas dataset Indian pine dataset
Class label SDC- VGG-16+ AHB_ Class label SDC-3DCNN VGG-16+ AHB_
3DCNN GoogleNet ZfNet+VGG19 GoogleNet ZfNet+VGG19
Sable 81.8 84.6 97.65 Autumn 86.5 86.7 98.65
Seedless Royal
Alphonse 86.4 87.4 96.58 Crimson 87.3 84.9 97.74
Lavallée
Lival 87.4 84.5 98.32 Itum4 84.7 87.6 98.54
Black 84.6 87.5 97.48 Itum5 86.9 83.5 98.78
Magic
Sugarone 82.6 86.4 98.25 Itum9 84.6 81.6 96.82
Superior
Seedless
Thompson 84.7 87.5 97.8 Vinyard_un 86.8 84.6 97.56
Seedless trained
Victoria 83.5 89.4 98.4 Vinyard_ve 86.4 89.7 98.23
rtical_trellis

The F1-score analysis for the Salinas and Indian Pine datasets is shown in Figures 9 and 10. The suggested AHB_
9
ZfNet+VGG19 is compared with the already existing SDC- pine dataset, SDC-3DCNN and VGG-16 +GoogleNet
3DCNN VGG-16 + GoogleNet. Different class names and methods achieved 86.4% and 82.4%, while the proposed
percentage values are shown on the X and Y axes, AHB_ ZfNet+VGG19 achieves 97.6% of F1-score, which is
accordingly. When contrasted, existing SDC-3DCNN and 11.2% and 15.2% better than existing methods.
VGG-16 +GoogleNet methods achieve 83.4% and 83% of F1- Tables 5 and 6 show the analysis of Salinas’s dataset and
scorerespectively, while the proposed AHB_ Indian pine dataset for the proposed AHB_ ZfNet+VGG19.
ZfNet+VGG19method achieves97.74% of F1-score for
Salinas dataset, which is 14.31% and 14.3% better than SDC-
3DCNN and VGG-16 +GoogleNet. While analyzing Indian

Table 5. Parametric analysis on Salinas’s dataset for the proposed AHB_ ZfNet+VGG19
Class label Accuracy Precision Recall F1-score
Sable Seedless 97.3 96.7 97.6 97.6
Alphonse Lavallée 99.7 97.6 98.9 97.6
Lival 96.8 98.5 97.4 97.6
Black Magic 99.6 97.6 95.8 96.8
Sugarone Superior 96.8 94.5 98.5
97.5
Seedless
Thompson 98.9 97.3 97.6
98.4
Seedless
Victoria 98.6 97.6 98.5 98.5
Average 98.27 97.67 97.41 97.74

Table 6. Parametric analysis on Indian pine dataset using the proposed AHB_ ZfNet+VGG19
Class label Accuracy Precision Recall F1-score
Autumn Royal 98.4 95.8 96.9 98.5
Crimson 97.6 96.7 97.8 96.7
Itum4 98.9 97.6 96.8 98.7
Itum5 98.5 98.1 94.9 95.9
Itum9 97.9 94.9 93.4 97.5
Vinyard_untrained 98.7 97.5 98.2 98.4
Vinyard_vertical_trellis 99.2 98.5 98.9 97.8
Average 98.45 97.1 97.41 97.6

4. CONCLUSION
researchers in managing disease outbreaks and maintaining crop
This research demonstrates the potential of hyper spectral health. The findings of this study contribute significantly to the
remote sensing technology for early detection of viral field of agricultural diagnostics and offer new avenues for
infections in grapevines, particularly in the case of the newly precision farming practices. Continued research in this area may
discovered DNA virus, grapevine vein-clearing virus lead to even more advanced and efficient disease detection
(GVCV). The study successfully identifies and categorizes methods, ultimately benefiting the grapevine industry and
infected grapevines during their initial asymptomatic stages, promoting sustainable and resilient agriculture
paving the way for timely intervention to prevent disease
spread across vineyards. By using hyper spectral photography REFERENCES
at the plant level and employing a statistical technique, the
research effectively distinguishes between healthy and [1] Sowmyalakshmi, R., Jayasankar, T., AyyemPiIllai, V.,
GVCV-infected grapevines based on their reflectance spectra Subramaniyan, K.,
patterns. The integration of the artificial hummingbird Pustokhina, Irina V., Pustokhin, Denis A., Shankar, K.
optimization technique aids in feature extraction, ensuring the (2021). An Optimal Classification Model for Rice Plant
selection of pertinent features and enhancing overall model Disease Detection. Computers, Materials & Continua,
classification accuracy. Moreover, the adoption of a 68(2): 1117–1128.
nondestructive method to calculate total chlorophyll content in https://doi.org/10.32604/cmc.2021.016825
grape leaves provides valuable insights into disease severity [2] Fang, Y., Ramasamy, R. P. (2015). Current and
assessment. The correlation between chlorophyll prospective methods for plant disease detection.
concentration and reflectance measurements in specific Biosensors, 5(3): 537–561.
spectral ranges offers a non-invasive means to gauge disease https://doi.org/10.3390/bios5030537
impact on grapevine health. The hybrid approach utilizing [3] Thanarajan, T., Alotaibi, Y., Rajendran, S., &
ZfNet+VGG19 for pixel-wise and image-wise classification of Nagappan, K. (2023). Improved wolf swarm
disease severity showcases the effectiveness of the proposed optimization with deep-learning-based movement
methodology in accurately identifying and categorizing analysis and self-regulated human activity
GVCV infections. Overall, this research establishes hyper recognition. AIMS Mathematics, 8(5), 12520-
spectral remote sensing as a promising tool for the early 12539.Chaerle, L., Van Der, S. D. (2000). Imaging
detection and monitoring of viral infections in grapevines, techniques and the early detection of plant stress. Trends
providing valuable support for viticulture farmers and in Plant Science, 5(11): 495–501.
10
 http://dx.doi.org/10.1016/S1360-1385(00)01781-7 Classification of Wheat Rust Diseases Using Deep
[4] Rumpf, T., Mahlein, A. K., Steiner, U., Oerke, E. C., Convolutional Neural Networks. In 2021 9th
Dehne, H. W., Plumer, L. (2010). Early detection and International Conference on Reliability, Infocom
classification of plant diseases with support vector Technologies and Optimization (Trends and Future
machines based on hyper spectral reflectance. Directions)(ICRITO), IEEE, 1-6.
Computers and Electronics in Agriculture, 74(1): 91– https://doi.org/10.1109/ICRITO51393.2021.9596133
99. https://doi.org/10.1016/j.compag.2010.06.009 [16] Jiang, Q., Wu, G., Tian, C., Li, N., Yang, H., Bai, Y.,
[5] Xie, C., He, Y. (2016). Spectrum and image texture Zhang, B. (2021). Hyper spectral imaging for early
features analysis for early blight disease detection on identification of strawberry leaves diseases with
eggplant leaves. Sensors, 16(5): 676. machine learning and spectral fingerprint
https://doi.org/10.3390/s16050676 features. Infrared Physics & Technology, 118: 103898.
[6] Mahlein, A. K., Rumpf, T., Welke, P., Dehne, H. W., https://doi.org/10.1016/j.infrared.2021.103898
Plumer, L., Steiner, U., Oerke, E. C. (2013). [17] Wang, D., Vinson, R., Holmes, M., Seibel, G., Bechar,
Development of spectral indices for detecting and A., Nof, S., Luo, Y., Tao, Y. (2018). Early tomato
identifying plant diseases. Remote Sensing of spotted wilt virus detection using hyper spectral imaging
Environment, 128: 21–30. technique and outlier removal auxiliary classifier
https://doi.org/10.1016/j.rse.2012.09.019 generative adversarial nets (OR-AC-GAN). In 2018
[7] Wahabzada, M., Mahlein, A. K., Bauckhage, C., ASABE Annual International Meeting (p. 1). American
Steiner, U., Oerke, E. C., Kersting, K. (2016). Plant Society of Agricultural and Biological Engineers.
phenotyping using probabilistic topic models: http://dx.doi.org/10.13031/aim.201800660.
Uncovering the hyper spectral language of plants. [18] Alharbi, M., Rajagopal, S.K., Rajendran, S., Alshahrani,
Scientific reports, 6(1): 22482. M. (2023). Plant disease classification based on
http://dx.doi.org/10.1038/srep22482 ConvLSTM U-Net with fully connected convolutional
[8] Rajagopal, S., Thanarajan, T., Alotaibi, Y., & layers. Traitement du Signal, Vol. 40, No. 1, pp. 157-
Alghamdi, S. (2023). Brain Tumor: Hybrid Feature 166. https://doi.org/10.18280/ts.400114.
Extraction Based on UNet and 3DCNN. Computer [19] Nguyen, C., Sagan, V., Maimaitiyiming, M.,
Systems Science & Engineering, 45(2).Govender, M., Maimaitijiang, M., Bhadra, S., & Kwasniewski, M. T.
Chetty, K., Bulcock, H. (2007). A review of hyper (2021). Early detection of plant viral disease using
spectral remote sensing and its application in hyperspectral imaging and deep learning. Sensors,
vegetation and water resource studies. Water Sa, 33(2): 21(3), 742.
145–151. http://dx.doi.org/10.4314/wsa.v33i2.49049 [20] Nguyen, C., Sagan, V., Maimaitiyiming, M.,
[9] Adam, E., Mutanga, O., Rugege, D. (2010). Maimaitijiang, M., Bhadra, S., & Kwasniewski, M. T.
Multispectral and hyper spectral remote sensing for (2021). Early detection of plant viral disease using
identification and mapping of wetland vegetation: a hyperspectral imaging and deep learning. Sensors,
review. Wetlands Ecology and Management, 18(3): 21(3), 742.
281–296. http://dx.doi.org/10.1007/s11273-009-9169-z [21] Sawyer, E., Laroche-Pinel, E., Flasco, M., Cooper, M.
[10] Narayanan, K. L., Krishnan, R. S., Robinson, Y. H., L., Corrales, B., Fuchs, M., & Brillante, L. (2023).
Julie, E. G., Vimal, S., Saravanan, V., Kaliappan, M. Phenotyping grapevine red blotch virus and grapevine
(2022). Banana plant disease classification using leafroll-associated viruses before and after symptom
hybrid convolutional neural network. Computational expression through machine-learning analysis of
Intelligence and Neuroscience, 2022. hyperspectral images. Frontiers in plant science, 14,
https://doi.org/10.1155/2022/9153699 1117869.
[11] Yakkundimath, R., Saunshi, G., Anami, B., Palaiah, S. [22] Gu, Q., Sheng, L., Zhang, T., Lu, Y., Zhang, Z.,
(2022). Classification of Rice Diseases using Zheng, K., ... & Zhou, H. (2019). Early detection of
Convolutional Neural Network Models. Journal of The tomato spotted wilt virus infection in tobacco using the
Institution of Engineers (India): Series B, 1-13. hyperspectral imaging technique and machine learning
https://ui.adsabs.harvard.edu/link_gateway/2022JIEIB. algorithms. Computers and Electronics in Agriculture,
103.1047Y/doi:10.1007/s40031-021-00704-4 167, 105066.
[12] Cao, Y., Yuan, P., Xu, H., Martínez-Ortega, J. F., [23] Soni, A., Dixit, Y., Reis, M. M., & Brightwell, G.
Feng, J., Zhai, Z. (2022). Detecting Asymptomatic (2022). Hyperspectral imaging and machine learning in
Infections of Rice Bacterial Leaf Blight Using Hyper food microbiology: Developments and challenges in
spectral Imaging and 3-Dimensional Convolutional detection of bacterial, fungal, and viral contaminants.
Neural Network with Spectral Dilated Comprehensive Reviews in Food Science and Food
Convolution. Frontiers in Plant Science, 13: 963170. Safety, 21(4), 3717-3745.
https://doi.org/10.3389%2Ffpls.2022.963170 [24] Palani, V., Thanarajan, T., Krishnamurthy, A., &
[13] Nalini, S., Krishnaraj, N., Jayasankar, T., Rajendran, S. (2023). Deep Learning Based
Vinothkumar, K., Francis Britto, A. S., Subramaniam, Compression with Classification Model on CMOS
K., Bharatiraja, C. (2021). Paddy Leaf Disease Image Sensors. Traitement du Signal, 40(3).
Detection Using an Optimized Deep Neural Network. [25] Palani, V., Alharbi, M., Alshahrani, M., & Rajendran,
Computers, Materials & Continua, 68(1): 1117–1128. S. (2023). Pixel Optimization Using Iterative Pixel
DOI:10.32604/cmc.2021.012431 Compression Algorithm for Complementary Metal
[14] Gui, J., Fei, J., Wu, Z., Fu, X., Diakite, A. (2021). Oxide Semiconductor Image Sensors. Traitement du
Grading method of soybean mosaic disease based on Signal, 40(2).
hyperspectral imaging technology. Information
Processing in Agriculture, 8(3): 380-385
https://doi.org/10.1016/j.inpa.2020.10.006
[15] Kukreja, V., Kumar, D. (2021). Automatic
11